[1/2] spi: rspi: Fix leaking of unused DMA descriptors

Commit Message

Geert Uytterhoeven July 11, 2014, 4:06 p.m. UTC

If dmaengine_prep_slave_sg() or dmaengine_submit() fail, we may leak
unused DMA descriptors.

As per Documentation/dmaengine.txt, once a DMA descriptor has been
obtained, it must be submitted. Hence:
  - First prepare and submit all DMA descriptors,
  - Prepare the SPI controller for DMA,
  - Start DMA by calling dma_async_issue_pending(),
  - Make sure to call dmaengine_terminate_all() on all descriptors that
    haven't completed.

Reported-by: Laurent Pinchart <[e-mail]>
Signed-off-by: Geert Uytterhoeven <[e-mail]>
---
 drivers/spi/spi-rspi.c | 94 +++++++++++++++++++++++++++++++-------------------
 1 file changed, 58 insertions(+), 36 deletions(-)

Comments

Mark Brown July 14, 2014, 6:58 p.m. UTC | #1

On Fri, Jul 11, 2014 at 06:06:41PM +0200, Geert Uytterhoeven wrote:
> If dmaengine_prep_slave_sg() or dmaengine_submit() fail, we may leak
> unused DMA descriptors.

This doesn't apply, I guess it may depend on some of your other DMA
related patches but it's not immediately obvious and it's not been sent
as part of a series.  Can you please check and/or clarify?

Geert Uytterhoeven Aug. 5, 2014, 2:34 p.m. UTC | #2

Hi Mark,

On Mon, Jul 14, 2014 at 8:58 PM, Mark Brown <[e-mail]> wrote:
> On Fri, Jul 11, 2014 at 06:06:41PM +0200, Geert Uytterhoeven wrote:
>> If dmaengine_prep_slave_sg() or dmaengine_submit() fail, we may leak
>> unused DMA descriptors.
>
> This doesn't apply, I guess it may depend on some of your other DMA
> related patches but it's not immediately obvious and it's not been sent
> as part of a series.  Can you please check and/or clarify?

Indeed, this (and the rspi one) depended on the two "Handle
dmaengine_prep_slave_single() failures gracefully" patches.

Unless you recover them from e.g. patchwork, I'll resend them with my
next batch.

Sorry for the fuzz.

Gr{oetje,eeting}s,

                        Geert

--
Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- [e-mail]

In personal conversations with technical people, I call myself a hacker. But
when I'm talking to journalists I just say "programmer" or something like that.
                                -- Linus Torvalds
--
To unsubscribe from this list: send the line "unsubscribe dmaengine" in
the body of a message to [e-mail]
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Patch

diff --git a/drivers/spi/spi-rspi.c b/drivers/spi/spi-rspi.c
index c850dfdfa9e3..ad87a98f8f68 100644
--- a/drivers/spi/spi-rspi.c
+++ b/drivers/spi/spi-rspi.c
@@ -472,25 +472,52 @@  static int rspi_dma_transfer(struct rspi_data *rspi, struct sg_table *tx,
 	dma_cookie_t cookie;
 	int ret;
 
-	if (tx) {
-		desc_tx = dmaengine_prep_slave_sg(rspi->master->dma_tx,
-					tx->sgl, tx->nents, DMA_TO_DEVICE,
-					DMA_PREP_INTERRUPT | DMA_CTRL_ACK);
-		if (!desc_tx)
-			goto no_dma;
-
-		irq_mask |= SPCR_SPTIE;
-	}
+	/* First prepare and submit the DMA request(s), as this may fail */
 	if (rx) {
 		desc_rx = dmaengine_prep_slave_sg(rspi->master->dma_rx,
 					rx->sgl, rx->nents, DMA_FROM_DEVICE,
 					DMA_PREP_INTERRUPT | DMA_CTRL_ACK);
-		if (!desc_rx)
-			goto no_dma;
+		if (!desc_rx) {
+			ret = -EAGAIN;
+			goto no_dma_rx;
+		}
+
+		desc_rx->callback = rspi_dma_complete;
+		desc_rx->callback_param = rspi;
+		cookie = dmaengine_submit(desc_rx);
+		if (dma_submit_error(cookie)) {
+			ret = cookie;
+			goto no_dma_rx;
+		}
 
