| Message ID | 1356636228-23096-2-git-send-email-lkundrak@v3.sk (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
Dear Lubomir Rintel, On Thu, 27 Dec 2012 20:23:48 +0100, Lubomir Rintel wrote: > dma_sync_single_for_cpu(dma_chan->device->dev, dest_dma, > MV_XOR_TEST_SIZE, DMA_FROM_DEVICE); > + dma_unmap_single(dma_chan->device->dev, dest_dma, MV_XOR_TEST_SIZE, > + DMA_FROM_DEVICE); I would assume that dma_unmap_single() implies a dma_sync_single_for_cpu() since you're unmapping the DMA buffer. So if you use dma_unmap_single(), I think you can remove the call to dma_sync_single_for_cpu(). > dma_sync_single_for_cpu(dma_chan->device->dev, dest_dma, > PAGE_SIZE, DMA_FROM_DEVICE); > + dma_unmap_page(dma_chan->device->dev, dest_dma, PAGE_SIZE, > + DMA_FROM_DEVICE); Ditto. Also, the mv_xor_memcpy_self_test() function not only dma_map_single() the destination buffer, but also the source buffer. So presumably, the source buffer should also be dma_unmap_single()'d. And for the mv_xor_xor_self_test() function, multiple source buffers are dma_map_page()'d, so they should all be dma_unmap_page()'d I guess, not only the destination buffer. Does that make sense? Thanks, Thomas
On Fri, 2013-01-04 at 17:10 +0100, Thomas Petazzoni wrote: > Dear Lubomir Rintel, > > On Thu, 27 Dec 2012 20:23:48 +0100, Lubomir Rintel wrote: > > > dma_sync_single_for_cpu(dma_chan->device->dev, dest_dma, > > MV_XOR_TEST_SIZE, DMA_FROM_DEVICE); > > + dma_unmap_single(dma_chan->device->dev, dest_dma, MV_XOR_TEST_SIZE, > > + DMA_FROM_DEVICE); > > I would assume that dma_unmap_single() implies a > dma_sync_single_for_cpu() since you're unmapping the DMA buffer. So if > you use dma_unmap_single(), I think you can remove the call to > dma_sync_single_for_cpu(). Yes, it indeed seems to be the case (comment for arm_dma_unmap_page(): "After this call, reads by the CPU to the buffer are guaranteed to see whatever the device wrote there.". > > dma_sync_single_for_cpu(dma_chan->device->dev, dest_dma, > > PAGE_SIZE, DMA_FROM_DEVICE); > > + dma_unmap_page(dma_chan->device->dev, dest_dma, PAGE_SIZE, > > + DMA_FROM_DEVICE); > > Ditto. > > Also, the mv_xor_memcpy_self_test() function not only dma_map_single() > the destination buffer, but also the source buffer. So presumably, the > source buffer should also be dma_unmap_single()'d. > > And for the mv_xor_xor_self_test() function, multiple source buffers > are dma_map_page()'d, so they should all be dma_unmap_page()'d I guess, > not only the destination buffer. Those get released by the mv_xor_run_tx_complete_actions() callback. I will follow up with an updated patch. Regards, Lubo
On Sun, 2013-01-13 at 14:18 +0100, Lubomir Rintel wrote: > On Fri, 2013-01-04 at 17:10 +0100, Thomas Petazzoni wrote: > > Dear Lubomir Rintel, > > > > On Thu, 27 Dec 2012 20:23:48 +0100, Lubomir Rintel wrote: > > > > > dma_sync_single_for_cpu(dma_chan->device->dev, dest_dma, > > > MV_XOR_TEST_SIZE, DMA_FROM_DEVICE); > > > + dma_unmap_single(dma_chan->device->dev, dest_dma, MV_XOR_TEST_SIZE, > > > + DMA_FROM_DEVICE); > > > > I would assume that dma_unmap_single() implies a > > dma_sync_single_for_cpu() since you're unmapping the DMA buffer. So if > > you use dma_unmap_single(), I think you can remove the call to > > dma_sync_single_for_cpu(). > > Yes, it indeed seems to be the case (comment for arm_dma_unmap_page(): > "After this call, reads by the CPU to the buffer are guaranteed to see > whatever the device wrote there.". > > > > dma_sync_single_for_cpu(dma_chan->device->dev, dest_dma, > > > PAGE_SIZE, DMA_FROM_DEVICE); > > > + dma_unmap_page(dma_chan->device->dev, dest_dma, PAGE_SIZE, > > > + DMA_FROM_DEVICE); > > > > Ditto. > > > > Also, the mv_xor_memcpy_self_test() function not only dma_map_single() > > the destination buffer, but also the source buffer. So presumably, the > > source buffer should also be dma_unmap_single()'d. > > > > And for the mv_xor_xor_self_test() function, multiple source buffers > > are dma_map_page()'d, so they should all be dma_unmap_page()'d I guess, > > not only the destination buffer. > > Those get released by the mv_xor_run_tx_complete_actions() callback. > > I will follow up with an updated patch. Actually, after cleaning up the sync a couple more selfcheck warnings popped up. I'd be very thankful if you could take a look (patches chained to this message).
