diff mbox

gpu/drm/ttm: Use mutex_lock_killable() for shrinker functions.

Message ID 201405292334.EAG00503.FLOOJFStHVQMFO@I-love.SAKURA.ne.jp (mailing list archive)
State New, archived
Headers show

Commit Message

Tetsuo Handa May 29, 2014, 2:34 p.m. UTC
Tetsuo Handa wrote:
> Konrad Rzeszutek Wilk wrote:
> > On Sat, May 24, 2014 at 11:22:09PM +0900, Tetsuo Handa wrote:
> > > Hello.
> > > 
> > > I tried to test whether it is OK (from point of view of reentrant) to use
> > > mutex_lock() or mutex_lock_killable() inside shrinker functions when shrinker
> > > functions do memory allocation, for drivers/gpu/drm/ttm/ttm_page_alloc_dma.c is
> > > doing memory allocation with mutex lock held inside ttm_dma_pool_shrink_scan().
> > > 
> > > If I compile a test module shown below which mimics extreme case of what
> > > ttm_dma_pool_shrink_scan() will do
> > 
> > And ttm_pool_shrink_scan.
> 
> I don't know why but ttm_pool_shrink_scan() does not take mutex.
> 
Well, it seems to me that ttm_pool_shrink_scan() not taking mutex is a bug
which could lead to stack overflow if kmalloc() in ttm_page_pool_free()
triggered recursion.

  shrink_slab()
  => ttm_pool_shrink_scan()
     => ttm_page_pool_free()
        => kmalloc(GFP_KERNEL)
           => shrink_slab()
              => ttm_pool_shrink_scan()
                 => ttm_page_pool_free()
                    => kmalloc(GFP_KERNEL)

Maybe shrink_slab() should be updated not to call same shrinker in parallel?

Also, it seems to me that ttm_dma_pool_shrink_scan() has potential division
by 0 bug as described below. Is this patch correct?
----------
>From 4a65744a300e14e5e202c5f13ba2759e1e797d29 Mon Sep 17 00:00:00 2001
From: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Date: Thu, 29 May 2014 18:25:42 +0900
Subject: [PATCH] gpu/drm/ttm: Use mutex_trylock() for shrinker functions.

I can observe that RHEL7 environment stalls with 100% CPU usage when a
certain type of memory pressure is given. While the shrinker functions
are called by shrink_slab() before the OOM killer is triggered, the stall
lasts for many minutes.

One of reasons of this stall is that
ttm_dma_pool_shrink_count()/ttm_dma_pool_shrink_scan() are called and
are blocked at mutex_lock(&_manager->lock). GFP_KERNEL allocation with
_manager->lock held causes someone (including kswapd) to deadlock when
these functions are called due to memory pressure. This patch changes
"mutex_lock();" to "if (!mutex_trylock()) return ...;" in order to
avoid deadlock.

At the same time, this patch fixes potential division by 0 due to
unconditionally doing "% _manager->npools". This is because
list_empty(&_manager->pools) being false does not guarantee that
_manager->npools != 0 after taking the _manager->lock because
_manager->npools is updated under the _manager->lock.

At the same time, this patch moves updating of start_pool variable
in order to avoid skipping when choosing a pool to shrink in
round-robin style. The start_pool is changed from "atomic_t" to
"unsigned int" because it is now updated under the _manager->lock.

Signed-off-by: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
Cc: stable <stable@kernel.org> [3.3+]
---
 drivers/gpu/drm/ttm/ttm_page_alloc_dma.c | 15 ++++++++++-----
 1 file changed, 10 insertions(+), 5 deletions(-)

