linux/drivers/dma/ti/edma.c

2749 lines
71 KiB
C
Raw Normal View History

/*
* TI EDMA DMA engine driver
*
* Copyright 2012 Texas Instruments
*
* This program is free software; you can redistribute it and/or
* modify it under the terms of the GNU General Public License as
* published by the Free Software Foundation version 2.
*
* This program is distributed "as is" WITHOUT ANY WARRANTY of any
* kind, whether express or implied; without even the implied warranty
* of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
* GNU General Public License for more details.
*/
#include <linux/dmaengine.h>
#include <linux/dma-mapping.h>
#include <linux/bitmap.h>
#include <linux/err.h>
#include <linux/init.h>
#include <linux/interrupt.h>
#include <linux/list.h>
#include <linux/module.h>
#include <linux/platform_device.h>
#include <linux/slab.h>
#include <linux/spinlock.h>
#include <linux/of.h>
#include <linux/of_dma.h>
#include <linux/of_irq.h>
#include <linux/of_address.h>
#include <linux/of_device.h>
#include <linux/pm_runtime.h>
#include <linux/platform_data/edma.h>
#include "../dmaengine.h"
#include "../virt-dma.h"
/* Offsets matching "struct edmacc_param" */
#define PARM_OPT 0x00
#define PARM_SRC 0x04
#define PARM_A_B_CNT 0x08
#define PARM_DST 0x0c
#define PARM_SRC_DST_BIDX 0x10
#define PARM_LINK_BCNTRLD 0x14
#define PARM_SRC_DST_CIDX 0x18
#define PARM_CCNT 0x1c
#define PARM_SIZE 0x20
/* Offsets for EDMA CC global channel registers and their shadows */
#define SH_ER 0x00 /* 64 bits */
#define SH_ECR 0x08 /* 64 bits */
#define SH_ESR 0x10 /* 64 bits */
#define SH_CER 0x18 /* 64 bits */
#define SH_EER 0x20 /* 64 bits */
#define SH_EECR 0x28 /* 64 bits */
#define SH_EESR 0x30 /* 64 bits */
#define SH_SER 0x38 /* 64 bits */
#define SH_SECR 0x40 /* 64 bits */
#define SH_IER 0x50 /* 64 bits */
#define SH_IECR 0x58 /* 64 bits */
#define SH_IESR 0x60 /* 64 bits */
#define SH_IPR 0x68 /* 64 bits */
#define SH_ICR 0x70 /* 64 bits */
#define SH_IEVAL 0x78
#define SH_QER 0x80
#define SH_QEER 0x84
#define SH_QEECR 0x88
#define SH_QEESR 0x8c
#define SH_QSER 0x90
#define SH_QSECR 0x94
#define SH_SIZE 0x200
/* Offsets for EDMA CC global registers */
#define EDMA_REV 0x0000
#define EDMA_CCCFG 0x0004
#define EDMA_QCHMAP 0x0200 /* 8 registers */
#define EDMA_DMAQNUM 0x0240 /* 8 registers (4 on OMAP-L1xx) */
#define EDMA_QDMAQNUM 0x0260
#define EDMA_QUETCMAP 0x0280
#define EDMA_QUEPRI 0x0284
#define EDMA_EMR 0x0300 /* 64 bits */
#define EDMA_EMCR 0x0308 /* 64 bits */
#define EDMA_QEMR 0x0310
#define EDMA_QEMCR 0x0314
#define EDMA_CCERR 0x0318
#define EDMA_CCERRCLR 0x031c
#define EDMA_EEVAL 0x0320
#define EDMA_DRAE 0x0340 /* 4 x 64 bits*/
#define EDMA_QRAE 0x0380 /* 4 registers */
#define EDMA_QUEEVTENTRY 0x0400 /* 2 x 16 registers */
#define EDMA_QSTAT 0x0600 /* 2 registers */
#define EDMA_QWMTHRA 0x0620
#define EDMA_QWMTHRB 0x0624
#define EDMA_CCSTAT 0x0640
#define EDMA_M 0x1000 /* global channel registers */
#define EDMA_ECR 0x1008
#define EDMA_ECRH 0x100C
#define EDMA_SHADOW0 0x2000 /* 4 shadow regions */
#define EDMA_PARM 0x4000 /* PaRAM entries */
#define PARM_OFFSET(param_no) (EDMA_PARM + ((param_no) << 5))
#define EDMA_DCHMAP 0x0100 /* 64 registers */
/* CCCFG register */
#define GET_NUM_DMACH(x) (x & 0x7) /* bits 0-2 */
#define GET_NUM_QDMACH(x) ((x & 0x70) >> 4) /* bits 4-6 */
#define GET_NUM_PAENTRY(x) ((x & 0x7000) >> 12) /* bits 12-14 */
#define GET_NUM_EVQUE(x) ((x & 0x70000) >> 16) /* bits 16-18 */
#define GET_NUM_REGN(x) ((x & 0x300000) >> 20) /* bits 20-21 */
#define CHMAP_EXIST BIT(24)
/* CCSTAT register */
#define EDMA_CCSTAT_ACTV BIT(4)
/*
* Max of 20 segments per channel to conserve PaRAM slots
* Also note that MAX_NR_SG should be atleast the no.of periods
* that are required for ASoC, otherwise DMA prep calls will
* fail. Today davinci-pcm is the only user of this driver and
* requires atleast 17 slots, so we setup the default to 20.
*/
#define MAX_NR_SG 20
#define EDMA_MAX_SLOTS MAX_NR_SG
#define EDMA_DESCRIPTORS 16
#define EDMA_CHANNEL_ANY -1 /* for edma_alloc_channel() */
#define EDMA_SLOT_ANY -1 /* for edma_alloc_slot() */
#define EDMA_CONT_PARAMS_ANY 1001
#define EDMA_CONT_PARAMS_FIXED_EXACT 1002
#define EDMA_CONT_PARAMS_FIXED_NOT_EXACT 1003
/*
* 64bit array registers are split into two 32bit registers:
* reg0: channel/event 0-31
* reg1: channel/event 32-63
*
* bit 5 in the channel number tells the array index (0/1)
* bit 0-4 (0x1f) is the bit offset within the register
*/
#define EDMA_REG_ARRAY_INDEX(channel) ((channel) >> 5)
#define EDMA_CHANNEL_BIT(channel) (BIT((channel) & 0x1f))
/* PaRAM slots are laid out like this */
struct edmacc_param {
u32 opt;
u32 src;
u32 a_b_cnt;
u32 dst;
u32 src_dst_bidx;
u32 link_bcntrld;
u32 src_dst_cidx;
u32 ccnt;
} __packed;
/* fields in edmacc_param.opt */
#define SAM BIT(0)
#define DAM BIT(1)
#define SYNCDIM BIT(2)
#define STATIC BIT(3)
#define EDMA_FWID (0x07 << 8)
#define TCCMODE BIT(11)
#define EDMA_TCC(t) ((t) << 12)
#define TCINTEN BIT(20)
#define ITCINTEN BIT(21)
#define TCCHEN BIT(22)
#define ITCCHEN BIT(23)
struct edma_pset {
u32 len;
dma_addr_t addr;
struct edmacc_param param;
};
struct edma_desc {
struct virt_dma_desc vdesc;
struct list_head node;
enum dma_transfer_direction direction;
int cyclic;
bool polled;
int absync;
int pset_nr;
struct edma_chan *echan;
int processed;
/*
* The following 4 elements are used for residue accounting.
*
* - processed_stat: the number of SG elements we have traversed
* so far to cover accounting. This is updated directly to processed
* during edma_callback and is always <= processed, because processed
* refers to the number of pending transfer (programmed to EDMA
* controller), where as processed_stat tracks number of transfers
* accounted for so far.
*
* - residue: The amount of bytes we have left to transfer for this desc
*
* - residue_stat: The residue in bytes of data we have covered
* so far for accounting. This is updated directly to residue
* during callbacks to keep it current.
*
* - sg_len: Tracks the length of the current intermediate transfer,
* this is required to update the residue during intermediate transfer
* completion callback.
*/
int processed_stat;
u32 sg_len;
u32 residue;
u32 residue_stat;
struct edma_pset pset[0];
};
struct edma_cc;
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
struct edma_tc {
struct device_node *node;
u16 id;
};
struct edma_chan {
struct virt_dma_chan vchan;
struct list_head node;
struct edma_desc *edesc;
struct edma_cc *ecc;
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
struct edma_tc *tc;
int ch_num;
bool alloced;
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
bool hw_triggered;
int slot[EDMA_MAX_SLOTS];
dma: edma: Find missed events and issue them In an effort to move to using Scatter gather lists of any size with EDMA as discussed at [1] instead of placing limitations on the driver, we work through the limitations of the EDMAC hardware to find missed events and issue them. The sequence of events that require this are: For the scenario where MAX slots for an EDMA channel is 3: SG1 -> SG2 -> SG3 -> SG4 -> SG5 -> SG6 -> Null The above SG list will have to be DMA'd in 2 sets: (1) SG1 -> SG2 -> SG3 -> Null (2) SG4 -> SG5 -> SG6 -> Null After (1) is succesfully transferred, the events from the MMC controller donot stop coming and are missed by the time we have setup the transfer for (2). So here, we catch the events missed as an error condition and issue them manually. In the second part of the patch, we make handle the NULL slot cases: For crypto IP, we continue to receive events even continuously in NULL slot, the setup of the next set of SG elements happens after the error handler executes. This is results in some recursion problems. Due to this, we continously receive error interrupts when we manually trigger an event from the error handler. We fix this, by first detecting if the Channel is currently transferring from a NULL slot or not, that's where the edma_read_slot in the error callback from interrupt handler comes in. With this we can determine if the set up of the next SG list has completed, and we manually trigger only in this case. If the setup has _not_ completed, we are still in NULL so we just set a missed flag and allow the manual triggerring to happen in edma_execute which will be eventually called. This fixes the above mentioned race conditions seen with the crypto drivers. [1] http://marc.info/?l=linux-omap&m=137416733628831&w=2 Signed-off-by: Joel Fernandes <joelf@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2013-08-30 07:05:43 +08:00
int missed;
struct dma_slave_config cfg;
};
struct edma_cc {
struct device *dev;
struct edma_soc_info *info;
void __iomem *base;
int id;
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
bool legacy_mode;
/* eDMA3 resource information */
unsigned num_channels;
unsigned num_qchannels;
unsigned num_region;
unsigned num_slots;
unsigned num_tc;
bool chmap_exist;
enum dma_event_q default_queue;
unsigned int ccint;
unsigned int ccerrint;
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
/*
* The slot_inuse bit for each PaRAM slot is clear unless the slot is
* in use by Linux or if it is allocated to be used by DSP.
*/
unsigned long *slot_inuse;
/*
* For tracking reserved channels used by DSP.
* If the bit is cleared, the channel is allocated to be used by DSP
* and Linux must not touch it.
