Some shortcomings introduced into pci_restore_state() by commit
26f41062f2 ("PCI: check for pci bar restore completion and retry")
have been fixed by recent commit ebfc5b802f ("PCI: Fix regression in
pci_restore_state(), v3"), but that commit treats all PCI devices as
those with Type 0 configuration headers.
That is not entirely correct, because Type 1 and Type 2 headers have
different layouts. In particular, the area occupied by BARs in Type 0
config headers contains the secondary status register in Type 1 ones and
it doesn't make sense to retry the restoration of that register even if
the value read back from it after a write is not the same as the written
one (it very well may be different).
For this reason, make pci_restore_state() only retry the restoration
of BARs for Type 0 config headers. This effectively makes it behave
as before commit 26f41062f2 for all header types except for Type 0.
Tested-by: Mikko Vinni <mmvinni@yahoo.com>
Signed-off-by: Rafael J. Wysocki <rjw@sisk.pl>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Commit 26f41062f2 ("PCI: check for pci bar restore completion and
retry") attempted to address problems with PCI BAR restoration on
systems where FLR had not been completed before pci_restore_state() was
called, but it did that in an utterly wrong way.
First off, instead of retrying the writes for the BAR registers only, it
did that for all of the PCI config space of the device, including the
status register (whose value after the write quite obviously need not be
the same as the written one). Second, it added arbitrary delay to
pci_restore_state() even for systems where the PCI config space
restoration was successful at first attempt. Finally, the mdelay(10) it
added to every iteration of the writing loop was way too much of a delay
for any reasonable device.
All of this actually caused resume failures for some devices on Mikko's
system.
To fix the regression, make pci_restore_state() only retry the writes
for BAR registers and only wait if the first read from the register
doesn't return the written value. Additionaly, make it wait for 1 ms,
instead of 10 ms, after every failing attempt to write into config
space.
Reported-by: Mikko Vinni <mmvinni@yahoo.com>
Signed-off-by: Rafael J. Wysocki <rjw@sisk.pl>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
There are PCIe devices on the market that report ARI support but
then fail to initialize correctly when ARI is actually used. This
leads to situations in which kernels 2.6.34 and newer fail to handle
systems where the previous kernels worked without any apparent
problems. Unfortunately, it is currently unknown how many such
devices are there.
For this reason, introduce a new kernel command line option,
pci=noari, allowing users to disable PCIe ARI altogether if they
see problems with PCIe device initialization.
Signed-off-by: Rafael J. Wysocki <rjw@sisk.pl>
Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
This isn't really a quirk; calling it directly from pci_add_device makes
more sense.
Signed-off-by: Yinghai Lu <yinghai@kernel.org>
Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Let the user could enable and disable with pci=realloc=on or pci=realloc=off
Also
1. move variable and functions near the place they are used.
2. change macro to function
3. change related functions and variable to static and _init
4. update parameter description accordingly.
This will let us add a config option to control default behavior, and
still allow the user to turn off automatic reallocation if it fails on
their platform until a permanent solution is found.
-v2: still honor pci=realloc, and treat it as pci=realloc=on
also use enum instead of ...
-v3: update kernel-paramenters.txt according to Jesse.
Signed-off-by: Yinghai Lu <yinghai@kernel.org>
Acked-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Only one user in driver/pci/pci.c, so we don't need to put it in global
pci.h
Reviewed-by: Bjorn Helgaas <bhelgaas@google.com>
Signed-off-by: Yinghai Lu <yinghai@kernel.org>
Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
On some OEM systems, pci_restore_state() is called while FLR has not yet
completed. As a result, PCI BAR register restore is not successful. This fix
reads back the restored value and compares it with saved value and re-tries 10
times before giving up.
Signed-off-by: Jean Guyader <jean.guyader@eu.citrix.com>
Signed-off-by: Eric Chanudet <eric.chanudet@citrix.com>
Signed-off-by: Allen Kay <allen.m.kay@intel.com>
Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
The use case of this is when a driver wants to call FLR when a device
is attached to it using the SysFS "bind" or "unbind" functionality.
