storvsc_queuecommand() maps the scatter/gather list using scsi_dma_map(),
which in a confidential VM allocates swiotlb bounce buffers. If the
I/O submission fails in storvsc_do_io(), the I/O is typically retried
by higher level code, but the bounce buffer memory is never freed.
The mostly like cause of I/O submission failure is a full VMBus
channel ring buffer, which is not uncommon under high I/O loads.
Eventually enough bounce buffer memory leaks that the confidential
VM can't do any I/O. The same problem can arise in a non-confidential
VM with kernel boot parameter swiotlb=force.
Fix this by doing scsi_dma_unmap() in the case of an I/O submission
error, which frees the bounce buffer memory.
Fixes: 743b237c3a7b ("scsi: storvsc: Add Isolation VM support for storvsc driver")
Signed-off-by: Michael Kelley <[email protected]>
---
drivers/scsi/storvsc_drv.c | 3 +++
1 file changed, 3 insertions(+)
diff --git a/drivers/scsi/storvsc_drv.c b/drivers/scsi/storvsc_drv.c
index 3c5b7e4..55d6fb4 100644
--- a/drivers/scsi/storvsc_drv.c
+++ b/drivers/scsi/storvsc_drv.c
@@ -1823,6 +1823,9 @@ static int storvsc_queuecommand(struct Scsi_Host *host, struct scsi_cmnd *scmnd)
ret = storvsc_do_io(dev, cmd_request, get_cpu());
put_cpu();
+ if (ret)
+ scsi_dma_unmap(scmnd);
+
if (ret == -EAGAIN) {
/* no more space */
ret = SCSI_MLQUEUE_DEVICE_BUSY;
--
1.8.3.1
> From: Michael Kelley (LINUX) <[email protected]>
> Sent: Sunday, December 4, 2022 11:53 AM
>
> storvsc_queuecommand() maps the scatter/gather list using scsi_dma_map(),
> which in a confidential VM allocates swiotlb bounce buffers. If the
> I/O submission fails in storvsc_do_io(), the I/O is typically retried
> by higher level code, but the bounce buffer memory is never freed.
> The mostly like cause of I/O submission failure is a full VMBus
> channel ring buffer, which is not uncommon under high I/O loads.
> Eventually enough bounce buffer memory leaks that the confidential
> VM can't do any I/O. The same problem can arise in a non-confidential
> VM with kernel boot parameter swiotlb=force.
>
> Fix this by doing scsi_dma_unmap() in the case of an I/O submission
> error, which frees the bounce buffer memory.
>
> Fixes: 743b237c3a7b ("scsi: storvsc: Add Isolation VM support for storvsc
> driver")
> Signed-off-by: Michael Kelley <[email protected]>
> ---
Reviewed-by: Dexuan Cui <[email protected]>
Tested-by: Dexuan Cui <[email protected]>
I hope this fix can be included in v6.1.
On 12/5/2022 3:52 AM, Michael Kelley wrote:
> storvsc_queuecommand() maps the scatter/gather list using scsi_dma_map(),
> which in a confidential VM allocates swiotlb bounce buffers. If the
> I/O submission fails in storvsc_do_io(), the I/O is typically retried
> by higher level code, but the bounce buffer memory is never freed.
> The mostly like cause of I/O submission failure is a full VMBus
> channel ring buffer, which is not uncommon under high I/O loads.
> Eventually enough bounce buffer memory leaks that the confidential
> VM can't do any I/O. The same problem can arise in a non-confidential
> VM with kernel boot parameter swiotlb=force.
>
> Fix this by doing scsi_dma_unmap() in the case of an I/O submission
> error, which frees the bounce buffer memory.
>
> Fixes: 743b237c3a7b ("scsi: storvsc: Add Isolation VM support for storvsc driver")
> Signed-off-by: Michael Kelley<[email protected]>
Nice catch! Thanks to fix this.
Reviewed-by: Tianyu Lan <[email protected]>
Michael,
> storvsc_queuecommand() maps the scatter/gather list using scsi_dma_map(),
> which in a confidential VM allocates swiotlb bounce buffers. If the
> I/O submission fails in storvsc_do_io(), the I/O is typically retried
> by higher level code, but the bounce buffer memory is never freed.
> The mostly like cause of I/O submission failure is a full VMBus
> channel ring buffer, which is not uncommon under high I/O loads.
> Eventually enough bounce buffer memory leaks that the confidential
> VM can't do any I/O. The same problem can arise in a non-confidential
> VM with kernel boot parameter swiotlb=force.
Applied to 6.2/scsi-staging, thanks!
--
Martin K. Petersen Oracle Linux Engineering