From f7da7782aba92593f7b82f03d2409a1c5f4db91b Mon Sep 17 00:00:00 2001 From: Lukas Wunner Date: Wed, 23 Jan 2019 09:26:00 +0100 Subject: [PATCH 1/2] dmaengine: bcm2835: Fix interrupt race on RT If IRQ handlers are threaded (either because CONFIG_PREEMPT_RT_BASE is enabled or "threadirqs" was passed on the command line) and if system load is sufficiently high that wakeup latency of IRQ threads degrades, SPI DMA transactions on the BCM2835 occasionally break like this: ks8851 spi0.0: SPI transfer timed out bcm2835-dma 3f007000.dma: DMA transfer could not be terminated ks8851 spi0.0 eth2: ks8851_rdfifo: spi_sync() failed The root cause is an assumption made by the DMA driver which is documented in a code comment in bcm2835_dma_terminate_all(): /* * Stop DMA activity: we assume the callback will not be called * after bcm_dma_abort() returns (even if it does, it will see * c->desc is NULL and exit.) */ That assumption falls apart if the IRQ handler bcm2835_dma_callback() is threaded: A client may terminate a descriptor and issue a new one before the IRQ handler had a chance to run. In fact the IRQ handler may miss an *arbitrary* number of descriptors. The result is the following race condition: 1. A descriptor finishes, its interrupt is deferred to the IRQ thread. 2. A client calls dma_terminate_async() which sets channel->desc = NULL. 3. The client issues a new descriptor. Because channel->desc is NULL, bcm2835_dma_issue_pending() immediately starts the descriptor. 4. Finally the IRQ thread runs and writes BCM2835_DMA_INT to the CS register to acknowledge the interrupt. This clears the ACTIVE flag, so the newly issued descriptor is paused in the middle of the transaction. Because channel->desc is not NULL, the IRQ thread finalizes the descriptor and tries to start the next one. I see two possible solutions: The first is to call synchronize_irq() in bcm2835_dma_issue_pending() to wait until the IRQ thread has finished before issuing a new descriptor. The downside of this approach is unnecessary latency if clients desire rapidly terminating and re-issuing descriptors and don't have any use for an IRQ callback. (The SPI TX DMA channel is a case in point.) A better alternative is to make the IRQ thread recognize that it has missed descriptors and avoid finalizing the newly issued descriptor. So first of all, set the ACTIVE flag when acknowledging the interrupt. This keeps a newly issued descriptor running. If the descriptor was finished, the channel remains idle despite the ACTIVE flag being set. However the ACTIVE flag can then no longer be used to check whether the channel is idle, so instead check whether the register containing the current control block address is zero and finalize the current descriptor only if so. That way, there is no impact on latency and throughput if the client doesn't care for the interrupt: Only minimal additional overhead is introduced for non-cyclic descriptors as one further MMIO read is necessary per interrupt to check for idleness of the channel. Cyclic descriptors are sped up slightly by removing one MMIO write per interrupt. Fixes: 96286b576690 ("dmaengine: Add support for BCM2835") Signed-off-by: Lukas Wunner Cc: stable@vger.kernel.org # v3.14+ Cc: Frank Pavlic Cc: Martin Sperl Cc: Florian Meier Cc: Clive Messer Cc: Matthias Reichl Tested-by: Stefan Wahren Acked-by: Florian Kauer Signed-off-by: Vinod Koul --- drivers/dma/bcm2835-dma.c | 33 ++++++++++++++++++--------------- 1 file changed, 18 insertions(+), 15 deletions(-) diff --git a/drivers/dma/bcm2835-dma.c b/drivers/dma/bcm2835-dma.c index 1a44c8086d77..0c3f5c71bb48 100644 --- a/drivers/dma/bcm2835-dma.c +++ b/drivers/dma/bcm2835-dma.c @@ -412,7 +412,12 @@ static int bcm2835_dma_abort(void __iomem *chan_base) long int timeout = 10000; cs = readl(chan_base + BCM2835_DMA_CS); - if (!