diff --git a/drivers/dma/mv_xor.c b/drivers/dma/mv_xor.c index 39387df..f642d8b 100644 --- a/drivers/dma/mv_xor.c +++ b/drivers/dma/mv_xor.c @@ -946,7 +946,7 @@ static int mv_xor_memcpy_self_test(struct mv_xor_chan *mv_chan) tx = mv_xor_prep_dma_memcpy(dma_chan, dest_dma, src_dma, MV_XOR_TEST_SIZE, DMA_COMPL_SRC_UNMAP_SINGLE | - DMA_COMPL_DEST_UNMAP_SINGLE); + DMA_COMPL_SKIP_DEST_UNMAP); cookie = mv_xor_tx_submit(tx); mv_xor_issue_pending(dma_chan); async_tx_ack(tx); @@ -962,6 +962,8 @@ static int mv_xor_memcpy_self_test(struct mv_xor_chan *mv_chan) dma_sync_single_for_cpu(dma_chan->device->dev, dest_dma, MV_XOR_TEST_SIZE, DMA_FROM_DEVICE); + dma_unmap_single(dma_chan->device->dev, dest_dma, MV_XOR_TEST_SIZE, + DMA_FROM_DEVICE); if (memcmp(src, dest, MV_XOR_TEST_SIZE)) { dev_err(dma_chan->device->dev, "Self-test copy failed compare, disabling\n"); @@ -1039,7 +1041,8 @@ mv_xor_xor_self_test(struct mv_xor_chan *mv_chan) 0, PAGE_SIZE, DMA_TO_DEVICE); tx = mv_xor_prep_dma_xor(dma_chan, dest_dma, dma_srcs, - MV_XOR_NUM_SRC_TEST, PAGE_SIZE, 0); + MV_XOR_NUM_SRC_TEST, PAGE_SIZE, + DMA_COMPL_SKIP_DEST_UNMAP); cookie = mv_xor_tx_submit(tx); mv_xor_issue_pending(dma_chan); @@ -1056,6 +1059,8 @@ mv_xor_xor_self_test(struct mv_xor_chan *mv_chan) dma_sync_single_for_cpu(dma_chan->device->dev, dest_dma, PAGE_SIZE, DMA_FROM_DEVICE); + dma_unmap_page(dma_chan->device->dev, dest_dma, PAGE_SIZE, + DMA_FROM_DEVICE); for (i = 0; i < (PAGE_SIZE / sizeof(u32)); i++) { u32 *ptr = page_address(dest); if (ptr[i] != cmp_word) {
DMA_CHECK was unhappy: mv_xor mv_xor.0: DMA-API: device driver tries to sync DMA memory it has not Also, the xor test was causing the mv_xor_slot_cleanup() to incorrectly dealocate the destination buffer. This is fixed as well, since the deallocation is done in mv_xor_probe() (with correct flags) now: mv_xor mv_xor.0: DMA-API: device driver frees DMA memory with different direction [device address=0x000000001dea4000] [size=4096 bytes] [mapped with DMA_FROM_DEVICE] [unmapped with DMA_BIDIRECTIONAL] Signed-off-by: Lubomir Rintel <lkundrak@v3.sk> --- drivers/dma/mv_xor.c | 9 +++++++-- 1 files changed, 7 insertions(+), 2 deletions(-)