From: Bean Huo <[email protected]>
There are two issues with the ufshcd_init_lrb() call in
ufshcd_init_cmd_priv():
- cmd->tag is set from inside scsi_mq_prep_fn() and hence is not yet set
when ufshcd_init_cmd_priv() is set.
- Inside ufshcd_init_cmd_priv() the tag can only be derived from the SCSI
command pointer if no scheduler has been associated with the UFS block
device. If no scheduler is associated with a block device, the
relationship between hctx->tags->static_rqs[] and rq->tag is static.
If a scheduler has been configured, a single tag can be associated
with a different struct request if a request is reallocated.
Fixes: 34656dda81ac ("ufs: Let the SCSI core allocate per-command UFS data")
v2-v3:
- delete ufshcd_init_cmd_priv()
- change commit message
Signed-off-by: Bean Huo <[email protected]>
---
drivers/scsi/ufs/ufshcd.c | 14 +++++---------
1 file changed, 5 insertions(+), 9 deletions(-)
diff --git a/drivers/scsi/ufs/ufshcd.c b/drivers/scsi/ufs/ufshcd.c
index e987fa3a77c7..dfb44e8cc9dd 100644
--- a/drivers/scsi/ufs/ufshcd.c
+++ b/drivers/scsi/ufs/ufshcd.c
@@ -2471,6 +2471,8 @@ static int ufshcd_queuecommand(struct Scsi_Host *host, struct scsi_cmnd *cmd)
BUG();
}
+ ufshcd_init_lrb(hba, lrbp, tag);
+
WARN_ON_ONCE(!ufshcd_is_scsi(cmd->request));
if (!down_read_trylock(&hba->clk_scaling_lock))
@@ -2707,6 +2709,7 @@ static int ufshcd_exec_dev_cmd(struct ufs_hba *hba,
init_completion(&wait);
lrbp = ufshcd_req_to_lrb(req);
+ ufshcd_init_lrb(hba, lrbp, tag);
err = ufshcd_compose_dev_cmd(hba, lrbp, cmd_type, tag);
if (unlikely(err))
goto out_put_tag;
@@ -3504,14 +3507,6 @@ static void ufshcd_host_memory_configure(struct ufs_hba *hba)
}
}
-static int ufshcd_init_cmd_priv(struct Scsi_Host *shost, struct scsi_cmnd *cmd)
-{
- struct ufs_hba *hba = shost_priv(shost);
-
- ufshcd_init_lrb(hba, scsi_cmd_priv(cmd), cmd->tag);
- return 0;
-}
-
/**
* ufshcd_dme_link_startup - Notify Unipro to perform link startup
* @hba: per adapter instance
@@ -4834,6 +4829,7 @@ static void __ufshcd_transfer_req_compl(struct ufs_hba *hba,
continue;
cmd = blk_mq_rq_to_pdu(req);
lrbp = scsi_cmd_priv(cmd);
+ ufshcd_init_lrb(hba, lrbp, index);
if (ufshcd_is_scsi(req)) {
ufshcd_add_command_trace(hba, req, "complete");
result = ufshcd_transfer_rsp_status(hba, lrbp);
@@ -5900,6 +5896,7 @@ static int ufshcd_issue_devman_upiu_cmd(struct ufs_hba *hba,
init_completion(&wait);
lrbp = ufshcd_req_to_lrb(req);
+ ufshcd_init_lrb(hba, lrbp, tag);
lrbp->sense_bufflen = 0;
lrbp->sense_buffer = NULL;
lrbp->task_tag = tag;
@@ -7180,7 +7177,6 @@ static struct scsi_host_template ufshcd_driver_template = {
.module = THIS_MODULE,
.name = UFSHCD,
.proc_name = UFSHCD,
- .init_cmd_priv = ufshcd_init_cmd_priv,
.queuecommand = ufshcd_queuecommand,
.slave_alloc = ufshcd_slave_alloc,
.slave_configure = ufshcd_slave_configure,
--
2.17.1
On 2020-03-09 09:10, [email protected] wrote:
> @@ -4834,6 +4829,7 @@ static void __ufshcd_transfer_req_compl(struct ufs_hba *hba,
> continue;
> cmd = blk_mq_rq_to_pdu(req);
> lrbp = scsi_cmd_priv(cmd);
> + ufshcd_init_lrb(hba, lrbp, index);
> if (ufshcd_is_scsi(req)) {
> ufshcd_add_command_trace(hba, req, "complete");
> result = ufshcd_transfer_rsp_status(hba, lrbp);
This ufshcd_init_lrb() call looks incorrect to me. I think that
ufshcd_init_lrb() should only be called before a request is submitted to
the UFS controller and also that ufshcd_init_lrb() should not be called
from the completion path.
