2024-01-04 01:26:01

by Kiwoong Kim

Subject: [PATCH v1] ufs: get target SQ entry within critical section

In IO centric scenarios, especially during a period that
many IO requests are submitted to a same HW queue at the same
time, it's found that one reqeust overwrote a SQ entry
that had been already occupied by another request submitted
in the past. And it eventually led to command timed-out
because one of two requests were overwritten, which could not
be completed.

[ 74.995185][ T176] exynos-ufs 17100000.ufs: ufshcd_abort: Device abort task at tag 30

Signed-off-by: Kiwoong Kim <[email protected]>
---
drivers/ufs/core/ufshcd.c | 3 ++-
1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/drivers/ufs/core/ufshcd.c b/drivers/ufs/core/ufshcd.c
index 7bc3fc4..da1a9c0 100644
--- a/drivers/ufs/core/ufshcd.c
+++ b/drivers/ufs/core/ufshcd.c
@@ -2199,9 +2199,10 @@ void ufshcd_send_command(struct ufs_hba *hba, unsigned int task_tag,
if (is_mcq_enabled(hba)) {
int utrd_size = sizeof(struct utp_transfer_req_desc);
struct utp_transfer_req_desc *src = lrbp->utr_descriptor_ptr;
- struct utp_transfer_req_desc *dest = hwq->sqe_base_addr + hwq->sq_tail_slot;
+ struct utp_transfer_req_desc *dest;

spin_lock(&hwq->sq_lock);
+ dest = hwq->sqe_base_addr + hwq->sq_tail_slot;
memcpy(dest, src, utrd_size);
ufshcd_inc_sq_tail(hwq);
spin_unlock(&hwq->sq_lock);
--
2.7.4

2024-01-07 16:03:10

by Bart Van Assche

Subject: Re: [PATCH v1] ufs: get target SQ entry within critical section

On 1/3/24 17:24, Kiwoong Kim wrote:
> In IO centric scenarios, especially during a period that
> many IO requests are submitted to a same HW queue at the same
> time, it's found that one reqeust overwrote a SQ entry
> that had been already occupied by another request submitted
> in the past. And it eventually led to command timed-out
> because one of two requests were overwritten, which could not
> be completed.
>
> [ 74.995185][ T176] exynos-ufs 17100000.ufs: ufshcd_abort: Device abort task at tag 30
>
> Signed-off-by: Kiwoong Kim <[email protected]>
> ---
> drivers/ufs/core/ufshcd.c | 3 ++-
> 1 file changed, 2 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/ufs/core/ufshcd.c b/drivers/ufs/core/ufshcd.c
> index 7bc3fc4..da1a9c0 100644
> --- a/drivers/ufs/core/ufshcd.c
> +++ b/drivers/ufs/core/ufshcd.c
> @@ -2199,9 +2199,10 @@ void ufshcd_send_command(struct ufs_hba *hba, unsigned int task_tag,
> if (is_mcq_enabled(hba)) {
> int utrd_size = sizeof(struct utp_transfer_req_desc);
> struct utp_transfer_req_desc *src = lrbp->utr_descriptor_ptr;
> - struct utp_transfer_req_desc *dest = hwq->sqe_base_addr + hwq->sq_tail_slot;
> + struct utp_transfer_req_desc *dest;
>
> spin_lock(&hwq->sq_lock);
> + dest = hwq->sqe_base_addr + hwq->sq_tail_slot;
> memcpy(dest, src, utrd_size);
> ufshcd_inc_sq_tail(hwq);
> spin_unlock(&hwq->sq_lock);

Is this perhaps a duplicate of patch "scsi: ufs: core: Let the sq_lock
protect sq_tail_slot access"? See also
https://lore.kernel.org/linux-scsi/[email protected]/#t

Thanks,

Bart.

2024-01-08 00:51:44

by Kiwoong Kim

Subject: RE: [PATCH v1] ufs: get target SQ entry within critical section

> > diff --git a/drivers/ufs/core/ufshcd.c b/drivers/ufs/core/ufshcd.c
> > index 7bc3fc4..da1a9c0 100644
> > --- a/drivers/ufs/core/ufshcd.c
> > +++ b/drivers/ufs/core/ufshcd.c
> > @@ -2199,9 +2199,10 @@ void ufshcd_send_command(struct ufs_hba *hba,
> unsigned int task_tag,
> > if (is_mcq_enabled(hba)) {
> > int utrd_size = sizeof(struct utp_transfer_req_desc);
> > struct utp_transfer_req_desc *src = lrbp->utr_descriptor_ptr;
> > - struct utp_transfer_req_desc *dest = hwq->sqe_base_addr +
> hwq->sq_tail_slot;
> > + struct utp_transfer_req_desc *dest;
> >
> > spin_lock(&hwq->sq_lock);
> > + dest = hwq->sqe_base_addr + hwq->sq_tail_slot;
> > memcpy(dest, src, utrd_size);
> > ufshcd_inc_sq_tail(hwq);
> > spin_unlock(&hwq->sq_lock);
>
> Is this perhaps a duplicate of patch "scsi: ufs: core: Let the sq_lock
> protect sq_tail_slot access"? See also https://lore.kernel.org/linux-
> scsi/[email protected]/#t

I didn’t see it. Thank you for letting me know.