| Message ID | 20200817222340.18042-1-yung-chuan.liao@linux.intel.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | soundwire: cadence: fix race condition between suspend and Slave device alerts | expand |
On 18-08-20, 06:23, Bard Liao wrote: > From: Pierre-Louis Bossart <pierre-louis.bossart@linux.intel.com> > > In system suspend stress cases, the SOF CI reports timeouts. The root > cause is that an alert is generated while the system suspends. The > interrupt handling generates transactions on the bus that will never > be handled because the interrupts are disabled in parallel. > > As a result, the transaction never completes and times out on resume. > This error doesn't seem too problematic since it happens in a work > queue, and the system recovers without issues. > > Nevertheless, this race condition should not happen. When doing a > system suspend, or when disabling interrupts, we should make sure the > current transaction can complete, and prevent new work from being > queued. > > BugLink: https://github.com/thesofproject/linux/issues/2344 > Signed-off-by: Pierre-Louis Bossart <pierre-louis.bossart@linux.intel.com> > Reviewed-by: Ranjani Sridharan <ranjani.sridharan@linux.intel.com> > Reviewed-by: Rander Wang <rander.wang@linux.intel.com> > Signed-off-by: Bard Liao <yung-chuan.liao@linux.intel.com> > --- > drivers/soundwire/cadence_master.c | 24 +++++++++++++++++++++++- > drivers/soundwire/cadence_master.h | 1 + > 2 files changed, 24 insertions(+), 1 deletion(-) > > diff --git a/drivers/soundwire/cadence_master.c b/drivers/soundwire/cadence_master.c > index 24eafe0aa1c3..1330ffc47596 100644 > --- a/drivers/soundwire/cadence_master.c > +++ b/drivers/soundwire/cadence_master.c > @@ -791,7 +791,16 @@ irqreturn_t sdw_cdns_irq(int irq, void *dev_id) > CDNS_MCP_INT_SLAVE_MASK, 0); > > int_status &= ~CDNS_MCP_INT_SLAVE_MASK; > - schedule_work(&cdns->work); > + > + /* > + * Deal with possible race condition between interrupt > + * handling and disabling interrupts on suspend. > + * > + * If the master is in the process of disabling > + * interrupts, don't schedule a workqueue > + */ > + if (cdns->interrupt_enabled) > + schedule_work(&cdns->work); would it not make sense to mask the interrupts first and then cancel the work? that way you are guaranteed that after this call you dont have interrupts and work scheduled? > } > > cdns_writel(cdns, CDNS_MCP_INTSTAT, int_status); > @@ -924,6 +933,19 @@ int sdw_cdns_enable_interrupt(struct sdw_cdns *cdns, bool state) > slave_state = cdns_readl(cdns, CDNS_MCP_SLAVE_INTSTAT1); > cdns_writel(cdns, CDNS_MCP_SLAVE_INTSTAT1, slave_state); > } > + cdns->interrupt_enabled = state; > + > + /* > + * Complete any on-going status updates before updating masks, > + * and cancel queued status updates. > + * > + * There could be a race with a new interrupt thrown before > + * the 3 mask updates below are complete, so in the interrupt > + * we use the 'interrupt_enabled' status to prevent new work > + * from being queued. > + */ > + if (!state) > + cancel_work_sync(&cdns->work); > > cdns_writel(cdns, CDNS_MCP_SLAVE_INTMASK0, slave_intmask0); > cdns_writel(cdns, CDNS_MCP_SLAVE_INTMASK1, slave_intmask1); > diff --git a/drivers/soundwire/cadence_master.h b/drivers/soundwire/cadence_master.h > index fdec62b912d3..4d1aab5b5ec2 100644 > --- a/drivers/soundwire/cadence_master.h > +++ b/drivers/soundwire/cadence_master.h > @@ -133,6 +133,7 @@ struct sdw_cdns { > > bool link_up; > unsigned int msg_count; > + bool interrupt_enabled; > > struct work_struct work; > > -- > 2.17.1
