* pm8001: locking in mpi_sata_completion() is broken
@ 2011-12-14 9:29 Dan Carpenter
2011-12-15 1:26 ` Jack Wang
0 siblings, 1 reply; 4+ messages in thread
From: Dan Carpenter @ 2011-12-14 9:29 UTC (permalink / raw)
To: jack_wang; +Cc: lindar_liu, linux-scsi
The locking in mpi_sata_completion() is broken. I'm not sure what was
intended at all. Here is what it does:
1876 mpi_sata_completion(struct pm8001_hba_info *pm8001_ha, void *piomb)
1877 {
1878 struct sas_task *t;
1879 struct pm8001_ccb_info *ccb;
1880 unsigned long flags = 0;
^^^^^^^^^^^^^^^^^^^^^^^
This is bogus. The flags should be set with irqsave. This implies a
knowledge of the internals which should be abstracted away.
[snip]
2006 case IO_OPEN_CNX_ERROR_IT_NEXUS_LOSS:
2007 PM8001_IO_DBG(pm8001_ha,
2008 pm8001_printk("IO_OPEN_CNX_ERROR_IT_NEXUS_LOSS\n"));
2009 ts->resp = SAS_TASK_COMPLETE;
2010 ts->stat = SAS_DEV_NO_RESPONSE;
2011 if (!t->uldd_task) {
2012 pm8001_handle_event(pm8001_ha,
2013 pm8001_dev,
2014 IO_OPEN_CNX_ERROR_IT_NEXUS_LOSS);
2015 ts->resp = SAS_TASK_UNDELIVERED;
2016 ts->stat = SAS_QUEUE_FULL;
2017 pm8001_ccb_task_free(pm8001_ha, t, ccb, tag);
2018 mb();/*in order to force CPU ordering*/
2019 spin_unlock_irqrestore(&pm8001_ha->lock, flags);
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
Calling irqrestore before we've done an irqsave. What are we trying to
restore?
2020 t->task_done(t);
2021 spin_lock_irqsave(&pm8001_ha->lock, flags);
2022 return;
2023 }
[snip]
2197 spin_unlock_irqrestore(&t->task_state_lock, flags);
2198 pm8001_ccb_task_free(pm8001_ha, t, ccb, tag);
2199 mb();/*ditto*/
2200 spin_unlock_irqrestore(&pm8001_ha->lock, flags);
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
We just called irqrestore so there is no need to restore it twice.
2201 t->task_done(t);
2202 spin_lock_irqsave(&pm8001_ha->lock, flags);
^^^^^
We're saving the irq status but we're at the end of the function so we
never use it again.
2203 }
2204 }
I've just picked out a couple examples, but basically the whole function
is like that.
regards,
dan carpenter
^ permalink raw reply [flat|nested] 4+ messages in thread
* RE: pm8001: locking in mpi_sata_completion() is broken
2011-12-14 9:29 pm8001: locking in mpi_sata_completion() is broken Dan Carpenter
@ 2011-12-15 1:26 ` Jack Wang
2011-12-15 6:15 ` Dan Carpenter
0 siblings, 1 reply; 4+ messages in thread
From: Jack Wang @ 2011-12-15 1:26 UTC (permalink / raw)
To: 'Dan Carpenter'; +Cc: lindar_liu, linux-scsi
RE: pm8001: locking in mpi_sata_completion() is broken
>
> The locking in mpi_sata_completion() is broken. I'm not sure what was
> intended at all. Here is what it does:
>
> 1876 mpi_sata_completion(struct pm8001_hba_info *pm8001_ha, void
*piomb)
> 1877 {
> 1878 struct sas_task *t;
> 1879 struct pm8001_ccb_info *ccb;
> 1880 unsigned long flags = 0;
> ^^^^^^^^^^^^^^^^^^^^^^^
> This is bogus. The flags should be set with irqsave. This implies a
> knowledge of the internals which should be abstracted away.
>
[Jack Wang]
Hi Dan,
Thanks for point out this. Other comments answer below.
> [snip]
>
> 2006 case IO_OPEN_CNX_ERROR_IT_NEXUS_LOSS:
> 2007 PM8001_IO_DBG(pm8001_ha,
> 2008
> pm8001_printk("IO_OPEN_CNX_ERROR_IT_NEXUS_LOSS\n"));
> 2009 ts->resp = SAS_TASK_COMPLETE;
> 2010 ts->stat = SAS_DEV_NO_RESPONSE;
> 2011 if (!t->uldd_task) {
> 2012 pm8001_handle_event(pm8001_ha,
> 2013 pm8001_dev,
> 2014 IO_OPEN_CNX_ERROR_IT_NEXUS_LOSS);
> 2015 ts->resp = SAS_TASK_UNDELIVERED;
> 2016 ts->stat = SAS_QUEUE_FULL;
> 2017 pm8001_ccb_task_free(pm8001_ha, t, ccb,
> tag);
> 2018 mb();/*in order to force CPU ordering*/
> 2019 spin_unlock_irqrestore(&pm8001_ha->lock,
> flags);
>
> ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
> Calling irqrestore before we've done an irqsave. What are we trying to
> restore?