*/
unsigned long *channels_mask;
struct dma_device dma_slave;
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
struct dma_device *dma_memcpy;
struct edma_chan *slave_chans;
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
struct edma_tc *tc_list;
int dummy_slot;
};
/* dummy param set used to (re)initialize parameter RAM slots */
static const struct edmacc_param dummy_paramset = {
.link_bcntrld = 0xffff,
.ccnt = 1,
};
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
#define EDMA_BINDING_LEGACY 0
#define EDMA_BINDING_TPCC 1
static const u32 edma_binding_type[] = {
[EDMA_BINDING_LEGACY] = EDMA_BINDING_LEGACY,
[EDMA_BINDING_TPCC] = EDMA_BINDING_TPCC,
};
static const struct of_device_id edma_of_ids[] = {
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
{
.compatible = "ti,edma3",
.data = &edma_binding_type[EDMA_BINDING_LEGACY],
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
},
{
.compatible = "ti,edma3-tpcc",
.data = &edma_binding_type[EDMA_BINDING_TPCC],
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
},
{}
};
MODULE_DEVICE_TABLE(of, edma_of_ids);
static const struct of_device_id edma_tptc_of_ids[] = {
{ .compatible = "ti,edma3-tptc", },
{}
};
MODULE_DEVICE_TABLE(of, edma_tptc_of_ids);
static inline unsigned int edma_read(struct edma_cc *ecc, int offset)
{
return (unsigned int)__raw_readl(ecc->base + offset);
}
static inline void edma_write(struct edma_cc *ecc, int offset, int val)
{
__raw_writel(val, ecc->base + offset);
}
static inline void edma_modify(struct edma_cc *ecc, int offset, unsigned and,
unsigned or)
{
unsigned val = edma_read(ecc, offset);
val &= and;
val |= or;
edma_write(ecc, offset, val);
}
static inline void edma_and(struct edma_cc *ecc, int offset, unsigned and)
{
unsigned val = edma_read(ecc, offset);
val &= and;
edma_write(ecc, offset, val);
}
static inline void edma_or(struct edma_cc *ecc, int offset, unsigned or)
{
unsigned val = edma_read(ecc, offset);
val |= or;
edma_write(ecc, offset, val);
}
static inline unsigned int edma_read_array(struct edma_cc *ecc, int offset,
int i)
{
return edma_read(ecc, offset + (i << 2));
}
static inline void edma_write_array(struct edma_cc *ecc, int offset, int i,
unsigned val)
{
edma_write(ecc, offset + (i << 2), val);
}
static inline void edma_modify_array(struct edma_cc *ecc, int offset, int i,
unsigned and, unsigned or)
{
edma_modify(ecc, offset + (i << 2), and, or);
}
static inline void edma_or_array(struct edma_cc *ecc, int offset, int i,
unsigned or)
{
edma_or(ecc, offset + (i << 2), or);
}
static inline void edma_or_array2(struct edma_cc *ecc, int offset, int i, int j,
unsigned or)
{
edma_or(ecc, offset + ((i * 2 + j) << 2), or);
}
static inline void edma_write_array2(struct edma_cc *ecc, int offset, int i,
int j, unsigned val)
{
edma_write(ecc, offset + ((i * 2 + j) << 2), val);
}
static inline unsigned int edma_shadow0_read(struct edma_cc *ecc, int offset)
{
return edma_read(ecc, EDMA_SHADOW0 + offset);
}
static inline unsigned int edma_shadow0_read_array(struct edma_cc *ecc,
int offset, int i)
{
return edma_read(ecc, EDMA_SHADOW0 + offset + (i << 2));
}
static inline void edma_shadow0_write(struct edma_cc *ecc, int offset,
unsigned val)
{
edma_write(ecc, EDMA_SHADOW0 + offset, val);
}
static inline void edma_shadow0_write_array(struct edma_cc *ecc, int offset,
int i, unsigned val)
{
edma_write(ecc, EDMA_SHADOW0 + offset + (i << 2), val);
}
static inline unsigned int edma_param_read(struct edma_cc *ecc, int offset,
int param_no)
{
return edma_read(ecc, EDMA_PARM + offset + (param_no << 5));
}
static inline void edma_param_write(struct edma_cc *ecc, int offset,
int param_no, unsigned val)
{
edma_write(ecc, EDMA_PARM + offset + (param_no << 5), val);
}
static inline void edma_param_modify(struct edma_cc *ecc, int offset,
int param_no, unsigned and, unsigned or)
{
edma_modify(ecc, EDMA_PARM + offset + (param_no << 5), and, or);
}
static inline void edma_param_and(struct edma_cc *ecc, int offset, int param_no,
unsigned and)
{
edma_and(ecc, EDMA_PARM + offset + (param_no << 5), and);
}
static inline void edma_param_or(struct edma_cc *ecc, int offset, int param_no,
unsigned or)
{
edma_or(ecc, EDMA_PARM + offset + (param_no << 5), or);
}
static void edma_assign_priority_to_queue(struct edma_cc *ecc, int queue_no,
int priority)
{
int bit = queue_no * 4;
edma_modify(ecc, EDMA_QUEPRI, ~(0x7 << bit), ((priority & 0x7) << bit));
}
static void edma_set_chmap(struct edma_chan *echan, int slot)
{
struct edma_cc *ecc = echan->ecc;
int channel = EDMA_CHAN_SLOT(echan->ch_num);
if (ecc->chmap_exist) {
slot = EDMA_CHAN_SLOT(slot);
edma_write_array(ecc, EDMA_DCHMAP, channel, (slot << 5));
}
}
static void edma_setup_interrupt(struct edma_chan *echan, bool enable)
{
struct edma_cc *ecc = echan->ecc;
int channel = EDMA_CHAN_SLOT(echan->ch_num);
int idx = EDMA_REG_ARRAY_INDEX(channel);
int ch_bit = EDMA_CHANNEL_BIT(channel);
if (enable) {
edma_shadow0_write_array(ecc, SH_ICR, idx, ch_bit);
edma_shadow0_write_array(ecc, SH_IESR, idx, ch_bit);
} else {
edma_shadow0_write_array(ecc, SH_IECR, idx, ch_bit);
}
}
/*
* paRAM slot management functions
*/
static void edma_write_slot(struct edma_cc *ecc, unsigned slot,
const struct edmacc_param *param)
{
slot = EDMA_CHAN_SLOT(slot);
if (slot >= ecc->num_slots)
return;
memcpy_toio(ecc->base + PARM_OFFSET(slot), param, PARM_SIZE);
}
static int edma_read_slot(struct edma_cc *ecc, unsigned slot,
struct edmacc_param *param)
{
slot = EDMA_CHAN_SLOT(slot);
if (slot >= ecc->num_slots)
return -EINVAL;
memcpy_fromio(param, ecc->base + PARM_OFFSET(slot), PARM_SIZE);
return 0;
}
/**
* edma_alloc_slot - allocate DMA parameter RAM
* @ecc: pointer to edma_cc struct
* @slot: specific slot to allocate; negative for "any unused slot"
*
* This allocates a parameter RAM slot, initializing it to hold a
* dummy transfer. Slots allocated using this routine have not been
* mapped to a hardware DMA channel, and will normally be used by
* linking to them from a slot associated with a DMA channel.
*
* Normal use is to pass EDMA_SLOT_ANY as the @slot, but specific
* slots may be allocated on behalf of DSP firmware.
*
* Returns the number of the slot, else negative errno.
*/
static int edma_alloc_slot(struct edma_cc *ecc, int slot)
{
if (slot >= 0) {
slot = EDMA_CHAN_SLOT(slot);
/* Requesting entry paRAM slot for a HW triggered channel. */
if (ecc->chmap_exist && slot < ecc->num_channels)
slot = EDMA_SLOT_ANY;
}
if (slot < 0) {
if (ecc->chmap_exist)
slot = 0;
else
slot = ecc->num_channels;
for (;;) {
slot = find_next_zero_bit(ecc->slot_inuse,
ecc->num_slots,
slot);
if (slot == ecc->num_slots)
return -ENOMEM;
if (!test_and_set_bit(slot, ecc->slot_inuse))
break;
}
} else if (slot >= ecc->num_slots) {
return -EINVAL;
} else if (test_and_set_bit(slot, ecc->slot_inuse)) {
return -EBUSY;
}
edma_write_slot(ecc, slot, &dummy_paramset);
return EDMA_CTLR_CHAN(ecc->id, slot);
}
static void edma_free_slot(struct edma_cc *ecc, unsigned slot)
{
slot = EDMA_CHAN_SLOT(slot);
if (slot >= ecc->num_slots)
return;
edma_write_slot(ecc, slot, &dummy_paramset);
clear_bit(slot, ecc->slot_inuse);
}
/**
* edma_link - link one parameter RAM slot to another
* @ecc: pointer to edma_cc struct
* @from: parameter RAM slot originating the link
* @to: parameter RAM slot which is the link target
*
* The originating slot should not be part of any active DMA transfer.
*/
static void edma_link(struct edma_cc *ecc, unsigned from, unsigned to)
{
if (unlikely(EDMA_CTLR(from) != EDMA_CTLR(to)))
dev_warn(ecc->dev, "Ignoring eDMA instance for linking\n");
from = EDMA_CHAN_SLOT(from);
to = EDMA_CHAN_SLOT(to);
if (from >= ecc->num_slots || to >= ecc->num_slots)
return;
edma_param_modify(ecc, PARM_LINK_BCNTRLD, from, 0xffff0000,
PARM_OFFSET(to));
}
/**
* edma_get_position - returns the current transfer point
* @ecc: pointer to edma_cc struct
* @slot: parameter RAM slot being examined
* @dst: true selects the dest position, false the source
*
* Returns the position of the current active slot
*/
static dma_addr_t edma_get_position(struct edma_cc *ecc, unsigned slot,
bool dst)
{
u32 offs;
slot = EDMA_CHAN_SLOT(slot);
offs = PARM_OFFSET(slot);
offs += dst ? PARM_DST : PARM_SRC;
return edma_read(ecc, offs);
}
/*
* Channels with event associations will be triggered by their hardware
* events, and channels without such associations will be triggered by
* software. (At this writing there is no interface for using software
* triggers except with channels that don't support hardware triggers.)
*/
static void edma_start(struct edma_chan *echan)
{
struct edma_cc *ecc = echan->ecc;
int channel = EDMA_CHAN_SLOT(echan->ch_num);
int idx = EDMA_REG_ARRAY_INDEX(channel);
int ch_bit = EDMA_CHANNEL_BIT(channel);
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
if (!echan->hw_triggered) {
/* EDMA channels without event association */
dev_dbg(ecc->dev, "ESR%d %08x\n", idx,
edma_shadow0_read_array(ecc, SH_ESR, idx));
edma_shadow0_write_array(ecc, SH_ESR, idx, ch_bit);
} else {
/* EDMA channel with event association */
dev_dbg(ecc->dev, "ER%d %08x\n", idx,
edma_shadow0_read_array(ecc, SH_ER, idx));
/* Clear any pending event or error */
edma_write_array(ecc, EDMA_ECR, idx, ch_bit);
edma_write_array(ecc, EDMA_EMCR, idx, ch_bit);
/* Clear any SER */
edma_shadow0_write_array(ecc, SH_SECR, idx, ch_bit);
edma_shadow0_write_array(ecc, SH_EESR, idx, ch_bit);
dev_dbg(ecc->dev, "EER%d %08x\n", idx,
edma_shadow0_read_array(ecc, SH_EER, idx));
}
}
static void edma_stop(struct edma_chan *echan)
{
struct edma_cc *ecc = echan->ecc;
int channel = EDMA_CHAN_SLOT(echan->ch_num);
int idx = EDMA_REG_ARRAY_INDEX(channel);
int ch_bit = EDMA_CHANNEL_BIT(channel);
edma_shadow0_write_array(ecc, SH_EECR, idx, ch_bit);
edma_shadow0_write_array(ecc, SH_ECR, idx, ch_bit);
edma_shadow0_write_array(ecc, SH_SECR, idx, ch_bit);
edma_write_array(ecc, EDMA_EMCR, idx, ch_bit);
/* clear possibly pending completion interrupt */
edma_shadow0_write_array(ecc, SH_ICR, idx, ch_bit);
dev_dbg(ecc->dev, "EER%d %08x\n", idx,
edma_shadow0_read_array(ecc, SH_EER, idx));
/* REVISIT: consider guarding against inappropriate event
* chaining by overwriting with dummy_paramset.
*/
}
/*
* Temporarily disable EDMA hardware events on the specified channel,
* preventing them from triggering new transfers
*/
static void edma_pause(struct edma_chan *echan)
{
int channel = EDMA_CHAN_SLOT(echan->ch_num);
edma_shadow0_write_array(echan->ecc, SH_EECR,
EDMA_REG_ARRAY_INDEX(channel),
EDMA_CHANNEL_BIT(channel));
}
/* Re-enable EDMA hardware events on the specified channel. */
static void edma_resume(struct edma_chan *echan)
{
int channel = EDMA_CHAN_SLOT(echan->ch_num);
edma_shadow0_write_array(echan->ecc, SH_EESR,
EDMA_REG_ARRAY_INDEX(channel),
EDMA_CHANNEL_BIT(channel));
}
static void edma_trigger_channel(struct edma_chan *echan)
{
struct edma_cc *ecc = echan->ecc;
int channel = EDMA_CHAN_SLOT(echan->ch_num);
int idx = EDMA_REG_ARRAY_INDEX(channel);
int ch_bit = EDMA_CHANNEL_BIT(channel);
edma_shadow0_write_array(ecc, SH_ESR, idx, ch_bit);
dev_dbg(ecc->dev, "ESR%d %08x\n", idx,
edma_shadow0_read_array(ecc, SH_ESR, idx));
}
static void edma_clean_channel(struct edma_chan *echan)
{
struct edma_cc *ecc = echan->ecc;
int channel = EDMA_CHAN_SLOT(echan->ch_num);
int idx = EDMA_REG_ARRAY_INDEX(channel);
int ch_bit = EDMA_CHANNEL_BIT(channel);
dev_dbg(ecc->dev, "EMR%d %08x\n", idx,
edma_read_array(ecc, EDMA_EMR, idx));
edma_shadow0_write_array(ecc, SH_ECR, idx, ch_bit);
/* Clear the corresponding EMR bits */
edma_write_array(ecc, EDMA_EMCR, idx, ch_bit);
/* Clear any SER */
edma_shadow0_write_array(ecc, SH_SECR, idx, ch_bit);
edma_write(ecc, EDMA_CCERRCLR, BIT(16) | BIT(1) | BIT(0));
}
/* Move channel to a specific event queue */
static void edma_assign_channel_eventq(struct edma_chan *echan,
enum dma_event_q eventq_no)
{
struct edma_cc *ecc = echan->ecc;
int channel = EDMA_CHAN_SLOT(echan->ch_num);
int bit = (channel & 0x7) * 4;
/* default to low priority queue */
if (eventq_no == EVENTQ_DEFAULT)
eventq_no = ecc->default_queue;
if (eventq_no >= ecc->num_tc)
return;
eventq_no &= 7;
edma_modify_array(ecc, EDMA_DMAQNUM, (channel >> 3), ~(0x7 << bit),
eventq_no << bit);
}
static int edma_alloc_channel(struct edma_chan *echan,
enum dma_event_q eventq_no)
{
struct edma_cc *ecc = echan->ecc;
int channel = EDMA_CHAN_SLOT(echan->ch_num);
if (!test_bit(echan->ch_num, ecc->channels_mask)) {
dev_err(ecc->dev, "Channel%d is reserved, can not be used!\n",
echan->ch_num);
return -EINVAL;
}
/* ensure access through shadow region 0 */
edma_or_array2(ecc, EDMA_DRAE, 0, EDMA_REG_ARRAY_INDEX(channel),
EDMA_CHANNEL_BIT(channel));
/* ensure no events are pending */
edma_stop(echan);
edma_setup_interrupt(echan, true);
edma_assign_channel_eventq(echan, eventq_no);
return 0;
}
static void edma_free_channel(struct edma_chan *echan)
{
/* ensure no events are pending */
edma_stop(echan);
/* REVISIT should probably take out of shadow region 0 */
edma_setup_interrupt(echan, false);
}
static inline struct edma_cc *to_edma_cc(struct dma_device *d)
{
return container_of(d, struct edma_cc, dma_slave);
}
static inline struct edma_chan *to_edma_chan(struct dma_chan *c)
{
return container_of(c, struct edma_chan, vchan.chan);
}
static inline struct edma_desc *to_edma_desc(struct dma_async_tx_descriptor *tx)
{
return container_of(tx, struct edma_desc, vdesc.tx);
}
static void edma_desc_free(struct virt_dma_desc *vdesc)
