dmaengine: cppi41: Fix oops in cppi41_runtime_resume
PATCH 1/1 dmaengine: cppi41: Fix oops in cppi41_runtime_resume
Commit fdea2d09b997 ("dmaengine: cppi41: Add basic PM runtime support") together with recent MUSB changes allowed USB and DMA on BeagleBone to idle when no cable is connected. But looks like few corner case issues still remain.
Looks like just by repluging USB cable about ten or so times on BeagleBone when configured in USB peripheral mode we can get warnings and eventually trigger an oops in cppi41 DMA:
WARNING: CPU: 0 PID: 14 at drivers/dma/cppi41.c:1154 cppi41_runtime_suspend+
x28/0x38 [cppi41]
...
WARNING: CPU: 0 PID: 14 at drivers/dma/cppi41.c:452 push_desc_queue+0x94/0x9c [cppi41]
...
Unable to handle kernel NULL pointer dereference at virtual address 00000104 pgd = c0004000 [00000104] *pgd=00000000 Internal error: Oops: 805 [#1] SMP ARM
...
[<bf0d92cc>] (cppi41_runtime_resume [cppi41]) from [<c0589838>] (__rpm_callback+0xc0/0x214)
[<c0589838>] (__rpm_callback) from [<c05899ac>] (rpm_callback+0x20/0x80) [<c05899ac>] (rpm_callback) from [<c0589460>] (rpm_resume+0x504/0x78c) [<c0589460>] (rpm_resume) from [<c058a1a0>] (pm_runtime_work+0x60/0xa8) [<c058a1a0>] (pm_runtime_work) from [<c0156120>] (process_one_work+0x2b4/0x808)
This is because of a race with PM runtime and cppi41_dma_issue_pending()
as reported by Alexandre Bailon <abailon-rdvid1DuHRBWk0Htik3J/[email protected]> in earlier set of patches. Based on mailing list discussions we however came to the conclusion that a different fix from Alexandre's fix is needed in order
to guarantee that DMA is really active when we try to use it.
To fix the issue, we need to add a driver specific flag as we otherwise can have -EINPROGRESS state set by PM runtime and can't rely on pm_runtime_active() to tell us when we can use the DMA.
For reference, this is also documented in Documentation/power/runtime_pm.txt in the example at the end of the file as pointed out by Grygorii Strashko <[email protected]>.
So let's fix the issue by introducing atomic_t active that gets toggled in runtime_suspend() and runtime_resume(). And then let's replace the pm_runtime_active() checks with atomic_read().
Note that we may also get transactions queued while in -EINPROGRESS state. So we need to check for new elements in cppi41_runtime_resume() by replacing list_for_each_entry_safe() with the commonly used test for while (!list_empty(&cdd->pending)).
--- drivers/dma/cppi41.c | 23 ++++++++++++++++++----- 1 file changed, 18 insertions(+), 5 deletions(-)
drivers/dma/cppi41.c +18 -5
@@ -1,3 +1,4 @@ +#include <linux/atomic.h> #include <linux/delay.h> #include <linux/dmaengine.h> #include <linux/dma-mapping.h>
@@ -153,6 +154,8 @@ struct cppi41_dd {/* context for suspend/resume */ unsigned int dma_tdfdq; + +atomic_t active; };#define FIST_COMPLETION_QUEUE 93
@@ -320,7 +323,8 @@ static irqreturn_t cppi41_irq(int irq, void *data)int error; error =pm_runtime_get (cdd->ddev.dev); -if (error <0 ) +if (((error != -EINPROGRESS ) && error <0 ) || + !atomic_read (&cdd->active))dev_err (cdd->ddev.dev,"%s pm runtime get: %i\n" , __func__, error);
@@ -482,7 +486,7 @@ static void cppi41_dma_issue_pending(struct dma_chan *chan)return ; } -if (likely (pm_runtime_active (cdd->ddev.dev))) +if (likely (atomic_read (&cdd->active)))push_desc_queue (c);else pending_desc (c);
@@ -1003,6 +1007,7 @@ static int cppi41_dma_probe(struct platform_device *pdev) cdd->ddev.dst_addr_widths =CPPI41_DMA_BUSWIDTHS ; cdd->ddev.residue_granularity =DMA_RESIDUE_GRANULARITY_BURST ; cdd->ddev.dev = dev; +atomic_set (&cdd->active,0 );INIT_LIST_HEAD (&cdd->ddev.channels); cpp41_dma_info.dma_cap = cdd->ddev.cap_mask;
@@ -1151,6 +1156,7 @@ static int __maybe_unused cppi41_runtime_suspend(struct device *dev) {struct cppi41_dd *cdd =dev_get_drvdata (dev); +atomic_set (&cdd->active,0 );WARN_ON (!list_empty (&cdd->pending));return 0 ;
@@ -1159,13 +1165,20 @@ static int __maybe_unused cppi41_runtime_suspend(struct device *dev)static int __maybe_unused cppi41_runtime_resume (struct device *dev) {struct cppi41_dd *cdd =dev_get_drvdata (dev); -struct cppi41_channel *c, *_c; +struct cppi41_channel *c;unsigned long flags; +atomic_set (&cdd->active,1 ); +spin_lock_irqsave (&cdd->lock, flags); -list_for_each_entry_safe (c, _c, &cdd->pending, node) { -push_desc_queue (c); +while (!list_empty (&cdd->pending)) { + c =list_first_entry (&cdd->pending, + struct cppi41_channel, + node);list_del (&c->node); +spin_unlock_irqrestore (&cdd->lock, flags); +push_desc_queue (c); +spin_lock_irqsave (&cdd->lock, flags); }spin_unlock_irqrestore (&cdd->lock, flags);
Signature · 5 lines
2.11.0 -- To unsubscribe from this list: send the line "unsubscribe linux-usb" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html
A message that is not in the archive (sent off-list or not yet received): 20170112213016.19367-1-tony-4v6yS6AI5VpBDgjK7y7TUQ@public.gmane.org
Quoted 94 lines · Tony Lindgren
Commit
fdea2d09b997("dmaengine: cppi41: Add basic PM runtime support") together with recent MUSB changes allowed USB and DMA on BeagleBone to idle when no cable is connected. But looks like few corner case issues still remain.Looks like just by repluging USB cable about ten or so times on BeagleBone when configured in USB peripheral mode we can get warnings and eventually trigger an oops in cppi41 DMA:
WARNING: CPU: 0 PID: 14 at drivers/dma/cppi41.c:1154 cppi41_runtime_suspend+x28/0x38 [cppi41]
...WARNING: CPU: 0 PID: 14 at drivers/dma/cppi41.c:452 push_desc_queue+0x94/0x9c [cppi41]...