(cs & BCM2835_DMA_ACTIVE)) + + /* + * A zero control block address means the channel is idle. + * (The ACTIVE flag in the CS register is not a reliable indicator.) + */ + if (!readl(chan_base + BCM2835_DMA_ADDR)) return 0; /* Write 0 to the active bit - Pause the DMA */ @@ -476,8 +481,15 @@ static irqreturn_t bcm2835_dma_callback(int irq, void *data) spin_lock_irqsave(&c->vc.lock, flags); - /* Acknowledge interrupt */ - writel(BCM2835_DMA_INT, c->chan_base + BCM2835_DMA_CS); + /* + * Clear the INT flag to receive further interrupts. Keep the channel + * active in case the descriptor is cyclic or in case the client has + * already terminated the descriptor and issued a new one. (May happen + * if this IRQ handler is threaded.) If the channel is finished, it + * will remain idle despite the ACTIVE flag being set. + */ + writel(BCM2835_DMA_INT | BCM2835_DMA_ACTIVE, + c->chan_base + BCM2835_DMA_CS); d = c->desc; @@ -485,11 +497,7 @@ static irqreturn_t bcm2835_dma_callback(int irq, void *data) if (d->cyclic) { /* call the cyclic callback */ vchan_cyclic_callback(&d->vd); - - /* Keep the DMA engine running */ - writel(BCM2835_DMA_ACTIVE, - c->chan_base + BCM2835_DMA_CS); - } else { + } else if (!readl(c->chan_base + BCM2835_DMA_ADDR)) { vchan_cookie_complete(&c->desc->vd); bcm2835_dma_start_desc(c); } @@ -789,11 +797,7 @@ static int bcm2835_dma_terminate_all(struct dma_chan *chan) list_del_init(&c->node); spin_unlock(&d->lock); - /* - * Stop DMA activity: we assume the callback will not be called - * after bcm_dma_abort() returns (even if it does, it will see - * c->desc is NULL and exit.) - */ + /* stop DMA activity */ if (c->desc) { vchan_terminate_vdesc(&c->desc->vd); c->desc = NULL; @@ -801,8 +805,7 @@ static int bcm2835_dma_terminate_all(struct dma_chan *chan) /* Wait for stopping */ while (--timeout) { - if (!(readl(c->chan_base + BCM2835_DMA_CS) & - BCM2835_DMA_ACTIVE)) + if (!readl(c->chan_base + BCM2835_DMA_ADDR)) break; cpu_relax(); From 9e528c799d17a4ac37d788c81440b50377dd592d Mon Sep 17 00:00:00 2001 From: Lukas Wunner Date: Wed, 23 Jan 2019 09:26:00 +0100 Subject: [PATCH 2/2] dmaengine: bcm2835: Fix abort of transactions There are multiple issues with bcm2835_dma_abort() (which is called on termination of a transaction): * The algorithm to abort the transaction first pauses the channel by clearing the ACTIVE flag in the CS register, then waits for the PAUSED flag to clear. Page 49 of the spec documents the latter as follows: "Indicates if the DMA is currently paused and not transferring data. This will occur if the active bit has been cleared [...]" https://www.raspberrypi.org/app/uploads/2012/02/BCM2835-ARM-Peripherals.pdf So the function is entering an infinite loop because it is waiting for PAUSED to clear which is always set due to the function having cleared the ACTIVE flag. The only thing that's saving it from itself is the upper bound of 10000 loop iterations. The code comment says that the intention is to "wait for any current AXI transfer to complete", so the author probably wanted to check the WAITING_FOR_OUTSTANDING_WRITES flag instead. Amend the function accordingly. * The CS register is only read at the beginning of the function. It needs to be read again after pausing the channel and before checking for outstanding writes, otherwise writes which were issued between the register read at the beginning of the function and pausing the channel may not be waited for. * The function seeks to abort the transfer by writing 0 to the NEXTCONBK register and setting the ABORT and ACTIVE flags. Thereby, the 0 in NEXTCONBK is sought to be loaded into the CONBLK_AD register. However experimentation has shown this approach to not work: The CONBLK_AD register remains the same as before and the CS register contains 0x00000030 (PAUSED | DREQ_STOPS_DMA). In