{
kfree(container_of(vdesc, struct edma_desc, vdesc));
}
/* Dispatch a queued descriptor to the controller (caller holds lock) */
static void edma_execute(struct edma_chan *echan)
{
struct edma_cc *ecc = echan->ecc;
struct virt_dma_desc *vdesc;
struct edma_desc *edesc;
struct device *dev = echan->vchan.chan.device->dev;
int i, j, left, nslots;
if (!echan->edesc) {
/* Setup is needed for the first transfer */
vdesc = vchan_next_desc(&echan->vchan);
if (!vdesc)
return;
list_del(&vdesc->node);
echan->edesc = to_edma_desc(&vdesc->tx);
}
edesc = echan->edesc;
/* Find out how many left */
left = edesc->pset_nr - edesc->processed;
nslots = min(MAX_NR_SG, left);
edesc->sg_len = 0;
/* Write descriptor PaRAM set(s) */
for (i = 0; i < nslots; i++) {
j = i + edesc->processed;
edma_write_slot(ecc, echan->slot[i], &edesc->pset[j].param);
edesc->sg_len += edesc->pset[j].len;
dev_vdbg(dev,
"\n pset[%d]:\n"
" chnum\t%d\n"
" slot\t%d\n"
" opt\t%08x\n"
" src\t%08x\n"
" dst\t%08x\n"
" abcnt\t%08x\n"
" ccnt\t%08x\n"
" bidx\t%08x\n"
" cidx\t%08x\n"
" lkrld\t%08x\n",
j, echan->ch_num, echan->slot[i],
edesc->pset[j].param.opt,
edesc->pset[j].param.src,
edesc->pset[j].param.dst,
edesc->pset[j].param.a_b_cnt,
edesc->pset[j].param.ccnt,
edesc->pset[j].param.src_dst_bidx,
edesc->pset[j].param.src_dst_cidx,
edesc->pset[j].param.link_bcntrld);
/* Link to the previous slot if not the last set */
if (i != (nslots - 1))
edma_link(ecc, echan->slot[i], echan->slot[i + 1]);
}
edesc->processed += nslots;
/*
* If this is either the last set in a set of SG-list transactions
* then setup a link to the dummy slot, this results in all future
* events being absorbed and that's OK because we're done
*/
if (edesc->processed == edesc->pset_nr) {
if (edesc->cyclic)
edma_link(ecc, echan->slot[nslots - 1], echan->slot[1]);
else
edma_link(ecc, echan->slot[nslots - 1],
echan->ecc->dummy_slot);
}
dma: edma: Find missed events and issue them In an effort to move to using Scatter gather lists of any size with EDMA as discussed at [1] instead of placing limitations on the driver, we work through the limitations of the EDMAC hardware to find missed events and issue them. The sequence of events that require this are: For the scenario where MAX slots for an EDMA channel is 3: SG1 -> SG2 -> SG3 -> SG4 -> SG5 -> SG6 -> Null The above SG list will have to be DMA'd in 2 sets: (1) SG1 -> SG2 -> SG3 -> Null (2) SG4 -> SG5 -> SG6 -> Null After (1) is succesfully transferred, the events from the MMC controller donot stop coming and are missed by the time we have setup the transfer for (2). So here, we catch the events missed as an error condition and issue them manually. In the second part of the patch, we make handle the NULL slot cases: For crypto IP, we continue to receive events even continuously in NULL slot, the setup of the next set of SG elements happens after the error handler executes. This is results in some recursion problems. Due to this, we continously receive error interrupts when we manually trigger an event from the error handler. We fix this, by first detecting if the Channel is currently transferring from a NULL slot or not, that's where the edma_read_slot in the error callback from interrupt handler comes in. With this we can determine if the set up of the next SG list has completed, and we manually trigger only in this case. If the setup has _not_ completed, we are still in NULL so we just set a missed flag and allow the manual triggerring to happen in edma_execute which will be eventually called. This fixes the above mentioned race conditions seen with the crypto drivers. [1] http://marc.info/?l=linux-omap&m=137416733628831&w=2 Signed-off-by: Joel Fernandes <joelf@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2013-08-30 07:05:43 +08:00
if (echan->missed) {
/*
* This happens due to setup times between intermediate
* transfers in long SG lists which have to be broken up into
* transfers of MAX_NR_SG
*/
dev_dbg(dev, "missed event on channel %d\n", echan->ch_num);
edma_clean_channel(echan);
edma_stop(echan);
edma_start(echan);
edma_trigger_channel(echan);
dma: edma: Find missed events and issue them In an effort to move to using Scatter gather lists of any size with EDMA as discussed at [1] instead of placing limitations on the driver, we work through the limitations of the EDMAC hardware to find missed events and issue them. The sequence of events that require this are: For the scenario where MAX slots for an EDMA channel is 3: SG1 -> SG2 -> SG3 -> SG4 -> SG5 -> SG6 -> Null The above SG list will have to be DMA'd in 2 sets: (1) SG1 -> SG2 -> SG3 -> Null (2) SG4 -> SG5 -> SG6 -> Null After (1) is succesfully transferred, the events from the MMC controller donot stop coming and are missed by the time we have setup the transfer for (2). So here, we catch the events missed as an error condition and issue them manually. In the second part of the patch, we make handle the NULL slot cases: For crypto IP, we continue to receive events even continuously in NULL slot, the setup of the next set of SG elements happens after the error handler executes. This is results in some recursion problems. Due to this, we continously receive error interrupts when we manually trigger an event from the error handler. We fix this, by first detecting if the Channel is currently transferring from a NULL slot or not, that's where the edma_read_slot in the error callback from interrupt handler comes in. With this we can determine if the set up of the next SG list has completed, and we manually trigger only in this case. If the setup has _not_ completed, we are still in NULL so we just set a missed flag and allow the manual triggerring to happen in edma_execute which will be eventually called. This fixes the above mentioned race conditions seen with the crypto drivers. [1] http://marc.info/?l=linux-omap&m=137416733628831&w=2 Signed-off-by: Joel Fernandes <joelf@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2013-08-30 07:05:43 +08:00
echan->missed = 0;
} else if (edesc->processed <= MAX_NR_SG) {
dev_dbg(dev, "first transfer starting on channel %d\n",
echan->ch_num);
edma_start(echan);
} else {
dev_dbg(dev, "chan: %d: completed %d elements, resuming\n",
echan->ch_num, edesc->processed);
edma_resume(echan);
dma: edma: Find missed events and issue them In an effort to move to using Scatter gather lists of any size with EDMA as discussed at [1] instead of placing limitations on the driver, we work through the limitations of the EDMAC hardware to find missed events and issue them. The sequence of events that require this are: For the scenario where MAX slots for an EDMA channel is 3: SG1 -> SG2 -> SG3 -> SG4 -> SG5 -> SG6 -> Null The above SG list will have to be DMA'd in 2 sets: (1) SG1 -> SG2 -> SG3 -> Null (2) SG4 -> SG5 -> SG6 -> Null After (1) is succesfully transferred, the events from the MMC controller donot stop coming and are missed by the time we have setup the transfer for (2). So here, we catch the events missed as an error condition and issue them manually. In the second part of the patch, we make handle the NULL slot cases: For crypto IP, we continue to receive events even continuously in NULL slot, the setup of the next set of SG elements happens after the error handler executes. This is results in some recursion problems. Due to this, we continously receive error interrupts when we manually trigger an event from the error handler. We fix this, by first detecting if the Channel is currently transferring from a NULL slot or not, that's where the edma_read_slot in the error callback from interrupt handler comes in. With this we can determine if the set up of the next SG list has completed, and we manually trigger only in this case. If the setup has _not_ completed, we are still in NULL so we just set a missed flag and allow the manual triggerring to happen in edma_execute which will be eventually called. This fixes the above mentioned race conditions seen with the crypto drivers. [1] http://marc.info/?l=linux-omap&m=137416733628831&w=2 Signed-off-by: Joel Fernandes <joelf@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2013-08-30 07:05:43 +08:00
}
}
static int edma_terminate_all(struct dma_chan *chan)
{
struct edma_chan *echan = to_edma_chan(chan);
unsigned long flags;
LIST_HEAD(head);
spin_lock_irqsave(&echan->vchan.lock, flags);
/*
* Stop DMA activity: we assume the callback will not be called
* after edma_dma() returns (even if it does, it will see
* echan->edesc is NULL and exit.)
*/
if (echan->edesc) {
edma_stop(echan);
/* Move the cyclic channel back to default queue */
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
if (!echan->tc && echan->edesc->cyclic)
edma_assign_channel_eventq(echan, EVENTQ_DEFAULT);
vchan_terminate_vdesc(&echan->edesc->vdesc);
echan->edesc = NULL;
}
vchan_get_all_descriptors(&echan->vchan, &head);
spin_unlock_irqrestore(&echan->vchan.lock, flags);
vchan_dma_desc_free_list(&echan->vchan, &head);
return 0;
}
static void edma_synchronize(struct dma_chan *chan)
{
struct edma_chan *echan = to_edma_chan(chan);
vchan_synchronize(&echan->vchan);
}
static int edma_slave_config(struct dma_chan *chan,
struct dma_slave_config *cfg)
{
struct edma_chan *echan = to_edma_chan(chan);
if (cfg->src_addr_width == DMA_SLAVE_BUSWIDTH_8_BYTES ||
cfg->dst_addr_width == DMA_SLAVE_BUSWIDTH_8_BYTES)
return -EINVAL;
if (cfg->src_maxburst > chan->device->max_burst ||
cfg->dst_maxburst > chan->device->max_burst)
return -EINVAL;
memcpy(&echan->cfg, cfg, sizeof(echan->cfg));
return 0;
}
static int edma_dma_pause(struct dma_chan *chan)
{
struct edma_chan *echan = to_edma_chan(chan);
if (!echan->edesc)
return -EINVAL;
edma_pause(echan);
return 0;
}
static int edma_dma_resume(struct dma_chan *chan)
{
struct edma_chan *echan = to_edma_chan(chan);
edma_resume(echan);
return 0;
}
/*
* A PaRAM set configuration abstraction used by other modes
* @chan: Channel who's PaRAM set we're configuring
* @pset: PaRAM set to initialize and setup.
* @src_addr: Source address of the DMA
* @dst_addr: Destination address of the DMA
* @burst: In units of dev_width, how much to send
* @dev_width: How much is the dev_width
* @dma_length: Total length of the DMA transfer
* @direction: Direction of the transfer
*/
static int edma_config_pset(struct dma_chan *chan, struct edma_pset *epset,
dma_addr_t src_addr, dma_addr_t dst_addr, u32 burst,
unsigned int acnt, unsigned int dma_length,
enum dma_transfer_direction direction)
{
struct edma_chan *echan = to_edma_chan(chan);
struct device *dev = chan->device->dev;
struct edmacc_param *param = &epset->param;
int bcnt, ccnt, cidx;
int src_bidx, dst_bidx, src_cidx, dst_cidx;
int absync;
/* src/dst_maxburst == 0 is the same case as src/dst_maxburst == 1 */
if (!burst)
burst = 1;
/*
* If the maxburst is equal to the fifo width, use
* A-synced transfers. This allows for large contiguous
* buffer transfers using only one PaRAM set.
*/
if (burst == 1) {
/*
* For the A-sync case, bcnt and ccnt are the remainder
* and quotient respectively of the division of:
* (dma_length / acnt) by (SZ_64K -1). This is so
* that in case bcnt over flows, we have ccnt to use.
* Note: In A-sync tranfer only, bcntrld is used, but it
* only applies for sg_dma_len(sg) >= SZ_64K.
* In this case, the best way adopted is- bccnt for the
* first frame will be the remainder below. Then for
* every successive frame, bcnt will be SZ_64K-1. This
* is assured as bcntrld = 0xffff in end of function.
*/
absync = false;
ccnt = dma_length / acnt / (SZ_64K - 1);
bcnt = dma_length / acnt - ccnt * (SZ_64K - 1);
/*
* If bcnt is non-zero, we have a remainder and hence an
* extra frame to transfer, so increment ccnt.
*/
if (bcnt)
ccnt++;
else
bcnt = SZ_64K - 1;
cidx = acnt;
} else {
/*
* If maxburst is greater than the fifo address_width,
* use AB-synced transfers where A count is the fifo
* address_width and B count is the maxburst. In this
* case, we are limited to transfers of C count frames
* of (address_width * maxburst) where C count is limited
* to SZ_64K-1. This places an upper bound on the length
* of an SG segment that can be handled.
*/
absync = true;
bcnt = burst;
ccnt = dma_length / (acnt * bcnt);
if (ccnt > (SZ_64K - 1)) {
dev_err(dev, "Exceeded max SG segment size\n");
return -EINVAL;
}
cidx = acnt * bcnt;
}
epset->len = dma_length;
if (direction == DMA_MEM_TO_DEV) {
src_bidx = acnt;
src_cidx = cidx;
dst_bidx = 0;
dst_cidx = 0;
epset->addr = src_addr;
} else if (direction == DMA_DEV_TO_MEM) {
src_bidx = 0;
src_cidx = 0;
dst_bidx = acnt;
dst_cidx = cidx;
epset->addr = dst_addr;
} else if (direction == DMA_MEM_TO_MEM) {
src_bidx = acnt;
src_cidx = cidx;
dst_bidx = acnt;
dst_cidx = cidx;
epset->addr = src_addr;
} else {
dev_err(dev, "%s: direction not implemented yet\n", __func__);
return -EINVAL;
}
param->opt = EDMA_TCC(EDMA_CHAN_SLOT(echan->ch_num));
/* Configure A or AB synchronized transfers */
if (absync)
param->opt |= SYNCDIM;
param->src = src_addr;
param->dst = dst_addr;
param->src_dst_bidx = (dst_bidx << 16) | src_bidx;
param->src_dst_cidx = (dst_cidx << 16) | src_cidx;
param->a_b_cnt = bcnt << 16 | acnt;
param->ccnt = ccnt;
/*
* Only time when (bcntrld) auto reload is required is for
* A-sync case, and in this case, a requirement of reload value
* of SZ_64K-1 only is assured. 'link' is initially set to NULL
* and then later will be populated by edma_execute.
