On 10/2/24 5:42 AM, Peter Wang (王信友) wrote:
This patch merely aligns with the approach of SDB mode
and does not involve the flow of scsi_done. Besides,
I don't see any issue with concurrency between
ufshcd_abort_one() calling ufshcd_try_to_abort_task()
and scsi_done(). Can you point out the specific flow where
the problem occurs? If there is one, shouldn't SDB mode
have the same issue?
Hi Peter,
Correct, my comment applies to both legacy mode and MCQ mode. From the
section in the UFS standard about ABORT TASK: "A response of FUNCTION
COMPLETE shall indicate that the command was aborted or was not in the
task set." In other words, if a command completes just before
ufshcd_try_to_abort_task() calls ufshcd_issue_tm_cmd(), then
ufshcd_try_to_abort_task() will call ufshcd_clear_cmd() for a command
that has already completed. In legacy mode, this call will succeed.
Hence, both ufshcd_compl_one_cqe() and ufshcd_abort_all() will call
ufshcd_release(hba). This will cause hba->clk_gating.active_reqs to be
decremented twice instead of only once. Do you agree that this can
happen and also that it should be prevented that this happens?
Thanks,
Bart.