 		irq_mask |= SPCR_SPRIE;
 	}
 
+	if (tx) {
+		desc_tx = dmaengine_prep_slave_sg(rspi->master->dma_tx,
+					tx->sgl, tx->nents, DMA_TO_DEVICE,
+					DMA_PREP_INTERRUPT | DMA_CTRL_ACK);
+		if (!desc_tx) {
+			ret = -EAGAIN;
+			goto no_dma_tx;
+		}
+
+		if (rx) {
+			/* No callback */
+			desc_tx->callback = NULL;
+		} else {
+			desc_tx->callback = rspi_dma_complete;
+			desc_tx->callback_param = rspi;
+		}
+		cookie = dmaengine_submit(desc_tx);
+		if (dma_submit_error(cookie)) {
+			ret = cookie;
+			goto no_dma_tx;
+		}
+
+		irq_mask |= SPCR_SPTIE;
+	}
+
 	/*
 	 * DMAC needs SPxIE, but if SPxIE is set, the IRQ routine will be
 	 * called. So, this driver disables the IRQ while DMA transfer.
@@ -503,34 +530,24 @@  static int rspi_dma_transfer(struct rspi_data *rspi, struct sg_table *tx,
 	rspi_enable_irq(rspi, irq_mask);
 	rspi->dma_callbacked = 0;
 
-	if (rx) {
-		desc_rx->callback = rspi_dma_complete;
-		desc_rx->callback_param = rspi;
-		cookie = dmaengine_submit(desc_rx);
-		if (dma_submit_error(cookie))
-			return cookie;
+	/* Now start DMA */
+	if (rx)
 		dma_async_issue_pending(rspi->master->dma_rx);
-	}
-	if (tx) {
-		if (rx) {
-			/* No callback */
-			desc_tx->callback = NULL;
-		} else {
-			desc_tx->callback = rspi_dma_complete;
-			desc_tx->callback_param = rspi;
-		}
-		cookie = dmaengine_submit(desc_tx);
-		if (dma_submit_error(cookie))
-			return cookie;
+	if (tx)
 		dma_async_issue_pending(rspi->master->dma_tx);
-	}
 
 	ret = wait_event_interruptible_timeout(rspi->wait,
 					       rspi->dma_callbacked, HZ);
 	if (ret > 0 && rspi->dma_callbacked)
 		ret = 0;
-	else if (!ret)
+	else if (!ret) {
+		dev_err(&rspi->master->dev, "DMA timeout\n");
 		ret = -ETIMEDOUT;
+		if (tx)
+			dmaengine_terminate_all(rspi->master->dma_tx);
+		if (rx)
+			dmaengine_terminate_all(rspi->master->dma_rx);
+	}
 
 	rspi_disable_irq(rspi, irq_mask);
 
@@ -541,11 +558,16 @@  static int rspi_dma_transfer(struct rspi_data *rspi, struct sg_table *tx,
 
 	return ret;
 
-no_dma:
-	pr_warn_once("%s %s: DMA not available, falling back to PIO\n",
-		     dev_driver_string(&rspi->master->dev),
-		     dev_name(&rspi->master->dev));
-	return -EAGAIN;
+no_dma_tx:
+	if (rx)
+		dmaengine_terminate_all(rspi->master->dma_rx);
+no_dma_rx:
+	if (ret == -EAGAIN) {
+		pr_warn_once("%s %s: DMA not available, falling back to PIO\n",
+			     dev_driver_string(&rspi->master->dev),
+			     dev_name(&rspi->master->dev));
+	}
+	return ret;
 }
 
 static void rspi_receive_init(const struct rspi_data *rspi)