On 8/19/20 4:06 AM, Vinod Koul wrote: > On 18-08-20, 06:23, Bard Liao wrote: >> From: Pierre-Louis Bossart <pierre-louis.bossart@linux.intel.com> >> >> In system suspend stress cases, the SOF CI reports timeouts. The root >> cause is that an alert is generated while the system suspends. The >> interrupt handling generates transactions on the bus that will never >> be handled because the interrupts are disabled in parallel. >> >> As a result, the transaction never completes and times out on resume. >> This error doesn't seem too problematic since it happens in a work >> queue, and the system recovers without issues. >> >> Nevertheless, this race condition should not happen. When doing a >> system suspend, or when disabling interrupts, we should make sure the >> current transaction can complete, and prevent new work from being >> queued. >> >> BugLink: https://github.com/thesofproject/linux/issues/2344 >> Signed-off-by: Pierre-Louis Bossart <pierre-louis.bossart@linux.intel.com> >> Reviewed-by: Ranjani Sridharan <ranjani.sridharan@linux.intel.com> >> Reviewed-by: Rander Wang <rander.wang@linux.intel.com> >> Signed-off-by: Bard Liao <yung-chuan.liao@linux.intel.com> >> --- >> drivers/soundwire/cadence_master.c | 24 +++++++++++++++++++++++- >> drivers/soundwire/cadence_master.h | 1 + >> 2 files changed, 24 insertions(+), 1 deletion(-) >> >> diff --git a/drivers/soundwire/cadence_master.c b/drivers/soundwire/cadence_master.c >> index 24eafe0aa1c3..1330ffc47596 100644 >> --- a/drivers/soundwire/cadence_master.c >> +++ b/drivers/soundwire/cadence_master.c >> @@ -791,7 +791,16 @@ irqreturn_t sdw_cdns_irq(int irq, void *dev_id) >> CDNS_MCP_INT_SLAVE_MASK, 0); >> >> int_status &= ~CDNS_MCP_INT_SLAVE_MASK; >> - schedule_work(&cdns->work); >> + >> + /* >> + * Deal with possible race condition between interrupt >> + * handling and disabling interrupts on suspend. >> + * >> + * If the master is in the process of disabling >> + * interrupts, don't schedule a workqueue >> + */ >> + if (cdns->interrupt_enabled) >> + schedule_work(&cdns->work); > > would it not make sense to mask the interrupts first and then cancel the > work? that way you are guaranteed that after this call you dont have > interrupts and work scheduled? cancel_work_sync() will either a) wait until the current work completes, or b) prevent a new one from starting. there's no way to really 'abort' a workqueue, 'cancel' means either complete or don't start. if you disable the interrupts then cancel the work, you have a risk of not letting the work complete if it already started (case a). The race is a) the interrupt thread (this function) starts b) the work is scheduled and starts c) the suspend handler starts and disables interrupts in [1] below. d) the work initiates transactions which will never complete since Cadence interrupts have been disabled. So the idea was that before disabling interrupts, the suspend handler changes the status, and then calls cancel_work_sync(). the status is also used to prevent a new work from being scheduled if you already know the suspend is on-going. The test on the status above is not strictly necessary, I believe the sequence is safe without it but it avoid starting a useless work. If you want to follow the flow it's better to start with what the suspend handler does below first, then look at how the interrupt thread might interfere. The diff format does not help, might be also easier to apply the patch and look at the rest of the code, e.g the 3 mask updates mentioned below are not included in the diff. > >> } >> >> cdns_writel(cdns, CDNS_MCP_INTSTAT, int_status); >> @@ -924,6 +933,19 @@ int sdw_cdns_enable_interrupt(struct sdw_cdns *cdns, bool state) >> slave_state = cdns_readl(cdns, CDNS_MCP_SLAVE_INTSTAT1); >> cdns_writel(cdns, CDNS_MCP_SLAVE_INTSTAT1, slave_state); >> } [1] >> + cdns->interrupt_enabled = state; >> + >> + /* >> + * Complete any on-going status updates before updating masks, >> + * and cancel queued status updates. >> + * >> + * There could be a race with a new interrupt thrown before >> + * the 3 mask updates below are complete, so in the interrupt >> + * we use the 'interrupt_enabled' status to prevent new work >> + * from being queued. >> + */ >> + if (!state) >> + cancel_work_sync(&cdns->work); >> >> cdns_writel(cdns, CDNS_MCP_SLAVE_INTMASK0, slave_intmask0); >> cdns_writel(cdns, CDNS_MCP_SLAVE_INTMASK1, slave_intmask1); >> diff --git a/drivers/soundwire/cadence_master.h b/drivers/soundwire/cadence_master.h >> index fdec62b912d3..4d1aab5b5ec2 100644 >> --- a/drivers/soundwire/cadence_master.h >> +++ b/drivers/soundwire/cadence_master.h >> @@ -133,6 +133,7 @@ struct sdw_cdns { >> >> bool link_up; >> unsigned int msg_count; >> + bool interrupt_enabled; >> >> struct work_struct work; >> >> -- >> 2.17.1 >
On 19-08-20, 07:51, Pierre-Louis Bossart wrote: > > > On 8/19/20 4:06 AM, Vinod Koul wrote: > > On 18-08-20, 06:23, Bard Liao wrote: > > > From: Pierre-Louis Bossart <pierre-louis.bossart@linux.intel.com> > > > > > > In system suspend stress cases, the SOF CI reports timeouts. The root > > > cause is that an alert is generated while the system suspends. The > > > interrupt handling generates transactions on the bus that will never > > > be handled because the interrupts are disabled in parallel. > > > > > > As a result, the transaction never completes and times out on resume. > > > This error doesn't seem too problematic since it happens in a work > > > queue, and the system recovers without issues. > > > > > > Nevertheless, this race condition should not happen. When doing a > > > system suspend, or when disabling interrupts, we should make sure the > > > current transaction can complete, and prevent new work from being > > > queued. > > > > > > BugLink: https://github.com/thesofproject/linux/issues/2344 > > > Signed-off-by: Pierre-Louis Bossart <pierre-louis.bossart@linux.intel.com> > > > Reviewed-by: Ranjani Sridharan <ranjani.sridharan@linux.intel.com> > > > Reviewed-by: Rander Wang <rander.wang@linux.intel.com> > > > Signed-off-by: Bard Liao <yung-chuan.liao@linux.intel.com> > > > --- > > > drivers/soundwire/cadence_master.c | 24 +++++++++++++++++++++++- > > > drivers/soundwire/cadence_master.h | 1 + > > > 2 files changed, 24 insertions(+), 1 deletion(-) > > > > > > diff --git a/drivers/soundwire/cadence_master.c b/drivers/soundwire/cadence_master.c > > > index 24eafe0aa1c3..1330ffc47596 100644 > > > --- a/drivers/soundwire/cadence_master.c > > > +++ b/drivers/soundwire/cadence_master.c > > > @@ -791,7 +791,16 @@ irqreturn_t sdw_cdns_irq(int irq, void *dev_id) > > > CDNS_MCP_INT_SLAVE_MASK, 0); > > > int_status &= ~CDNS_MCP_INT_SLAVE_MASK; > > > - schedule_work(&cdns->work); > > > + > > > + /* > > > + * Deal with possible race condition between interrupt > > > + * handling and disabling interrupts on suspend. > > > + * > > > + * If the master is in the process of disabling > > > + * interrupts, don't schedule a workqueue > > > + */ > > > + if (cdns->interrupt_enabled) > > > + schedule_work(&cdns->work); > > > > would it not make sense to mask the interrupts first and then cancel the > > work? that way you are guaranteed that after this call you dont have > > interrupts and work scheduled? > > cancel_work_sync() will either > a) wait until the current work completes, or > b) prevent a new one from starting. > > there's no way to really 'abort' a workqueue, 'cancel' means either complete > or don't start. Quite right, as that is how everyone deals with it. Stop the irq from firing first and then wait until work is cancelled or completed, hence cancel_work_sync() > if you disable the interrupts then cancel the work, you have a risk of not > letting the work complete if it already started (case a). > > The race is > a) the interrupt thread (this function) starts > b) the work is scheduled and starts > c) the suspend handler starts and disables interrupts in [1] below. > d) the work initiates transactions which will never complete since Cadence > interrupts have been disabled. Would it not be better to let work handle the case of interrupts disabled and not initiates transactions which wont complete here? That sounds more reasonable to do rather than complete the work which anyone doesn't matter as you are suspending > So the idea was that before disabling interrupts, the suspend handler > changes the status, and then calls cancel_work_sync(). the status is also > used to prevent a new work from being scheduled if you already know the > suspend is on-going. The test on the status above is not strictly necessary, > I believe the sequence is safe without it but it avoid starting a useless > work. > > If you want to follow the flow it's better to start with what the suspend > handler does below first, then look at how the interrupt thread might > interfere. The diff format does not help, might be also