Comments

Konrad Rzeszutek Wilk May 30, 2014, 4:08 p.m. UTC | #1
On Thu, May 29, 2014 at 11:34:59PM +0900, Tetsuo Handa wrote:
> Tetsuo Handa wrote:
> > Konrad Rzeszutek Wilk wrote:
> > > On Sat, May 24, 2014 at 11:22:09PM +0900, Tetsuo Handa wrote:
> > > > Hello.
> > > > 
> > > > I tried to test whether it is OK (from point of view of reentrant) to use
> > > > mutex_lock() or mutex_lock_killable() inside shrinker functions when shrinker
> > > > functions do memory allocation, for drivers/gpu/drm/ttm/ttm_page_alloc_dma.c is
> > > > doing memory allocation with mutex lock held inside ttm_dma_pool_shrink_scan().
> > > > 
> > > > If I compile a test module shown below which mimics extreme case of what
> > > > ttm_dma_pool_shrink_scan() will do
> > > 
> > > And ttm_pool_shrink_scan.
> > 
> > I don't know why but ttm_pool_shrink_scan() does not take mutex.
> > 
> Well, it seems to me that ttm_pool_shrink_scan() not taking mutex is a bug
> which could lead to stack overflow if kmalloc() in ttm_page_pool_free()
> triggered recursion.
> 
>   shrink_slab()
>   => ttm_pool_shrink_scan()
>      => ttm_page_pool_free()
>         => kmalloc(GFP_KERNEL)
>            => shrink_slab()
>               => ttm_pool_shrink_scan()
>                  => ttm_page_pool_free()
>                     => kmalloc(GFP_KERNEL)
> 
> Maybe shrink_slab() should be updated not to call same shrinker in parallel?
> 
> Also, it seems to me that ttm_dma_pool_shrink_scan() has potential division
> by 0 bug as described below. Is this patch correct?

Looks OK. I would need to test it first. Could you send both patches
to me please so I can just test them and queue them up together?

Thank you!
> ----------
> >From 4a65744a300e14e5e202c5f13ba2759e1e797d29 Mon Sep 17 00:00:00 2001
> From: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
> Date: Thu, 29 May 2014 18:25:42 +0900
> Subject: [PATCH] gpu/drm/ttm: Use mutex_trylock() for shrinker functions.
> 
> I can observe that RHEL7 environment stalls with 100% CPU usage when a
> certain type of memory pressure is given. While the shrinker functions
> are called by shrink_slab() before the OOM killer is triggered, the stall
> lasts for many minutes.
> 
> One of reasons of this stall is that
> ttm_dma_pool_shrink_count()/ttm_dma_pool_shrink_scan() are called and
> are blocked at mutex_lock(&_manager->lock). GFP_KERNEL allocation with
> _manager->lock held causes someone (including kswapd) to deadlock when
> these functions are called due to memory pressure. This patch changes
> "mutex_lock();" to "if (!mutex_trylock()) return ...;" in order to
> avoid deadlock.
> 
> At the same time, this patch fixes potential division by 0 due to
> unconditionally doing "% _manager->npools". This is because
> list_empty(&_manager->pools) being false does not guarantee that
> _manager->npools != 0 after taking the _manager->lock because
> _manager->npools is updated under the _manager->lock.
> 
> At the same time, this patch moves updating of start_pool variable
> in order to avoid skipping when choosing a pool to shrink in
> round-robin style. The start_pool is changed from "atomic_t" to
> "unsigned int" because it is now updated under the _manager->lock.
> 
> Signed-off-by: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
> Cc: stable <stable@kernel.org> [3.3+]
> ---
>  drivers/gpu/drm/ttm/ttm_page_alloc_dma.c | 15 ++++++++++-----
>  1 file changed, 10 insertions(+), 5 deletions(-)
> 
> diff --git a/drivers/gpu/drm/ttm/ttm_page_alloc_dma.c b/drivers/gpu/drm/ttm/ttm_page_alloc_dma.c
> index fb8259f..5e332b4 100644
> --- a/drivers/gpu/drm/ttm/ttm_page_alloc_dma.c
> +++ b/drivers/gpu/drm/ttm/ttm_page_alloc_dma.c
> @@ -1004,9 +1004,9 @@ EXPORT_SYMBOL_GPL(ttm_dma_unpopulate);
>  static unsigned long
>  ttm_dma_pool_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
>  {
> -	static atomic_t start_pool = ATOMIC_INIT(0);
> +	static unsigned int start_pool;
>  	unsigned idx = 0;
> -	unsigned pool_offset = atomic_add_return(1, &start_pool);
> +	unsigned pool_offset;
>  	unsigned shrink_pages = sc->nr_to_scan;
>  	struct device_pools *p;
>  	unsigned long freed = 0;
> @@ -1014,8 +1014,11 @@ ttm_dma_pool_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
>  	if (list_empty(&_manager->pools))
>  		return SHRINK_STOP;
>  
> -	mutex_lock(&_manager->lock);
> -	pool_offset = pool_offset % _manager->npools;
> +	if (!mutex_trylock(&_manager->lock))
> +		return SHRINK_STOP;
> +	if (!_manager->npools)
> +		goto out;
> +	pool_offset = ++start_pool % _manager->npools;
>  	list_for_each_entry(p, &_manager->pools, pools) {
>  		unsigned nr_free;
>  
> @@ -1034,6 +1037,7 @@ ttm_dma_pool_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
>  			 p->pool->dev_name, p->pool->name, current->pid,
>  			 nr_free, shrink_pages);
>  	}
> +out:
>  	mutex_unlock(&_manager->lock);
>  	return freed;
>  }
> @@ -1044,7 +1048,8 @@ ttm_dma_pool_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
>  	struct device_pools *p;
>  	unsigned long count = 0;
>  
> -	mutex_lock(&_manager->lock);
> +	if (!mutex_trylock(&_manager->lock))
> +		return 0;
>  	list_for_each_entry(p, &_manager->pools, pools)
>  		count += p->pool->npages_free;
>  	mutex_unlock(&_manager->lock);
> -- 
> 1.8.3.1
diff mbox

