drm/msm/gem: Drop obj lock in msm_gem_free_object()
PATCH 1/1 drm/msm/gem: Drop obj lock in msm_gem_free_object()
From: Rob Clark <[email protected]>
The only reason we grabbed the lock was to satisfy a bunch of places that WARN_ON() if called without the lock held. But this angers lockdep which doesn't realize no one else can be holding the lock by the time we end up destroying the object (and sees what would otherwise be a locking inversion between reservation_ww_class_mutex and fs_reclaim).
--- drivers/gpu/drm/msm/msm_gem.c | 8 -------- drivers/gpu/drm/msm/msm_gem.h | 14 +++++++++++++- 2 files changed, 13 insertions(+), 9 deletions(-)
drivers/gpu/drm/msm/msm_gem.c +0 -8
@@ -1020,8 +1020,6 @@ static void msm_gem_free_object(struct drm_gem_object *obj)list_del (&msm_obj->mm_list);mutex_unlock (&priv->mm_lock); -msm_gem_lock (obj); -/* object should not be on active list: */ GEM_WARN_ON (is_active (msm_obj));
@@ -1037,17 +1035,11 @@ static void msm_gem_free_object(struct drm_gem_object *obj)put_iova_vmas (obj); -/* dma_buf_detach() grabs resv lock, so we need to unlock -* prior to drm_prime_gem_destroy -*/ -msm_gem_unlock (obj); -drm_prime_gem_destroy (obj, msm_obj->sgt); }else {msm_gem_vunmap (obj);put_pages (obj);put_iova_vmas (obj); -msm_gem_unlock (obj); }drm_gem_object_release (obj);
drivers/gpu/drm/msm/msm_gem.h +13 -1
@@ -229,7 +229,19 @@ msm_gem_unlock(struct drm_gem_object *obj)static inline bool msm_gem_is_locked (struct drm_gem_object *obj) { -return dma_resv_is_locked (obj->resv); +/* +* Destroying the object is a special case.. msm_gem_free_object() +* calls many things that WARN_ON if the obj lock is not held. But +* acquiring the obj lock in msm_gem_free_object() can cause a +* locking order inversion between reservation_ww_class_mutex and +* fs_reclaim. +* +* This deadlock is not actually possible, because no one should +* be already holding the lock when msm_gem_free_object() is called. +* Unfortunately lockdep is not aware of this detail. So when the +* refcount drops to zero, we pretend it is already locked. +*/ +return dma_resv_is_locked (obj->resv) || (kref_read (&obj->refcount) ==0 ); }static inline bool is_active (struct msm_gem_object *msm_obj)
Signature · 1 line
2.36.1
Quoting Rob Clark (2022-06-13 13:50:32)
Quoted 12 lines
drivers/gpu/drm/msm/msm_gem.h +3
-1@@ -229,7 +229,19 @@ msm_gem_unlock(struct drm_gem_object *obj)static inline bool msm_gem_is_locked (struct drm_gem_object *obj) { -return dma_resv_is_locked (obj->resv); +/* +* Destroying the object is a special case.. msm_gem_free_object() +* calls many things that WARN_ON if the obj lock is not held. But
drivers/gpu/drm/msm/msm_gem.h msm_gem_unlock(struct drm_gem_object *obj) · line 244
Quoted 10 lines
drivers/gpu/drm/msm/msm_gem.h +10
-0+* acquiring the obj lock in msm_gem_free_object() can cause a +* locking order inversion between reservation_ww_class_mutex and +* fs_reclaim. +* +* This deadlock is not actually possible, because no one should +* be already holding the lock when msm_gem_free_object() is called. +* Unfortunately lockdep is not aware of this detail. So when the +* refcount drops to zero, we pretend it is already locked. +*/ +return dma_resv_is_locked (obj->resv) || (kref_read (&obj->refcount) ==0 );
Instead of modifying this function can we push down the fact that this function is being called from the free path and skip checking this condition in that case? Or add some "_locked/free_path" wrappers that skip the lock assertion? That would make it clearer to understand while reading the code that it is locked when it is asserted to be locked, and that we don't care when we're freeing because all references to the object are gone.