Unable to handle kernel NULL pointer dereference at virtual address 00000104 pgd = c0004000 [00000104] *pgd=00000000 Internal error: Oops: 805 [#1] SMP ARM...
[<bf0d92cc>] (cppi41_runtime_resume[cppi41]) from [<c0589838>] (__rpm_callback+0xc0/0x214)
[<c0589838>] (__rpm_callback) from [<c05899ac>] (rpm_callback+0x20/0x80) [<c05899ac>] (rpm_callback) from [<c0589460>] (rpm_resume+0x504/0x78c) [<c0589460>] (rpm_resume) from [<c058a1a0>] (pm_runtime_work+0x60/0xa8) [<c058a1a0>] (pm_runtime_work) from [<c0156120>] (process_one_work+0x2b4/0x808)This is because of a race with PM runtime and
cppi41_dma_issue_pending()
as reported by Alexandre Bailon <abailon-rdvid1DuHRBWk0Htik3J/[email protected]> in earlier set of patches. Based on mailing list discussions we however came to the conclusion that a different fix from Alexandre's fix is needed in order
to guarantee that DMA is really active when we try to use it.To fix the issue, we need to add a driver specific flag as we otherwise can have
-EINPROGRESSstate set by PM runtime and can't rely onpm_runtime_active()to tell us when we can use the DMA.For reference, this is also documented in
Documentation/power/runtime_pm.txtin the example at the end of the file as pointed out by Grygorii Strashko <[email protected]>.So let's fix the issue by introducing
atomic_tactive that gets toggled inruntime_suspend()andruntime_resume(). And then let's replace thepm_runtime_active()checks withatomic_read().Note that we may also get transactions queued while in
-EINPROGRESSstate. So we need to check for new elements incppi41_runtime_resume()by replacinglist_for_each_entry_safe()with the commonly used test for while (!list_empty(&cdd->pending)).Fixes fdea2d09b997 ("dmaengine: cppi41: Add basic PM runtime support")Cc Andy Shevchenko <[email protected]>Cc Bin Liu <[email protected]>Cc Grygorii Strashko <[email protected]>Cc Kevin Hilman <khilman-rdvid1DuHRBWk0Htik3J/[email protected]>Cc Patrick Titiano <ptitiano-rdvid1DuHRBWk0Htik3J/[email protected]>Cc Sergei Shtylyov <sergei.shtylyov-M4DtvfQ/[email protected]>Signed-off-by Tony Lindgren <[email protected]>--- drivers/dma/cppi41.c | 23 ++++++++++++++++++----- 1 file changed, 18 insertions(+), 5 deletions(-)drivers/dma/cppi41.c +5
-1@@ -1,3 +1,4 @@ +#include <linux/atomic.h> #include <linux/delay.h> #include <linux/dmaengine.h> #include <linux/dma-mapping.h> @@ -153,6 +154,8 @@ struct cppi41_dd {/* context for suspend/resume */ unsigned int dma_tdfdq; + +atomic_t active; };#define FIST_COMPLETION_QUEUE 93@@ -320,7 +323,8 @@ static irqreturn_t cppi41_irq(int irq, void *data)int error; error =pm_runtime_get (cdd->ddev.dev); -if (error <0 ) +if (((error != -EINPROGRESS ) && error <0 ) || + !atomic_read (&cdd->active))dev_err (cdd->ddev.dev,"%s pm runtime get: %i\n" , __func__, error);
drivers/dma/cppi41.c static void cppi41_dma_issue_pending(struct dma_chan *chan) · line 488
Quoted 3 lines
drivers/dma/cppi41.c +0
-0@@ -482,7 +486,7 @@ static void cppi41_dma_issue_pending(struct dma_chan *chan)return ; }
Sry, but even if it looks better it might still race with PM runtime :(
drivers/dma/cppi41.c static void cppi41_dma_issue_pending(struct dma_chan *chan) · line 491
Quoted 4 lines
drivers/dma/cppi41.c +1
-1-if (likely (pm_runtime_active (cdd->ddev.dev))) +if (likely (atomic_read (&cdd->active)))push_desc_queue (c);else
- CPU is here (-EINPROGRESS and active == 0) and then preempted
- PM runtime will finish cppi41_runtime_resume and clean-up pending descs
- CPU return here and adds desc to the pending queue
- oops
Am I wrong?
Quoted 42 lines
drivers/dma/cppi41.c +12
-3pending_desc (c);@@ -1003,6 +1007,7 @@ static int cppi41_dma_probe(struct platform_device *pdev) cdd->ddev.dst_addr_widths =CPPI41_DMA_BUSWIDTHS ; cdd->ddev.residue_granularity =DMA_RESIDUE_GRANULARITY_BURST ; cdd->ddev.dev = dev; +atomic_set (&cdd->active,0 );INIT_LIST_HEAD (&cdd->ddev.channels); cpp41_dma_info.dma_cap = cdd->ddev.cap_mask;@@ -1151,6 +1156,7 @@ static int __maybe_unused cppi41_runtime_suspend(struct device *dev) {struct cppi41_dd *cdd =dev_get_drvdata (dev); +atomic_set (&cdd->active,0 );WARN_ON (!list_empty (&cdd->pending));return 0 ;@@ -1159,13 +1165,20 @@ static int __maybe_unused cppi41_runtime_suspend(struct device *dev)static int __maybe_unused cppi41_runtime_resume (struct device *dev) {struct cppi41_dd *cdd =dev_get_drvdata (dev); -struct cppi41_channel *c, *_c; +struct cppi41_channel *c;unsigned long flags; +atomic_set (&cdd->active,1 ); +spin_lock_irqsave (&cdd->lock, flags); -list_for_each_entry_safe (c, _c, &cdd->pending, node) { -push_desc_queue (c); +while (!list_empty (&cdd->pending)) { + c =list_first_entry (&cdd->pending, + struct cppi41_channel, + node);list_del (&c->node); +spin_unlock_irqrestore (&cdd->lock, flags); +push_desc_queue (c); +spin_lock_irqsave (&cdd->lock, flags); }spin_unlock_irqrestore (&cdd->lock, flags);
Signature · 6 lines
regards, -grygorii -- To unsubscribe from this list: send the line "unsubscribe linux-usb" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html
A message that is not in the archive (sent off-list or not yet received): 927792da-2e90-b2ae-1206-8fcb504d7551-l0cyMroinI0@public.gmane.org
Quoted 15 lines · Grygorii Strashko
On 01/12/2017 03:30 PM, Tony Lindgren wrote:
Sry, but even if it looks better it might still race with PM runtime :(
drivers/dma/cppi41.c static void cppi41_dma_issue_pending(struct dma_chan *chan) · line 491
Quoted 4 lines
drivers/dma/cppi41.c +1
-1-if (likely (pm_runtime_active (cdd->ddev.dev))) +if (likely (atomic_read (&cdd->active)))push_desc_queue (c);else - CPU is here (
-EINPROGRESSand active == 0) and then preempted
- PM runtime will finishcppi41_runtime_resumeand clean-up pending descs
- CPU return here and adds desc to the pending queue
- oops
Quoted 1 line
Am I wrong?
We had cppi41_dma_issue_pending() getting called from atomic contex and cppi41_runtime_resume() getting preempted where cppi41_dma_issue_pending() would add to the queue.
If what you're saying can happen in some cases, then we again see the WARN_ON(!list_empty(&cdd->pending)) in cppi41_runtime_suspend(). At least so far I have no longer been able to make happen today.
Regards,
Tony
--
To unsubscribe from this list: send the line "unsubscribe linux-usb" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html
A message that is not in the archive (sent off-list or not yet received): [email protected]
Quoted 16 lines · Tony Lindgren
Quoted 15 lines · Grygorii Strashko
On 01/12/2017 03:30 PM, Tony Lindgren wrote:
Sry, but even if it looks better it might still race with PM runtime :(
drivers/dma/cppi41.c static void cppi41_dma_issue_pending(struct dma_chan *chan) · line 491
Quoted 4 lines
drivers/dma/cppi41.c +1
-1-if (likely (pm_runtime_active (cdd->ddev.dev))) +if (likely (atomic_read (&cdd->active)))push_desc_queue (c);else - CPU is here (
-EINPROGRESSand active == 0) and then preempted
- PM runtime will finishcppi41_runtime_resumeand clean-up pending descs
- CPU return here and adds desc to the pending queue
- oops
Quoted 5 lines
Quoted 1 line
Am I wrong?
We had
cppi41_dma_issue_pending()getting called from atomic contex andcppi41_runtime_resume()getting preempted wherecppi41_dma_issue_pending()would add to the queue.
Again, I can be mistaken but cppi41_configure_channel() seems not atomic. cppi41_configure_channel()->dma_async_issue_pending()
+ documentation says "This function can be called in an interrupt context"
And definitely it will be preemptive on RT :(
Quoted 4 lines
If what you're saying can happen in some cases, then we again see the WARN_ON(!list_empty(&cdd->pending)) in
cppi41_runtime_suspend(). At least so far I have no longer been able to make happen today.
Signature · 6 lines
regards, -grygorii -- To unsubscribe from this list: send the line "unsubscribe linux-usb" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html
A message that is not in the archive (sent off-list or not yet received): 1c8967b7-d59b-e53d-feeb-80c71464fb94-l0cyMroinI0@public.gmane.org
Quoted 25 lines · Grygorii Strashko
Quoted 16 lines · Tony Lindgren
Quoted 15 lines · Grygorii Strashko
On 01/12/2017 03:30 PM, Tony Lindgren wrote:
Sry, but even if it looks better it might still race with PM runtime :(
Review
Quoted 4 lines
(quoted diff) +1
-1- if (likely(pm_runtime_active(cdd->ddev.dev))) + if (likely(atomic_read(&cdd->active))) push_desc_queue(c); else- CPU is here (
-EINPROGRESSand active == 0) and then preempted
- PM runtime will finishcppi41_runtime_resumeand clean-up pending descs
- CPU return here and adds desc to the pending queue
- oopsQuoted 5 lines
Quoted 1 line
Am I wrong?
We had
cppi41_dma_issue_pending()getting called from atomic contex andcppi41_runtime_resume()getting preempted wherecppi41_dma_issue_pending()would add to the queue.
Quoted 5 lines
Again, I can be mistaken but
cppi41_configure_channel()seems not atomic. cppi41_configure_channel()->dma_async_issue_pending()
+ documentation says "This function can be called in an interrupt context"And definitely it will be preemptive on RT :(
Hmm OK. So are you thinking we should add a spinlock around the test in cppi41_dma_issue_pending() and when modifying cdd->active?
Something like:
spin_lock_irqsave(&cdd->lock, flags);
if (likely(cdd->active))
push_desc_queue(c); else list_add_tail(&c->node, &cdd->pending);
spin_unlock_irqrestore(&cdd->lock, flags);
Or do you have something better in mind?
