Message ID | 20230330085357.2653599-5-damien.lemoal@opensource.wdc.com (mailing list archive) |
---|---|
State | Superseded |
Headers | show |
Series | PCI endpoint fixes and improvements | expand |
On Thu, Mar 30, 2023 at 05:53:44PM +0900, Damien Le Moal wrote: > pci_epf_test_data_transfer() and pci_epf_test_dma_callback() are not > handling DMA transfer completion correctly, leading to completion > notifications to the RC side that are too early. This problem can be > detected when the RC side is running an IOMMU with messages such as: > > pci-endpoint-test 0000:0b:00.0: AMD-Vi: Event logged [IO_PAGE_FAULT > domain=0x001c address=0xfff00000 flags=0x0000] > > When running the pcitest.sh tests: the address used for a previous > test transfer generates the above error while the next test transfer is > running. > > Fix this by testing the dma transfer status in > pci_epf_test_dma_callback() and notifying the completion only when the > transfer status is DMA_COMPLETE or DMA_ERROR. Furthermore, in > pci_epf_test_data_transfer(), be paranoid and check again the transfer > status and always call dmaengine_terminate_sync() before returning. > > Fixes: 8353813c88ef ("PCI: endpoint: Enable DMA tests for endpoints with DMA capabilities") > Cc: stable@vger.kernel.org > Signed-off-by: Damien Le Moal <damien.lemoal@opensource.wdc.com> Reviewed-by: Manivannan Sadhasivam <mani@kernel.org> - Mani > --- > drivers/pci/endpoint/functions/pci-epf-test.c | 38 +++++++++++++------ > 1 file changed, 27 insertions(+), 11 deletions(-) > > diff --git a/drivers/pci/endpoint/functions/pci-epf-test.c b/drivers/pci/endpoint/functions/pci-epf-test.c > index d65419735d2e..dbea6eb0dee7 100644 > --- a/drivers/pci/endpoint/functions/pci-epf-test.c > +++ b/drivers/pci/endpoint/functions/pci-epf-test.c > @@ -54,6 +54,9 @@ struct pci_epf_test { > struct delayed_work cmd_handler; > struct dma_chan *dma_chan_tx; > struct dma_chan *dma_chan_rx; > + struct dma_chan *transfer_chan; > + dma_cookie_t transfer_cookie; > + enum dma_status transfer_status; > struct completion transfer_complete; > bool dma_supported; > bool dma_private; > @@ -85,8 +88,14 @@ static size_t bar_size[] = { 512, 512, 1024, 16384, 131072, 1048576 }; > static void pci_epf_test_dma_callback(void *param) > { > struct pci_epf_test *epf_test = param; > - > - complete(&epf_test->transfer_complete); > + struct dma_tx_state state; > + > + epf_test->transfer_status = > + dmaengine_tx_status(epf_test->transfer_chan, > + epf_test->transfer_cookie, &state); > + if (epf_test->transfer_status == DMA_COMPLETE || > + epf_test->transfer_status == DMA_ERROR) > + complete(&epf_test->transfer_complete); > } > > /** > @@ -120,7 +129,6 @@ static int pci_epf_test_data_transfer(struct pci_epf_test *epf_test, > struct dma_async_tx_descriptor *tx; > struct dma_slave_config sconf = {}; > struct device *dev = &epf->dev; > - dma_cookie_t cookie; > int ret; > > if (IS_ERR_OR_NULL(chan)) { > @@ -152,25 +160,33 @@ static int pci_epf_test_data_transfer(struct pci_epf_test *epf_test, > } > > reinit_completion(&epf_test->transfer_complete); > + epf_test->transfer_chan = chan; > tx->callback = pci_epf_test_dma_callback; > tx->callback_param = epf_test; > - cookie = tx->tx_submit(tx); > + epf_test->transfer_cookie = tx->tx_submit(tx); > > - ret = dma_submit_error(cookie); > + ret = dma_submit_error(epf_test->transfer_cookie); > if (ret) { > - dev_err(dev, "Failed to do DMA tx_submit %d\n", cookie); > - return -EIO; > + dev_err(dev, "Failed to do DMA tx_submit %d\n", ret); > + goto terminate; > } > > dma_async_issue_pending(chan); > ret = wait_for_completion_interruptible(&epf_test->transfer_complete); > if (ret < 0) { > - dmaengine_terminate_sync(chan); > - dev_err(dev, "DMA wait_for_completion_timeout\n"); > - return -ETIMEDOUT; > + dev_err(dev, "DMA wait_for_completion interrupted\n"); > + goto terminate; > } > > - return 0; > + if (epf_test->transfer_status == DMA_ERROR) { > + dev_err(dev, "DMA transfer failed\n"); > + ret = -EIO; > + } > + > +terminate: > + dmaengine_terminate_sync(chan); > + > + return ret; > } > > struct epf_dma_filter { > -- > 2.39.2 >
