scsi_error: do not queue pointless abort workqueue functions
PATCH 1/1 scsi_error: do not queue pointless abort workqueue functions
If a host template doesn't implement the .eh_abort_handler() there is no point in queueing the abort workqueue function; all it does is invoking SCSI EH anyway.
So return 'FAILED' from scsi_abort_command() if the .eh_abort_handler() is not implemented and save us from having to wait for the abort workqueue function to complete.
--- drivers/scsi/scsi_error.c | 5 +++++ 1 file changed, 5 insertions(+)
drivers/scsi/scsi_error.c +5 -0
@@ -242,6 +242,11 @@ scsi_abort_command(struct scsi_cmnd *scmd)return FAILED ; } +if (!shost->hostt->eh_abort_handler) { +/* No abort handler, fail command directly */ +return FAILED ; + } +spin_lock_irqsave (shost->host_lock, flags);if (shost->eh_deadline !=-1 && !shost->last_reset) shost->last_reset = jiffies;
Signature · 1 line
2.35.3
Re: [External] : [PATCH] scsi_error: do not queue pointless abort workqueue functions
Quoted 6 lines · Hannes Reinecke
If a host template doesn't implement the .eh_abort_handler() there is no point in queueing the abort workqueue function; all it does is invoking SCSI EH anyway.
So return 'FAILED' fromscsi_abort_command()if the .eh_abort_handler() is not implemented and save us from having to wait for the abort workqueue function to complete.
Do we ever use shost->tmf_work_q in this case? Doesn't seem much point in allocating it, apart from keeping the code simpler
Quoted 4 lines
Cc Niklas Cassel <[email protected]>Cc Damien Le Moal <[email protected]>Cc John Garry <[email protected]>
That's someone else :)
Quoted 10 lines
Signed-off-by Hannes Reinecke <[email protected]>--- drivers/scsi/scsi_error.c | 5 +++++ 1 file changed, 5 insertions(+)drivers/scsi/scsi_error.c +0
-0
drivers/scsi/scsi_error.c scsi_abort_command(struct scsi_cmnd *scmd) · line 245
Quoted 4 lines
drivers/scsi/scsi_error.c +1
-0@@ -242,6 +242,11 @@ scsi_abort_command(struct scsi_cmnd *scmd)return FAILED ; } +if (!shost->hostt->eh_abort_handler) {
nit: no need for {}, but maybe better put comment above the check if removing it. However maybe it's also a bit obvious comment.
Quoted 7 lines
drivers/scsi/scsi_error.c +4
-0+/* No abort handler, fail command directly */ +return FAILED ; + } +spin_lock_irqsave (shost->host_lock, flags);if (shost->eh_deadline !=-1 && !shost->last_reset) shost->last_reset = jiffies;
Quoted 34 lines · Hannes Reinecke
If a host template doesn't implement the .eh_abort_handler() there is no point in queueing the abort workqueue function; all it does is invoking SCSI EH anyway.
So return 'FAILED' fromscsi_abort_command()if the .eh_abort_handler() is not implemented and save us from having to wait for the abort workqueue function to complete.Cc Niklas Cassel <[email protected]>Cc Damien Le Moal <[email protected]>Cc John Garry <[email protected]>Signed-off-by Hannes Reinecke <[email protected]>--- drivers/scsi/scsi_error.c | 5 +++++ 1 file changed, 5 insertions(+)drivers/scsi/scsi_error.c +5
-0@@ -242,6 +242,11 @@ scsi_abort_command(struct scsi_cmnd *scmd)return FAILED ; } +if (!shost->hostt->eh_abort_handler) { +/* No abort handler, fail command directly */ +return FAILED ; + } +spin_lock_irqsave (shost->host_lock, flags);if (shost->eh_deadline !=-1 && !shost->last_reset) shost->last_reset = jiffies;Signature · 1 line
2.35.3
Re: [External] : [PATCH] scsi_error: do not queue pointless abort workqueue functions
In reply to John GarryQuoted 10 lines · John Garry
Quoted 6 lines · Hannes Reinecke
If a host template doesn't implement the .eh_abort_handler() there is no point in queueing the abort workqueue function; all it does is invoking SCSI EH anyway.