The call chain when a user does "bind" looks as so:
echo "0000:01.07.0" > /sys/bus/pci/drivers/XXXX/bind
and ends up calling:
driver_bind:
device_lock(dev); <=== TAKES LOCK
XXXX_probe:
.. pci_enable_device()
...__pci_reset_function(), which calls
pci_dev_reset(dev, 0):
if (!0) {
device_lock(dev) <==== DEADLOCK
The __pci_reset_function_locked function allows the the drivers
'probe' function to call the "pci_reset_function" while still holding
the driver mutex lock.
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Fix new kernel-doc warnings:
Warning(drivers/pci/pci.c:2811): No description found for parameter 'dev'
Warning(drivers/pci/pci.c:2811): Excess function parameter 'pdev' description in 'pci_intx_mask_supported'
Warning(drivers/pci/pci.c:2894): No description found for parameter 'dev'
Warning(drivers/pci/pci.c:2894): Excess function parameter 'pdev' description in 'pci_check_and_mask_intx'
Warning(drivers/pci/pci.c:2908): No description found for parameter 'dev'
Warning(drivers/pci/pci.c:2908): Excess function parameter 'pdev' description in 'pci_check_and_unmask_intx'
Signed-off-by: Randy Dunlap <rdunlap@xenotime.net>
Cc: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
During S3 or S4 resume or PCI reset, ATS regs aren't restored correctly.
This patch enables ATS at the device state restore if PCI device has ATS
capability.
Signed-off-by: Xudong Hao <xudong.hao@intel.com>
Signed-off-by: Xiantao Zhang <xiantao.zhang@intel.com>
Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
When the runtime PM is activated on PCI, if a device switches state
frequently (e.g. an EHCI controller with autosuspending USB devices
connected) the PCI configuration traces might be very verbose in the
kernel log. Let's guard those traces with DEBUG condition.
Acked-by: "Rafael J. Wysocki" <rjw@sisk.pl>
Signed-off-by: Vincent Palatin <vpalatin@chromium.org>
Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
The latency timer is read-only and hardwired to zero for all PCIe
devices, both Type 0 and Type 1, so don't bother trying to update it
and cluttering the dmesg log with meaningless "setting latency timer
to 64" messages.
Signed-off-by: Myron Stowe <myron.stowe@redhat.com>
Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
The 'latency timer' of PCI devices, both Type 0 and Type 1,
is setup in architecture-specific code [see: 'pcibios_set_master()'].
There are two approaches being taken by all the architectures - check
if the 'latency timer' is currently set between 16 and 255 and if not
bring it within bounds, or, do nothing (and then there is the
gratuitously different PA-RISC implementation).
There is nothing architecture-specific about PCI's 'latency timer' so
this patch pulls its setup functionality up into the PCI core by
creating a generic 'pcibios_set_master()' function using the '__weak'
attribute which can be used by all architectures as a default which,
if necessary, can then be over-ridden by architecture-specific code.
No functional change.
Signed-off-by: Myron Stowe <myron.stowe@redhat.com>
Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
These new PCI services allow to probe for 2.3-compliant INTx masking
support and then use the feature from PCI interrupt handlers. The
services are properly synchronized with concurrent config space access
via sysfs or on device reset.
This enables generic PCI device drivers like uio_pci_generic or KVM's
device assignment to implement the necessary kernel-side IRQ handling
without any knowledge about device-specific interrupt status and control
registers.
Acked-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>
Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
pci_block_user_cfg_access was designed for the use case that a single
context, the IPR driver, temporarily delays user space accesses to the
config space via sysfs. This assumption became invalid by the time
pci_dev_reset was added as locking instance. Today, if you run two loops
in parallel that reset the same device via sysfs, you end up with a
kernel BUG as pci_block_user_cfg_access detect the broken assumption.
This reworks the pci_block_user_cfg_access to a sleeping service
pci_cfg_access_lock and an atomic-compatible variant called
pci_cfg_access_trylock. The former not only blocks user space access as
before but also waits if access was already locked. The latter service
just returns false in this case, allowing the caller to resolve the
conflict instead of raising a BUG.