On 2023/03/30 17:53, Damien Le Moal wrote: > pci_epf_test_data_transfer() and pci_epf_test_dma_callback() are not > handling DMA transfer completion correctly, leading to completion > notifications to the RC side that are too early. This problem can be > detected when the RC side is running an IOMMU with messages such as: > > pci-endpoint-test 0000:0b:00.0: AMD-Vi: Event logged [IO_PAGE_FAULT > domain=0x001c address=0xfff00000 flags=0x0000] > > When running the pcitest.sh tests: the address used for a previous > test transfer generates the above error while the next test transfer is > running. > > Fix this by testing the dma transfer status in > pci_epf_test_dma_callback() and notifying the completion only when the > transfer status is DMA_COMPLETE or DMA_ERROR. Furthermore, in > pci_epf_test_data_transfer(), be paranoid and check again the transfer > status and always call dmaengine_terminate_sync() before returning. > > Fixes: 8353813c88ef ("PCI: endpoint: Enable DMA tests for endpoints with DMA capabilities") > Cc: stable@vger.kernel.org > Signed-off-by: Damien Le Moal <damien.lemoal@opensource.wdc.com> > --- > drivers/pci/endpoint/functions/pci-epf-test.c | 38 +++++++++++++------ > 1 file changed, 27 insertions(+), 11 deletions(-) > > diff --git a/drivers/pci/endpoint/functions/pci-epf-test.c b/drivers/pci/endpoint/functions/pci-epf-test.c > index d65419735d2e..dbea6eb0dee7 100644 > --- a/drivers/pci/endpoint/functions/pci-epf-test.c > +++ b/drivers/pci/endpoint/functions/pci-epf-test.c > @@ -54,6 +54,9 @@ struct pci_epf_test { > struct delayed_work cmd_handler; > struct dma_chan *dma_chan_tx; > struct dma_chan *dma_chan_rx; > + struct dma_chan *transfer_chan; > + dma_cookie_t transfer_cookie; > + enum dma_status transfer_status; > struct completion transfer_complete; > bool dma_supported; > bool dma_private; > @@ -85,8 +88,14 @@ static size_t bar_size[] = { 512, 512, 1024, 16384, 131072, 1048576 }; > static void pci_epf_test_dma_callback(void *param) > { > struct pci_epf_test *epf_test = param; > - > - complete(&epf_test->transfer_complete); > + struct dma_tx_state state; > + > + epf_test->transfer_status = > + dmaengine_tx_status(epf_test->transfer_chan, > + epf_test->transfer_cookie, &state); > + if (epf_test->transfer_status == DMA_COMPLETE || > + epf_test->transfer_status == DMA_ERROR) > + complete(&epf_test->transfer_complete); > } > > /** > @@ -120,7 +129,6 @@ static int pci_epf_test_data_transfer(struct pci_epf_test *epf_test, > struct dma_async_tx_descriptor *tx; > struct dma_slave_config sconf = {}; > struct device *dev = &epf->dev; > - dma_cookie_t cookie; > int ret; > > if (IS_ERR_OR_NULL(chan)) { > @@ -152,25 +160,33 @@ static int pci_epf_test_data_transfer(struct pci_epf_test *epf_test, > } > > reinit_completion(&epf_test->transfer_complete); > + epf_test->transfer_chan = chan; > tx->callback = pci_epf_test_dma_callback; > tx->callback_param = epf_test; > - cookie = tx->tx_submit(tx); > + epf_test->transfer_cookie = tx->tx_submit(tx); How about changing the code to use dmaengine_submit() API instead of calling a raw function pointer? > > - ret = dma_submit_error(cookie); > + ret = dma_submit_error(epf_test->transfer_cookie); > if (ret) { > - dev_err(dev, "Failed to do DMA tx_submit %d\n", cookie); > - return -EIO; > + dev_err(dev, "Failed to do DMA tx_submit %d\n", ret); > + goto terminate; > } > > dma_async_issue_pending(chan); > ret = wait_for_completion_interruptible(&epf_test->transfer_complete); > if (ret < 0) { > - dmaengine_terminate_sync(chan); > - dev_err(dev, "DMA wait_for_completion_timeout\n"); > - return -ETIMEDOUT; > + dev_err(dev, "DMA wait_for_completion interrupted\n"); > + goto terminate; > } > > - return 0; > + if (epf_test->transfer_status == DMA_ERROR) { > + dev_err(dev, "DMA transfer failed\n"); > + ret = -EIO; > + } > + > +terminate: > + dmaengine_terminate_sync(chan); > + > + return ret; > } > > struct epf_dma_filter { Best, Shunsuke.