other words, the control block is not aborted but merely paused and it will be resumed once the next DMA transaction is started. That is absolutely not the desired behavior. A simpler approach is to set the channel's RESET flag instead. This reliably zeroes the NEXTCONBK as well as the CS register. It requires less code and only a single MMIO write. This is also what popular user space DMA drivers do, e.g.: https://github.com/metachris/RPIO/blob/master/source/c_pwm/pwm.c Note that the spec is contradictory whether the NEXTCONBK register is writeable at all. On the one hand, page 41 claims: "The value loaded into the NEXTCONBK register can be overwritten so that the linked list of Control Block data structures can be dynamically altered. However it is only safe to do this when the DMA is paused." On the other hand, page 40 specifies: "Only three registers in each channel's register set are directly writeable (CS, CONBLK_AD and DEBUG). The other registers (TI, SOURCE_AD, DEST_AD, TXFR_LEN, STRIDE & NEXTCONBK), are automatically loaded from a Control Block data structure held in external memory." Fixes: 96286b576690 ("dmaengine: Add support for BCM2835") Signed-off-by: Lukas Wunner Cc: stable@vger.kernel.org # v3.14+ Cc: Frank Pavlic Cc: Martin Sperl Cc: Florian Meier Cc: Clive Messer Cc: Matthias Reichl Tested-by: Stefan Wahren Acked-by: Florian Kauer Signed-off-by: Vinod Koul --- drivers/dma/bcm2835-dma.c | 41 +++++++++------------------------------ 1 file changed, 9 insertions(+), 32 deletions(-) diff --git a/drivers/dma/bcm2835-dma.c b/drivers/dma/bcm2835-dma.c index 0c3f5c71bb48..ae10f5614f95 100644 --- a/drivers/dma/bcm2835-dma.c +++ b/drivers/dma/bcm2835-dma.c @@ -406,13 +406,11 @@ static void bcm2835_dma_fill_cb_chain_with_sg( } } -static int bcm2835_dma_abort(void __iomem *chan_base) +static int bcm2835_dma_abort(struct bcm2835_chan *c) { - unsigned long cs; + void __iomem *chan_base = c->chan_base; long int timeout = 10000; - cs = readl(chan_base + BCM2835_DMA_CS); - /* * A zero control block address means the channel is idle. * (The ACTIVE flag in the CS register is not a reliable indicator.) @@ -424,25 +422,16 @@ static int bcm2835_dma_abort(void __iomem *chan_base) writel(0, chan_base + BCM2835_DMA_CS); /* Wait for any current AXI transfer to complete */ - while ((cs & BCM2835_DMA_ISPAUSED) && --timeout) { + while ((readl(chan_base + BCM2835_DMA_CS) & + BCM2835_DMA_WAITING_FOR_WRITES) && --timeout) cpu_relax(); - cs = readl(chan_base + BCM2835_DMA_CS); - } - /* We'll un-pause when we set of our next DMA */ + /* Peripheral might be stuck and fail to signal AXI write responses */ if (!timeout) - return -ETIMEDOUT; - - if (!(cs & BCM2835_DMA_ACTIVE)) - return 0; - - /* Terminate the control block chain */ - writel(0, chan_base + BCM2835_DMA_NEXTCB); - - /* Abort the whole DMA */ - writel(BCM2835_DMA_ABORT | BCM2835_DMA_ACTIVE, - chan_base + BCM2835_DMA_CS); + dev_err(c->vc.chan.device->dev, + "failed to complete outstanding writes\n"); + writel(BCM2835_DMA_RESET, chan_base + BCM2835_DMA_CS); return 0; } @@ -787,7 +776,6 @@ static int bcm2835_dma_terminate_all(struct dma_chan *chan) struct bcm2835_chan *c = to_bcm2835_dma_chan(chan); struct bcm2835_dmadev *d = to_bcm2835_dma_dev(c->vc.chan.device); unsigned long flags; - int timeout = 10000; LIST_HEAD(head); spin_lock_irqsave(&c->vc.lock, flags); @@ -801,18 +789,7 @@ static int bcm2835_dma_terminate_all(struct dma_chan *chan) if (c->desc) { vchan_terminate_vdesc(&c->desc->vd); c->desc = NULL; - bcm2835_dma_abort(c->chan_base); - - /* Wait for stopping */ - while (--timeout) { - if (!readl(c->chan_base + BCM2835_DMA_ADDR)) - break; - - cpu_relax(); - } - - if (!timeout) - dev_err(d->ddev.dev, "DMA transfer could not be terminated\n"); + bcm2835_dma_abort(c); } vchan_get_all_descriptors(&c->vc, &head);