Adaptions of the ipr driver were originally written by Brian King.
Acked-by: Brian King <brking@linux.vnet.ibm.com>
Acked-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>
Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
I noticed that hotplug of one setup does not work with recent change in
pci tree.
After checking the bridge conf setup, I noticed that the bridges get
assigned but do not get enabled.
The reason is the following commit, while simply ignores bridge
resources when enabling a pci device:
| commit bbef98ab0f
| Author: Ram Pai <linuxram@us.ibm.com>
| Date: Sun Nov 6 10:33:10 2011 +0800
|
| PCI: defer enablement of SRIOV BARS
|...
| NOTE: Note, there is subtle change in the pci_enable_device() API. Any
| driver that depends on SRIOV BARS to be enabled in pci_enable_device()
| can fail.
Put back bridge resource and ROM resource checking to fix the problem.
That should fix regression like BIOS does not assign correct resource to
bridge.
Discussion can be found at:
http://www.spinics.net/lists/linux-pci/msg12874.html
Signed-off-by: Yinghai Lu <yinghai@kernel.org>
Acked-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
During test of one IB card with guest VM, found that, msi is not
initialized properly.
It turns out __write_msi_msg will do nothing if device current_state is
not PCI_D0. And, that pci device does not have pm_cap in guest VM.
There is an error in setting of power state to PCI_D0 in
pci_enable_device(), but error is not returned for this. Following is
code flow:
pci_enable_device() --> __pci_enable_device_flags() -->
do_pci_enable_device() --> pci_set_power_state() -->
__pci_start_power_transition()
We have following condition inside __pci_start_power_transition():
if (platform_pci_power_manageable(dev)) {
error = platform_pci_set_power_state(dev, state);
if (!error)
pci_update_current_state(dev, state);
} else {
error = -ENODEV;
/* Fall back to PCI_D0 if native PM is not supported */
if (!dev->pm_cap)
dev->current_state = PCI_D0;
}
Here, from platform_pci_set_power_state(), acpi_pci_set_power_state() is
getting called and that is failing with ENODEV because of following
condition:
if (!handle || ACPI_SUCCESS(acpi_get_handle(handle, "_EJ0",&tmp)))
return -ENODEV;
Because of that, pci_update_current_state() is not getting called.
With this patch, if device power state can not be set via
platform_pci_set_power_state and that device does not have native pm
support, then PCI device power state will be set to PCI_D0.
-v2: This also reverts 47e9037ac1, as it's
not needed after this change.
Acked-by: "Rafael J. Wysocki" <rjw@sisk.pl>
Signed-off-by: Ajaykumar Hotchandani<ajaykumar.hotchandani@oracle.com>
Signed-off-by: Yinghai Lu<yinghai.lu@oracle.com>
Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
All the PCI BARs of a device are enabled when the device is enabled
using pci_enable_device(). This unnecessarily enables SRIOV BARs of the
device.
On some platforms, which do not support SRIOV as yet, the
pci_enable_device() fails to enable the device if its SRIOV BARs are not
allocated resources correctly.
The following patch fixes the above problem. The SRIOV BARs are now
enabled when IOV capability of the device is enabled in sriov_enable().
NOTE: Note, there is subtle change in the pci_enable_device() API. Any
driver that depends on SRIOV BARS to be enabled in pci_enable_device()
can fail.
The patch has been touch tested on power and x86 platform.
Tested-by: Michael Wang <wangyun@linux.vnet.ibm.com>
Signed-off-by: Ram Pai <linuxram@us.ibm.com>
Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
When configuring the PCIe settings for "performance", we allow parents
to have a larger Max Payload Size than children and rely on children
Max Read Request Size to not be larger than their own MPS to avoid
having the host bridge generate responses they can't cope with.
However, various drivers in Linux call pci_set_readrq() with arbitrary
values, assuming this to be a simple performance tweak. This breaks
under our "performance" configuration.
Fix that by making sure the value programmed by pcie_set_readrq() is
never larger than the configured MPS for that device.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Jon Mason <mason@myri.com>
Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
The land of PCI power management is a land of sorrow and ugliness,
especially in the area of signaling events by devices. There are
devices that set their PME Status bits, but don't really bother
to send a PME message or assert PME#. There are hardware vendors
who don't connect PME# lines to the system core logic (they know
who they are). There are PCI Express Root Ports that don't bother
to trigger interrupts when they receive PME messages from the devices
below. There are ACPI BIOSes that forget to provide _PRW methods for
devices capable of signaling wakeup. Finally, there are BIOSes that
do provide _PRW methods for such devices, but then don't bother to
call Notify() for those devices from the corresponding _Lxx/_Exx
GPE-handling methods. In all of these cases the kernel doesn't have
a chance to receive a proper notification that it should wake up a
device, so devices stay in low-power states forever. Worse yet, in
some cases they continuously send PME Messages that are silently
ignored, because the kernel simply doesn't know that it should clear
the device's PME Status bit.
This problem was first observed for "parallel" (non-Express) PCI
devices on add-on cards and Matthew Garrett addressed it by adding
code that polls PME Status bits of such devices, if they are enabled
to signal PME, to the kernel. Recently, however, it has turned out
that PCI Express devices are also affected by this issue and that it
is not limited to add-on devices, so it seems necessary to extend
the PME polling to all PCI devices, including PCI Express and planar
ones. Still, it would be wasteful to poll the PME Status bits of
devices that are known to receive proper PME notifications, so make
the kernel (1) poll the PME Status bits of all PCI and PCIe devices
enabled to signal PME and (2) disable the PME Status polling for
devices for which correct PME notifications are received.
Tested-by: Sarah Sharp <sarah.a.sharp@linux.intel.com>
Signed-off-by: Rafael J. Wysocki <rjw@sisk.pl>
Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Add the ability to disable PCI-E MPS turning and using the BIOS
configured MPS defaults. Due to the number of issues recently
discovered on some x86 chipsets, make this the default behavior.
Also, add the option for peer to peer DMA MPS configuration. Peer to
peer DMA is outside the scope of this patch, but MPS configuration could
prevent it from working by having the MPS on one root port different
than the MPS on another. To work around this, simply make the system
wide MPS the smallest possible value (128B).
Signed-off-by: Jon Mason <mason@myri.com>
Acked-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Modifying the Maximum Read Request Size to 0 (value of 128Bytes) has
massive negative ramifications on some devices. Without knowing which
devices have this issue, do not modify from the default value when
walking the PCI-E bus in pcie_bus_safe mode. Also, make pcie_bus_safe
the default procedure.
Tested-by: Sven Schnelle <svens@stackframe.org>
Tested-by: Simon Kirby <sim@hostway.ca>
Tested-by: Stephen M. Cameron <scameron@beardog.cce.hp.com>
Reported-and-tested-by: Eric Dumazet <eric.dumazet@gmail.com>
Reported-and-tested-by: Niels Ole Salscheider <niels_ole@salscheider-online.de>
References: https://bugzilla.kernel.org/show_bug.cgi?id=42162
Signed-off-by: Jon Mason <mason@myri.com>
Acked-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Fix new kernel-doc warning in pci.c:
Warning(drivers/pci/pci.c:3259): No description found for parameter 'mps'
Warning(drivers/pci/pci.c:3259): Excess function parameter 'rq' description in 'pcie_set_mps'
Signed-off-by: Randy Dunlap <rdunlap@xenotime.net>
Cc: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
On a given PCI-E fabric, each device, bridge, and root port can have a
different PCI-E maximum payload size. There is a sizable performance
boost for having the largest possible maximum payload size on each PCI-E
device. However, if improperly configured, fatal bus errors can occur.
Thus, it is important to ensure that PCI-E payloads sends by a device
are never larger than the MPS setting of all devices on the way to the
destination.
This can be achieved two ways:
- A conservative approach is to use the smallest common denominator of
the entire tree below a root complex for every device on that fabric.