*/
param->link_bcntrld = 0xffffffff;
return absync;
}
static struct dma_async_tx_descriptor *edma_prep_slave_sg(
struct dma_chan *chan, struct scatterlist *sgl,
unsigned int sg_len, enum dma_transfer_direction direction,
unsigned long tx_flags, void *context)
{
struct edma_chan *echan = to_edma_chan(chan);
struct device *dev = chan->device->dev;
struct edma_desc *edesc;
dma_addr_t src_addr = 0, dst_addr = 0;
enum dma_slave_buswidth dev_width;
u32 burst;
struct scatterlist *sg;
int i, nslots, ret;
if (unlikely(!echan || !sgl || !sg_len))
return NULL;
if (direction == DMA_DEV_TO_MEM) {
src_addr = echan->cfg.src_addr;
dev_width = echan->cfg.src_addr_width;
burst = echan->cfg.src_maxburst;
} else if (direction == DMA_MEM_TO_DEV) {
dst_addr = echan->cfg.dst_addr;
dev_width = echan->cfg.dst_addr_width;
burst = echan->cfg.dst_maxburst;
} else {
dev_err(dev, "%s: bad direction: %d\n", __func__, direction);
return NULL;
}
if (dev_width == DMA_SLAVE_BUSWIDTH_UNDEFINED) {
dev_err(dev, "%s: Undefined slave buswidth\n", __func__);
return NULL;
}
treewide: Use struct_size() for kmalloc()-family One of the more common cases of allocation size calculations is finding the size of a structure that has a zero-sized array at the end, along with memory for some number of elements for that array. For example: struct foo { int stuff; void *entry[]; }; instance = kmalloc(sizeof(struct foo) + sizeof(void *) * count, GFP_KERNEL); Instead of leaving these open-coded and prone to type mistakes, we can now use the new struct_size() helper: instance = kmalloc(struct_size(instance, entry, count), GFP_KERNEL); This patch makes the changes for kmalloc()-family (and kvmalloc()-family) uses. It was done via automatic conversion with manual review for the "CHECKME" non-standard cases noted below, using the following Coccinelle script: // pkey_cache = kmalloc(sizeof *pkey_cache + tprops->pkey_tbl_len * // sizeof *pkey_cache->table, GFP_KERNEL); @@ identifier alloc =~ "kmalloc|kzalloc|kvmalloc|kvzalloc"; expression GFP; identifier VAR, ELEMENT; expression COUNT; @@ - alloc(sizeof(*VAR) + COUNT * sizeof(*VAR->ELEMENT), GFP) + alloc(struct_size(VAR, ELEMENT, COUNT), GFP) // mr = kzalloc(sizeof(*mr) + m * sizeof(mr->map[0]), GFP_KERNEL); @@ identifier alloc =~ "kmalloc|kzalloc|kvmalloc|kvzalloc"; expression GFP; identifier VAR, ELEMENT; expression COUNT; @@ - alloc(sizeof(*VAR) + COUNT * sizeof(VAR->ELEMENT[0]), GFP) + alloc(struct_size(VAR, ELEMENT, COUNT), GFP) // Same pattern, but can't trivially locate the trailing element name, // or variable name. @@ identifier alloc =~ "kmalloc|kzalloc|kvmalloc|kvzalloc"; expression GFP; expression SOMETHING, COUNT, ELEMENT; @@ - alloc(sizeof(SOMETHING) + COUNT * sizeof(ELEMENT), GFP) + alloc(CHECKME_struct_size(&SOMETHING, ELEMENT, COUNT), GFP) Signed-off-by: Kees Cook <keescook@chromium.org>
2018-05-09 04:45:50 +08:00
edesc = kzalloc(struct_size(edesc, pset, sg_len), GFP_ATOMIC);
if (!edesc)
return NULL;
edesc->pset_nr = sg_len;
edesc->residue = 0;
edesc->direction = direction;
edesc->echan = echan;
/* Allocate a PaRAM slot, if needed */
nslots = min_t(unsigned, MAX_NR_SG, sg_len);
for (i = 0; i < nslots; i++) {
if (echan->slot[i] < 0) {
echan->slot[i] =
edma_alloc_slot(echan->ecc, EDMA_SLOT_ANY);
if (echan->slot[i] < 0) {
kfree(edesc);
dev_err(dev, "%s: Failed to allocate slot\n",
__func__);
return NULL;
}
}
}
/* Configure PaRAM sets for each SG */
for_each_sg(sgl, sg, sg_len, i) {
/* Get address for each SG */
if (direction == DMA_DEV_TO_MEM)
dst_addr = sg_dma_address(sg);
else
src_addr = sg_dma_address(sg);
ret = edma_config_pset(chan, &edesc->pset[i], src_addr,
dst_addr, burst, dev_width,
sg_dma_len(sg), direction);
if (ret < 0) {
kfree(edesc);
return NULL;
}
edesc->absync = ret;
edesc->residue += sg_dma_len(sg);
if (i == sg_len - 1)
/* Enable completion interrupt */
edesc->pset[i].param.opt |= TCINTEN;
else if (!((i+1) % MAX_NR_SG))
/*
* Enable early completion interrupt for the
* intermediateset. In this case the driver will be
* notified when the paRAM set is submitted to TC. This
* will allow more time to set up the next set of slots.
*/
edesc->pset[i].param.opt |= (TCINTEN | TCCMODE);
}
edesc->residue_stat = edesc->residue;
return vchan_tx_prep(&echan->vchan, &edesc->vdesc, tx_flags);
}
static struct dma_async_tx_descriptor *edma_prep_dma_memcpy(
struct dma_chan *chan, dma_addr_t dest, dma_addr_t src,
size_t len, unsigned long tx_flags)
{
int ret, nslots;
struct edma_desc *edesc;
struct device *dev = chan->device->dev;
struct edma_chan *echan = to_edma_chan(chan);
unsigned int width, pset_len, array_size;
if (unlikely(!echan || !len))
return NULL;
/* Align the array size (acnt block) with the transfer properties */
switch (__ffs((src | dest | len))) {
case 0:
array_size = SZ_32K - 1;
break;
case 1:
array_size = SZ_32K - 2;
break;
default:
array_size = SZ_32K - 4;
break;
}
if (len < SZ_64K) {
/*
* Transfer size less than 64K can be handled with one paRAM
* slot and with one burst.
* ACNT = length
*/
width = len;
pset_len = len;
nslots = 1;
} else {
/*
* Transfer size bigger than 64K will be handled with maximum of
* two paRAM slots.
* slot1: (full_length / 32767) times 32767 bytes bursts.
* ACNT = 32767, length1: (full_length / 32767) * 32767
* slot2: the remaining amount of data after slot1.
* ACNT = full_length - length1, length2 = ACNT
*
* When the full_length is multibple of 32767 one slot can be
* used to complete the transfer.
*/
width = array_size;
pset_len = rounddown(len, width);
/* One slot is enough for lengths multiple of (SZ_32K -1) */
if (unlikely(pset_len == len))
nslots = 1;
else
nslots = 2;
}
treewide: Use struct_size() for kmalloc()-family One of the more common cases of allocation size calculations is finding the size of a structure that has a zero-sized array at the end, along with memory for some number of elements for that array. For example: struct foo { int stuff; void *entry[]; }; instance = kmalloc(sizeof(struct foo) + sizeof(void *) * count, GFP_KERNEL); Instead of leaving these open-coded and prone to type mistakes, we can now use the new struct_size() helper: instance = kmalloc(struct_size(instance, entry, count), GFP_KERNEL); This patch makes the changes for kmalloc()-family (and kvmalloc()-family) uses. It was done via automatic conversion with manual review for the "CHECKME" non-standard cases noted below, using the following Coccinelle script: // pkey_cache = kmalloc(sizeof *pkey_cache + tprops->pkey_tbl_len * // sizeof *pkey_cache->table, GFP_KERNEL); @@ identifier alloc =~ "kmalloc|kzalloc|kvmalloc|kvzalloc"; expression GFP; identifier VAR, ELEMENT; expression COUNT; @@ - alloc(sizeof(*VAR) + COUNT * sizeof(*VAR->ELEMENT), GFP) + alloc(struct_size(VAR, ELEMENT, COUNT), GFP) // mr = kzalloc(sizeof(*mr) + m * sizeof(mr->map[0]), GFP_KERNEL); @@ identifier alloc =~ "kmalloc|kzalloc|kvmalloc|kvzalloc"; expression GFP; identifier VAR, ELEMENT; expression COUNT; @@ - alloc(sizeof(*VAR) + COUNT * sizeof(VAR->ELEMENT[0]), GFP) + alloc(struct_size(VAR, ELEMENT, COUNT), GFP) // Same pattern, but can't trivially locate the trailing element name, // or variable name. @@ identifier alloc =~ "kmalloc|kzalloc|kvmalloc|kvzalloc"; expression GFP; expression SOMETHING, COUNT, ELEMENT; @@ - alloc(sizeof(SOMETHING) + COUNT * sizeof(ELEMENT), GFP) + alloc(CHECKME_struct_size(&SOMETHING, ELEMENT, COUNT), GFP) Signed-off-by: Kees Cook <keescook@chromium.org>
2018-05-09 04:45:50 +08:00
edesc = kzalloc(struct_size(edesc, pset, nslots), GFP_ATOMIC);
if (!edesc)
return NULL;
edesc->pset_nr = nslots;
edesc->residue = edesc->residue_stat = len;
edesc->direction = DMA_MEM_TO_MEM;
edesc->echan = echan;
ret = edma_config_pset(chan, &edesc->pset[0], src, dest, 1,
width, pset_len, DMA_MEM_TO_MEM);
if (ret < 0) {
kfree(edesc);
return NULL;
}
edesc->absync = ret;
edesc->pset[0].param.opt |= ITCCHEN;
if (nslots == 1) {
/* Enable transfer complete interrupt if requested */
if (tx_flags & DMA_PREP_INTERRUPT)
edesc->pset[0].param.opt |= TCINTEN;
} else {
/* Enable transfer complete chaining for the first slot */
edesc->pset[0].param.opt |= TCCHEN;
if (echan->slot[1] < 0) {
echan->slot[1] = edma_alloc_slot(echan->ecc,
EDMA_SLOT_ANY);
if (echan->slot[1] < 0) {
kfree(edesc);
dev_err(dev, "%s: Failed to allocate slot\n",
__func__);
return NULL;
}
}
dest += pset_len;
src += pset_len;
pset_len = width = len % array_size;
ret = edma_config_pset(chan, &edesc->pset[1], src, dest, 1,
width, pset_len, DMA_MEM_TO_MEM);
if (ret < 0) {
kfree(edesc);
return NULL;
}
edesc->pset[1].param.opt |= ITCCHEN;
/* Enable transfer complete interrupt if requested */
if (tx_flags & DMA_PREP_INTERRUPT)
edesc->pset[1].param.opt |= TCINTEN;
}
if (!(tx_flags & DMA_PREP_INTERRUPT))
edesc->polled = true;
return vchan_tx_prep(&echan->vchan, &edesc->vdesc, tx_flags);
}
static struct dma_async_tx_descriptor *
edma_prep_dma_interleaved(struct dma_chan *chan,
struct dma_interleaved_template *xt,
unsigned long tx_flags)
{
struct device *dev = chan->device->dev;
struct edma_chan *echan = to_edma_chan(chan);
struct edmacc_param *param;
struct edma_desc *edesc;
size_t src_icg, dst_icg;
int src_bidx, dst_bidx;
/* Slave mode is not supported */
if (is_slave_direction(xt->dir))
return NULL;
if (xt->frame_size != 1 || xt->numf == 0)
return NULL;
if (xt->sgl[0].size > SZ_64K || xt->numf > SZ_64K)
return NULL;
src_icg = dmaengine_get_src_icg(xt, &xt->sgl[0]);
if (src_icg) {
src_bidx = src_icg + xt->sgl[0].size;
} else if (xt->src_inc) {
src_bidx = xt->sgl[0].size;
} else {
dev_err(dev, "%s: SRC constant addressing is not supported\n",
__func__);
return NULL;
}
dst_icg = dmaengine_get_dst_icg(xt, &xt->sgl[0]);
if (dst_icg) {
dst_bidx = dst_icg + xt->sgl[0].size;
} else if (xt->dst_inc) {
dst_bidx = xt->sgl[0].size;
} else {
dev_err(dev, "%s: DST constant addressing is not supported\n",
__func__);
return NULL;
}
if (src_bidx > SZ_64K || dst_bidx > SZ_64K)
return NULL;
edesc = kzalloc(struct_size(edesc, pset, 1), GFP_ATOMIC);
if (!edesc)
return NULL;
edesc->direction = DMA_MEM_TO_MEM;
edesc->echan = echan;
edesc->pset_nr = 1;
param = &edesc->pset[0].param;
param->src = xt->src_start;
param->dst = xt->dst_start;
param->a_b_cnt = xt->numf << 16 | xt->sgl[0].size;
param->ccnt = 1;
param->src_dst_bidx = (dst_bidx << 16) | src_bidx;
param->src_dst_cidx = 0;
param->opt = EDMA_TCC(EDMA_CHAN_SLOT(echan->ch_num));
param->opt |= ITCCHEN;
/* Enable transfer complete interrupt if requested */
if (tx_flags & DMA_PREP_INTERRUPT)
param->opt |= TCINTEN;
else
edesc->polled = true;
return vchan_tx_prep(&echan->vchan, &edesc->vdesc, tx_flags);
}
static struct dma_async_tx_descriptor *edma_prep_dma_cyclic(
struct dma_chan *chan, dma_addr_t buf_addr, size_t buf_len,
size_t period_len, enum dma_transfer_direction direction,
unsigned long tx_flags)
{
struct edma_chan *echan = to_edma_chan(chan);
struct device *dev = chan->device->dev;
struct edma_desc *edesc;
dma_addr_t src_addr, dst_addr;
enum dma_slave_buswidth dev_width;
bool use_intermediate = false;
u32 burst;
int i, ret, nslots;
if (unlikely(!echan || !buf_len || !period_len))
return NULL;
if (direction == DMA_DEV_TO_MEM) {
src_addr = echan->cfg.src_addr;
dst_addr = buf_addr;
dev_width = echan->cfg.src_addr_width;
burst = echan->cfg.src_maxburst;
} else if (direction == DMA_MEM_TO_DEV) {
src_addr = buf_addr;
dst_addr = echan->cfg.dst_addr;
dev_width = echan->cfg.dst_addr_width;
burst = echan->cfg.dst_maxburst;
} else {
dev_err(dev, "%s: bad direction: %d\n", __func__, direction);
return NULL;
}
if (dev_width == DMA_SLAVE_BUSWIDTH_UNDEFINED) {
dev_err(dev, "%s: Undefined slave buswidth\n", __func__);
return NULL;
}
if (unlikely(buf_len % period_len)) {
dev_err(dev, "Period should be multiple of Buffer length\n");
return NULL;
}
nslots = (buf_len / period_len) + 1;
/*
* Cyclic DMA users such as audio cannot tolerate delays introduced
* by cases where the number of periods is more than the maximum
* number of SGs the EDMA driver can handle at a time. For DMA types
* such as Slave SGs, such delays are tolerable and synchronized,
* but the synchronization is difficult to achieve with Cyclic and
* cannot be guaranteed, so we error out early.