Regards,
Tony
--
To unsubscribe from this list: send the line "unsubscribe linux-usb" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html
A message that is not in the archive (sent off-list or not yet received): [email protected]
Quoted 26 lines · Tony Lindgren
Quoted 25 lines · Grygorii Strashko
Quoted 21 lines · Tony Lindgren
Quoted 15 lines · Grygorii Strashko
On 01/12/2017 03:30 PM, Tony Lindgren wrote:
Sry, but even if it looks better it might still race with PM runtime :(
Review
Quoted 4 lines
(quoted diff) +1
-1- if (likely(pm_runtime_active(cdd->ddev.dev))) + if (likely(atomic_read(&cdd->active))) push_desc_queue(c); else- CPU is here (
-EINPROGRESSand active == 0) and then preempted
- PM runtime will finishcppi41_runtime_resumeand clean-up pending descs
- CPU return here and adds desc to the pending queue
- oopsQuoted 1 line
Am I wrong?
We had
cppi41_dma_issue_pending()getting called from atomic contex andcppi41_runtime_resume()getting preempted wherecppi41_dma_issue_pending()would add to the queue.
Quoted 8 lines
Quoted 5 lines
Again, I can be mistaken but
cppi41_configure_channel()seems not atomic. cppi41_configure_channel()->dma_async_issue_pending()
+ documentation says "This function can be called in an interrupt context"And definitely it will be preemptive on RT :(
Hmm OK. So are you thinking we should add a spinlock around the test in
cppi41_dma_issue_pending()and when modifyingcdd->active?
in general yes.
Quoted 11 lines
Something like:
spin_lock_irqsave(&cdd->lock, flags);
if (likely(cdd->active))push_desc_queue(c); else list_add_tail(&c->node, &cdd->pending);
spin_unlock_irqrestore(&cdd->lock, flags);Or do you have something better in mind?
I've come up with smth like below. *Not tested*
But may be it can be done more simpler.
drivers/dma/cppi41.c +28 -26
@@ -153,6 +153,7 @@ struct cppi41_dd {/* context for suspend/resume */ unsigned int dma_tdfdq; +int is_suspended; };#define FIST_COMPLETION_QUEUE 93
@@ -317,12 +318,6 @@ static irqreturn_t cppi41_irq(int irq, void *data)while (val) {u32 desc, len; -int error; - - error =pm_runtime_get (cdd->ddev.dev); -if (error <0 ) -dev_err (cdd->ddev.dev,"%s pm runtime get: %i\n" , - __func__, error); q_num =__fls (val); val &= ~(1 << q_num);
@@ -343,9 +338,6 @@ static irqreturn_t cppi41_irq(int irq, void *data) c->residue =pd_trans_len (c->desc->pd6) - len;dma_cookie_complete (&c->txd);dmaengine_desc_get_callback_invoke (&c->txd,NULL ); - -pm_runtime_mark_last_busy (cdd->ddev.dev); -pm_runtime_put_autosuspend (cdd->ddev.dev); } }return IRQ_HANDLED ;
@@ -457,38 +449,36 @@ static void push_desc_queue(struct cppi41_channel *c)cppi_writel (reg, cdd->qmgr_mem +QMGR_QUEUE_D (c->q_num)); } -static void pending_desc (struct cppi41_channel *c) +static void cppi41_dma_issue_pending (struct dma_chan *chan) { +struct cppi41_channel *c =to_cpp41_chan (chan);struct cppi41_dd *cdd = c->cdd; +int error; +struct cppi41_channel *_c;unsigned long flags;spin_lock_irqsave (&cdd->lock, flags);list_add_tail (&c->node, &cdd->pending); -spin_unlock_irqrestore (&cdd->lock, flags); -} - -static void cppi41_dma_issue_pending (struct dma_chan *chan) -{ -struct cppi41_channel *c =to_cpp41_chan (chan); -struct cppi41_dd *cdd = c->cdd; -int error; error =pm_runtime_get (cdd->ddev.dev); -if ((error != -EINPROGRESS ) && error <0 ) { +if (error <0 ) {pm_runtime_put_noidle (cdd->ddev.dev);dev_err (cdd->ddev.dev,"Failed to pm_runtime_get: %i\n" , error); - +spin_unlock_irqrestore (&cdd->lock, flags);return ; } -if (likely (pm_runtime_active (cdd->ddev.dev))) -push_desc_queue (c); -else -pending_desc (c); +if (!cdd->is_suspended) { +list_for_each_entry_safe (c, _c, &cdd->pending, node) { +push_desc_queue (c); +list_del (&c->node); + }; + }pm_runtime_mark_last_busy (cdd->ddev.dev);pm_runtime_put_autosuspend (cdd->ddev.dev); +spin_lock_irqsave (&cdd->lock, flags); }static u32 get_host_pd0 (u32 length)
@@ -1150,10 +1140,20 @@ static int __maybe_unused cppi41_resume(struct device *dev)static int __maybe_unused cppi41_runtime_suspend (struct device *dev) {struct cppi41_dd *cdd =dev_get_drvdata (dev); +int ret; +unsigned long flags; -WARN_ON (!list_empty (&cdd->pending)); +spin_lock_irqsave (&cdd->lock, flags); +if (!list_empty (&cdd->pending)) { +WARN_ON (!list_empty (&cdd->pending)); + ret = -EBUSY ; + }else { +/* ... suspend the device ... */ + cdd->is_suspended =1 ; + } +spin_unlock_irqrestore (&cdd->lock, flags); -return 0 ; +return ret; }static int __maybe_unused cppi41_runtime_resume (struct device *dev)
