| Message ID | 379b841f-210f-41dc-a44c-f1dc3197e10f@kernel.dk (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | ublk_drv: set DMA alignment mask to 3 | expand |
On Sat, May 11, 2024 at 08:40:57AM -0600, Jens Axboe wrote: > By default, this will be 511, as that's the block layer default. But > drivers these days can support memory alignments that aren't tied to > the sector sizes, instead just being limited by what the DMA engine > supports. An example is NVMe, where it's generally set to a 32-bit or > 64-bit boundary. As ublk itself doesn't really care, just set it low > enough that we don't run into issues with NVMe where the required > O_DIRECT memory alignment is now more restrictive on ublk than it is > on the underlying device. > > This was triggered by spurious -EINVAL returns on O_DIRECT IO on a > setup with ublk managing NVMe devices, which previously worked just > fine on the NVMe device itself. With the alignment relaxed, the test > works fine. > > Signed-off-by: Jens Axboe <axboe@kernel.dk> It should be triggered since DIO DMA alignment is relaxed: Reviewed-by: Ming Lei <ming.lei@redhat.com> Thanks, Ming
diff --git a/drivers/block/ublk_drv.c b/drivers/block/ublk_drv.c index 851c78913de2..292fa2bdd77d 100644 --- a/drivers/block/ublk_drv.c +++ b/drivers/block/ublk_drv.c @@ -2176,6 +2176,7 @@ static int ublk_ctrl_start_dev(struct ublk_device *ub, struct io_uring_cmd *cmd) .max_hw_sectors = p->max_sectors, .chunk_sectors = p->chunk_sectors, .virt_boundary_mask = p->virt_boundary_mask, + .dma_alignment = 3, }; struct gendisk *disk;
By default, this will be 511, as that's the block layer default. But drivers these days can support memory alignments that aren't tied to the sector sizes, instead just being limited by what the DMA engine supports. An example is NVMe, where it's generally set to a 32-bit or 64-bit boundary. As ublk itself doesn't really care, just set it low enough that we don't run into issues with NVMe where the required O_DIRECT memory alignment is now more restrictive on ublk than it is on the underlying device. This was triggered by spurious -EINVAL returns on O_DIRECT IO on a setup with ublk managing NVMe devices, which previously worked just fine on the NVMe device itself. With the alignment relaxed, the test works fine. Signed-off-by: Jens Axboe <axboe@kernel.dk> ---