*/
if (nslots > MAX_NR_SG) {
/*
* If the burst and period sizes are the same, we can put
* the full buffer into a single period and activate
* intermediate interrupts. This will produce interrupts
* after each burst, which is also after each desired period.
*/
if (burst == period_len) {
period_len = buf_len;
nslots = 2;
use_intermediate = true;
} else {
return NULL;
}
}
treewide: Use struct_size() for kmalloc()-family One of the more common cases of allocation size calculations is finding the size of a structure that has a zero-sized array at the end, along with memory for some number of elements for that array. For example: struct foo { int stuff; void *entry[]; }; instance = kmalloc(sizeof(struct foo) + sizeof(void *) * count, GFP_KERNEL); Instead of leaving these open-coded and prone to type mistakes, we can now use the new struct_size() helper: instance = kmalloc(struct_size(instance, entry, count), GFP_KERNEL); This patch makes the changes for kmalloc()-family (and kvmalloc()-family) uses. It was done via automatic conversion with manual review for the "CHECKME" non-standard cases noted below, using the following Coccinelle script: // pkey_cache = kmalloc(sizeof *pkey_cache + tprops->pkey_tbl_len * // sizeof *pkey_cache->table, GFP_KERNEL); @@ identifier alloc =~ "kmalloc|kzalloc|kvmalloc|kvzalloc"; expression GFP; identifier VAR, ELEMENT; expression COUNT; @@ - alloc(sizeof(*VAR) + COUNT * sizeof(*VAR->ELEMENT), GFP) + alloc(struct_size(VAR, ELEMENT, COUNT), GFP) // mr = kzalloc(sizeof(*mr) + m * sizeof(mr->map[0]), GFP_KERNEL); @@ identifier alloc =~ "kmalloc|kzalloc|kvmalloc|kvzalloc"; expression GFP; identifier VAR, ELEMENT; expression COUNT; @@ - alloc(sizeof(*VAR) + COUNT * sizeof(VAR->ELEMENT[0]), GFP) + alloc(struct_size(VAR, ELEMENT, COUNT), GFP) // Same pattern, but can't trivially locate the trailing element name, // or variable name. @@ identifier alloc =~ "kmalloc|kzalloc|kvmalloc|kvzalloc"; expression GFP; expression SOMETHING, COUNT, ELEMENT; @@ - alloc(sizeof(SOMETHING) + COUNT * sizeof(ELEMENT), GFP) + alloc(CHECKME_struct_size(&SOMETHING, ELEMENT, COUNT), GFP) Signed-off-by: Kees Cook <keescook@chromium.org>
2018-05-09 04:45:50 +08:00
edesc = kzalloc(struct_size(edesc, pset, nslots), GFP_ATOMIC);
if (!edesc)
return NULL;
edesc->cyclic = 1;
edesc->pset_nr = nslots;
edesc->residue = edesc->residue_stat = buf_len;
edesc->direction = direction;
edesc->echan = echan;
dev_dbg(dev, "%s: channel=%d nslots=%d period_len=%zu buf_len=%zu\n",
__func__, echan->ch_num, nslots, period_len, buf_len);
for (i = 0; i < nslots; i++) {
/* Allocate a PaRAM slot, if needed */
if (echan->slot[i] < 0) {
echan->slot[i] =
edma_alloc_slot(echan->ecc, EDMA_SLOT_ANY);
if (echan->slot[i] < 0) {
kfree(edesc);
dev_err(dev, "%s: Failed to allocate slot\n",
__func__);
return NULL;
}
}
if (i == nslots - 1) {
memcpy(&edesc->pset[i], &edesc->pset[0],
sizeof(edesc->pset[0]));
break;
}
ret = edma_config_pset(chan, &edesc->pset[i], src_addr,
dst_addr, burst, dev_width, period_len,
direction);
if (ret < 0) {
kfree(edesc);
return NULL;
}
if (direction == DMA_DEV_TO_MEM)
dst_addr += period_len;
else
src_addr += period_len;
dev_vdbg(dev, "%s: Configure period %d of buf:\n", __func__, i);
dev_vdbg(dev,
"\n pset[%d]:\n"
" chnum\t%d\n"
" slot\t%d\n"
" opt\t%08x\n"
" src\t%08x\n"
" dst\t%08x\n"
" abcnt\t%08x\n"
" ccnt\t%08x\n"
" bidx\t%08x\n"
" cidx\t%08x\n"
" lkrld\t%08x\n",
i, echan->ch_num, echan->slot[i],
edesc->pset[i].param.opt,
edesc->pset[i].param.src,
edesc->pset[i].param.dst,
edesc->pset[i].param.a_b_cnt,
edesc->pset[i].param.ccnt,
edesc->pset[i].param.src_dst_bidx,
edesc->pset[i].param.src_dst_cidx,
edesc->pset[i].param.link_bcntrld);
edesc->absync = ret;
/*
* Enable period interrupt only if it is requested
*/
if (tx_flags & DMA_PREP_INTERRUPT) {
edesc->pset[i].param.opt |= TCINTEN;
/* Also enable intermediate interrupts if necessary */
if (use_intermediate)
edesc->pset[i].param.opt |= ITCINTEN;
}
}
/* Place the cyclic channel to highest priority queue */
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
if (!echan->tc)
edma_assign_channel_eventq(echan, EVENTQ_0);
return vchan_tx_prep(&echan->vchan, &edesc->vdesc, tx_flags);
}
static void edma_completion_handler(struct edma_chan *echan)
{
struct device *dev = echan->vchan.chan.device->dev;
struct edma_desc *edesc;
spin_lock(&echan->vchan.lock);
edesc = echan->edesc;
if (edesc) {
if (edesc->cyclic) {
vchan_cyclic_callback(&edesc->vdesc);
spin_unlock(&echan->vchan.lock);
return;
} else if (edesc->processed == edesc->pset_nr) {
edesc->residue = 0;
edma_stop(echan);
vchan_cookie_complete(&edesc->vdesc);
echan->edesc = NULL;
dev_dbg(dev, "Transfer completed on channel %d\n",
echan->ch_num);
} else {
dev_dbg(dev, "Sub transfer completed on channel %d\n",
echan->ch_num);
edma_pause(echan);
/* Update statistics for tx_status */
edesc->residue -= edesc->sg_len;
edesc->residue_stat = edesc->residue;
edesc->processed_stat = edesc->processed;
}
edma_execute(echan);
}
spin_unlock(&echan->vchan.lock);
}
/* eDMA interrupt handler */
static irqreturn_t dma_irq_handler(int irq, void *data)
{
struct edma_cc *ecc = data;
int ctlr;
u32 sh_ier;
u32 sh_ipr;
u32 bank;
ctlr = ecc->id;
if (ctlr < 0)
return IRQ_NONE;
dev_vdbg(ecc->dev, "dma_irq_handler\n");
sh_ipr = edma_shadow0_read_array(ecc, SH_IPR, 0);
if (!sh_ipr) {
sh_ipr = edma_shadow0_read_array(ecc, SH_IPR, 1);
if (!sh_ipr)
return IRQ_NONE;
sh_ier = edma_shadow0_read_array(ecc, SH_IER, 1);
bank = 1;
} else {
sh_ier = edma_shadow0_read_array(ecc, SH_IER, 0);
bank = 0;
}
do {
u32 slot;
u32 channel;
slot = __ffs(sh_ipr);
sh_ipr &= ~(BIT(slot));
if (sh_ier & BIT(slot)) {
channel = (bank << 5) | slot;
/* Clear the corresponding IPR bits */
edma_shadow0_write_array(ecc, SH_ICR, bank, BIT(slot));
edma_completion_handler(&ecc->slave_chans[channel]);
}
} while (sh_ipr);
edma_shadow0_write(ecc, SH_IEVAL, 1);
return IRQ_HANDLED;
}
static void edma_error_handler(struct edma_chan *echan)
{
struct edma_cc *ecc = echan->ecc;
struct device *dev = echan->vchan.chan.device->dev;
struct edmacc_param p;
int err;
if (!echan->edesc)
return;
spin_lock(&echan->vchan.lock);
dma: edma: Find missed events and issue them In an effort to move to using Scatter gather lists of any size with EDMA as discussed at [1] instead of placing limitations on the driver, we work through the limitations of the EDMAC hardware to find missed events and issue them. The sequence of events that require this are: For the scenario where MAX slots for an EDMA channel is 3: SG1 -> SG2 -> SG3 -> SG4 -> SG5 -> SG6 -> Null The above SG list will have to be DMA'd in 2 sets: (1) SG1 -> SG2 -> SG3 -> Null (2) SG4 -> SG5 -> SG6 -> Null After (1) is succesfully transferred, the events from the MMC controller donot stop coming and are missed by the time we have setup the transfer for (2). So here, we catch the events missed as an error condition and issue them manually. In the second part of the patch, we make handle the NULL slot cases: For crypto IP, we continue to receive events even continuously in NULL slot, the setup of the next set of SG elements happens after the error handler executes. This is results in some recursion problems. Due to this, we continously receive error interrupts when we manually trigger an event from the error handler. We fix this, by first detecting if the Channel is currently transferring from a NULL slot or not, that's where the edma_read_slot in the error callback from interrupt handler comes in. With this we can determine if the set up of the next SG list has completed, and we manually trigger only in this case. If the setup has _not_ completed, we are still in NULL so we just set a missed flag and allow the manual triggerring to happen in edma_execute which will be eventually called. This fixes the above mentioned race conditions seen with the crypto drivers. [1] http://marc.info/?l=linux-omap&m=137416733628831&w=2 Signed-off-by: Joel Fernandes <joelf@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2013-08-30 07:05:43 +08:00
err = edma_read_slot(ecc, echan->slot[0], &p);
/*
* Issue later based on missed flag which will be sure
* to happen as:
* (1) we finished transmitting an intermediate slot and
* edma_execute is coming up.
* (2) or we finished current transfer and issue will
* call edma_execute.
*
* Important note: issuing can be dangerous here and
* lead to some nasty recursion when we are in a NULL
* slot. So we avoid doing so and set the missed flag.