So return 'FAILED' fromscsi_abort_command()if the .eh_abort_handler() is not implemented and save us from having to wait for the abort workqueue function to complete.Do we ever use
shost->tmf_work_qin this case? Doesn't seem much point in allocating it, apart from keeping the code simpler
Actually, no. Guess we can skip allocating it.
Quoted 7 lines
Quoted 4 lines
Cc Niklas Cassel <[email protected]>Cc Damien Le Moal <[email protected]>Cc John Garry <[email protected]>That's someone else :)
Oh. Sorry, John :-)
Quoted 5 lines
Quoted 5 lines
Signed-off-by Hannes Reinecke <[email protected]>---
drivers/scsi/scsi_error.c| 5 +++++1 file changed, 5 insertions(+)
Quoted 11 lines
drivers/scsi/scsi_error.c scsi_abort_command(struct scsi_cmnd *scmd) · line 244
Quoted 8 lines
drivers/scsi/scsi_error.c +1
-0@@ -242,6 +242,11 @@ scsi_abort_command(struct scsi_cmnd *scmd) return FAILED; } + if (!shost->hostt->eh_abort_handler) {nit: no need for {}, but maybe better put comment above the check if removing it. However maybe it's also a bit obvious comment.
Yeah, will do.
Quoted 8 lines
drivers/scsi/scsi_error.c scsi_abort_command(struct scsi_cmnd *scmd) · line 252
Quoted 7 lines
drivers/scsi/scsi_error.c +4
-0+ /* No abort handler, fail command directly */ + return FAILED; + } + spin_lock_irqsave(shost->host_lock, flags); if (shost->eh_deadline != -1 && !shost->last_reset) shost->last_reset = jiffies;
Cheers,
Hannes
Signature · 4 lines
Dr. Hannes Reinecke Kernel Storage Architect [email protected] +49 911 74053 688 SUSE Software Solutions Germany GmbH, Maxfeldstr. 5, 90409 Nürnberg HRB 36809 (AG Nürnberg), GF: Felix Imendörffer
Quoted 20 lines · Hannes Reinecke
If a host template doesn't implement the .eh_abort_handler() there is no point in queueing the abort workqueue function; all it does is invoking SCSI EH anyway.
So return 'FAILED' fromscsi_abort_command()if the .eh_abort_handler() is not implemented and save us from having to wait for the abort workqueue function to complete.Cc Niklas Cassel <[email protected]>Cc Damien Le Moal <[email protected]>Cc John Garry <[email protected]>Signed-off-by Hannes Reinecke <[email protected]>--- drivers/scsi/scsi_error.c | 5 +++++ 1 file changed, 5 insertions(+)drivers/scsi/scsi_error.c +0
-0
drivers/scsi/scsi_error.c scsi_abort_command(struct scsi_cmnd *scmd) · line 249
Quoted 8 lines
drivers/scsi/scsi_error.c +5
-0@@ -242,6 +242,11 @@ scsi_abort_command(struct scsi_cmnd *scmd)return FAILED ; } +if (!shost->hostt->eh_abort_handler) { +/* No abort handler, fail command directly */ +return FAILED ; + } +
Hello Hannes,
is there any reason why you didn't put this before the preceding if (scmd->eh_eflags & SCSI_EH_ABORT_SCHEDULED) {
if statement?
Quoted 6 lines
drivers/scsi/scsi_error.c +0
-1spin_lock_irqsave (shost->host_lock, flags);if (shost->eh_deadline !=-1 && !shost->last_reset) shost->last_reset = jiffies; --2.35.3
After some additional testing with this patch, I did notice that it does introduce a behavioural change from libata perspective:
Before this patch, for libata, scmd_eh_abort_handler() would get called, and we would come into this statement:
rtn = scsi_try_to_abort_cmd(shost->hostt, scmd);
if (rtn != SUCCESS) {
SCSI_LOG_ERROR_RECOVERY(3,
scmd_printk(KERN_INFO, scmd,
"cmd abort %s\n",
(rtn == FAST_IO_FAIL) ?
"not send" : "failed"));
goto out;
}Which jumps to:
out:
spin_lock_irqsave(shost->host_lock, flags); list_del_init(&scmd->eh_entry); spin_unlock_irqrestore(shost->host_lock, flags); scsi_eh_scmd_add(scmd);
So scsi_eh_scmd_add() would be called.
After this patch, scsi_abort_command() will return FAILED instead of SUCCESS, which means that scsi_timeout() instead enters this if statement:
if (scsi_abort_command(scmd) != SUCCESS) {
set_host_byte(scmd, DID_TIME_OUT);
scsi_eh_scmd_add(scmd);
}Which means that scmds reaching libata .eh_strategy_handler() now has host_byte DID_TIME_OUT set, while before this patch, that was not the case.
I guess we could simply clear the host_byte in libata's .eh_strategy_handler() (and that is actually what we do).
I just want to understand how it is meant to work.
Looking back at the code to when libata first started to use blk_abort_request()/scsi_req_abort_cmd(), DID_TIME_OUT was only set if scsi_eh_scmd_add() failed:
https://github.com/torvalds/linux/blob/7b70fc039824bc7303e4007a5f758f832de56611/drivers/scsi/scsi_error.c#L181
Martin then moved the set_host_byte(scmd, DID_TIME_OUT)
to be done regardless if scsi_eh_scmd_add() failed or not in:18a4d0a22ed6 ("[SCSI] Handle disk devices which can not process medium access commands") without really explaining why.
Then in your commit:2171b6d08bf8 ("scsi: make scsi_eh_scmd_add() always succeed")
you changed scsi_times_out() to:
+ if (host->hostt->no_async_abort ||
+ scsi_abort_command(scmd) != SUCCESS) {
+ set_host_byte(scmd, DID_TIME_OUT);
+ scsi_eh_scmd_add(scmd, SCSI_EH_CANCEL_CMD);
+ }
And it also changed scmd_eh_abort_handler():
- if (!scsi_eh_scmd_add(scmd, 0)) {
- SCSI_LOG_ERROR_RECOVERY(3,
- scmd_printk(KERN_WARNING, scmd,
- "terminate aborted command\n"));
- set_host_byte(scmd, DID_TIME_OUT);
- scsi_finish_command(scmd);
- }
+ scsi_eh_scmd_add(scmd, 0);
So for libata, which did not set host->hostt->no_async_abort, scsi_abort_command() would return SUCCESS, so we would not go into that if statement. Instead we would have the current behavior where scmd_eh_abort_handler() fails, and does a goto out; to add the scsi_eh_scmd_add() (without setting DID_TIME_OUT).
Should perhaps scmd_eh_abort_handler() perhaps set DID_TIME_OUT unconditionally, to match the code before the change?
To me, it is not really clear how the SCSI code is meant to behave.
I think if the timeout has actually triggered, because the timer expired, it makes sense that scsi_timeout() sets DID_TIME_OUT.
But if e.g. libata called blk_abort_request() to abort the command before the timer actually expired, I'm not sure.
For ata part, it does not really matter, because currently, libata always overwrites the scmd->result anyway. However, there might be other LLDD where this change actually do matter.
(For the curious, libata's own way of detecting if the command actually was a timeout from scsi_timeout() or if it was an aborted command works like this:
If QCFLAG_FAILED is not set in .eh_strategy_handler, then libata EH does not own the QC, so it was scsi_timeout() that won the race, without libata ever aborting the command. It then sets qc->err_mask = AC_ERR_TIMEOUT).
So libata currently never looks at host_byte(), it always overwrites it, and it instead uses its own way of detecting that a timeout occured, if so, it freezes the port (disables IRQs) and resets the controller, and increases smcd->allowed, such that the command is retried.)
Kind regards,
Niklas