Thanks,
Bart.
Hi, Bart
> Subject: [EXT] Re: [PATCH v3 1/1] scsi: ufs: fix LRB pointer incorrect initialization
> issue
>
> On 2020-03-09 09:10, [email protected] wrote:
> > @@ -4834,6 +4829,7 @@ static void __ufshcd_transfer_req_compl(struct
> ufs_hba *hba,
> > continue;
> > cmd = blk_mq_rq_to_pdu(req);
> > lrbp = scsi_cmd_priv(cmd);
> > + ufshcd_init_lrb(hba, lrbp, index);
> > if (ufshcd_is_scsi(req)) {
> > ufshcd_add_command_trace(hba, req, "complete");
> > result = ufshcd_transfer_rsp_status(hba, lrbp);
>
> This ufshcd_init_lrb() call looks incorrect to me. I think that
> ufshcd_init_lrb() should only be called before a request is submitted to the UFS
> controller and also that ufshcd_init_lrb() should not be called from the
> completion path.
>
__ufshcd_transfer_req_compl()
ufshcd_transfer_rsp_status() will access lrbp->ucd_rsp_ptr.
Without calling ufshcd_init_lrb() here, there will be an error.
//Bean
> Thanks,
>
> Bart.
On 2020-03-10 00:53, Bean Huo (beanhuo) wrote:
> Hi, Bart
>
>> Subject: [EXT] Re: [PATCH v3 1/1] scsi: ufs: fix LRB pointer incorrect initialization
>> issue
>>
>> On 2020-03-09 09:10, [email protected] wrote:
>>> @@ -4834,6 +4829,7 @@ static void __ufshcd_transfer_req_compl(struct
>> ufs_hba *hba,
>>> continue;
>>> cmd = blk_mq_rq_to_pdu(req);
>>> lrbp = scsi_cmd_priv(cmd);
>>> + ufshcd_init_lrb(hba, lrbp, index);
>>> if (ufshcd_is_scsi(req)) {
>>> ufshcd_add_command_trace(hba, req, "complete");
>>> result = ufshcd_transfer_rsp_status(hba, lrbp);
>>
>> This ufshcd_init_lrb() call looks incorrect to me. I think that
>> ufshcd_init_lrb() should only be called before a request is submitted to the UFS
>> controller and also that ufshcd_init_lrb() should not be called from the
>> completion path.
>
> __ufshcd_transfer_req_compl()
> ufshcd_transfer_rsp_status() will access lrbp->ucd_rsp_ptr.
> Without calling ufshcd_init_lrb() here, there will be an error.
Hi Bean,
I think that ufshcd_init_lrb() should only be called from the code that
prepares a command before it is submitted and not from the command
completion path. Because v5.6-rc5 has already been released, there is
not that much time left until the merge window opens. I think it is less
risky to revert commit 34656dda81ac ("scsi: ufs: Let the SCSI core
allocate per-command UFS data") than to proceed with the above patch. Do
you want to submit a revert or do you perhaps want me to do that?
Thanks,
Bart.