This means for example that having a 128 bytes MPS USB controller on one
leg of a switch will dramatically reduce performances of a video card or
10GE adapter on another leg of that same switch.
It also means that any hierarchy supporting hotplug slots (including
expresscard or thunderbolt I suppose, dbl check that) will have to be
entirely clamped to 128 bytes since we cannot predict what will be
plugged into those slots, and we cannot change the MPS on a "live"
system.
- A more optimal way is possible, if it falls within a couple of
constraints:
* The top-level host bridge will never generate packets larger than the
smallest TLP (or if it can be controlled independently from its MPS at
least)
* The device will never generate packets larger than MPS (which can be
configured via MRRS)
* No support of direct PCI-E <-> PCI-E transfers between devices without
some additional code to specifically deal with that case
Then we can use an approach that basically ignores downstream requests
and focuses exclusively on upstream requests. In that case, all we need
to care about is that a device MPS is no larger than its parent MPS,
which allows us to keep all switches/bridges to the max MPS supported by
their parent and eventually the PHB.
In this case, your USB controller would no longer "starve" your 10GE
Ethernet and your hotplug slots won't affect your global MPS.
Additionally, the hotplugged devices themselves can be configured to a
larger MPS up to the value configured in the hotplug bridge.
To choose between the two available options, two PCI kernel boot args
have been added to the PCI calls. "pcie_bus_safe" will provide the
former behavior, while "pcie_bus_perf" will perform the latter behavior.
By default, the latter behavior is used.
NOTE: due to the location of the enablement, each arch will need to add
calls to this function. This patch only enables x86.
This patch includes a number of changes recommended by Benjamin
Herrenschmidt.
Tested-by: Jordan_Hargrave@dell.com
Signed-off-by: Jon Mason <mason@myri.com>
Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
When setting the PCI-E MRRS, pcie_set_readrq queries the current
settings via a pci_read_config_word call but writes the modified result
via a pci_write_config_dword. This results in writing 16 more bits than
were queried.
Also, the function description comment is slightly incorrect.
Signed-off-by: Jon Mason <jdmason@kudzu.us>
Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
The function pci_enable_ari() may mistakenly set the downstream port
of a v1 PCIe switch in ARI Forwarding mode. This is a PCIe v2 feature,
and with an SR-IOV device on that switch port believing the switch above
is ARI capable it may attempt to use functions 8-255, translating into
invalid (non-zero) device numbers for that bus. This has been seen
to cause Completion Timeouts and general misbehaviour including hangs
and panics.
Cc: stable@kernel.org
Acked-by: Don Dutile <ddutile@redhat.com>
Tested-by: Don Dutile <ddutile@redhat.com>
Signed-off-by: Chris Wright <chrisw@sous-sol.org>
Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Multiple attempts to dynamically reallocate pci resources have
unfortunately lead to regressions. Though we continue to fix the
regressions and fine tune the dynamic-reallocation behavior, we have not
reached a acceptable state yet.
This patch provides a interim solution. It disables dynamic reallocation
by default, but adds the ability to enable it through pci=realloc kernel
command line parameter.
Tested-by: Oliver Hartkopp <socketcan@hartkopp.net>
Signed-off-by: Ram Pai <linuxram@us.ibm.com>
Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/jbarnes/pci-2.6:
x86/PCI/ACPI: fix type mismatch
PCI: fix new kernel-doc warning
PCI: Fix warning in drivers/pci/probe.c on sparc64
When I added 3448a19da4
I forgot about the special uv handling code for this, so this
patch fixes it up.