easier to apply the > patch and look at the rest of the code, e.g the 3 mask updates mentioned > below are not included in the diff. > > > > > > } > > > cdns_writel(cdns, CDNS_MCP_INTSTAT, int_status); > > > @@ -924,6 +933,19 @@ int sdw_cdns_enable_interrupt(struct sdw_cdns *cdns, bool state) > > > slave_state = cdns_readl(cdns, CDNS_MCP_SLAVE_INTSTAT1); > > > cdns_writel(cdns, CDNS_MCP_SLAVE_INTSTAT1, slave_state); > > > } > > [1] > > > > + cdns->interrupt_enabled = state; > > > + > > > + /* > > > + * Complete any on-going status updates before updating masks, > > > + * and cancel queued status updates. > > > + * > > > + * There could be a race with a new interrupt thrown before > > > + * the 3 mask updates below are complete, so in the interrupt > > > + * we use the 'interrupt_enabled' status to prevent new work > > > + * from being queued. > > > + */ > > > + if (!state) > > > + cancel_work_sync(&cdns->work); > > > cdns_writel(cdns, CDNS_MCP_SLAVE_INTMASK0, slave_intmask0); > > > cdns_writel(cdns, CDNS_MCP_SLAVE_INTMASK1, slave_intmask1); > > > diff --git a/drivers/soundwire/cadence_master.h b/drivers/soundwire/cadence_master.h > > > index fdec62b912d3..4d1aab5b5ec2 100644 > > > --- a/drivers/soundwire/cadence_master.h > > > +++ b/drivers/soundwire/cadence_master.h > > > @@ -133,6 +133,7 @@ struct sdw_cdns { > > > bool link_up; > > > unsigned int msg_count; > > > + bool interrupt_enabled; > > > struct work_struct work; > > > -- > > > 2.17.1 > >
>> cancel_work_sync() will either >> a) wait until the current work completes, or >> b) prevent a new one from starting. >> >> there's no way to really 'abort' a workqueue, 'cancel' means either complete >> or don't start. > > Quite right, as that is how everyone deals with it. Stop the irq from > firing first and then wait until work is cancelled or completed, hence > cancel_work_sync() No, this cannot work... The work queue in progress will initiate transactions which would never complete because the interrupts are disabled. >> if you disable the interrupts then cancel the work, you have a risk of not >> letting the work complete if it already started (case a). >> >> The race is >> a) the interrupt thread (this function) starts >> b) the work is scheduled and starts >> c) the suspend handler starts and disables interrupts in [1] below. >> d) the work initiates transactions which will never complete since Cadence >> interrupts have been disabled. > > Would it not be better to let work handle the case of interrupts > disabled and not initiates transactions which wont complete here? That > sounds more reasonable to do rather than complete the work which anyone > doesn't matter as you are suspending A in-progress workqueue has no notion that interrupts are disabled, nor that the device is in the process of suspending. It writes into a fifo and blocks with wait_for_completion(). the complete() is done in an interrupt thread, triggered when the RX Fifo reaches a watermark. So if you disable interrupts, the complete() never happens. The safe way to do things with the current code is to let the workqueue complete, then disable interrupts. We only disable interrupts on suspend, we don't need to test if interrupts are enabled for every single byte that's transmitted on the bus. Not to mention that this additional test would be racy as hell and require yet another synchronization primitive making the code more complicated. So yes, the current transactions will complete and will be ignored, but it's a lot better than trying to prevent these transactions from happening with extra layers of complexity that will impact *every* transaction. BTW I looked at another alternative based on the msg lock, so that interrupts cannot be disabled while a message is being sent. However because a workqueue may send multiple messages, e.g. to read registers are non-contiguous locations, there is no way to guarantee what happens how messages and interrupt disabling are scheduled, so there'd still be a case of a workqueue not completing and being stuck on a mutex_lock(), not so good either. In short, this is the simplest way to fix the timeout on resume.