Here's a totally untested patch to show the idea. The comment about pretending the lock is held can be put in msm_gem_free_object() to clarify why it's OK to call the locked variants of the functions.
---8<---
drivers/gpu/drm/msm/msm_gem.c +29 -24
@@ -346,13 +346,11 @@ static void del_vma(struct msm_gem_vma *vma)* mapping. */ static void -put_iova_spaces (struct drm_gem_object *obj,bool close) +put_iova_spaces_locked (struct drm_gem_object *obj,bool close) {struct msm_gem_object *msm_obj =to_msm_bo (obj);struct msm_gem_vma *vma; -GEM_WARN_ON (!msm_gem_is_locked (obj)); -list_for_each_entry (vma, &msm_obj->vmas, list) {if (vma->aspace) {msm_gem_purge_vma (vma->aspace, vma);
@@ -362,18 +360,28 @@ put_iova_spaces(struct drm_gem_object *obj, bool close) } } -/* Called with msm_obj locked */ +static void put_iova_spaces (struct drm_gem_object *obj,bool close) +{ +GEM_WARN_ON (!msm_gem_is_locked (obj)); +put_iova_spaces_locked (obj, close); +} + +/* Called with msm_obj locked or on the free path */ static void -put_iova_vmas (struct drm_gem_object *obj) +put_iova_vmas_locked (struct drm_gem_object *obj) {struct msm_gem_object *msm_obj =to_msm_bo (obj);struct msm_gem_vma *vma, *tmp; -GEM_WARN_ON (!msm_gem_is_locked (obj)); - -list_for_each_entry_safe (vma, tmp, &msm_obj->vmas, list) { +list_for_each_entry_safe (vma, tmp, &msm_obj->vmas, list)del_vma (vma); - } +} + +static void +put_iova_vmas (struct drm_gem_object *obj) +{ +GEM_WARN_ON (!msm_gem_is_locked (obj)); +put_iova_vmas_locked (obj); } static struct msm_gem_vma *get_vma_locked(struct drm_gem_object *obj,
@@ -795,12 +803,10 @@ void msm_gem_evict(struct drm_gem_object *obj)update_inactive (msm_obj); } -void msm_gem_vunmap (struct drm_gem_object *obj) +static void msm_gem_vunmap_locked (struct drm_gem_object *obj) {struct msm_gem_object *msm_obj =to_msm_bo (obj); -GEM_WARN_ON (!msm_gem_is_locked (obj)); -if (!msm_obj->vaddr ||GEM_WARN_ON (!is_vunmapable (msm_obj)))return ;
@@ -808,6 +814,12 @@ void msm_gem_vunmap(struct drm_gem_object *obj) msm_obj->vaddr =NULL ; } +void msm_gem_vunmap (struct drm_gem_object *obj) +{ +GEM_WARN_ON (!msm_gem_is_locked (obj)); +msm_gem_vunmap_locked (obj); +} +void msm_gem_active_get (struct drm_gem_object *obj,struct msm_gpu *gpu) {struct msm_gem_object *msm_obj =to_msm_bo (obj);
@@ -1021,12 +1033,11 @@ void msm_gem_free_object(struct drm_gem_object *obj)list_del (&msm_obj->mm_list);mutex_unlock (&priv->mm_lock); -msm_gem_lock (obj); -/* object should not be on active list: */ GEM_WARN_ON (is_active (msm_obj)); -put_iova_spaces (obj, true); +put_iova_spaces_locked (obj, true); +if (obj->import_attach) {GEM_WARN_ON (msm_obj->vaddr);