@@ -1163,6 +1163,8 @@ static int __maybe_unused cppi41_runtime_resume(struct device *dev)unsigned long flags;spin_lock_irqsave (&cdd->lock, flags); + cdd->is_suspended =0 ; +list_for_each_entry_safe (c, _c, &cdd->pending, node) {push_desc_queue (c);list_del (&c->node);
Signature · 6 lines
regards, -grygorii -- To unsubscribe from this list: send the line "unsubscribe linux-usb" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html
A message that is not in the archive (sent off-list or not yet received): aafe7afb-6d8a-2fef-acdd-bd331151d16a-l0cyMroinI0@public.gmane.org
Quoted 40 lines · Grygorii Strashko
drivers/dma/cppi41.c +7
-15@@ -457,38 +449,36 @@ static void push_desc_queue(struct cppi41_channel *c)cppi_writel (reg, cdd->qmgr_mem +QMGR_QUEUE_D (c->q_num)); } -static void pending_desc (struct cppi41_channel *c) +static void cppi41_dma_issue_pending (struct dma_chan *chan) { +struct cppi41_channel *c =to_cpp41_chan (chan);struct cppi41_dd *cdd = c->cdd; +int error; +struct cppi41_channel *_c;unsigned long flags;spin_lock_irqsave (&cdd->lock, flags);list_add_tail (&c->node, &cdd->pending); -spin_unlock_irqrestore (&cdd->lock, flags); -} - -static void cppi41_dma_issue_pending (struct dma_chan *chan) -{ -struct cppi41_channel *c =to_cpp41_chan (chan); -struct cppi41_dd *cdd = c->cdd; -int error; error =pm_runtime_get (cdd->ddev.dev); -if ((error != -EINPROGRESS ) && error <0 ) { +if (error <0 ) {pm_runtime_put_noidle (cdd->ddev.dev);dev_err (cdd->ddev.dev,"Failed to pm_runtime_get: %i\n" , error); - +spin_unlock_irqrestore (&cdd->lock, flags);return ; } -if (likely (pm_runtime_active (cdd->ddev.dev))) -push_desc_queue (c); -else -pending_desc (c); +if (!cdd->is_suspended) {
drivers/dma/cppi41.c static void push_desc_queue(struct cppi41_channel *c) · line 482
Quoted 10 lines
drivers/dma/cppi41.c +6
-0+list_for_each_entry_safe (c, _c, &cdd->pending, node) { +push_desc_queue (c); +list_del (&c->node); + }; + }pm_runtime_mark_last_busy (cdd->ddev.dev);pm_runtime_put_autosuspend (cdd->ddev.dev); +spin_lock_irqsave (&cdd->lock, flags); }
So always add to the queue no matter, then always flush the queue directly if active? Yeah that might work :)
Don't we need spinlock in the list_for_each_entry_safe() to prevent it racing with runtime_resume() though?
Tony
--
To unsubscribe from this list: send the line "unsubscribe linux-usb" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html
A message that is not in the archive (sent off-list or not yet received): [email protected]
Quoted 55 lines · Tony Lindgren
Quoted 40 lines · Grygorii Strashko
(quoted diff) +7
-15@@ -457,38 +449,36 @@ static void push_desc_queue(struct cppi41_channel *c) cppi_writel(reg, cdd->qmgr_mem + QMGR_QUEUE_D(c->q_num)); } -static void pending_desc(struct cppi41_channel *c) +static void cppi41_dma_issue_pending(struct dma_chan *chan) { + struct cppi41_channel *c = to_cpp41_chan(chan); struct cppi41_dd *cdd = c->cdd; + int error; + struct cppi41_channel *_c; unsigned long flags; spin_lock_irqsave(&cdd->lock, flags); list_add_tail(&c->node, &cdd->pending); - spin_unlock_irqrestore(&cdd->lock, flags); -} - -static void cppi41_dma_issue_pending(struct dma_chan *chan) -{ - struct cppi41_channel *c = to_cpp41_chan(chan); - struct cppi41_dd *cdd = c->cdd; - int error; error = pm_runtime_get(cdd->ddev.dev); - if ((error != -EINPROGRESS) && error < 0) { + if (error < 0) { pm_runtime_put_noidle(cdd->ddev.dev); dev_err(cdd->ddev.dev, "Failed to pm_runtime_get: %i\n", error); - + spin_unlock_irqrestore(&cdd->lock, flags); return; } - if (likely(pm_runtime_active(cdd->ddev.dev))) - push_desc_queue(c); - else - pending_desc(c); + if (!cdd->is_suspended) {Review static void push_desc_queue(struct cppi41_channel *c) · line 482
Quoted 10 lines
(quoted diff) +6
-0+ list_for_each_entry_safe(c, _c, &cdd->pending, node) { + push_desc_queue(c); + list_del(&c->node); + }; + } pm_runtime_mark_last_busy(cdd->ddev.dev); pm_runtime_put_autosuspend(cdd->ddev.dev); + spin_lock_irqsave(&cdd->lock, flags); }So always add to the queue no matter, then always flush the queue directly if active? Yeah that might work :)
Quoted 2 lines
Don't we need spinlock in the
list_for_each_entry_safe()to prevent it racing withruntime_resume()though?
I could not apply for me as looks like your mail server replaced tabs with spaces it seems :(
But anyways here's your basic idea plugged into my recent patch and it seems to work. I maybe have missed something though while reading so please check.
The pm_runtime_get/mark_last_busy/put_autosuspend and WARN_ON() we want to keep in cppi41_irq() at least for now :)
And I'm thinking we must also callcppi41_run_queue() with spinlock held to prevent out of order triggering of the DMA transfers.
Does this look OK to you?
