Re: [PATCH V7 2/3] scsi: ufs: Fix a race condition between ufshcd_abort and eh_work

From: Stanley Chu
Date: Thu Dec 03 2020 - 00:20:06 EST


On Thu, 2020-12-03 at 13:11 +0800, Can Guo wrote:
> On 2020-12-03 12:51, Stanley Chu wrote:
> > On Thu, 2020-12-03 at 12:01 +0800, Can Guo wrote:
> >> On 2020-12-03 10:21, Stanley Chu wrote:
> >> > On Wed, 2020-12-02 at 04:04 -0800, Can Guo wrote:
> >> >> In current task abort routine, if task abort happens to the device
> >> >> W-LU,
> >> >> the code directly jumps to ufshcd_eh_host_reset_handler() to perform a
> >> >> full reset and restore then returns FAIL or SUCCESS. Commands sent to
> >> >> the
> >> >> device W-LU are most likely the SSU cmds sent during UFS PM
> >> >> operations. If
> >> >> such SSU cmd enters task abort routine, when
> >> >> ufshcd_eh_host_reset_handler()
> >> >> flushes eh_work, it will get stuck there since err_handler is
> >> >> serialized
> >> >> with PM operations.
> >> >>
> >> >> In order to unblock above call path, we merely clean up the lrb taken
> >> >> by
> >> >> this cmd, queue the eh_work and return SUCCESS. Once the cmd is
> >> >> aborted,
> >> >> the PM operation which sends out the cmd just errors out, then
> >> >> err_handler
> >> >> shall be able to proceed with the full reset and restore.
> >> >>
> >> >> In this scenario, the cmd is aborted even before it is actually
> >> >> cleared by
> >> >> HW, set the lrb->in_use flag to prevent subsequent cmds, including
> >> >> SCSI
> >> >> cmds and dev cmds, from taking the lrb released from abort. The flag
> >> >> shall
> >> >> evetually be cleared in __ufshcd_transfer_req_compl() invoked by the
> >> >> full
> >> >> reset and restore from err_handler.
> >> >>
> >> >> Reviewed-by: Asutosh Das <asutoshd@xxxxxxxxxxxxxx>
> >> >> Signed-off-by: Can Guo <cang@xxxxxxxxxxxxxx>
> >> >> ---
> >> >> drivers/scsi/ufs/ufshcd.c | 60
> >> >> +++++++++++++++++++++++++++++++++++++----------
> >> >> drivers/scsi/ufs/ufshcd.h | 2 ++
> >> >> 2 files changed, 49 insertions(+), 13 deletions(-)
> >> >>
> >> >> diff --git a/drivers/scsi/ufs/ufshcd.c b/drivers/scsi/ufs/ufshcd.c
> >> >> index f0bb3fc..26c1fa0 100644
> >> >> --- a/drivers/scsi/ufs/ufshcd.c
> >> >> +++ b/drivers/scsi/ufs/ufshcd.c
> >> >> @@ -2539,6 +2539,14 @@ static int ufshcd_queuecommand(struct Scsi_Host
> >> >> *host, struct scsi_cmnd *cmd)
> >> >> (hba->clk_gating.state != CLKS_ON));
> >> >>
> >> >> lrbp = &hba->lrb[tag];
> >> >> + if (unlikely(lrbp->in_use)) {
> >> >> + if (hba->pm_op_in_progress)
> >> >> + set_host_byte(cmd, DID_BAD_TARGET);
> >> >> + else
> >> >> + err = SCSI_MLQUEUE_HOST_BUSY;
> >> >> + ufshcd_release(hba);
> >> >> + goto out;
> >> >> + }
> >> >>
> >> >> WARN_ON(lrbp->cmd);
> >> >> lrbp->cmd = cmd;
> >> >> @@ -2781,6 +2789,11 @@ static int ufshcd_exec_dev_cmd(struct ufs_hba
> >> >> *hba,
> >> >>
> >> >> init_completion(&wait);
> >> >> lrbp = &hba->lrb[tag];
> >> >> + if (unlikely(lrbp->in_use)) {
> >> >> + err = -EBUSY;
> >> >> + goto out;
> >> >> + }
> >> >> +
> >> >> WARN_ON(lrbp->cmd);
> >> >> err = ufshcd_compose_dev_cmd(hba, lrbp, cmd_type, tag);
> >> >> if (unlikely(err))
> >> >> @@ -2797,6 +2810,7 @@ static int ufshcd_exec_dev_cmd(struct ufs_hba
> >> >> *hba,