*/
if (err || (p.a_b_cnt == 0 && p.ccnt == 0)) {
dev_dbg(dev, "Error on null slot, setting miss\n");
echan->missed = 1;
} else {
dma: edma: Find missed events and issue them In an effort to move to using Scatter gather lists of any size with EDMA as discussed at [1] instead of placing limitations on the driver, we work through the limitations of the EDMAC hardware to find missed events and issue them. The sequence of events that require this are: For the scenario where MAX slots for an EDMA channel is 3: SG1 -> SG2 -> SG3 -> SG4 -> SG5 -> SG6 -> Null The above SG list will have to be DMA'd in 2 sets: (1) SG1 -> SG2 -> SG3 -> Null (2) SG4 -> SG5 -> SG6 -> Null After (1) is succesfully transferred, the events from the MMC controller donot stop coming and are missed by the time we have setup the transfer for (2). So here, we catch the events missed as an error condition and issue them manually. In the second part of the patch, we make handle the NULL slot cases: For crypto IP, we continue to receive events even continuously in NULL slot, the setup of the next set of SG elements happens after the error handler executes. This is results in some recursion problems. Due to this, we continously receive error interrupts when we manually trigger an event from the error handler. We fix this, by first detecting if the Channel is currently transferring from a NULL slot or not, that's where the edma_read_slot in the error callback from interrupt handler comes in. With this we can determine if the set up of the next SG list has completed, and we manually trigger only in this case. If the setup has _not_ completed, we are still in NULL so we just set a missed flag and allow the manual triggerring to happen in edma_execute which will be eventually called. This fixes the above mentioned race conditions seen with the crypto drivers. [1] http://marc.info/?l=linux-omap&m=137416733628831&w=2 Signed-off-by: Joel Fernandes <joelf@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2013-08-30 07:05:43 +08:00
/*
* The slot is already programmed but the event got
* missed, so its safe to issue it here.
dma: edma: Find missed events and issue them In an effort to move to using Scatter gather lists of any size with EDMA as discussed at [1] instead of placing limitations on the driver, we work through the limitations of the EDMAC hardware to find missed events and issue them. The sequence of events that require this are: For the scenario where MAX slots for an EDMA channel is 3: SG1 -> SG2 -> SG3 -> SG4 -> SG5 -> SG6 -> Null The above SG list will have to be DMA'd in 2 sets: (1) SG1 -> SG2 -> SG3 -> Null (2) SG4 -> SG5 -> SG6 -> Null After (1) is succesfully transferred, the events from the MMC controller donot stop coming and are missed by the time we have setup the transfer for (2). So here, we catch the events missed as an error condition and issue them manually. In the second part of the patch, we make handle the NULL slot cases: For crypto IP, we continue to receive events even continuously in NULL slot, the setup of the next set of SG elements happens after the error handler executes. This is results in some recursion problems. Due to this, we continously receive error interrupts when we manually trigger an event from the error handler. We fix this, by first detecting if the Channel is currently transferring from a NULL slot or not, that's where the edma_read_slot in the error callback from interrupt handler comes in. With this we can determine if the set up of the next SG list has completed, and we manually trigger only in this case. If the setup has _not_ completed, we are still in NULL so we just set a missed flag and allow the manual triggerring to happen in edma_execute which will be eventually called. This fixes the above mentioned race conditions seen with the crypto drivers. [1] http://marc.info/?l=linux-omap&m=137416733628831&w=2 Signed-off-by: Joel Fernandes <joelf@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2013-08-30 07:05:43 +08:00
*/
dev_dbg(dev, "Missed event, TRIGGERING\n");
edma_clean_channel(echan);
edma_stop(echan);
edma_start(echan);
edma_trigger_channel(echan);
}
spin_unlock(&echan->vchan.lock);
}
static inline bool edma_error_pending(struct edma_cc *ecc)
{
if (edma_read_array(ecc, EDMA_EMR, 0) ||
edma_read_array(ecc, EDMA_EMR, 1) ||
edma_read(ecc, EDMA_QEMR) || edma_read(ecc, EDMA_CCERR))
return true;
return false;
}
/* eDMA error interrupt handler */
static irqreturn_t dma_ccerr_handler(int irq, void *data)
{
struct edma_cc *ecc = data;
int i, j;
int ctlr;
unsigned int cnt = 0;
unsigned int val;
ctlr = ecc->id;
if (ctlr < 0)
return IRQ_NONE;
dev_vdbg(ecc->dev, "dma_ccerr_handler\n");
if (!edma_error_pending(ecc)) {
/*
* The registers indicate no pending error event but the irq
* handler has been called.
* Ask eDMA to re-evaluate the error registers.
*/
dev_err(ecc->dev, "%s: Error interrupt without error event!\n",
__func__);
edma_write(ecc, EDMA_EEVAL, 1);
return IRQ_NONE;
}
while (1) {
/* Event missed register(s) */
for (j = 0; j < 2; j++) {
unsigned long emr;
val = edma_read_array(ecc, EDMA_EMR, j);
if (!val)
continue;
dev_dbg(ecc->dev, "EMR%d 0x%08x\n", j, val);
emr = val;
for (i = find_next_bit(&emr, 32, 0); i < 32;
i = find_next_bit(&emr, 32, i + 1)) {
int k = (j << 5) + i;
/* Clear the corresponding EMR bits */
edma_write_array(ecc, EDMA_EMCR, j, BIT(i));
/* Clear any SER */
edma_shadow0_write_array(ecc, SH_SECR, j,
BIT(i));
edma_error_handler(&ecc->slave_chans[k]);
}
dma: edma: Find missed events and issue them In an effort to move to using Scatter gather lists of any size with EDMA as discussed at [1] instead of placing limitations on the driver, we work through the limitations of the EDMAC hardware to find missed events and issue them. The sequence of events that require this are: For the scenario where MAX slots for an EDMA channel is 3: SG1 -> SG2 -> SG3 -> SG4 -> SG5 -> SG6 -> Null The above SG list will have to be DMA'd in 2 sets: (1) SG1 -> SG2 -> SG3 -> Null (2) SG4 -> SG5 -> SG6 -> Null After (1) is succesfully transferred, the events from the MMC controller donot stop coming and are missed by the time we have setup the transfer for (2). So here, we catch the events missed as an error condition and issue them manually. In the second part of the patch, we make handle the NULL slot cases: For crypto IP, we continue to receive events even continuously in NULL slot, the setup of the next set of SG elements happens after the error handler executes. This is results in some recursion problems. Due to this, we continously receive error interrupts when we manually trigger an event from the error handler. We fix this, by first detecting if the Channel is currently transferring from a NULL slot or not, that's where the edma_read_slot in the error callback from interrupt handler comes in. With this we can determine if the set up of the next SG list has completed, and we manually trigger only in this case. If the setup has _not_ completed, we are still in NULL so we just set a missed flag and allow the manual triggerring to happen in edma_execute which will be eventually called. This fixes the above mentioned race conditions seen with the crypto drivers. [1] http://marc.info/?l=linux-omap&m=137416733628831&w=2 Signed-off-by: Joel Fernandes <joelf@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2013-08-30 07:05:43 +08:00
}
val = edma_read(ecc, EDMA_QEMR);
if (val) {
dev_dbg(ecc->dev, "QEMR 0x%02x\n", val);
/* Not reported, just clear the interrupt reason. */
edma_write(ecc, EDMA_QEMCR, val);
edma_shadow0_write(ecc, SH_QSECR, val);
}
val = edma_read(ecc, EDMA_CCERR);
if (val) {
dev_warn(ecc->dev, "CCERR 0x%08x\n", val);
/* Not reported, just clear the interrupt reason. */
edma_write(ecc, EDMA_CCERRCLR, val);
}
if (!edma_error_pending(ecc))
break;
cnt++;
if (cnt > 10)
break;
}
edma_write(ecc, EDMA_EEVAL, 1);
return IRQ_HANDLED;
}
/* Alloc channel resources */
static int edma_alloc_chan_resources(struct dma_chan *chan)
{
struct edma_chan *echan = to_edma_chan(chan);
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
struct edma_cc *ecc = echan->ecc;
struct device *dev = ecc->dev;
enum dma_event_q eventq_no = EVENTQ_DEFAULT;
int ret;
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
if (echan->tc) {
eventq_no = echan->tc->id;
} else if (ecc->tc_list) {
/* memcpy channel */
echan->tc = &ecc->tc_list[ecc->info->default_queue];
eventq_no = echan->tc->id;
}
ret = edma_alloc_channel(echan, eventq_no);
if (ret)
return ret;
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
echan->slot[0] = edma_alloc_slot(ecc, echan->ch_num);
if (echan->slot[0] < 0) {
dev_err(dev, "Entry slot allocation failed for channel %u\n",
EDMA_CHAN_SLOT(echan->ch_num));
ret = echan->slot[0];
goto err_slot;
}
/* Set up channel -> slot mapping for the entry slot */
edma_set_chmap(echan, echan->slot[0]);
echan->alloced = true;
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
dev_dbg(dev, "Got eDMA channel %d for virt channel %d (%s trigger)\n",
EDMA_CHAN_SLOT(echan->ch_num), chan->chan_id,
echan->hw_triggered ? "HW" : "SW");
return 0;
err_slot:
edma_free_channel(echan);
return ret;
}
/* Free channel resources */
static void edma_free_chan_resources(struct dma_chan *chan)
{
struct edma_chan *echan = to_edma_chan(chan);
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
struct device *dev = echan->ecc->dev;
int i;
/* Terminate transfers */
edma_stop(echan);
vchan_free_chan_resources(&echan->vchan);
/* Free EDMA PaRAM slots */
for (i = 0; i < EDMA_MAX_SLOTS; i++) {
if (echan->slot[i] >= 0) {
edma_free_slot(echan->ecc, echan->slot[i]);
echan->slot[i] = -1;
}
}
/* Set entry slot to the dummy slot */
edma_set_chmap(echan, echan->ecc->dummy_slot);
/* Free EDMA channel */
if (echan->alloced) {
edma_free_channel(echan);
echan->alloced = false;
}
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
echan->tc = NULL;
echan->hw_triggered = false;
dev_dbg(dev, "Free eDMA channel %d for virt channel %d\n",
EDMA_CHAN_SLOT(echan->ch_num), chan->chan_id);
}
/* Send pending descriptor to hardware */
static void edma_issue_pending(struct dma_chan *chan)
{
struct edma_chan *echan = to_edma_chan(chan);
unsigned long flags;
spin_lock_irqsave(&echan->vchan.lock, flags);
if (vchan_issue_pending(&echan->vchan) && !echan->edesc)
edma_execute(echan);
spin_unlock_irqrestore(&echan->vchan.lock, flags);
}
/*
* This limit exists to avoid a possible infinite loop when waiting for proof
* that a particular transfer is completed. This limit can be hit if there
* are large bursts to/from slow devices or the CPU is never able to catch
* the DMA hardware idle. On an AM335x transfering 48 bytes from the UART
* RX-FIFO, as many as 55 loops have been seen.
*/
#define EDMA_MAX_TR_WAIT_LOOPS 1000
static u32 edma_residue(struct edma_desc *edesc)
{
bool dst = edesc->direction == DMA_DEV_TO_MEM;
int loop_count = EDMA_MAX_TR_WAIT_LOOPS;
struct edma_chan *echan = edesc->echan;
struct edma_pset *pset = edesc->pset;
dma_addr_t done, pos, pos_old;
int channel = EDMA_CHAN_SLOT(echan->ch_num);
int idx = EDMA_REG_ARRAY_INDEX(channel);
int ch_bit = EDMA_CHANNEL_BIT(channel);
int event_reg;
int i;
/*
* We always read the dst/src position from the first RamPar
* pset. That's the one which is active now.
*/
pos = edma_get_position(echan->ecc, echan->slot[0], dst);
/*
* "pos" may represent a transfer request that is still being
* processed by the EDMACC or EDMATC. We will busy wait until
* any one of the situations occurs:
* 1. while and event is pending for the channel
* 2. a position updated
* 3. we hit the loop limit
*/
if (is_slave_direction(edesc->direction))
event_reg = SH_ER;
else
event_reg = SH_ESR;
pos_old = pos;
while (edma_shadow0_read_array(echan->ecc, event_reg, idx) & ch_bit) {
pos = edma_get_position(echan->ecc, echan->slot[0], dst);
if (pos != pos_old)
break;
if (!--loop_count) {
dev_dbg_ratelimited(echan->vchan.chan.device->dev,
"%s: timeout waiting for PaRAM update\n",
__func__);
break;
}
cpu_relax();
}
/*
* Cyclic is simple. Just subtract pset[0].addr from pos.
*
* We never update edesc->residue in the cyclic case, so we
* can tell the remaining room to the end of the circular
* buffer.
*/
if (edesc->cyclic) {
done = pos - pset->addr;
edesc->residue_stat = edesc->residue - done;
return edesc->residue_stat;
}
/*
* If the position is 0, then EDMA loaded the closing dummy slot, the
* transfer is completed
*/
if (!pos)
return 0;
/*
* For SG operation we catch up with the last processed
* status.
*/
pset += edesc->processed_stat;
for (i = edesc->processed_stat; i < edesc->processed; i++, pset++) {
/*
* If we are inside this pset address range, we know
* this is the active one. Get the current delta and
* stop walking the psets.