@@ -1036,19 +1047,13 @@ void msm_gem_free_object(struct drm_gem_object *obj)*/ kvfree (msm_obj->pages); -put_iova_vmas (obj); - -/* dma_buf_detach() grabs resv lock, so we need to unlock -* prior to drm_prime_gem_destroy -*/ -msm_gem_unlock (obj); +put_iova_vmas_locked (obj);drm_prime_gem_destroy (obj, msm_obj->sgt); }else { -msm_gem_vunmap (obj); +msm_gem_vunmap_locked (obj);put_pages (obj); -put_iova_vmas (obj); -msm_gem_unlock (obj); +put_iova_vmas_locked (obj); }drm_gem_object_release (obj);
drivers/gpu/drm/msm/msm_gem.h +0 -0
@@ -253,7 +253,6 @@ static inline bool is_purgeable(struct
msm_gem_object *msm_obj)
static inline bool is_vunmapable(struct msm_gem_object *msm_obj) {
- GEM_WARN_ON(!msm_gem_is_locked(&msm_obj->base));
return (msm_obj->vmap_count == 0) && msm_obj->vaddr;
}
(quoted diff) +0 -0
b/drivers/gpu/drm/msm/msm_gem_shrinker.c
index 086dacf2f26a..afff3a79e925 100644
drivers/gpu/drm/msm/msm_gem_shrinker.c +1 -0
@@ -175,6 +175,7 @@ static const int vmap_shrink_limit = 15;static bool vmap_shrink (struct msm_gem_object *msm_obj) { +GEM_WARN_ON (!msm_gem_is_locked (&msm_obj->base));if (!is_vunmapable (msm_obj))return false;
Quoted 25 lines · Stephen Boyd
Quoting Rob Clark (2022-06-13 13:50:32)
Quoted 12 lines
drivers/gpu/drm/msm/msm_gem.h +3
-1@@ -229,7 +229,19 @@ msm_gem_unlock(struct drm_gem_object *obj)static inline bool msm_gem_is_locked (struct drm_gem_object *obj) { -return dma_resv_is_locked (obj->resv); +/* +* Destroying the object is a special case.. msm_gem_free_object() +* calls many things that WARN_ON if the obj lock is not held. But drivers/gpu/drm/msm/msm_gem.h msm_gem_unlock(struct drm_gem_object *obj) · line 244
Quoted 10 lines
drivers/gpu/drm/msm/msm_gem.h +10
-0+* acquiring the obj lock in msm_gem_free_object() can cause a +* locking order inversion between reservation_ww_class_mutex and +* fs_reclaim. +* +* This deadlock is not actually possible, because no one should +* be already holding the lock when msm_gem_free_object() is called. +* Unfortunately lockdep is not aware of this detail. So when the +* refcount drops to zero, we pretend it is already locked. +*/ +return dma_resv_is_locked (obj->resv) || (kref_read (&obj->refcount) ==0 );
Quoted 7 lines
Instead of modifying this function can we push down the fact that this function is being called from the free path and skip checking this condition in that case? Or add some "_locked/free_path" wrappers that skip the lock assertion? That would make it clearer to understand while reading the code that it is locked when it is asserted to be locked, and that we don't care when we're freeing because all references to the object are gone.