Regards,
Tony
8< -----------------------
drivers/dma/cppi41.c +33 -16
@@ -153,6 +153,8 @@ struct cppi41_dd {/* context for suspend/resume */ unsigned int dma_tdfdq; + +bool is_suspended; };#define FIST_COMPLETION_QUEUE 93
@@ -316,11 +318,12 @@ static irqreturn_t cppi41_irq(int irq, void *data)__iormb ();while (val) { +unsigned long flags;u32 desc, len;int error; error =pm_runtime_get (cdd->ddev.dev); -if (error <0 ) +if ((error != -EINPROGRESS ) && error <0 )dev_err (cdd->ddev.dev,"%s pm runtime get: %i\n" , __func__, error);
@@ -340,6 +343,11 @@ static irqreturn_t cppi41_irq(int irq, void *data)else len =pd_trans_len (c->desc->pd0); +/* This warning should never trigger */ +spin_lock_irqsave (&cdd->lock, flags); +WARN_ON (cdd->is_suspended); +spin_unlock_irqrestore (&cdd->lock, flags); + c->residue =pd_trans_len (c->desc->pd6) - len;dma_cookie_complete (&c->txd);dmaengine_desc_get_callback_invoke (&c->txd,NULL );
@@ -457,20 +465,26 @@ static void push_desc_queue(struct cppi41_channel *c)cppi_writel (reg, cdd->qmgr_mem +QMGR_QUEUE_D (c->q_num)); } -static void pending_desc (struct cppi41_channel *c) +/* +* Caller must hold cdd->lock to prevent push_desc_queue() +* getting called out of order. We have both cppi41_dma_issue_pending() +* and cppi41_runtime_resume() call this function. +*/ +static void cppi41_run_queue (struct cppi41_dd *cdd) { -struct cppi41_dd *cdd = c->cdd; -unsigned long flags; +struct cppi41_channel *c, *_c; -spin_lock_irqsave (&cdd->lock, flags); -list_add_tail (&c->node, &cdd->pending); -spin_unlock_irqrestore (&cdd->lock, flags); +list_for_each_entry_safe (c, _c, &cdd->pending, node) { +push_desc_queue (c); +list_del (&c->node); + } }static void cppi41_dma_issue_pending (struct dma_chan *chan) {struct cppi41_channel *c =to_cpp41_chan (chan);struct cppi41_dd *cdd = c->cdd; +unsigned long flags;int error; error =pm_runtime_get (cdd->ddev.dev);
@@ -482,10 +496,11 @@ static void cppi41_dma_issue_pending(struct dma_chan *chan)return ; } -if (likely (pm_runtime_active (cdd->ddev.dev))) -push_desc_queue (c); -else -pending_desc (c); +spin_lock_irqsave (&cdd->lock, flags); +list_add_tail (&c->node, &cdd->pending); +if (!cdd->is_suspended) +cppi41_run_queue (cdd); +spin_unlock_irqrestore (&cdd->lock, flags);pm_runtime_mark_last_busy (cdd->ddev.dev);pm_runtime_put_autosuspend (cdd->ddev.dev);
@@ -1150,6 +1165,11 @@ static int __maybe_unused cppi41_resume(struct device *dev)static int __maybe_unused cppi41_runtime_suspend (struct device *dev) {struct cppi41_dd *cdd =dev_get_drvdata (dev); +unsigned long flags; + +spin_lock_irqsave (&cdd->lock, flags); + cdd->is_suspended = true; +spin_unlock_irqrestore (&cdd->lock, flags);WARN_ON (!list_empty (&cdd->pending));
@@ -1159,14 +1179,11 @@ static int __maybe_unused cppi41_runtime_suspend(struct device *dev)static int __maybe_unused cppi41_runtime_resume (struct device *dev) {struct cppi41_dd *cdd =dev_get_drvdata (dev); -struct cppi41_channel *c, *_c;unsigned long flags;spin_lock_irqsave (&cdd->lock, flags); -list_for_each_entry_safe (c, _c, &cdd->pending, node) { -push_desc_queue (c); -list_del (&c->node); - } + cdd->is_suspended = false; +cppi41_run_queue (cdd);spin_unlock_irqrestore (&cdd->lock, flags);return 0 ;
--
To unsubscribe from this list: send the line "unsubscribe linux-usb" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html
A message that is not in the archive (sent off-list or not yet received): [email protected]
Quoted 16 lines · Tony Lindgren
drivers/dma/cppi41.c +2
-1@@ -316,11 +318,12 @@ static irqreturn_t cppi41_irq(int irq, void *data)__iormb ();while (val) { +unsigned long flags;u32 desc, len;int error; error =pm_runtime_get (cdd->ddev.dev); -if (error <0 ) +if ((error != -EINPROGRESS ) && error <0 )dev_err (cdd->ddev.dev,"%s pm runtime get: %i\n" , __func__, error);@@ -340,6 +343,11 @@ static irqreturn_t cppi41_irq(int irq, void *data)else
drivers/dma/cppi41.c static irqreturn_t cppi41_irq(int irq, void *data) · line 353
Quoted 10 lines
drivers/dma/cppi41.c +5
-0len =pd_trans_len (c->desc->pd0); +/* This warning should never trigger */ +spin_lock_irqsave (&cdd->lock, flags); +WARN_ON (cdd->is_suspended); +spin_unlock_irqrestore (&cdd->lock, flags); + c->residue =pd_trans_len (c->desc->pd6) - len;dma_cookie_complete (&c->txd);dmaengine_desc_get_callback_invoke (&c->txd,NULL );
Hmm this check needs to be before cppi41_pop_desc() already as that does cppi_readl(). And the spinlocks here don't help anything, we just want to see a warning if we hit a bug somewhere else.