*/
if (pos >= pset->addr && pos < pset->addr + pset->len)
return edesc->residue_stat - (pos - pset->addr);
/* Otherwise mark it done and update residue_stat. */
edesc->processed_stat++;
edesc->residue_stat -= pset->len;
}
return edesc->residue_stat;
}
/* Check request completion status */
static enum dma_status edma_tx_status(struct dma_chan *chan,
dma_cookie_t cookie,
struct dma_tx_state *txstate)
{
struct edma_chan *echan = to_edma_chan(chan);
struct dma_tx_state txstate_tmp;
enum dma_status ret;
unsigned long flags;
ret = dma_cookie_status(chan, cookie, txstate);
if (ret == DMA_COMPLETE)
return ret;
/* Provide a dummy dma_tx_state for completion checking */
if (!txstate)
txstate = &txstate_tmp;
spin_lock_irqsave(&echan->vchan.lock, flags);
if (echan->edesc && echan->edesc->vdesc.tx.cookie == cookie) {
txstate->residue = edma_residue(echan->edesc);
} else {
struct virt_dma_desc *vdesc = vchan_find_desc(&echan->vchan,
cookie);
if (vdesc)
txstate->residue = to_edma_desc(&vdesc->tx)->residue;
else
txstate->residue = 0;
}
/*
* Mark the cookie completed if the residue is 0 for non cyclic
* transfers
*/
if (ret != DMA_COMPLETE && !txstate->residue &&
echan->edesc && echan->edesc->polled &&
echan->edesc->vdesc.tx.cookie == cookie) {
edma_stop(echan);
vchan_cookie_complete(&echan->edesc->vdesc);
echan->edesc = NULL;
edma_execute(echan);
ret = DMA_COMPLETE;
}
spin_unlock_irqrestore(&echan->vchan.lock, flags);
return ret;
}
static bool edma_is_memcpy_channel(int ch_num, s32 *memcpy_channels)
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
{
if (!memcpy_channels)
return false;
while (*memcpy_channels != -1) {
if (*memcpy_channels == ch_num)
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
return true;
memcpy_channels++;
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
}
return false;
}
#define EDMA_DMA_BUSWIDTHS (BIT(DMA_SLAVE_BUSWIDTH_1_BYTE) | \
BIT(DMA_SLAVE_BUSWIDTH_2_BYTES) | \
BIT(DMA_SLAVE_BUSWIDTH_3_BYTES) | \
BIT(DMA_SLAVE_BUSWIDTH_4_BYTES))
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
static void edma_dma_init(struct edma_cc *ecc, bool legacy_mode)
{
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
struct dma_device *s_ddev = &ecc->dma_slave;
struct dma_device *m_ddev = NULL;
s32 *memcpy_channels = ecc->info->memcpy_channels;
int i, j;
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
dma_cap_zero(s_ddev->cap_mask);
dma_cap_set(DMA_SLAVE, s_ddev->cap_mask);
dma_cap_set(DMA_CYCLIC, s_ddev->cap_mask);
if (ecc->legacy_mode && !memcpy_channels) {
dev_warn(ecc->dev,
"Legacy memcpy is enabled, things might not work\n");
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
dma_cap_set(DMA_MEMCPY, s_ddev->cap_mask);
dma_cap_set(DMA_INTERLEAVE, s_ddev->cap_mask);
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
s_ddev->device_prep_dma_memcpy = edma_prep_dma_memcpy;
s_ddev->device_prep_interleaved_dma = edma_prep_dma_interleaved;
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
s_ddev->directions = BIT(DMA_MEM_TO_MEM);
}
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
s_ddev->device_prep_slave_sg = edma_prep_slave_sg;
s_ddev->device_prep_dma_cyclic = edma_prep_dma_cyclic;
s_ddev->device_alloc_chan_resources = edma_alloc_chan_resources;
s_ddev->device_free_chan_resources = edma_free_chan_resources;
s_ddev->device_issue_pending = edma_issue_pending;
s_ddev->device_tx_status = edma_tx_status;
s_ddev->device_config = edma_slave_config;
s_ddev->device_pause = edma_dma_pause;
s_ddev->device_resume = edma_dma_resume;
s_ddev->device_terminate_all = edma_terminate_all;
s_ddev->device_synchronize = edma_synchronize;
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
s_ddev->src_addr_widths = EDMA_DMA_BUSWIDTHS;
s_ddev->dst_addr_widths = EDMA_DMA_BUSWIDTHS;
s_ddev->directions |= (BIT(DMA_DEV_TO_MEM) | BIT(DMA_MEM_TO_DEV));
s_ddev->residue_granularity = DMA_RESIDUE_GRANULARITY_BURST;
s_ddev->max_burst = SZ_32K - 1; /* CIDX: 16bit signed */
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
s_ddev->dev = ecc->dev;
INIT_LIST_HEAD(&s_ddev->channels);
if (memcpy_channels) {
m_ddev = devm_kzalloc(ecc->dev, sizeof(*m_ddev), GFP_KERNEL);
if (!m_ddev) {
dev_warn(ecc->dev, "memcpy is disabled due to OoM\n");
memcpy_channels = NULL;
goto ch_setup;
}
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
ecc->dma_memcpy = m_ddev;
dma_cap_zero(m_ddev->cap_mask);
dma_cap_set(DMA_MEMCPY, m_ddev->cap_mask);
dma_cap_set(DMA_INTERLEAVE, m_ddev->cap_mask);
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
m_ddev->device_prep_dma_memcpy = edma_prep_dma_memcpy;
m_ddev->device_prep_interleaved_dma = edma_prep_dma_interleaved;
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
m_ddev->device_alloc_chan_resources = edma_alloc_chan_resources;
m_ddev->device_free_chan_resources = edma_free_chan_resources;
m_ddev->device_issue_pending = edma_issue_pending;
m_ddev->device_tx_status = edma_tx_status;
m_ddev->device_config = edma_slave_config;
m_ddev->device_pause = edma_dma_pause;
m_ddev->device_resume = edma_dma_resume;
m_ddev->device_terminate_all = edma_terminate_all;
m_ddev->device_synchronize = edma_synchronize;
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
m_ddev->src_addr_widths = EDMA_DMA_BUSWIDTHS;
m_ddev->dst_addr_widths = EDMA_DMA_BUSWIDTHS;
m_ddev->directions = BIT(DMA_MEM_TO_MEM);
m_ddev->residue_granularity = DMA_RESIDUE_GRANULARITY_BURST;
m_ddev->dev = ecc->dev;
INIT_LIST_HEAD(&m_ddev->channels);
} else if (!ecc->legacy_mode) {
dev_info(ecc->dev, "memcpy is disabled\n");
}
ch_setup:
for (i = 0; i < ecc->num_channels; i++) {
struct edma_chan *echan = &ecc->slave_chans[i];
echan->ch_num = EDMA_CTLR_CHAN(ecc->id, i);
echan->ecc = ecc;
echan->vchan.desc_free = edma_desc_free;
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
if (m_ddev && edma_is_memcpy_channel(i, memcpy_channels))
vchan_init(&echan->vchan, m_ddev);
else
vchan_init(&echan->vchan, s_ddev);
INIT_LIST_HEAD(&echan->node);
for (j = 0; j < EDMA_MAX_SLOTS; j++)
echan->slot[j] = -1;
}
}
static int edma_setup_from_hw(struct device *dev, struct edma_soc_info *pdata,
struct edma_cc *ecc)
{
int i;
u32 value, cccfg;
s8 (*queue_priority_map)[2];
/* Decode the eDMA3 configuration from CCCFG register */
cccfg = edma_read(ecc, EDMA_CCCFG);
value = GET_NUM_REGN(cccfg);
ecc->num_region = BIT(value);
value = GET_NUM_DMACH(cccfg);
ecc->num_channels = BIT(value + 1);
value = GET_NUM_QDMACH(cccfg);
ecc->num_qchannels = value * 2;
value = GET_NUM_PAENTRY(cccfg);
ecc->num_slots = BIT(value + 4);
value = GET_NUM_EVQUE(cccfg);
ecc->num_tc = value + 1;
ecc->chmap_exist = (cccfg & CHMAP_EXIST) ? true : false;
dev_dbg(dev, "eDMA3 CC HW configuration (cccfg: 0x%08x):\n", cccfg);
dev_dbg(dev, "num_region: %u\n", ecc->num_region);
dev_dbg(dev, "num_channels: %u\n", ecc->num_channels);
dev_dbg(dev, "num_qchannels: %u\n", ecc->num_qchannels);
dev_dbg(dev, "num_slots: %u\n", ecc->num_slots);
dev_dbg(dev, "num_tc: %u\n", ecc->num_tc);
dev_dbg(dev, "chmap_exist: %s\n", ecc->chmap_exist ? "yes" : "no");
/* Nothing need to be done if queue priority is provided */
if (pdata->queue_priority_mapping)
return 0;
/*
* Configure TC/queue priority as follows:
* Q0 - priority 0
* Q1 - priority 1
* Q2 - priority 2
* ...
* The meaning of priority numbers: 0 highest priority, 7 lowest
* priority. So Q0 is the highest priority queue and the last queue has
* the lowest priority.
*/
queue_priority_map = devm_kcalloc(dev, ecc->num_tc + 1, sizeof(s8),
GFP_KERNEL);
if (!queue_priority_map)
return -ENOMEM;
for (i = 0; i < ecc->num_tc; i++) {
queue_priority_map[i][0] = i;
queue_priority_map[i][1] = i;
}
queue_priority_map[i][0] = -1;
queue_priority_map[i][1] = -1;
pdata->queue_priority_mapping = queue_priority_map;
/* Default queue has the lowest priority */
pdata->default_queue = i - 1;
return 0;
}
#if IS_ENABLED(CONFIG_OF)
static int edma_xbar_event_map(struct device *dev, struct edma_soc_info *pdata,
size_t sz)
{
const char pname[] = "ti,edma-xbar-event-map";
struct resource res;
void __iomem *xbar;
s16 (*xbar_chans)[2];
size_t nelm = sz / sizeof(s16);
u32 shift, offset, mux;
int ret, i;
xbar_chans = devm_kcalloc(dev, nelm + 2, sizeof(s16), GFP_KERNEL);
if (!xbar_chans)
return -ENOMEM;
ret = of_address_to_resource(dev->of_node, 1, &res);
if (ret)
return -ENOMEM;
xbar = devm_ioremap(dev, res.start, resource_size(&res));
if (!xbar)
return -ENOMEM;
ret = of_property_read_u16_array(dev->of_node, pname, (u16 *)xbar_chans,
nelm);
if (ret)
return -EIO;
/* Invalidate last entry for the other user of this mess */
nelm >>= 1;
xbar_chans[nelm][0] = -1;
xbar_chans[nelm][1] = -1;
for (i = 0; i < nelm; i++) {
shift = (xbar_chans[i][1] & 0x03) << 3;
offset = xbar_chans[i][1] & 0xfffffffc;
mux = readl(xbar + offset);
mux &= ~(0xff << shift);
mux |= xbar_chans[i][0] << shift;
writel(mux, (xbar + offset));
}
pdata->xbar_chans = (const s16 (*)[2]) xbar_chans;
return 0;
}
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
static struct edma_soc_info *edma_setup_info_from_dt(struct device *dev,
bool legacy_mode)
{
struct edma_soc_info *info;
struct property *prop;
int sz, ret;
info = devm_kzalloc(dev, sizeof(struct edma_soc_info), GFP_KERNEL);
if (!info)
return ERR_PTR(-ENOMEM);
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
if (legacy_mode) {
prop = of_find_property(dev->of_node, "ti,edma-xbar-event-map",
&sz);
if (prop) {
ret = edma_xbar_event_map(dev, info, sz);
if (ret)
return ERR_PTR(ret);
}
return info;
}
/* Get the list of channels allocated to be used for memcpy */
prop = of_find_property(dev->of_node, "ti,edma-memcpy-channels", &sz);
if (prop) {
const char pname[] = "ti,edma-memcpy-channels";
size_t nelm = sz / sizeof(s32);
s32 *memcpy_ch;
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
memcpy_ch = devm_kcalloc(dev, nelm + 1, sizeof(s32),
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
GFP_KERNEL);
if (!memcpy_ch)
return ERR_PTR(-ENOMEM);
ret = of_property_read_u32_array(dev->of_node, pname,
(u32 *)memcpy_ch, nelm);
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
if (ret)
return ERR_PTR(ret);
memcpy_ch[nelm] = -1;
info->memcpy_channels = memcpy_ch;
}
prop = of_find_property(dev->of_node, "ti,edma-reserved-slot-ranges",
&sz);
if (prop) {
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
const char pname[] = "ti,edma-reserved-slot-ranges";
u32 (*tmp)[2];
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
s16 (*rsv_slots)[2];
size_t nelm = sz / sizeof(*tmp);
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
struct edma_rsv_info *rsv_info;
int i;
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
if (!nelm)
return info;
tmp = kcalloc(nelm, sizeof(*tmp), GFP_KERNEL);
if (!tmp)
return ERR_PTR(-ENOMEM);
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
rsv_info = devm_kzalloc(dev, sizeof(*rsv_info), GFP_KERNEL);
if (!rsv_info) {
kfree(tmp);
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
return ERR_PTR(-ENOMEM);
}
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
rsv_slots = devm_kcalloc(dev, nelm + 1, sizeof(*rsv_slots),
GFP_KERNEL);
if (!rsv_slots) {
kfree(tmp);
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
return ERR_PTR(-ENOMEM);
}
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
ret = of_property_read_u32_array(dev->of_node, pname,
(u32 *)tmp, nelm * 2);
if (ret) {
kfree(tmp);
return ERR_PTR(ret);
}
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
for (i = 0; i < nelm; i++) {
rsv_slots[i][0] = tmp[i][0];
rsv_slots[i][1] = tmp[i][1];
}
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
rsv_slots[nelm][0] = -1;
rsv_slots[nelm][1] = -1;
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
info->rsv = rsv_info;
info->rsv->rsv_slots = (const s16 (*)[2])rsv_slots;
kfree(tmp);
}
return info;
}
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
static struct dma_chan *of_edma_xlate(struct of_phandle_args *dma_spec,
struct of_dma *ofdma)
{
struct edma_cc *ecc = ofdma->of_dma_data;
struct dma_chan *chan = NULL;
struct edma_chan *echan;
int i;
if (!ecc || dma_spec->args_count < 1)
return NULL;
for (i = 0; i < ecc->num_channels; i++) {
echan = &ecc->slave_chans[i];
if (echan->ch_num == dma_spec->args[0]) {
chan = &echan->vchan.chan;
break;
}
}
if (!chan)
return NULL;
if (echan->ecc->legacy_mode && dma_spec->args_count == 1)
goto out;
if (!echan->ecc->legacy_mode && dma_spec->args_count == 2 &&
dma_spec->args[1] < echan->ecc->num_tc) {
echan->tc = &echan->ecc->tc_list[dma_spec->args[1]];
goto out;
}
return NULL;
out:
/* The channel is going to be used as HW synchronized */
echan->hw_triggered = true;
return dma_get_slave_channel(chan);
}
#else
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
static struct edma_soc_info *edma_setup_info_from_dt(struct device *dev,
bool legacy_mode)
{
return ERR_PTR(-EINVAL);
}