On 21-08-20, 10:17, Pierre-Louis Bossart wrote: > > > > > cancel_work_sync() will either > > > a) wait until the current work completes, or > > > b) prevent a new one from starting. > > > > > > there's no way to really 'abort' a workqueue, 'cancel' means either complete > > > or don't start. > > > > Quite right, as that is how everyone deals with it. Stop the irq from > > firing first and then wait until work is cancelled or completed, hence > > cancel_work_sync() > > No, this cannot work... The work queue in progress will initiate > transactions which would never complete because the interrupts are disabled. Okay that makes sense then > > > if you disable the interrupts then cancel the work, you have a risk of not > > > letting the work complete if it already started (case a). > > > > > > The race is > > > a) the interrupt thread (this function) starts > > > b) the work is scheduled and starts > > > c) the suspend handler starts and disables interrupts in [1] below. > > > d) the work initiates transactions which will never complete since Cadence > > > interrupts have been disabled. > > > > Would it not be better to let work handle the case of interrupts > > disabled and not initiates transactions which wont complete here? That > > sounds more reasonable to do rather than complete the work which anyone > > doesn't matter as you are suspending > > A in-progress workqueue has no notion that interrupts are disabled, nor that > the device is in the process of suspending. It writes into a fifo and blocks > with wait_for_completion(). the complete() is done in an interrupt thread, > triggered when the RX Fifo reaches a watermark. > > So if you disable interrupts, the complete() never happens. > > The safe way to do things with the current code is to let the workqueue > complete, then disable interrupts. > > We only disable interrupts on suspend, we don't need to test if interrupts > are enabled for every single byte that's transmitted on the bus. Not to > mention that this additional test would be racy as hell and require yet > another synchronization primitive making the code more complicated. > > So yes, the current transactions will complete and will be ignored, but it's > a lot better than trying to prevent these transactions from happening with > extra layers of complexity that will impact *every* transaction. > > BTW I looked at another alternative based on the msg lock, so that > interrupts cannot be disabled while a message is being sent. However because > a workqueue may send multiple messages, e.g. to read registers are > non-contiguous locations, there is no way to guarantee what happens how > messages and interrupt disabling are scheduled, so there'd still be a case > of a workqueue not completing and being stuck on a mutex_lock(), not so good > either. > > In short, this is the simplest way to fix the timeout on resume. Is this timeout for suspend or resume? Somehow I was under the assumption that it is former? Or is the result seen on resume? Rereading the race describe above in steps, I think this should be handled in step c above. Btw is that suspend or runtime suspend which causes this? Former would be bigger issue as we should not have work running when we return from suspend call. Latter should be dealt with anyway as device might be off after suspend.
> Is this timeout for suspend or resume? Somehow I was under the > assumption that it is former? Or is the result seen on resume? > > Rereading the race describe above in steps, I think this should be > handled in step c above. Btw is that suspend or runtime suspend which > causes this? Former would be bigger issue as we should not have work > running when we return from suspend call. Latter should be dealt with > anyway as device might be off after suspend. This happens with a system suspend. Because we disable the interrupts, the workqueue never completes, and we have a timeout on system resume. That's why we want to prevent the workqueue from starting, or let it complete, but not have this zombie state where we suspend but there's still a wait for completion that times out later. The point here is really making sure the workqueue is not used before suspend.