Acked-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Acked-by: Ingo Molnar
Signed-off-by: Dave Airlie <airlied@redhat.com>
Fix pci.c kernel-doc warnings:
Warning(drivers/pci/pci.c:3292): No description found for parameter 'flags'
Warning(drivers/pci/pci.c:3292): Excess function parameter 'change_bridge_flags' description in 'pci_set_vga_state'
Signed-off-by: Randy Dunlap <randy.dunlap@oracle.com>
Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
* 'drm-core-next' of git://git.kernel.org/pub/scm/linux/kernel/git/airlied/drm-2.6: (169 commits)
drivers/gpu/drm/radeon/atom.c: fix warning
drm/radeon/kms: bump kms version number
drm/radeon/kms: properly set num banks for fusion asics
drm/radeon/kms/atom: move dig phy init out of modesetting
drm/radeon/kms/cayman: fix typo in register mask
drm/radeon/kms: fix typo in spread spectrum code
drm/radeon/kms: fix tile_config value reported to userspace on cayman.
drm/radeon/kms: fix incorrect comparison in cayman setup code.
drm/radeon/kms: add wait idle ioctl for eg->cayman
drm/radeon/cayman: setup hdp to invalidate and flush when asked
drm/radeon/evergreen/btc/fusion: setup hdp to invalidate and flush when asked
agp/uninorth: Fix lockups with radeon KMS and >1x.
drm/radeon/kms: the SS_Id field in the LCD table if for LVDS only
drm/radeon/kms: properly set the CLK_REF bit for DCE3 devices
drm/radeon/kms: fixup eDP connector handling
drm/radeon/kms: bail early for eDP in hotplug callback
drm/radeon/kms: simplify hotplug handler logic
drm/radeon/kms: rewrite DP handling
drm/radeon/kms/atom: add support for setting DP panel mode
drm/radeon/kms: atombios.h updates for DP panel mode
...
For KVM device assignment, we'd like to save off the state of a device
prior to passing it to the guest and restore it later. We also want
to allow pci_reset_funciton() to be called while the device is owned
by the guest. This however overwrites and invalidates the struct pci_dev
buffers, so we can't just manually call save and restore. Add generic
interfaces for the saved state to be stored and reloaded back into
struct pci_dev at a later time.
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
This will allow us to store and load it later.
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Latency tolerance reporting allows devices to send messages to the root
complex indicating their latency tolerance for snooped & unsnooped
memory transactions. Add support for enabling & disabling this
feature, along with a routine to set the max latencies a device should
send upstream.
Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
OBFF (optimized buffer flush/fill), where supported, can help improve
energy efficiency by giving devices information about when interrupts
and other activity will have a reduced power impact. It requires
support from both the device and system (i.e. not only does the device
need to respond to OBFF messages, but the platform must be capable of
generating and routing them to the end point).
Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Add support to allow drivers to enable/disable ID-based ordering. Where
supported, ID-based ordering can significantly improve the latency of
individual requests by preventing them from queueing up behind unrelated
traffic.
Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
The pci_pm_reset() function is not a very nice interface due to its
limitations and conditional behavior (e.g. it doesn't affect devices
in low-power states), but it cannot be simply dropped, because
existing device drivers may depend on it. However, its behavior and
limitations should be well documented, so add an appropriate
kerneldoc comment to it.
Signed-off-by: Rafael J. Wysocki <rjw@sisk.pl>
Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
So in a lot of modern systems, a GPU will always be below a parent bridge that won't share with any other GPUs. This means VGA arbitration on those GPUs can be controlled by using the bridge routing instead of io/mem decodes.
The problem is locating which GPUs share which upstream bridges. This patch attempts to identify all the GPUs which can be controlled via bridges, and ones that can't. This patch endeavours to work out the bridge sharing semantics.
When disabling GPUs via a bridge, it doesn't do irq callbacks or touch the io/mem decodes for the gpu.
Signed-off-by: Dave Airlie <airlied@redhat.com>
v3 -> v2: Moved ASPM enabling logic to pci_set_power_state()
v2 -> v1: Preserved the logic in pci_raw_set_power_state()
: Added ASPM enabling logic after scanning Root Bridge
: http://marc.info/?l=linux-pci&m=130046996216391&w=2
v1 : http://marc.info/?l=linux-pci&m=130013164703283&w=2
The assumption made in commit 41cd766b06
(PCI: Don't enable aspm before drivers have had a chance to veto it) that
pci_enable_device() will result in re-configuring ASPM when aspm_policy is
POWERSAVE is no longer valid. This is due to commit
97c145f7c8 (PCI: read current power state
at enable time) which resets dev->current_state to D0. Due to this the
call to pcie_aspm_pm_state_change() is never made. Note the equality check
(below) that returns early:
./drivers/pci/pci.c: pci_raw_set_pci_power_state()
546 /* Check if we're already there */
547 if (dev->current_state == state)
548 return 0;
Therefore OSPM never configures the PCIe links for ASPM to turn them "on".