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
static struct dma_chan *of_edma_xlate(struct of_phandle_args *dma_spec,
struct of_dma *ofdma)
{
return NULL;
}
#endif
static bool edma_filter_fn(struct dma_chan *chan, void *param);
static int edma_probe(struct platform_device *pdev)
{
struct edma_soc_info *info = pdev->dev.platform_data;
s8 (*queue_priority_mapping)[2];
const s16 (*reserved)[2];
int i, irq;
char *irq_name;
struct resource *mem;
struct device_node *node = pdev->dev.of_node;
struct device *dev = &pdev->dev;
struct edma_cc *ecc;
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
bool legacy_mode = true;
int ret;
if (node) {
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
const struct of_device_id *match;
match = of_match_node(edma_of_ids, node);
if (match && (*(u32 *)match->data) == EDMA_BINDING_TPCC)
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
legacy_mode = false;
info = edma_setup_info_from_dt(dev, legacy_mode);
if (IS_ERR(info)) {
dev_err(dev, "failed to get DT data\n");
return PTR_ERR(info);
}
}
if (!info)
return -ENODEV;
ret = dma_set_mask_and_coherent(dev, DMA_BIT_MASK(32));
if (ret)
return ret;
ecc = devm_kzalloc(dev, sizeof(*ecc), GFP_KERNEL);
if (!ecc)
return -ENOMEM;
ecc->dev = dev;
ecc->id = pdev->id;
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
ecc->legacy_mode = legacy_mode;
/* When booting with DT the pdev->id is -1 */
if (ecc->id < 0)
ecc->id = 0;
mem = platform_get_resource_byname(pdev, IORESOURCE_MEM, "edma3_cc");
if (!mem) {
dev_dbg(dev, "mem resource not found, using index 0\n");
mem = platform_get_resource(pdev, IORESOURCE_MEM, 0);
if (!mem) {
dev_err(dev, "no mem resource?\n");
return -ENODEV;
}
}
ecc->base = devm_ioremap_resource(dev, mem);
if (IS_ERR(ecc->base))
return PTR_ERR(ecc->base);
platform_set_drvdata(pdev, ecc);
pm_runtime_enable(dev);
ret = pm_runtime_get_sync(dev);
if (ret < 0) {
dev_err(dev, "pm_runtime_get_sync() failed\n");
pm_runtime_disable(dev);
return ret;
}
/* Get eDMA3 configuration from IP */
ret = edma_setup_from_hw(dev, info, ecc);
if (ret)
goto err_disable_pm;
/* Allocate memory based on the information we got from the IP */
ecc->slave_chans = devm_kcalloc(dev, ecc->num_channels,
sizeof(*ecc->slave_chans), GFP_KERNEL);
ecc->slot_inuse = devm_kcalloc(dev, BITS_TO_LONGS(ecc->num_slots),
sizeof(unsigned long), GFP_KERNEL);
ecc->channels_mask = devm_kcalloc(dev,
BITS_TO_LONGS(ecc->num_channels),
sizeof(unsigned long), GFP_KERNEL);
if (!ecc->slave_chans || !ecc->slot_inuse || !ecc->channels_mask) {
ret = -ENOMEM;
goto err_disable_pm;
}
/* Mark all channels available initially */
bitmap_fill(ecc->channels_mask, ecc->num_channels);
ecc->default_queue = info->default_queue;
if (info->rsv) {
/* Set the reserved slots in inuse list */
reserved = info->rsv->rsv_slots;
if (reserved) {
for (i = 0; reserved[i][0] != -1; i++)
bitmap_set(ecc->slot_inuse, reserved[i][0],
reserved[i][1]);
}
/* Clear channels not usable for Linux */
reserved = info->rsv->rsv_chans;
if (reserved) {
for (i = 0; reserved[i][0] != -1; i++)
bitmap_clear(ecc->channels_mask, reserved[i][0],
reserved[i][1]);
}
}
for (i = 0; i < ecc->num_slots; i++) {
/* Reset only unused - not reserved - paRAM slots */
if (!test_bit(i, ecc->slot_inuse))
edma_write_slot(ecc, i, &dummy_paramset);
}
irq = platform_get_irq_byname(pdev, "edma3_ccint");
if (irq < 0 && node)
irq = irq_of_parse_and_map(node, 0);
if (irq >= 0) {
irq_name = devm_kasprintf(dev, GFP_KERNEL, "%s_ccint",
dev_name(dev));
ret = devm_request_irq(dev, irq, dma_irq_handler, 0, irq_name,
ecc);
if (ret) {
dev_err(dev, "CCINT (%d) failed --> %d\n", irq, ret);
goto err_disable_pm;
}
ecc->ccint = irq;
}
irq = platform_get_irq_byname(pdev, "edma3_ccerrint");
if (irq < 0 && node)
irq = irq_of_parse_and_map(node, 2);
if (irq >= 0) {
irq_name = devm_kasprintf(dev, GFP_KERNEL, "%s_ccerrint",
dev_name(dev));
ret = devm_request_irq(dev, irq, dma_ccerr_handler, 0, irq_name,
ecc);
if (ret) {
dev_err(dev, "CCERRINT (%d) failed --> %d\n", irq, ret);
goto err_disable_pm;
}
ecc->ccerrint = irq;
}
ecc->dummy_slot = edma_alloc_slot(ecc, EDMA_SLOT_ANY);
if (ecc->dummy_slot < 0) {
dev_err(dev, "Can't allocate PaRAM dummy slot\n");
ret = ecc->dummy_slot;
goto err_disable_pm;
}
queue_priority_mapping = info->queue_priority_mapping;
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
if (!ecc->legacy_mode) {
int lowest_priority = 0;
unsigned int array_max;
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
struct of_phandle_args tc_args;
ecc->tc_list = devm_kcalloc(dev, ecc->num_tc,
sizeof(*ecc->tc_list), GFP_KERNEL);
if (!ecc->tc_list) {
ret = -ENOMEM;
goto err_reg1;
}
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
for (i = 0;; i++) {
ret = of_parse_phandle_with_fixed_args(node, "ti,tptcs",
1, i, &tc_args);
if (ret || i == ecc->num_tc)
break;
ecc->tc_list[i].node = tc_args.np;
ecc->tc_list[i].id = i;
queue_priority_mapping[i][1] = tc_args.args[0];
if (queue_priority_mapping[i][1] > lowest_priority) {
lowest_priority = queue_priority_mapping[i][1];
info->default_queue = i;
}
}
/* See if we have optional dma-channel-mask array */
array_max = DIV_ROUND_UP(ecc->num_channels, BITS_PER_TYPE(u32));
ret = of_property_read_variable_u32_array(node,
"dma-channel-mask",
(u32 *)ecc->channels_mask,
1, array_max);
if (ret > 0 && ret != array_max)
dev_warn(dev, "dma-channel-mask is not complete.\n");
else if (ret == -EOVERFLOW || ret == -ENODATA)
dev_warn(dev,
"dma-channel-mask is out of range or empty\n");
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
}
/* Event queue priority mapping */
for (i = 0; queue_priority_mapping[i][0] != -1; i++)
edma_assign_priority_to_queue(ecc, queue_priority_mapping[i][0],
queue_priority_mapping[i][1]);
edma_write_array2(ecc, EDMA_DRAE, 0, 0, 0x0);
edma_write_array2(ecc, EDMA_DRAE, 0, 1, 0x0);
edma_write_array(ecc, EDMA_QRAE, 0, 0x0);
ecc->info = info;
/* Init the dma device and channels */
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
edma_dma_init(ecc, legacy_mode);
for (i = 0; i < ecc->num_channels; i++) {
/* Do not touch reserved channels */
if (!test_bit(i, ecc->channels_mask))
continue;
/* Assign all channels to the default queue */
edma_assign_channel_eventq(&ecc->slave_chans[i],
info->default_queue);
/* Set entry slot to the dummy slot */
edma_set_chmap(&ecc->slave_chans[i], ecc->dummy_slot);
}
ecc->dma_slave.filter.map = info->slave_map;
ecc->dma_slave.filter.mapcnt = info->slavecnt;
ecc->dma_slave.filter.fn = edma_filter_fn;
ret = dma_async_device_register(&ecc->dma_slave);
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
if (ret) {
dev_err(dev, "slave ddev registration failed (%d)\n", ret);
goto err_reg1;
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
}
if (ecc->dma_memcpy) {
ret = dma_async_device_register(ecc->dma_memcpy);
if (ret) {
dev_err(dev, "memcpy ddev registration failed (%d)\n",
ret);
dma_async_device_unregister(&ecc->dma_slave);
goto err_reg1;
}
}
if (node)
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
of_dma_controller_register(node, of_edma_xlate, ecc);
dev_info(dev, "TI EDMA DMA engine driver\n");
return 0;
err_reg1:
edma_free_slot(ecc, ecc->dummy_slot);
err_disable_pm:
pm_runtime_put_sync(dev);
pm_runtime_disable(dev);
return ret;
}
static void edma_cleanupp_vchan(struct dma_device *dmadev)
{
struct edma_chan *echan, *_echan;
list_for_each_entry_safe(echan, _echan,
&dmadev->channels, vchan.chan.device_node) {
list_del(&echan->vchan.chan.device_node);
tasklet_kill(&echan->vchan.task);
}
}
static int edma_remove(struct platform_device *pdev)
{
struct device *dev = &pdev->dev;
struct edma_cc *ecc = dev_get_drvdata(dev);
devm_free_irq(dev, ecc->ccint, ecc);
devm_free_irq(dev, ecc->ccerrint, ecc);
edma_cleanupp_vchan(&ecc->dma_slave);
if (dev->of_node)
of_dma_controller_free(dev->of_node);
dma_async_device_unregister(&ecc->dma_slave);
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
if (ecc->dma_memcpy)
dma_async_device_unregister(ecc->dma_memcpy);
edma_free_slot(ecc, ecc->dummy_slot);
pm_runtime_put_sync(dev);
pm_runtime_disable(dev);
return 0;
}
#ifdef CONFIG_PM_SLEEP
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
static int edma_pm_suspend(struct device *dev)
{
struct edma_cc *ecc = dev_get_drvdata(dev);
struct edma_chan *echan = ecc->slave_chans;
int i;
for (i = 0; i < ecc->num_channels; i++) {
if (echan[i].alloced)
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
edma_setup_interrupt(&echan[i], false);
}
return 0;
}
static int edma_pm_resume(struct device *dev)
{
struct edma_cc *ecc = dev_get_drvdata(dev);
struct edma_chan *echan = ecc->slave_chans;
int i;
s8 (*queue_priority_mapping)[2];
/* re initialize dummy slot to dummy param set */
edma_write_slot(ecc, ecc->dummy_slot, &dummy_paramset);
queue_priority_mapping = ecc->info->queue_priority_mapping;
/* Event queue priority mapping */
for (i = 0; queue_priority_mapping[i][0] != -1; i++)
edma_assign_priority_to_queue(ecc, queue_priority_mapping[i][0],
queue_priority_mapping[i][1]);
for (i = 0; i < ecc->num_channels; i++) {
if (echan[i].alloced) {
/* ensure access through shadow region 0 */
edma_or_array2(ecc, EDMA_DRAE, 0,
EDMA_REG_ARRAY_INDEX(i),
EDMA_CHANNEL_BIT(i));
edma_setup_interrupt(&echan[i], true);
/* Set up channel -> slot mapping for the entry slot */
edma_set_chmap(&echan[i], echan[i].slot[0]);
}
}
return 0;
}
#endif
static const struct dev_pm_ops edma_pm_ops = {
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
SET_LATE_SYSTEM_SLEEP_PM_OPS(edma_pm_suspend, edma_pm_resume)
};
static struct platform_driver edma_driver = {
.probe = edma_probe,
.remove = edma_remove,
.driver = {
.name = "edma",
.pm = &edma_pm_ops,
.of_match_table = edma_of_ids,
},
};
static int edma_tptc_probe(struct platform_device *pdev)
{
pm_runtime_enable(&pdev->dev);
return pm_runtime_get_sync(&pdev->dev);
}
static struct platform_driver edma_tptc_driver = {
.probe = edma_tptc_probe,
.driver = {
.name = "edma3-tptc",
.of_match_table = edma_tptc_of_ids,
},
};
static bool edma_filter_fn(struct dma_chan *chan, void *param)
{
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
bool match = false;
if (chan->device->dev->driver == &edma_driver.driver) {
struct edma_chan *echan = to_edma_chan(chan);
unsigned ch_req = *(unsigned *)param;
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
if (ch_req == echan->ch_num) {
/* The channel is going to be used as HW synchronized */
echan->hw_triggered = true;
match = true;
}
}
dmaengine: edma: New device tree binding With the old binding and driver architecture we had many issues: No way to assign eDMA channels to event queues, thus not able to tune the system by moving specific DMA channels to low/high priority servicing. We moved the cyclic channels to high priority within the code, but that was just a workaround to this issue. Memcopy was fundamentally broken: even if the driver scanned the DT/devices in the booted system for direct DMA users (which is not effective when the events are going through a crossbar) and created a map of 'used' channels, this information was not really usable. Since via dmaengien API the eDMA driver will be called with _some_ channel number, we would try to request this channel when any channel is requested for memcpy. By luck we got channel which is not used by any device most of the time so things worked, but if a device would have been using the given channel, but not requested it, the memcpy channel would have been waiting for HW event. The old code had the am33xx/am43xx DMA event router handling embedded. This should have been done in a separate driver since it is not part of the actual eDMA IP. There were no way to 'lock' PaRAM slots to be used by the DSP for example when booting with DT. In DT boot the edma node used more than one hwmod which is not a good practice and the kernel prints warning because of this. With the new bindings and the changes in the driver we can: - No regression with Legacy binding and non DT boot - DMA channels can be assigned to any TC (to set priority) - PaRAM slots can be reserved for other cores to use - Dynamic power management for CC and TCs, if only TC0 is used all other TC can be powered down for example Signed-off-by: Peter Ujfalusi <peter.ujfalusi@ti.com> Signed-off-by: Vinod Koul <vinod.koul@intel.com>
2015-10-16 15:18:10 +08:00
return match;
}
static int edma_init(void)
{
int ret;
ret = platform_driver_register(&edma_tptc_driver);
if (ret)
return ret;
return platform_driver_register(&edma_driver);
}
subsys_initcall(edma_init);
static void __exit edma_exit(void)
{
platform_driver_unregister(&edma_driver);
platform_driver_unregister(&edma_tptc_driver);
}
module_exit(edma_exit);
MODULE_AUTHOR("Matt Porter <matt.porter@linaro.org>");
MODULE_DESCRIPTION("TI EDMA DMA engine driver");
MODULE_LICENSE("GPL v2");