Patch

diff --git a/drivers/gpu/drm/ttm/ttm_page_alloc_dma.c b/drivers/gpu/drm/ttm/ttm_page_alloc_dma.c
index fb8259f..5e332b4 100644
--- a/drivers/gpu/drm/ttm/ttm_page_alloc_dma.c
+++ b/drivers/gpu/drm/ttm/ttm_page_alloc_dma.c
@@ -1004,9 +1004,9 @@  EXPORT_SYMBOL_GPL(ttm_dma_unpopulate);
 static unsigned long
 ttm_dma_pool_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
 {
-	static atomic_t start_pool = ATOMIC_INIT(0);
+	static unsigned int start_pool;
 	unsigned idx = 0;
-	unsigned pool_offset = atomic_add_return(1, &start_pool);
+	unsigned pool_offset;
 	unsigned shrink_pages = sc->nr_to_scan;
 	struct device_pools *p;
 	unsigned long freed = 0;
@@ -1014,8 +1014,11 @@  ttm_dma_pool_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
 	if (list_empty(&_manager->pools))
 		return SHRINK_STOP;
 
-	mutex_lock(&_manager->lock);
-	pool_offset = pool_offset % _manager->npools;
+	if (!mutex_trylock(&_manager->lock))
+		return SHRINK_STOP;
+	if (!_manager->npools)
+		goto out;
+	pool_offset = ++start_pool % _manager->npools;
 	list_for_each_entry(p, &_manager->pools, pools) {
 		unsigned nr_free;
 
@@ -1034,6 +1037,7 @@  ttm_dma_pool_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
 			 p->pool->dev_name, p->pool->name, current->pid,
 			 nr_free, shrink_pages);
 	}
+out:
 	mutex_unlock(&_manager->lock);
 	return freed;
 }
@@ -1044,7 +1048,8 @@  ttm_dma_pool_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
 	struct device_pools *p;
 	unsigned long count = 0;
 
-	mutex_lock(&_manager->lock);
+	if (!mutex_trylock(&_manager->lock))
+		return 0;
 	list_for_each_entry(p, &_manager->pools, pools)
 		count += p->pool->npages_free;
 	mutex_unlock(&_manager->lock);