Fix it by configuring ASPM from the pci_enable_device() code path. This
also allows a driver such as the e1000e networking driver a chance to
disable ASPM (L0s, L1), if need be, prior to enabling the device. A
driver may perform this action if the device is known to mis-behave
wrt ASPM.
Signed-off-by: Naga Chumbalkar <nagananda.chumbalkar@hp.com>
Acked-by: Rafael J. Wysocki <rjw@sisk.pl>
Cc: Matthew Garrett <mjg59@srcf.ucam.org>
Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Make wakeup events be reported by the PCI subsystem before attempting to
resume devices or queuing up runtime resume requests for them, because
wakeup events should be reported as soon as they have been detected.
Signed-off-by: Rafael J. Wysocki <rjw@sisk.pl>
Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
After recent changes related to wakeup events pm_wakeup_event()
automatically checks if the given device is configured to signal wakeup,
so pci_wakeup_event() may be a static inline function calling
pm_wakeup_event() directly.
Signed-off-by: Rafael J. Wysocki <rjw@sisk.pl>
Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
pci_restore_state only ever returns 0, thus there is no benefit in
having it return any value. Also, a large majority of the callers do
not check the return code of pci_restore_state. Make the
pci_restore_state a void return and avoid the overhead.
Acked-by: Mauro Carvalho Chehab <mchehab@redhat.com>
Signed-off-by: Jon Mason <jon.mason@exar.com>
Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
When we enable a PCI device, we avoid doing a lot of the initial setup
work if the device's enable count is non-zero. If we don't fetch the
power state though, we may later fail to set up MSI due to the unknown
status. So pick it up before we short circuit the rest due to a
pre-existing enable or mismatched enable/disable pair (as happens with
VGA devices, which are special in a special way).
Tested-by: Jesse Brandeburg <jesse.brandeburg@gmail.com>
Reported-by: Dave Airlie <airlied@linux.ie>
Tested-by: Dave Airlie <airlied@linux.ie>
Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Not all hardware vendors hook up the PME line for legacy PCI devices,
meaning that wakeup events get lost. The only way around this is to poll
the devices to see if their state has changed, so add support for doing
that on legacy PCI devices that aren't part of the core chipset.
Acked-by: Rafael J. Wysocki <rjw@sisk.pl>
Signed-off-by: Matthew Garrett <mjg@redhat.com>
Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Indent the branch of an if.
The semantic match that finds this problem is as follows:
(http://coccinelle.lip6.fr/)
// <smpl>
@r disable braces4@
position p1,p2;
statement S1,S2;
@@
(
if (...) { ... }
|
if (...) S1@p1 S2@p2
)
@script:python@
p1 << r.p1;
p2 << r.p2;
@@
if (p1[0].column == p2[0].column):
cocci.print_main("branch",p1)
cocci.print_secs("after",p2)
// </smpl>
Signed-off-by: Julia Lawall <julia@diku.dk>
Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
* 'linux-next' of git://git.kernel.org/pub/scm/linux/kernel/git/jbarnes/pci-2.6: (30 commits)
PCI: update for owner removal from struct device_attribute
PCI: Fix warnings when CONFIG_DMI unset
PCI: Do not run NVidia quirks related to MSI with MSI disabled
x86/PCI: use for_each_pci_dev()
PCI: use for_each_pci_dev()
PCI: MSI: Restore read_msi_msg_desc(); add get_cached_msi_msg_desc()
PCI: export SMBIOS provided firmware instance and label to sysfs
PCI: Allow read/write access to sysfs I/O port resources
x86/PCI: use host bridge _CRS info on ASRock ALiveSATA2-GLAN
PCI: remove unused HAVE_ARCH_PCI_SET_DMA_MAX_SEGMENT_{SIZE|BOUNDARY}
PCI: disable mmio during bar sizing
PCI: MSI: Remove unsafe and unnecessary hardware access
PCI: Default PCIe ASPM control to on and require !EMBEDDED to disable
PCI: kernel oops on access to pci proc file while hot-removal
PCI: pci-sysfs: remove casts from void*
ACPI: Disable ASPM if the platform won't provide _OSC control for PCIe
PCI hotplug: make sure child bridges are enabled at hotplug time
PCI hotplug: shpchp: Removed check for hotplug of display devices
PCI hotplug: pciehp: Fixed return value sign for pciehp_unconfigure_device
PCI: Don't enable aspm before drivers have had a chance to veto it
...