> >> >>
> >> >> err = ufshcd_wait_for_dev_cmd(hba, lrbp, timeout);
> >> >>
> >> >> +out:
> >> >> ufshcd_add_query_upiu_trace(hba, tag,
> >> >> err ? "query_complete_err" : "query_complete");
> >> >>
> >> >> @@ -4929,9 +4943,11 @@ static void __ufshcd_transfer_req_compl(struct
> >> >> ufs_hba *hba,
> >> >> struct scsi_cmnd *cmd;
> >> >> int result;
> >> >> int index;
> >> >> + bool update_scaling = false;
> >> >>
> >> >> for_each_set_bit(index, &completed_reqs, hba->nutrs) {
> >> >> lrbp = &hba->lrb[index];
> >> >> + lrbp->in_use = false;
> >> >> lrbp->compl_time_stamp = ktime_get();
> >> >> cmd = lrbp->cmd;
> >> >> if (cmd) {
> >> >> @@ -4944,15 +4960,17 @@ static void __ufshcd_transfer_req_compl(struct
> >> >> ufs_hba *hba,
> >> >> /* Do not touch lrbp after scsi done */
> >> >> cmd->scsi_done(cmd);
> >> >> __ufshcd_release(hba);
> >> >> + update_scaling = true;
> >> >> } else if (lrbp->command_type == UTP_CMD_TYPE_DEV_MANAGE ||
> >> >> lrbp->command_type == UTP_CMD_TYPE_UFS_STORAGE) {
> >> >> if (hba->dev_cmd.complete) {
> >> >> ufshcd_add_command_trace(hba, index,
> >> >> "dev_complete");
> >> >> complete(hba->dev_cmd.complete);
> >> >> + update_scaling = true;
> >> >> }
> >> >> }
> >> >> - if (ufshcd_is_clkscaling_supported(hba))
> >> >> + if (ufshcd_is_clkscaling_supported(hba) && update_scaling)
> >> >> hba->clk_scaling.active_reqs--;
> >> >> }
> >> >>
> >> >> @@ -6374,8 +6392,12 @@ static int ufshcd_issue_devman_upiu_cmd(struct
> >> >> ufs_hba *hba,
> >> >>
> >> >> init_completion(&wait);
> >> >> lrbp = &hba->lrb[tag];
> >> >> - WARN_ON(lrbp->cmd);
> >> >> + if (unlikely(lrbp->in_use)) {
> >> >> + err = -EBUSY;
> >> >> + goto out;
> >> >> + }
> >> >>
> >> >> + WARN_ON(lrbp->cmd);
> >> >> lrbp->cmd = NULL;
> >> >> lrbp->sense_bufflen = 0;
> >> >> lrbp->sense_buffer = NULL;
> >> >> @@ -6447,6 +6469,7 @@ static int ufshcd_issue_devman_upiu_cmd(struct
> >> >> ufs_hba *hba,
> >> >> }
> >> >> }
> >> >>
> >> >> +out:
> >> >> blk_put_request(req);
> >> >> out_unlock:
> >> >> up_read(&hba->clk_scaling_lock);
> >> >> @@ -6696,16 +6719,6 @@ static int ufshcd_abort(struct scsi_cmnd *cmd)
> >> >> BUG();
> >> >> }
> >> >>
> >> >> - /*
> >> >> - * Task abort to the device W-LUN is illegal. When this command
> >> >> - * will fail, due to spec violation, scsi err handling next step
> >> >> - * will be to send LU reset which, again, is a spec violation.
> >> >> - * To avoid these unnecessary/illegal step we skip to the last error
> >> >> - * handling stage: reset and restore.
> >> >> - */
> >> >> - if (lrbp->lun == UFS_UPIU_UFS_DEVICE_WLUN)
> >> >> - return ufshcd_eh_host_reset_handler(cmd);
> >> >> -
> >> >> ufshcd_hold(hba, false);
> >> >> reg = ufshcd_readl(hba, REG_UTP_TRANSFER_REQ_DOOR_BELL);
> >> >> /* If command is already aborted/completed, return SUCCESS */
> >> >> @@ -6726,7 +6739,7 @@ static int ufshcd_abort(struct scsi_cmnd *cmd)
> >> >> * to reduce repeated printouts. For other aborted requests only
> >> >> print
> >> >> * basic details.
> >> >> */
> >> >> - scsi_print_command(hba->lrb[tag].cmd);
> >> >> + scsi_print_command(cmd);
> >> >> if (!hba->req_abort_count) {
> >> >> ufshcd_update_reg_hist(&hba->ufs_stats.task_abort, 0);
