| Message ID | 20181113185712.10063-1-keith.busch@intel.com (mailing list archive) |
|---|---|
| State | Superseded |
| Headers | show |
| Series | [1/2] scsi: Do not rely on blk-mq for double completions | expand |
On 11/13/18 11:57 AM, Keith Busch wrote: > The scsi timeout error handling had been directly updating the request > state to prevent a natural completion and error handling from completing > the same request twice. Fix this layering violation by having scsi > control the fate of its commands with scsi owned flags rather than > use blk-mq's. > > diff --git a/drivers/scsi/scsi_lib.c b/drivers/scsi/scsi_lib.c > index 61babcb269ab..c680171ca201 100644 > --- a/drivers/scsi/scsi_lib.c > +++ b/drivers/scsi/scsi_lib.c > @@ -1635,8 +1635,18 @@ static blk_status_t scsi_mq_prep_fn(struct request *req) > > static void scsi_mq_done(struct scsi_cmnd *cmd) > { > + if (test_and_set_bit(__SCMD_COMPLETE, &cmd->flags)) > + return; > trace_scsi_dispatch_cmd_done(cmd); > blk_mq_complete_request(cmd->request); > + > +#ifdef CONFIG_FAIL_IO_TIMEOUT > + /* > + * Clearing complete here serves only to allow the desired recovery to > + * escalate on blk_rq_should_fake_timeout()'s error injection. > + */ > + clear_bit(__SCMD_COMPLETE, &cmd->flags); > +#endif > } We could have this be: static void scsi_mq_done(struct scsi_cmnd *cmd) { if (test_and_set_bit(__SCMD_COMPLETE, &cmd->flags)) return; trace_scsi_dispatch_cmd_done(cmd); if (blk_mq_complete_request(cmd->request)) { /* * Clearing complete here serves only to allow the * desired recovery to escalate on * blk_rq_should_fake_timeout()'s error injection. */ clear_bit(__SCMD_COMPLETE, &cmd->flags); } } with bool blk_mq_complete_request(struct request *rq) { if (unlikely(blk_should_fake_timeout(rq->q))) return true; __blk_mq_complete_request(rq); return false; } and not have this CONFIG_FAIL_IO_TIMEOUT dependency, but that'd be a bit more expensive. Was going to suggest a request flag, but the request is gone at this point. So that won't really work... I'm with your solution as well, fwiw.
On Tue, Nov 13, 2018 at 12:20:46PM -0700, Jens Axboe wrote: > On 11/13/18 11:57 AM, Keith Busch wrote: > > static void scsi_mq_done(struct scsi_cmnd *cmd) > > { > > + if (test_and_set_bit(__SCMD_COMPLETE, &cmd->flags)) > > + return; > > trace_scsi_dispatch_cmd_done(cmd); > > blk_mq_complete_request(cmd->request); > > + > > +#ifdef CONFIG_FAIL_IO_TIMEOUT > > + /* > > + * Clearing complete here serves only to allow the desired recovery to > > + * escalate on blk_rq_should_fake_timeout()'s error injection. > > + */ > > + clear_bit(__SCMD_COMPLETE, &cmd->flags); > > +#endif > > } > > We could have this be: > > static void scsi_mq_done(struct scsi_cmnd *cmd) > { > if (test_and_set_bit(__SCMD_COMPLETE, &cmd->flags)) > return; > trace_scsi_dispatch_cmd_done(cmd); > > if (blk_mq_complete_request(cmd->request)) { > /* > * Clearing complete here serves only to allow the > * desired recovery to escalate on > * blk_rq_should_fake_timeout()'s error injection. > */ > clear_bit(__SCMD_COMPLETE, &cmd->flags); > } > } > > with > > bool blk_mq_complete_request(struct request *rq) > { > if (unlikely(blk_should_fake_timeout(rq->q))) > return true; > __blk_mq_complete_request(rq); > return false; > } > > and not have this CONFIG_FAIL_IO_TIMEOUT dependency, but that'd be a bit > more expensive. I was trying to avoid every cost no matter how negligable (those are the only types of costs left as far as I can see), but I think your proposal might actually be necessary: if a timeout wasn't faked, clearing the completion flag unconditionally might have a problem with a real timeout racing with the real completion. :(
On 11/13/18 12:45 PM, Keith Busch wrote: > On Tue, Nov 13, 2018 at 12:20:46PM -0700, Jens Axboe wrote: >> On 11/13/18 11:57 AM, Keith Busch wrote: >>> static void scsi_mq_done(struct scsi_cmnd *cmd) >>> { >>> + if (test_and_set_bit(__SCMD_COMPLETE, &cmd->flags)) >>> + return; >>> trace_scsi_dispatch_cmd_done(cmd); >>> blk_mq_complete_request(cmd->request); >>> + >>> +#ifdef CONFIG_FAIL_IO_TIMEOUT >>> + /* >>> + * Clearing complete here serves only to allow the desired recovery to >>> + * escalate on blk_rq_should_fake_timeout()'s error injection. >>> + */ >>> + clear_bit(__SCMD_COMPLETE, &cmd->flags); >>> +#endif >>> } >> >> We could have this be: >> >> static void scsi_mq_done(struct scsi_cmnd *cmd) >> { >> if (test_and_set_bit(__SCMD_COMPLETE, &cmd->flags)) >> return; >> trace_scsi_dispatch_cmd_done(cmd); >> >> if (blk_mq_complete_request(cmd->request)) { >> /* >> * Clearing complete here serves only to allow the >> * desired recovery to escalate on >> * blk_rq_should_fake_timeout()'s error injection. >> */ >> clear_bit(__SCMD_COMPLETE, &cmd->flags); >> } >> } >> >> with >> >> bool blk_mq_complete_request(struct request *rq) >> { >> if (unlikely(blk_should_fake_timeout(rq->q))) >> return true; >> __blk_mq_complete_request(rq); >> return false; >> } >> >> and not have this CONFIG_FAIL_IO_TIMEOUT dependency, but that'd be a bit >> more expensive. > > I was trying to avoid every cost no matter how negligable (those are the > only types of costs left as far as I can see), but I think your proposal > might actually be necessary: if a timeout wasn't faked, clearing the > completion flag unconditionally might have a problem with a real timeout > racing with the real completion. :( Yep, I think you are right... No way around it then. Are you going to resend it with the fix?