Dne 28. 08. 20 v 17:14 Pierre-Louis Bossart napsal(a): > > > >> Is this timeout for suspend or resume? Somehow I was under the >> assumption that it is former? Or is the result seen on resume? >> >> Rereading the race describe above in steps, I think this should be >> handled in step c above. Btw is that suspend or runtime suspend which >> causes this? Former would be bigger issue as we should not have work >> running when we return from suspend call. Latter should be dealt with >> anyway as device might be off after suspend. > > This happens with a system suspend. Because we disable the interrupts, > the workqueue never completes, and we have a timeout on system resume. > > That's why we want to prevent the workqueue from starting, or let it > complete, but not have this zombie state where we suspend but there's > still a wait for completion that times out later. The point here is > really making sure the workqueue is not used before suspend. > Vinod, there is no acceptance progress on this. The patch is really straight and for the Intel controller. They know what they're doing. I would apply this. The code can be refined at anytime. It's a fix. I tested it and I can confirm, that it fixes the issue. It's a vital patch for 5.10 to enable finally SoundWire drivers for the Intel hardware. Acked-by: Jaroslav Kysela <perex@perex.cz>
On 08-09-20, 13:58, Jaroslav Kysela wrote: > Dne 28. 08. 20 v 17:14 Pierre-Louis Bossart napsal(a): > > > > > > > >> Is this timeout for suspend or resume? Somehow I was under the > >> assumption that it is former? Or is the result seen on resume? > >> > >> Rereading the race describe above in steps, I think this should be > >> handled in step c above. Btw is that suspend or runtime suspend which > >> causes this? Former would be bigger issue as we should not have work > >> running when we return from suspend call. Latter should be dealt with > >> anyway as device might be off after suspend. > > > > This happens with a system suspend. Because we disable the interrupts, > > the workqueue never completes, and we have a timeout on system resume. > > > > That's why we want to prevent the workqueue from starting, or let it > > complete, but not have this zombie state where we suspend but there's > > still a wait for completion that times out later. The point here is > > really making sure the workqueue is not used before suspend. > > > > Vinod, there is no acceptance progress on this. The patch is really straight > and for the Intel controller. They know what they're doing. I would apply > this. The code can be refined at anytime. It's a fix. I tested it and I can > confirm, that it fixes the issue. It's a vital patch for 5.10 to enable > finally SoundWire drivers for the Intel hardware. I do feel that there is something else going on, but not able to pin point, anyway this fixes the issue so I am applying it now > > Acked-by: Jaroslav Kysela <perex@perex.cz> Thanks for ack
diff --git a/drivers/soundwire/cadence_master.c b/drivers/soundwire/cadence_master.c index 24eafe0aa1c3..1330ffc47596 100644 --- a/drivers/soundwire/cadence_master.c +++ b/drivers/soundwire/cadence_master.c @@ -791,7 +791,16 @@ irqreturn_t sdw_cdns_irq(int irq, void *dev_id) CDNS_MCP_INT_SLAVE_MASK, 0); int_status &= ~CDNS_MCP_INT_SLAVE_MASK; - schedule_work(&cdns->work); + + /* + * Deal with possible race condition between interrupt + * handling and disabling interrupts on suspend. + * + * If the master is in the process of disabling + * interrupts, don't schedule a workqueue + */ + if (cdns->interrupt_enabled) + schedule_work(&cdns->work); } cdns_writel(cdns, CDNS_MCP_INTSTAT, int_status); @@ -924,6 +933,19 @@ int sdw_cdns_enable_interrupt(struct sdw_cdns *cdns, bool state) slave_state = cdns_readl(cdns, CDNS_MCP_SLAVE_INTSTAT1); cdns_writel(cdns, CDNS_MCP_SLAVE_INTSTAT1, slave_state); } + cdns->interrupt_enabled = state; + + /* + * Complete any on-going status updates before updating masks, + * and cancel queued status updates. + * + * There could be a race with a new interrupt thrown before + * the 3 mask updates below are complete, so in the interrupt + * we use the 'interrupt_enabled' status to prevent new work + * from being queued. + */ + if (!state) + cancel_work_sync(&cdns->work); cdns_writel(cdns, CDNS_MCP_SLAVE_INTMASK0, slave_intmask0); cdns_writel(cdns, CDNS_MCP_SLAVE_INTMASK1, slave_intmask1); diff --git a/drivers/soundwire/cadence_master.h b/drivers/soundwire/cadence_master.h index fdec62b912d3..4d1aab5b5ec2 100644 --- a/drivers/soundwire/cadence_master.h +++ b/drivers/soundwire/cadence_master.h @@ -133,6 +133,7 @@ struct sdw_cdns { bool link_up; unsigned int msg_count; + bool interrupt_enabled; struct work_struct work;