> >> >> ufshcd_print_host_regs(hba);
> >> >> @@ -6745,6 +6758,27 @@ static int ufshcd_abort(struct scsi_cmnd *cmd)
> >> >> goto cleanup;
> >> >> }
> >> >>
> >> >> + /*
> >> >> + * Task abort to the device W-LUN is illegal. When this command
> >> >> + * will fail, due to spec violation, scsi err handling next step
> >> >> + * will be to send LU reset which, again, is a spec violation.
> >> >> + * To avoid these unnecessary/illegal steps, first we clean up
> >> >> + * the lrb taken by this cmd and mark the lrb as in_use, then
> >> >> + * queue the eh_work and bail.
> >> >> + */
> >> >> + if (lrbp->lun == UFS_UPIU_UFS_DEVICE_WLUN) {
> >> >> + spin_lock_irqsave(host->host_lock, flags);
> >> >> + if (lrbp->cmd) {
> >> >> + __ufshcd_transfer_req_compl(hba, (1UL << tag));
> >> >> + __set_bit(tag, &hba->outstanding_reqs);
> >> >> + lrbp->in_use = true;
> >> >> + hba->force_reset = true;
> >> >> + ufshcd_schedule_eh_work(hba);
> >> >
> >> > ufshcd_schedule_eh_work() will set hba->ufshcd_state as
> >> > UFSHCD_STATE_EH_SCHEDULED_FATAL. While in this state,
> >> > ufshcd_queuecommand() will set_host_byte(DID_BAD_TARGET) which is
> >> > similar as what you would like to do in this patch.
> >> >
> >> > Is this enough for avoiding reusing tag issue? Just wonder if
> >> > lrpb->in_use flag is really required to be added.
> >>
> >> Hi Stanley,
> >>
> >> Thanks for the discussion.
> >>
> >> To be accurate, it is to prevent lrb from being re-used, not the
> >> tag.
> >> Block layer and/or scsi layer can re-use the tag right after
> >> we abort the cmd, but the lrb is empty since we cleared it from
> >> abort path and we need to make sure the lrb stays empty before the
> >> full reset and restore happens.
> >
> > What is the definition of "empty" here?
>
> Yes, it means lrb->cmd is NULL.
>
> >
> > If it means lrb->cmd shall be empty (to not invoking scsi_done again),
> > then the hba->ufshcd_state check in ufshcd_queuecommend() will also
> > clear the re-used lrb->cmd if ufshcd_state is in
> > UFSHCD_STATE_EH_SCHEDULED_FATAL case.
>
> Yes, but that would have a race condition - say a cmd re-uses the tag
> and
> occupies the lrb in queuecommand, at the same time the real completion
> IRQ
> of previous cmd comes (althrough we aborted it due to it timed out, but
> the
> completion IRQ can come at any time before the full reset and restore is
> performed), then the new cmd will be wrongly completed by IRQ handler
> from
> __ufshcd_transfer_req_compl() which is actually fired by the last cmd,
> so
> we need to block it even before the cmd takes the lrb - before
>
> lrbp->cmd = cmd;

I see.
Thanks for the explanation.

>
> >
> > However ufshcd_state cannot protect other paths now, for example,
> > ufshcd_exec_dev_cmd(), so lrbp->in_use may be required for this usage,
> > or ufshcd_state check can be added to help.
> >
> > BTW, would you also need to consider ufshcd_issue_devman_upiu_cmd()
> > that
> > is another possible path to re-use lrb?
>
> In this change, there are checks of lrb->in_use in
> ufshcd_queuecommand(),
> ufshcd_exec_dev_cmd() and ufshcd_issue_devman_upiu_cmd().

Oops I missed it. Sorry for this noise.

Feel free to add
Reviewed-by: Stanley Chu <stanley.chu@xxxxxxxxxxxx>