In 2.6.34, we transformed the PCI DMA API into the generic device
mode. The PCI DMA API is just the wrapper of the DMA API.
So we don't need HAVE_ARCH_PCI_SET_DMA_MAX_SEGMENT_SIZE or
HAVE_ARCH_PCI_SET_DMA_SEGMENT_BOUNDARY (which enable architectures to
have the own implementations). Both haven't been used anyway.
Signed-off-by: FUJITA Tomonori <fujita.tomonori@lab.ntt.co.jp>
Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
One of the arguments during the suspend blockers discussion was that
the mainline kernel didn't contain any mechanisms making it possible
to avoid races between wakeup and system suspend.
Generally, there are two problems in that area. First, if a wakeup
event occurs exactly when /sys/power/state is being written to, it
may be delivered to user space right before the freezer kicks in, so
the user space consumer of the event may not be able to process it
before the system is suspended. Second, if a wakeup event occurs
after user space has been frozen, it is not generally guaranteed that
the ongoing transition of the system into a sleep state will be
aborted.
To address these issues introduce a new global sysfs attribute,
/sys/power/wakeup_count, associated with a running counter of wakeup
events and three helper functions, pm_stay_awake(), pm_relax(), and
pm_wakeup_event(), that may be used by kernel subsystems to control
the behavior of this attribute and to request the PM core to abort
system transitions into a sleep state already in progress.
The /sys/power/wakeup_count file may be read from or written to by
user space. Reads will always succeed (unless interrupted by a
signal) and return the current value of the wakeup events counter.
Writes, however, will only succeed if the written number is equal to
the current value of the wakeup events counter. If a write is
successful, it will cause the kernel to save the current value of the
wakeup events counter and to abort the subsequent system transition
into a sleep state if any wakeup events are reported after the write
has returned.
[The assumption is that before writing to /sys/power/state user space
will first read from /sys/power/wakeup_count. Next, user space
consumers of wakeup events will have a chance to acknowledge or
veto the upcoming system transition to a sleep state. Finally, if
the transition is allowed to proceed, /sys/power/wakeup_count will
be written to and if that succeeds, /sys/power/state will be written
to as well. Still, if any wakeup events are reported to the PM core
by kernel subsystems after that point, the transition will be
aborted.]
Additionally, put a wakeup events counter into struct dev_pm_info and
make these per-device wakeup event counters available via sysfs,
so that it's possible to check the activity of various wakeup event
sources within the kernel.
To illustrate how subsystems can use pm_wakeup_event(), make the
low-level PCI runtime PM wakeup-handling code use it.
Signed-off-by: Rafael J. Wysocki <rjw@sisk.pl>
Acked-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Acked-by: Greg Kroah-Hartman <gregkh@suse.de>
Acked-by: markgross <markgross@thegnar.org>
Reviewed-by: Alan Stern <stern@rowland.harvard.edu>
virtio-pci resets the device at startup by writing to the status
register, but this does not clear the pci config space,
specifically msi enable status which affects register
layout.
This breaks things like kdump when they try to use e.g. virtio-blk.
Fix by forcing msi off at startup. Since pci.c already has
a routine to do this, we export and use it instead of duplicating code.
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Tested-by: Vivek Goyal <vgoyal@redhat.com>
Acked-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Cc: linux-pci@vger.kernel.org
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Cc: stable@kernel.org