diff --git a/drivers/scsi/scsi_error.c b/drivers/scsi/scsi_error.c index dd338a8cd275..6156f45c2c80 100644 --- a/drivers/scsi/scsi_error.c +++ b/drivers/scsi/scsi_error.c @@ -204,6 +204,9 @@ scsi_abort_command(struct scsi_cmnd *scmd) shost->last_reset = jiffies; spin_unlock_irqrestore(shost->host_lock, flags); + if (test_and_set_bit(__SCMD_COMPLETE, &scmd->flags)) + return SUCCESS; + scmd->eh_eflags |= SCSI_EH_ABORT_SCHEDULED; SCSI_LOG_ERROR_RECOVERY(3, scmd_printk(KERN_INFO, scmd, "abort scheduled\n")); @@ -296,20 +299,6 @@ enum blk_eh_timer_return scsi_times_out(struct request *req) rtn = host->hostt->eh_timed_out(scmd); if (rtn == BLK_EH_DONE) { - /* - * For blk-mq, we must set the request state to complete now - * before sending the request to the scsi error handler. This - * will prevent a use-after-free in the event the LLD manages - * to complete the request before the error handler finishes - * processing this timed out request. - * - * If the request was already completed, then the LLD beat the - * time out handler from transferring the request to the scsi - * error handler. In that case we can return immediately as no - * further action is required. - */ - if (!blk_mq_mark_complete(req)) - return rtn; if (scsi_abort_command(scmd) != SUCCESS) { set_host_byte(scmd, DID_TIME_OUT); scsi_eh_scmd_add(scmd); diff --git a/drivers/scsi/scsi_lib.c b/drivers/scsi/scsi_lib.c index 61babcb269ab..c680171ca201 100644 --- a/drivers/scsi/scsi_lib.c +++ b/drivers/scsi/scsi_lib.c @@ -1635,8 +1635,18 @@ static blk_status_t scsi_mq_prep_fn(struct request *req) static void scsi_mq_done(struct scsi_cmnd *cmd) { + if (test_and_set_bit(__SCMD_COMPLETE, &cmd->flags)) + return; trace_scsi_dispatch_cmd_done(cmd); blk_mq_complete_request(cmd->request); + +#ifdef CONFIG_FAIL_IO_TIMEOUT + /* + * Clearing complete here serves only to allow the desired recovery to + * escalate on blk_rq_should_fake_timeout()'s error injection. + */ + clear_bit(__SCMD_COMPLETE, &cmd->flags); +#endif } static void scsi_mq_put_budget(struct blk_mq_hw_ctx *hctx) @@ -1701,6 +1711,7 @@ static blk_status_t scsi_queue_rq(struct blk_mq_hw_ctx *hctx, goto out_dec_host_busy; req->rq_flags |= RQF_DONTPREP; } else { + cmd->flags &= ~SCMD_COMPLETE; blk_mq_start_request(req); } diff --git a/include/scsi/scsi_cmnd.h b/include/scsi/scsi_cmnd.h index d6fd2aba0380..ded7c7194a28 100644 --- a/include/scsi/scsi_cmnd.h +++ b/include/scsi/scsi_cmnd.h @@ -58,6 +58,9 @@ struct scsi_pointer { #define SCMD_TAGGED (1 << 0) #define SCMD_UNCHECKED_ISA_DMA (1 << 1) #define SCMD_INITIALIZED (1 << 2) + +#define __SCMD_COMPLETE 3 +#define SCMD_COMPLETE (1 << __SCMD_COMPLETE) /* flags preserved across unprep / reprep */ #define SCMD_PRESERVED_FLAGS (SCMD_UNCHECKED_ISA_DMA | SCMD_INITIALIZED) @@ -144,7 +147,7 @@ struct scsi_cmnd { * to be at an address < 16Mb). */ int result; /* Status code from lower level driver */ - int flags; /* Command flags */ + unsigned long flags; /* Command flags */ unsigned char tag; /* SCSI-II queued command tag */ };
The scsi timeout error handling had been directly updating the request state to prevent a natural completion and error handling from completing the same request twice. Fix this layering violation by having scsi control the fate of its commands with scsi owned flags rather than use blk-mq's. Signed-off-by: Keith Busch <keith.busch@intel.com> --- drivers/scsi/scsi_error.c | 17 +++-------------- drivers/scsi/scsi_lib.c | 11 +++++++++++ include/scsi/scsi_cmnd.h | 5 ++++- 3 files changed, 18 insertions(+), 15 deletions(-)