On 4/4/23 18:47, Shunsuke Mie wrote: >> @@ -120,7 +129,6 @@ static int pci_epf_test_data_transfer(struct pci_epf_test *epf_test, >> struct dma_async_tx_descriptor *tx; >> struct dma_slave_config sconf = {}; >> struct device *dev = &epf->dev; >> - dma_cookie_t cookie; >> int ret; >> >> if (IS_ERR_OR_NULL(chan)) { >> @@ -152,25 +160,33 @@ static int pci_epf_test_data_transfer(struct pci_epf_test *epf_test, >> } >> >> reinit_completion(&epf_test->transfer_complete); >> + epf_test->transfer_chan = chan; >> tx->callback = pci_epf_test_dma_callback; >> tx->callback_param = epf_test; >> - cookie = tx->tx_submit(tx); >> + epf_test->transfer_cookie = tx->tx_submit(tx); > > How about changing the code to use dmaengine_submit() API instead of > calling a raw function pointer? This is done in patch 5 of the series.
2023年4月4日(火) 19:17 Damien Le Moal <dlemoal@fastmail.com>: > > On 4/4/23 18:47, Shunsuke Mie wrote: > >> @@ -120,7 +129,6 @@ static int pci_epf_test_data_transfer(struct pci_epf_test *epf_test, > >> struct dma_async_tx_descriptor *tx; > >> struct dma_slave_config sconf = {}; > >> struct device *dev = &epf->dev; > >> - dma_cookie_t cookie; > >> int ret; > >> > >> if (IS_ERR_OR_NULL(chan)) { > >> @@ -152,25 +160,33 @@ static int pci_epf_test_data_transfer(struct pci_epf_test *epf_test, > >> } > >> > >> reinit_completion(&epf_test->transfer_complete); > >> + epf_test->transfer_chan = chan; > >> tx->callback = pci_epf_test_dma_callback; > >> tx->callback_param = epf_test; > >> - cookie = tx->tx_submit(tx); > >> + epf_test->transfer_cookie = tx->tx_submit(tx); > > > > How about changing the code to use dmaengine_submit() API instead of > > calling a raw function pointer? > > This is done in patch 5 of the series. Sorry, I missed it. >
On 4/4/23 19:16, Damien Le Moal wrote: > On 4/4/23 18:47, Shunsuke Mie wrote: >>> @@ -120,7 +129,6 @@ static int pci_epf_test_data_transfer(struct pci_epf_test *epf_test, >>> struct dma_async_tx_descriptor *tx; >>> struct dma_slave_config sconf = {}; >>> struct device *dev = &epf->dev; >>> - dma_cookie_t cookie; >>> int ret; >>> >>> if (IS_ERR_OR_NULL(chan)) { >>> @@ -152,25 +160,33 @@ static int pci_epf_test_data_transfer(struct pci_epf_test *epf_test, >>> } >>> >>> reinit_completion(&epf_test->transfer_complete); >>> + epf_test->transfer_chan = chan; >>> tx->callback = pci_epf_test_dma_callback; >>> tx->callback_param = epf_test; >>> - cookie = tx->tx_submit(tx); >>> + epf_test->transfer_cookie = tx->tx_submit(tx); >> >> How about changing the code to use dmaengine_submit() API instead of >> calling a raw function pointer? > > This is done in patch 5 of the series. Bjorn, My apologies for not replying to the series cover letter as I do not have it locally (due to WD on-going network issues). This is a ping about this series. Can we get it queued ? (I do not see it in PCI next tree)
