[PATCH v2] drm/msm: Drop mm_lock in scan loop

[PATCH v2] drm/msm: Drop mm_lock in scan loop

[Rob Clark]

From: Rob Clark <robdclark@chromium.org>

lock_stat + mmm_donut[1] say that this reduces contention on mm_lock
significantly (~350x lower waittime-max, and ~100x lower waittime-avg)

[1] https://chromium.googlesource.com/chromiumos/platform/microbenchmarks/+/refs/heads/main/mmm_donut.py

Signed-off-by: Rob Clark <robdclark@chromium.org>
---
 drivers/gpu/drm/msm/msm_drv.h          |  3 +-
 drivers/gpu/drm/msm/msm_gem.c          |  2 +-
 drivers/gpu/drm/msm/msm_gem_shrinker.c | 48 ++++++++++++++++++++++----
 3 files changed, 45 insertions(+), 8 deletions(-)

diff --git a/drivers/gpu/drm/msm/msm_drv.h b/drivers/gpu/drm/msm/msm_drv.h
index c84e6f84cb6d..d8d64d34e6e3 100644
--- a/drivers/gpu/drm/msm/msm_drv.h
+++ b/drivers/gpu/drm/msm/msm_drv.h
@@ -184,7 +184,8 @@ struct msm_drm_private {
 	/**
 	 * Lists of inactive GEM objects.  Every bo is either in one of the
 	 * inactive lists (depending on whether or not it is shrinkable) or
-	 * gpu->active_list (for the gpu it is active on[1])
+	 * gpu->active_list (for the gpu it is active on[1]), or transiently
+	 * on a temporary list as the shrinker is running.
 	 *
 	 * These lists are protected by mm_lock (which should be acquired
 	 * before per GEM object lock).  One should *not* hold mm_lock in
diff --git a/drivers/gpu/drm/msm/msm_gem.c b/drivers/gpu/drm/msm/msm_gem.c
index 2ecf7f1cef25..75cea5b801da 100644
--- a/drivers/gpu/drm/msm/msm_gem.c
+++ b/drivers/gpu/drm/msm/msm_gem.c
@@ -719,7 +719,7 @@ void msm_gem_purge(struct drm_gem_object *obj)
 	put_iova_vmas(obj);
 
 	msm_obj->madv = __MSM_MADV_PURGED;
-	mark_unpurgable(msm_obj);
+	update_inactive(msm_obj);
 
 	drm_vma_node_unmap(&obj->vma_node, dev->anon_inode->i_mapping);
 	drm_gem_free_mmap_offset(obj);
diff --git a/drivers/gpu/drm/msm/msm_gem_shrinker.c b/drivers/gpu/drm/msm/msm_gem_shrinker.c
index f3e948af01c5..33a49641ef30 100644
--- a/drivers/gpu/drm/msm/msm_gem_shrinker.c
+++ b/drivers/gpu/drm/msm/msm_gem_shrinker.c
@@ -22,26 +22,62 @@ msm_gem_shrinker_scan(struct shrinker *shrinker, struct shrink_control *sc)
 {
 	struct msm_drm_private *priv =
 		container_of(shrinker, struct msm_drm_private, shrinker);
-	struct msm_gem_object *msm_obj;
+	struct list_head still_in_list;
 	unsigned long freed = 0;
 
+	INIT_LIST_HEAD(&still_in_list);
+
 	mutex_lock(&priv->mm_lock);
 
-	list_for_each_entry(msm_obj, &priv->inactive_dontneed, mm_list) {
-		if (freed >= sc->nr_to_scan)
+	while (freed < sc->nr_to_scan) {
+		struct msm_gem_object *msm_obj = list_first_entry_or_null(
+				&priv->inactive_dontneed, typeof(*msm_obj), mm_list);
+
+		if (!msm_obj)
 			break;
-		/* Use trylock, because we cannot block on a obj that
-		 * might be trying to acquire mm_lock
+
+		list_move_tail(&msm_obj->mm_list, &still_in_list);
+
+		/*
+		 * If it is in the process of being freed, msm_gem_free_object
+		 * can be blocked on mm_lock waiting to remove it.  So just
+		 * skip it.
 		 */
-		if (!msm_gem_trylock(&msm_obj->base))
+		if (!kref_get_unless_zero(&msm_obj->base.refcount))
 			continue;
+
+		/*
+		 * Now that we own a reference, we can drop mm_lock for the
+		 * rest of the loop body, to reduce contention with the
+		 * retire_submit path (which could make more objects purgable)
+		 */
+
+		mutex_unlock(&priv->mm_lock);
+
+		/*
+		 * Note that this still needs to be trylock, since we can
+		 * hit shrinker in response to trying to get backing pages
+		 * for this obj (ie. while it's lock is already held)
+		 */
+		if (!msm_gem_trylock(&msm_obj->base))
+			goto tail;
+
 		if (is_purgeable(msm_obj)) {
+			/*
+			 * This will move the obj out of still_in_list to
+			 * the purged list
+			 */
 			msm_gem_purge(&msm_obj->base);
 			freed += msm_obj->base.size >> PAGE_SHIFT;
 		}
 		msm_gem_unlock(&msm_obj->base);
+
+tail:
+		drm_gem_object_put(&msm_obj->base);
+		mutex_lock(&priv->mm_lock);
 	}
 
+	list_splice_tail(&still_in_list, &priv->inactive_dontneed);
 	mutex_unlock(&priv->mm_lock);
 
 	if (freed > 0) {
-- 
2.30.2

Re: [PATCH v2] drm/msm: Drop mm_lock in scan loop

[Doug Anderson]

Hi,

On Fri, Apr 2, 2021 at 2:08 PM Rob Clark <robdclark@gmail.com> wrote:
>
> From: Rob Clark <robdclark@chromium.org>
>
> lock_stat + mmm_donut[1] say that this reduces contention on mm_lock
> significantly (~350x lower waittime-max, and ~100x lower waittime-avg)
>
> [1] https://chromium.googlesource.com/chromiumos/platform/microbenchmarks/+/refs/heads/main/mmm_donut.py
>
> Signed-off-by: Rob Clark <robdclark@chromium.org>
> ---
>  drivers/gpu/drm/msm/msm_drv.h          |  3 +-
>  drivers/gpu/drm/msm/msm_gem.c          |  2 +-
>  drivers/gpu/drm/msm/msm_gem_shrinker.c | 48 ++++++++++++++++++++++----
>  3 files changed, 45 insertions(+), 8 deletions(-)

Looks good to me now!

Reviewed-by: Douglas Anderson <dianders@chromium.org>

Re: [PATCH v2] drm/msm: Drop mm_lock in scan loop

[patchwork-bot+linux-arm-msm]

Hello:

This patch was applied to qcom/linux.git (refs/heads/for-next):

On Fri,  2 Apr 2021 14:12:26 -0700 you wrote:
> From: Rob Clark <robdclark@chromium.org>
> 
> lock_stat + mmm_donut[1] say that this reduces contention on mm_lock
> significantly (~350x lower waittime-max, and ~100x lower waittime-avg)
> 
> [1] https://chromium.googlesource.com/chromiumos/platform/microbenchmarks/+/refs/heads/main/mmm_donut.py
> 
> [...]

Here is the summary with links:
  - [v2] drm/msm: Drop mm_lock in scan loop
    https://git.kernel.org/qcom/c/25ed38b3ed26

You are awesome, thank you!
--
Deet-doot-dot, I am a bot.
https://korg.docs.kernel.org/patchwork/pwbot.html