Regards,
Tony
--
To unsubscribe from this list: send the line "unsubscribe linux-usb" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html
Quoted 56 lines · Tony Lindgren
Quoted 55 lines · Tony Lindgren
Quoted 40 lines · Grygorii Strashko
(quoted diff) +7
-15@@ -457,38 +449,36 @@ static void push_desc_queue(struct cppi41_channel *c) cppi_writel(reg, cdd->qmgr_mem + QMGR_QUEUE_D(c->q_num)); } -static void pending_desc(struct cppi41_channel *c) +static void cppi41_dma_issue_pending(struct dma_chan *chan) { + struct cppi41_channel *c = to_cpp41_chan(chan); struct cppi41_dd *cdd = c->cdd; + int error; + struct cppi41_channel *_c; unsigned long flags; spin_lock_irqsave(&cdd->lock, flags); list_add_tail(&c->node, &cdd->pending); - spin_unlock_irqrestore(&cdd->lock, flags); -} - -static void cppi41_dma_issue_pending(struct dma_chan *chan) -{ - struct cppi41_channel *c = to_cpp41_chan(chan); - struct cppi41_dd *cdd = c->cdd; - int error; error = pm_runtime_get(cdd->ddev.dev); - if ((error != -EINPROGRESS) && error < 0) { + if (error < 0) { pm_runtime_put_noidle(cdd->ddev.dev); dev_err(cdd->ddev.dev, "Failed to pm_runtime_get: %i\n", error); - + spin_unlock_irqrestore(&cdd->lock, flags); return; } - if (likely(pm_runtime_active(cdd->ddev.dev))) - push_desc_queue(c); - else - pending_desc(c); + if (!cdd->is_suspended) {Review static void push_desc_queue(struct cppi41_channel *c) · line 482
Quoted 10 lines
(quoted diff) +6
-0+ list_for_each_entry_safe(c, _c, &cdd->pending, node) { + push_desc_queue(c); + list_del(&c->node); + }; + } pm_runtime_mark_last_busy(cdd->ddev.dev); pm_runtime_put_autosuspend(cdd->ddev.dev); + spin_lock_irqsave(&cdd->lock, flags); }So always add to the queue no matter, then always flush the queue directly if active? Yeah that might work :)
Quoted 5 lines
Quoted 2 lines
Don't we need spinlock in the
list_for_each_entry_safe()to prevent it racing withruntime_resume()though?I could not apply for me as looks like your mail server replaced tabs with spaces it seems :(
Sorry.
Quoted 7 lines
But anyways here's your basic idea plugged into my recent patch and it seems to work. I maybe have missed something though while reading so please check.
The pm_runtime_get/mark_last_busy/put_autosuspend and
WARN_ON()we want to keep incppi41_irq()at least for now :)
As per my understanding and testing it looks like might be enough to have just pm_runtime_mark_last_busy(cdd->ddev.dev); in cppi41_irq(). But it can be left as is and yes - over all idea is that irq should not be triggered if device is Idle.
Quoted 6 lines
And I'm thinking we must also
callcppi41_run_queue()with spinlock held to prevent out of order triggering of the DMA transfers.Does this look OK to you?
I think yes. My current version is mostly similar to yours.
Quoted 16 lines
8< -----------------------
drivers/dma/cppi41.c +2
-0@@ -153,6 +153,8 @@ struct cppi41_dd {/* context for suspend/resume */ unsigned int dma_tdfdq; + +bool is_suspended; };#define FIST_COMPLETION_QUEUE 93@@ -316,11 +318,12 @@ static irqreturn_t cppi41_irq(int irq, void *data)
[..]
Quoted 4 lines
drivers/dma/cppi41.c +0
-0pm_runtime_mark_last_busy (cdd->ddev.dev);pm_runtime_put_autosuspend (cdd->ddev.dev);
drivers/dma/cppi41.c static int __maybe_unused cppi41_resume(struct device *dev) · line 1174
Quoted 10 lines
drivers/dma/cppi41.c +5
-0@@ -1150,6 +1165,11 @@ static int __maybe_unused cppi41_resume(struct device *dev)static int __maybe_unused cppi41_runtime_suspend (struct device *dev) {struct cppi41_dd *cdd =dev_get_drvdata (dev); +unsigned long flags; + +spin_lock_irqsave (&cdd->lock, flags); + cdd->is_suspended = true; +spin_unlock_irqrestore (&cdd->lock, flags);WARN_ON (!list_empty (&cdd->pending));
Shouldn't we check list_empty() under spin lock?
Quoted 19 lines
drivers/dma/cppi41.c +2
-5@@ -1159,14 +1179,11 @@ static int __maybe_unused cppi41_runtime_suspend(struct device *dev)static int __maybe_unused cppi41_runtime_resume (struct device *dev) {struct cppi41_dd *cdd =dev_get_drvdata (dev); -struct cppi41_channel *c, *_c;unsigned long flags;spin_lock_irqsave (&cdd->lock, flags); -list_for_each_entry_safe (c, _c, &cdd->pending, node) { -push_desc_queue (c); -list_del (&c->node); - } + cdd->is_suspended = false; +cppi41_run_queue (cdd);spin_unlock_irqrestore (&cdd->lock, flags);return 0 ;
Signature · 6 lines
regards, -grygorii -- To unsubscribe from this list: send the line "unsubscribe linux-usb" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html
A message that is not in the archive (sent off-list or not yet received): ba31f412-1b92-c4db-a1f4-82448e5ecf18-l0cyMroinI0@public.gmane.org
Quoted 5 lines · Grygorii Strashko
Quoted 4 lines · Tony Lindgren
But anyways here's your basic idea plugged into my recent patch and it seems to work. I maybe have missed something though while reading so please check.
Quoted 7 lines
Quoted 2 lines
The pm_runtime_get/mark_last_busy/put_autosuspend and
WARN_ON()we want to keep incppi41_irq()at least for now :)As per my understanding and testing it looks like might be enough to have just
pm_runtime_mark_last_busy(cdd->ddev.dev); incppi41_irq(). But it can be left as is and yes - over all idea is that irq should not be triggered if device is Idle.
OK yeah kicking the autoidle timeout is needed here.
Quoted 6 lines
Quoted 4 lines
And I'm thinking we must also
callcppi41_run_queue()with spinlock held to prevent out of order triggering of the DMA transfers.Does this look OK to you?
I think yes. My current version is mostly similar to yours.
OK will update description and repost shortly.
Quoted 10 lines
Quoted 10 lines
drivers/dma/cppi41.c +5
-0@@ -1150,6 +1165,11 @@ static int __maybe_unused cppi41_resume(struct device *dev)static int __maybe_unused cppi41_runtime_suspend (struct device *dev) {struct cppi41_dd *cdd =dev_get_drvdata (dev); +unsigned long flags; + +spin_lock_irqsave (&cdd->lock, flags); + cdd->is_suspended = true; +spin_unlock_irqrestore (&cdd->lock, flags);
Quoted 3 lines
drivers/dma/cppi41.c static int __maybe_unused cppi41_resume(struct device *dev) · line 1174
Quoted 1 line
drivers/dma/cppi41.c +0
-0WARN_ON (!list_empty (&cdd->pending));Shouldn't we check
list_empty()under spin lock?
Yeah let's do that.
Regards,
Tony
--
To unsubscribe from this list: send the line "unsubscribe linux-usb" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html
Hello!
Quoted 6 lines · Tony Lindgren
Commit
fdea2d09b997("dmaengine: cppi41: Add basic PM runtime support") together with recent MUSB changes allowed USB and DMA on BeagleBone to idle when no cable is connected. But looks like few corner case issues still remain.Looks like just by repluging USB cable about ten or so times on BeagleBone
Re-plugging.
Quoted 24 lines
when configured in USB peripheral mode we can get warnings and eventually trigger an oops in cppi41 DMA:
WARNING: CPU: 0 PID: 14 at drivers/dma/cppi41.c:1154 cppi41_runtime_suspend+x28/0x38 [cppi41]
...WARNING: CPU: 0 PID: 14 at drivers/dma/cppi41.c:452 push_desc_queue+0x94/0x9c [cppi41]...
Unable to handle kernel NULL pointer dereference at virtual address 00000104 pgd = c0004000 [00000104] *pgd=00000000 Internal error: Oops: 805 [#1] SMP ARM...
[<bf0d92cc>] (cppi41_runtime_resume[cppi41]) from [<c0589838>] (__rpm_callback+0xc0/0x214)
[<c0589838>] (__rpm_callback) from [<c05899ac>] (rpm_callback+0x20/0x80) [<c05899ac>] (rpm_callback) from [<c0589460>] (rpm_resume+0x504/0x78c) [<c0589460>] (rpm_resume) from [<c058a1a0>] (pm_runtime_work+0x60/0xa8) [<c058a1a0>] (pm_runtime_work) from [<c0156120>] (process_one_work+0x2b4/0x808)
Quoted 8 lines
This is because of a race with PM runtime and
cppi41_dma_issue_pending()
as reported by Alexandre Bailon <abailon-rdvid1DuHRBWk0Htik3J/[email protected]> in earlier set of patches. Based on mailing list discussions we however came to the conclusion that a different fix from Alexandre's fix is needed in order
to guarantee that DMA is really active when we try to use it.To fix the issue, we need to add a driver specific flag as we otherwise can have
-EINPROGRESSstate set by PM runtime and can't rely on
Runtime PM.
Quoted 31 lines
pm_runtime_active()to tell us when we can use the DMA.For reference, this is also documented in
Documentation/power/runtime_pm.txtin the example at the end of the file as pointed out by Grygorii Strashko <[email protected]>.So let's fix the issue by introducing
atomic_tactive that gets toggled inruntime_suspend()andruntime_resume(). And then let's replace thepm_runtime_active()checks withatomic_read().Note that we may also get transactions queued while in
-EINPROGRESSstate. So we need to check for new elements incppi41_runtime_resume()by replacinglist_for_each_entry_safe()with the commonly used test for while (!list_empty(&cdd->pending)).Fixes fdea2d09b997 ("dmaengine: cppi41: Add basic PM runtime support")Cc Andy Shevchenko <[email protected]>Cc Bin Liu <[email protected]>Cc Grygorii Strashko <[email protected]>Cc Kevin Hilman <khilman-rdvid1DuHRBWk0Htik3J/[email protected]>Cc Patrick Titiano <ptitiano-rdvid1DuHRBWk0Htik3J/[email protected]>Cc Sergei Shtylyov <sergei.shtylyov-M4DtvfQ/[email protected]>Signed-off-by Tony Lindgren <[email protected]>--- drivers/dma/cppi41.c | 23 ++++++++++++++++++----- 1 file changed, 18 insertions(+), 5 deletions(-)drivers/dma/cppi41.c +0
-0
[...]
drivers/dma/cppi41.c static irqreturn_t cppi41_irq(int irq, void *data) · line 326
Quoted 6 lines
drivers/dma/cppi41.c +1
-1@@ -320,7 +323,8 @@ static irqreturn_t cppi41_irq(int irq, void *data)int error; error =pm_runtime_get (cdd->ddev.dev); -if (error <0 ) + if (((error != -EINPROGRESS) && error < 0) ||
Hum, the innermost parens are inconsistent: around != but not around <.
[...]
MBR, Sergei
--
To unsubscribe from this list: send the line "unsubscribe linux-usb" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html
A message that is not in the archive (sent off-list or not yet received): 17c98bce-b8fd-33d9-f9f1-65f114306eed-M4DtvfQ/[email protected]
Quoted 8 lines · Sergei Shtylyov
drivers/dma/cppi41.c static irqreturn_t cppi41_irq(int irq, void *data) · line 326
Quoted 6 lines
drivers/dma/cppi41.c +1
-1@@ -320,7 +323,8 @@ static irqreturn_t cppi41_irq(int irq, void *data)int error; error =pm_runtime_get (cdd->ddev.dev); -if (error <0 ) + if (((error != -EINPROGRESS) && error < 0) ||Hum, the innermost parens are inconsistent: around != but not around <.
Oops just noticed I missed this one in the latest version, will update for the next revision. The typos I already fixed.
Regards,
Tony
--
To unsubscribe from this list: send the line "unsubscribe linux-usb" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html