[Jack Wang]
We hold spin_lock_irqsave when we process interrupt, so here we need restore
it to do task done clean.
>
> 2020 t->task_done(t);
> 2021 spin_lock_irqsave(&pm8001_ha->lock,
> flags);
> 2022 return;
> 2023 }
>
> [snip]
>
> 2197 spin_unlock_irqrestore(&t->task_state_lock,
> flags);
> 2198 pm8001_ccb_task_free(pm8001_ha, t, ccb, tag);
> 2199 mb();/*ditto*/
> 2200 spin_unlock_irqrestore(&pm8001_ha->lock, flags);
> ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
> We just called irqrestore so there is no need to restore it twice.
>
> 2201 t->task_done(t);
> 2202 spin_lock_irqsave(&pm8001_ha->lock, flags);
> ^^^^^
> We're saving the irq status but we're at the end of the function so we
> never use it again.
>
> 2203 }
> 2204 }
>
> I've just picked out a couple examples, but basically the whole function
> is like that.
[Jack Wang]
Others are basically the same, what we want is make the lock balance.
Thanks.
>
> regards,
> dan carpenter
> --
> To unsubscribe from this list: send the line "unsubscribe linux-scsi" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at http://vger.kernel.org/majordomo-info.html
^ permalink raw reply [flat|nested] 4+ messages in thread
* Re: pm8001: locking in mpi_sata_completion() is broken
2011-12-15 1:26 ` Jack Wang
@ 2011-12-15 6:15 ` Dan Carpenter
2011-12-15 9:03 ` Jack Wang
0 siblings, 1 reply; 4+ messages in thread
From: Dan Carpenter @ 2011-12-15 6:15 UTC (permalink / raw)
To: Jack Wang; +Cc: lindar_liu, linux-scsi
[-- Attachment #1: Type: text/plain, Size: 945 bytes --]
On Thu, Dec 15, 2011 at 09:26:46AM +0800, Jack Wang wrote:
> > 2018 mb();/*in order to force CPU ordering*/
> > 2019 spin_unlock_irqrestore(&pm8001_ha->lock,
> > flags);
> >
> > ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
> > Calling irqrestore before we've done an irqsave. What are we trying to
> > restore?
> [Jack Wang]
> We hold spin_lock_irqsave when we process interrupt, so here we need restore
> it to do task done clean.
You're not restoring a previous saved state, flags is set to zero
here. You're calling:
spin_unlock_irqrestore(&pm8001_ha->lock, 0);
The whole function is badly broken. mpi_sata_event() has similar
problems.
With spin_lock_irqsave() and spin_unlock_irqrestore() you first have
to save the state to "flags" before you can restore it. Doing it
the other way round doesn't make any sort sense.
regards,
dan carpenter
[-- Attachment #2: Digital signature --]
[-- Type: application/pgp-signature, Size: 836 bytes --]
^ permalink raw reply [flat|nested] 4+ messages in thread
* RE: pm8001: locking in mpi_sata_completion() is broken
2011-12-15 6:15 ` Dan Carpenter
@ 2011-12-15 9:03 ` Jack Wang
0 siblings, 0 replies; 4+ messages in thread
From: Jack Wang @ 2011-12-15 9:03 UTC (permalink / raw)
To: 'Dan Carpenter'; +Cc: lindar_liu, linux-scsi
So could you send a patch to not set the flags to 0, or I post a patch to do
this later.
Thanks
Jack
>
> On Thu, Dec 15, 2011 at 09:26:46AM +0800, Jack Wang wrote:
> > > 2018 mb();/*in order to force CPU
ordering*/
> > > 2019
> spin_unlock_irqrestore(&pm8001_ha->lock,
> > > flags);
> > >
> > > ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
> > > Calling irqrestore before we've done an irqsave. What are we trying
to
> > > restore?
> > [Jack Wang]
> > We hold spin_lock_irqsave when we process interrupt, so here we need
restore
> > it to do task done clean.
>
>
> You're not restoring a previous saved state, flags is set to zero
> here. You're calling:
> spin_unlock_irqrestore(&pm8001_ha->lock, 0);
>
> The whole function is badly broken. mpi_sata_event() has similar
> problems.
>
> With spin_lock_irqsave() and spin_unlock_irqrestore() you first have
> to save the state to "flags" before you can restore it. Doing it
> the other way round doesn't make any sort sense.
>
> regards,
> dan carpenter
^ permalink raw reply [flat|nested] 4+ messages in thread
end of thread, other threads:[~2011-12-15 9:04 UTC | newest]
Thread overview: 4+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2011-12-14 9:29 pm8001: locking in mpi_sata_completion() is broken Dan Carpenter
2011-12-15 1:26 ` Jack Wang
2011-12-15 6:15 ` Dan Carpenter
2011-12-15 9:03 ` Jack Wang
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox