Re: [PATCH V2] libcfs: Fix a sleep-in-atomic bug in cfs_wi_deschedule

[Oleg Drokin]

Hello!

On May 31, 2017, at 3:57 AM, Jia-Ju Bai wrote:

> The driver may sleep under a spin lock, and the function call path is:
> cfs_wi_deschedule (acquire the lock by spin_lock)
>  LASSERT
>lbug_with_loc
>  libcfs_debug_dumplog
>schedule and kthread_run --> may sleep
> 
> To fix it, all "LASSERT" is placed out of the spin_lock and spin_unlock.
> 
> Signed-off-by: Jia-Ju Bai 
> ---
> drivers/staging/lustre/lnet/libcfs/workitem.c |   12 ++--
> 1 file changed, 6 insertions(+), 6 deletions(-)
> 
> diff --git a/drivers/staging/lustre/lnet/libcfs/workitem.c 
> b/drivers/staging/lustre/lnet/libcfs/workitem.c
> index dbc2a9b..9c530cf 100644
> --- a/drivers/staging/lustre/lnet/libcfs/workitem.c
> +++ b/drivers/staging/lustre/lnet/libcfs/workitem.c
> @@ -140,6 +140,10 @@ struct cfs_wi_sched {
> 
>   LASSERT(!in_interrupt()); /* because we use plain spinlock */
>   LASSERT(!sched->ws_stopping);
> + if (wi->wi_scheduled) {
> + LASSERT(!list_empty(>wi_list));
> + LASSERT(sched->ws_nscheduled > 0);
> + }

I don't think you can do this,
this was under spinlock because those values could change from a different 
thread
and we do need to look at them all together.

You are correct that LASSET/LBUG might schedule to dump a debug log
into a file and even if not it does sleep indefinitely after that.
But in reality the default option is "panic_on_lbug=1" which simply converts
LASSERT() into panic().

This is certainly not a normal condition and as such I think we can leave the 
code
as is.

Thanks.


> 
>   /*
>* return 0 if it's running already, otherwise return 1, which
> @@ -151,18 +155,14 @@ struct cfs_wi_sched {
>   rc = !(wi->wi_running);
> 
>   if (wi->wi_scheduled) { /* cancel pending schedules */
> - LASSERT(!list_empty(>wi_list));
>   list_del_init(>wi_list);
> -
> - LASSERT(sched->ws_nscheduled > 0);
>   sched->ws_nscheduled--;
> -
>   wi->wi_scheduled = 0;
>   }
> 
> - LASSERT(list_empty(>wi_list));
> -
>   spin_unlock(>ws_lock);
> +
> + LASSERT(list_empty(>wi_list));
>   return rc;
> }
> EXPORT_SYMBOL(cfs_wi_deschedule);
> -- 
> 1.7.9.5
> 

Re: [PATCH V2] libcfs: Fix a sleep-in-atomic bug in cfs_wi_deschedule

[Oleg Drokin]

Hello!

On May 31, 2017, at 3:57 AM, Jia-Ju Bai wrote:

> The driver may sleep under a spin lock, and the function call path is:
> cfs_wi_deschedule (acquire the lock by spin_lock)
>  LASSERT
>lbug_with_loc
>  libcfs_debug_dumplog
>schedule and kthread_run --> may sleep
> 
> To fix it, all "LASSERT" is placed out of the spin_lock and spin_unlock.
> 
> Signed-off-by: Jia-Ju Bai 
> ---
> drivers/staging/lustre/lnet/libcfs/workitem.c |   12 ++--
> 1 file changed, 6 insertions(+), 6 deletions(-)
> 
> diff --git a/drivers/staging/lustre/lnet/libcfs/workitem.c 
> b/drivers/staging/lustre/lnet/libcfs/workitem.c
> index dbc2a9b..9c530cf 100644
> --- a/drivers/staging/lustre/lnet/libcfs/workitem.c
> +++ b/drivers/staging/lustre/lnet/libcfs/workitem.c
> @@ -140,6 +140,10 @@ struct cfs_wi_sched {
> 
>   LASSERT(!in_interrupt()); /* because we use plain spinlock */
>   LASSERT(!sched->ws_stopping);
> + if (wi->wi_scheduled) {
> + LASSERT(!list_empty(>wi_list));
> + LASSERT(sched->ws_nscheduled > 0);
> + }

I don't think you can do this,
this was under spinlock because those values could change from a different 
thread
and we do need to look at them all together.

You are correct that LASSET/LBUG might schedule to dump a debug log
into a file and even if not it does sleep indefinitely after that.
But in reality the default option is "panic_on_lbug=1" which simply converts
LASSERT() into panic().

This is certainly not a normal condition and as such I think we can leave the 
code
as is.

Thanks.


> 
>   /*
>* return 0 if it's running already, otherwise return 1, which
> @@ -151,18 +155,14 @@ struct cfs_wi_sched {
>   rc = !(wi->wi_running);
> 
>   if (wi->wi_scheduled) { /* cancel pending schedules */
> - LASSERT(!list_empty(>wi_list));
>   list_del_init(>wi_list);
> -
> - LASSERT(sched->ws_nscheduled > 0);
>   sched->ws_nscheduled--;
> -
>   wi->wi_scheduled = 0;
>   }
> 
> - LASSERT(list_empty(>wi_list));
> -
>   spin_unlock(>ws_lock);
> +
> + LASSERT(list_empty(>wi_list));
>   return rc;
> }
> EXPORT_SYMBOL(cfs_wi_deschedule);
> -- 
> 1.7.9.5
> 

[PATCH V2] libcfs: Fix a sleep-in-atomic bug in cfs_wi_deschedule

[Jia-Ju Bai]

The driver may sleep under a spin lock, and the function call path is:
cfs_wi_deschedule (acquire the lock by spin_lock)
  LASSERT
lbug_with_loc
  libcfs_debug_dumplog
schedule and kthread_run --> may sleep

To fix it, all "LASSERT" is placed out of the spin_lock and spin_unlock.

Signed-off-by: Jia-Ju Bai 
---
 drivers/staging/lustre/lnet/libcfs/workitem.c |   12 ++--
 1 file changed, 6 insertions(+), 6 deletions(-)

diff --git a/drivers/staging/lustre/lnet/libcfs/workitem.c 
b/drivers/staging/lustre/lnet/libcfs/workitem.c
index dbc2a9b..9c530cf 100644
--- a/drivers/staging/lustre/lnet/libcfs/workitem.c
+++ b/drivers/staging/lustre/lnet/libcfs/workitem.c
@@ -140,6 +140,10 @@ struct cfs_wi_sched {
 
LASSERT(!in_interrupt()); /* because we use plain spinlock */
LASSERT(!sched->ws_stopping);
+   if (wi->wi_scheduled) {
+   LASSERT(!list_empty(>wi_list));
+   LASSERT(sched->ws_nscheduled > 0);
+   }
 
/*
 * return 0 if it's running already, otherwise return 1, which
@@ -151,18 +155,14 @@ struct cfs_wi_sched {
rc = !(wi->wi_running);
 
if (wi->wi_scheduled) { /* cancel pending schedules */
-   LASSERT(!list_empty(>wi_list));
list_del_init(>wi_list);
-
-   LASSERT(sched->ws_nscheduled > 0);
sched->ws_nscheduled--;
-
wi->wi_scheduled = 0;
}
 
-   LASSERT(list_empty(>wi_list));
-
spin_unlock(>ws_lock);
+
+   LASSERT(list_empty(>wi_list));
return rc;
 }
 EXPORT_SYMBOL(cfs_wi_deschedule);
-- 
1.7.9.5

[PATCH V2] libcfs: Fix a sleep-in-atomic bug in cfs_wi_deschedule

[Jia-Ju Bai]

The driver may sleep under a spin lock, and the function call path is:
cfs_wi_deschedule (acquire the lock by spin_lock)
  LASSERT
lbug_with_loc
  libcfs_debug_dumplog
schedule and kthread_run --> may sleep

To fix it, all "LASSERT" is placed out of the spin_lock and spin_unlock.

Signed-off-by: Jia-Ju Bai 
---
 drivers/staging/lustre/lnet/libcfs/workitem.c |   12 ++--
 1 file changed, 6 insertions(+), 6 deletions(-)

diff --git a/drivers/staging/lustre/lnet/libcfs/workitem.c 
b/drivers/staging/lustre/lnet/libcfs/workitem.c
index dbc2a9b..9c530cf 100644
--- a/drivers/staging/lustre/lnet/libcfs/workitem.c
+++ b/drivers/staging/lustre/lnet/libcfs/workitem.c
@@ -140,6 +140,10 @@ struct cfs_wi_sched {
 
LASSERT(!in_interrupt()); /* because we use plain spinlock */
LASSERT(!sched->ws_stopping);
+   if (wi->wi_scheduled) {
+   LASSERT(!list_empty(>wi_list));
+   LASSERT(sched->ws_nscheduled > 0);
+   }
 
/*
 * return 0 if it's running already, otherwise return 1, which
@@ -151,18 +155,14 @@ struct cfs_wi_sched {
rc = !(wi->wi_running);
 
if (wi->wi_scheduled) { /* cancel pending schedules */
-   LASSERT(!list_empty(>wi_list));
list_del_init(>wi_list);
-
-   LASSERT(sched->ws_nscheduled > 0);
sched->ws_nscheduled--;
-
wi->wi_scheduled = 0;
}
 
-   LASSERT(list_empty(>wi_list));
-
spin_unlock(>ws_lock);
+
+   LASSERT(list_empty(>wi_list));
return rc;
 }
 EXPORT_SYMBOL(cfs_wi_deschedule);
-- 
1.7.9.5