On Thu, Apr 13, 2023 at 10:50:03AM +0900, Damien Le Moal wrote: > On 4/4/23 19:16, Damien Le Moal wrote: > > On 4/4/23 18:47, Shunsuke Mie wrote: > >>> @@ -120,7 +129,6 @@ static int pci_epf_test_data_transfer(struct pci_epf_test *epf_test, > >>> struct dma_async_tx_descriptor *tx; > >>> struct dma_slave_config sconf = {}; > >>> struct device *dev = &epf->dev; > >>> - dma_cookie_t cookie; > >>> int ret; > >>> > >>> if (IS_ERR_OR_NULL(chan)) { > >>> @@ -152,25 +160,33 @@ static int pci_epf_test_data_transfer(struct pci_epf_test *epf_test, > >>> } > >>> > >>> reinit_completion(&epf_test->transfer_complete); > >>> + epf_test->transfer_chan = chan; > >>> tx->callback = pci_epf_test_dma_callback; > >>> tx->callback_param = epf_test; > >>> - cookie = tx->tx_submit(tx); > >>> + epf_test->transfer_cookie = tx->tx_submit(tx); > >> > >> How about changing the code to use dmaengine_submit() API instead of > >> calling a raw function pointer? > > > > This is done in patch 5 of the series. > > Bjorn, > > My apologies for not replying to the series cover letter as I do not have it > locally (due to WD on-going network issues). > > This is a ping about this series. Can we get it queued ? > (I do not see it in PCI next tree) Lorenzo and Krzysztof take care of the endpoint subsystem, so I'll wait for them. In the meantime I'll send a couple typo fixes and a cursory review for trivial issue.s Bjorn
On Thu, Apr 13, 2023 at 10:50:03AM +0900, Damien Le Moal wrote: > My apologies for not replying to the series cover letter as I do not have it > locally (due to WD on-going network issues). Oh, I forgot to mention that in this situation where a message is lost for some reason (or if I want to reply to a thread that I didn't receive directly), lore is a great resource. For example, here's the cover letter on lore: https://lore.kernel.org/r/20230330085357.2653599-1-damien.lemoal@opensource.wdcom and there's a link to download the entire thread as an mbox file. Then it's easy to uncompress it, point mutt at it, and respond. There's also a sample "git send-email" line to respond, but I've never actually used that. Might not fit your workflow, but just FYI in case it's ever useful. Bjorn
On Thu, Mar 30, 2023 at 05:53:44PM +0900, Damien Le Moal wrote: > pci_epf_test_data_transfer() and pci_epf_test_dma_callback() are not > handling DMA transfer completion correctly, leading to completion > notifications to the RC side that are too early. This problem can be > detected when the RC side is running an IOMMU with messages such as: > > pci-endpoint-test 0000:0b:00.0: AMD-Vi: Event logged [IO_PAGE_FAULT > domain=0x001c address=0xfff00000 flags=0x0000] Indent quoted material two spaces. This looks like the original message was a single line. I'd keep it that way in the commit log to make it easier to grep rot. > When running the pcitest.sh tests: the address used for a previous > test transfer generates the above error while the next test transfer is > running. > > Fix this by testing the dma transfer status in > pci_epf_test_dma_callback() and notifying the completion only when the > transfer status is DMA_COMPLETE or DMA_ERROR. Furthermore, in > pci_epf_test_data_transfer(), be paranoid and check again the transfer > status and always call dmaengine_terminate_sync() before returning. s/the dma transfer/the DMA transfer/
diff --git a/drivers/pci/endpoint/functions/pci-epf-test.c b/drivers/pci/endpoint/functions/pci-epf-test.c index d65419735d2e..dbea6eb0dee7 100644 --- a/drivers/pci/endpoint/functions/pci-epf-test.c +++ b/drivers/pci/endpoint/functions/pci-epf-test.c @@ -54,6 +54,9 @@ struct pci_epf_test { struct delayed_work cmd_handler; struct dma_chan *dma_chan_tx; struct dma_chan *dma_chan_rx; + struct dma_chan *transfer_chan; + dma_cookie_t transfer_cookie; + enum dma_status transfer_status; struct completion transfer_complete; bool dma_supported; bool dma_private; @@ -85,8 +88,14 @@ static size_t bar_size[] = { 512, 512, 1024, 16384, 131072, 1048576 }; static void pci_epf_test_dma_callback(void *param) { struct pci_epf_test *epf_test = param; - - complete(&epf_test->transfer_complete); + struct dma_tx_state state; + + epf_test->transfer_status = + dmaengine_tx_status(epf_test->transfer_chan, + epf_test->transfer_cookie, &state); + if (epf_test->transfer_status == DMA_COMPLETE || + epf_test->transfer_status == DMA_ERROR) + complete(&epf_test->transfer_complete); } /** @@ -120,7 +129,6 @@ static int pci_epf_test_data_transfer(struct pci_epf_test *epf_test, struct dma_async_tx_descriptor *tx; struct dma_slave_config sconf = {}; struct device *dev = &epf->dev; - dma_cookie_t cookie; int ret; if (IS_ERR_OR_NULL(chan)) { @@ -152,25 +160,33 @@ static int pci_epf_test_data_transfer(struct pci_epf_test *epf_test, } reinit_completion(&epf_test->transfer_complete); + epf_test->transfer_chan = chan; tx->callback = pci_epf_test_dma_callback; tx->callback_param = epf_test; - cookie = tx->tx_submit(tx); + epf_test->transfer_cookie = tx->tx_submit(tx); - ret = dma_submit_error(cookie); + ret = dma_submit_error(epf_test->transfer_cookie); if (ret) { - dev_err(dev, "Failed to do DMA tx_submit %d\n", cookie); - return -EIO; + dev_err(dev, "Failed to do DMA tx_submit %d\n", ret); + goto terminate; } dma_async_issue_pending(chan); ret = wait_for_completion_interruptible(&epf_test->transfer_complete); if (ret < 0) { - dmaengine_terminate_sync(chan); - dev_err(dev, "DMA wait_for_completion_timeout\n"); - return -ETIMEDOUT; + dev_err(dev, "DMA wait_for_completion interrupted\n"); + goto terminate; } - return 0; + if (epf_test->transfer_status == DMA_ERROR) { + dev_err(dev, "DMA transfer failed\n"); + ret = -EIO; + } + +terminate: + dmaengine_terminate_sync(chan); + + return ret; } struct epf_dma_filter {
pci_epf_test_data_transfer() and pci_epf_test_dma_callback() are not handling DMA transfer completion correctly, leading to completion notifications to the RC side that are too early. This problem can be detected when the RC side is running an IOMMU with messages such as: pci-endpoint-test 0000:0b:00.0: AMD-Vi: Event logged [IO_PAGE_FAULT domain=0x001c address=0xfff00000 flags=0x0000] When running the pcitest.sh tests: the address used for a previous test transfer generates the above error while the next test transfer is running. Fix this by testing the dma transfer status in pci_epf_test_dma_callback() and notifying the completion only when the transfer status is DMA_COMPLETE or DMA_ERROR. Furthermore, in pci_epf_test_data_transfer(), be paranoid and check again the transfer status and always call dmaengine_terminate_sync() before returning. Fixes: 8353813c88ef ("PCI: endpoint: Enable DMA tests for endpoints with DMA capabilities") Cc: stable@vger.kernel.org Signed-off-by: Damien Le Moal <damien.lemoal@opensource.wdc.com> --- drivers/pci/endpoint/functions/pci-epf-test.c | 38 +++++++++++++------ 1 file changed, 27 insertions(+), 11 deletions(-)