that was my earlier attempt, and I wasn't too happy with the result. And then I realized if refcount==0 then by definition we aren't racing with anyone else ;-)
Quoted 25 lines
Here's a totally untested patch to show the idea. The comment about pretending the lock is held can be put in
msm_gem_free_object()to clarify why it's OK to call the locked variants of the functions.---8<---
drivers/gpu/drm/msm/msm_gem.c +1
-3@@ -346,13 +346,11 @@ static void del_vma(struct msm_gem_vma *vma)* mapping. */ static void -put_iova_spaces (struct drm_gem_object *obj,bool close) +put_iova_spaces_locked (struct drm_gem_object *obj,bool close) {struct msm_gem_object *msm_obj =to_msm_bo (obj);struct msm_gem_vma *vma; -GEM_WARN_ON (!msm_gem_is_locked (obj)); -list_for_each_entry (vma, &msm_obj->vmas, list) {if (vma->aspace) {msm_gem_purge_vma (vma->aspace, vma);
drivers/gpu/drm/msm/msm_gem.c put_iova_spaces(struct drm_gem_object *obj, bool close) · line 367
Quoted 9 lines
drivers/gpu/drm/msm/msm_gem.c +5
-1@@ -362,18 +360,28 @@ put_iova_spaces(struct drm_gem_object *obj, bool close) } } -/* Called with msm_obj locked */ +static void put_iova_spaces (struct drm_gem_object *obj,bool close) +{ +GEM_WARN_ON (!msm_gem_is_locked (obj)); +put_iova_spaces_locked (obj, close); +}
they are both _locked paths ;-)
But in general I think the parallel code paths make things more confusing about what is the right thing to call. And I would like to put more WARN_ON(!locked()) in the gem_vma code, to make it clear that the obj lock is protecting things there as well.. which, with this approach would turn into parallel code paths multiple levels deep
BR,
-R
Quoted 115 lines
drivers/gpu/drm/msm/msm_gem.c +23
-20+ +/* Called with msm_obj locked or on the free path */ static void -put_iova_vmas (struct drm_gem_object *obj) +put_iova_vmas_locked (struct drm_gem_object *obj) {struct msm_gem_object *msm_obj =to_msm_bo (obj);struct msm_gem_vma *vma, *tmp; -GEM_WARN_ON (!msm_gem_is_locked (obj)); - -list_for_each_entry_safe (vma, tmp, &msm_obj->vmas, list) { +list_for_each_entry_safe (vma, tmp, &msm_obj->vmas, list)del_vma (vma); - } +} + +static void +put_iova_vmas (struct drm_gem_object *obj) +{ +GEM_WARN_ON (!msm_gem_is_locked (obj)); +put_iova_vmas_locked (obj); } static struct msm_gem_vma *get_vma_locked(struct drm_gem_object *obj,@@ -795,12 +803,10 @@ void msm_gem_evict(struct drm_gem_object *obj)update_inactive (msm_obj); } -void msm_gem_vunmap (struct drm_gem_object *obj) +static void msm_gem_vunmap_locked (struct drm_gem_object *obj) {struct msm_gem_object *msm_obj =to_msm_bo (obj); -GEM_WARN_ON (!msm_gem_is_locked (obj)); -if (!msm_obj->vaddr ||GEM_WARN_ON (!is_vunmapable (msm_obj)))return ;@@ -808,6 +814,12 @@ void msm_gem_vunmap(struct drm_gem_object *obj) msm_obj->vaddr =NULL ; } +void msm_gem_vunmap (struct drm_gem_object *obj) +{ +GEM_WARN_ON (!msm_gem_is_locked (obj)); +msm_gem_vunmap_locked (obj); +} +void msm_gem_active_get (struct drm_gem_object *obj,struct msm_gpu *gpu) {struct msm_gem_object *msm_obj =to_msm_bo (obj);@@ -1021,12 +1033,11 @@ void msm_gem_free_object(struct drm_gem_object *obj)list_del (&msm_obj->mm_list);mutex_unlock (&priv->mm_lock); -msm_gem_lock (obj); -/* object should not be on active list: */ GEM_WARN_ON (is_active (msm_obj)); -put_iova_spaces (obj, true); +put_iova_spaces_locked (obj, true); +if (obj->import_attach) {GEM_WARN_ON (msm_obj->vaddr);@@ -1036,19 +1047,13 @@ void msm_gem_free_object(struct drm_gem_object *obj)*/ kvfree (msm_obj->pages); -put_iova_vmas (obj); - -/* dma_buf_detach() grabs resv lock, so we need to unlock -* prior to drm_prime_gem_destroy -*/ -msm_gem_unlock (obj); +put_iova_vmas_locked (obj);drm_prime_gem_destroy (obj, msm_obj->sgt); }else { -msm_gem_vunmap (obj); +msm_gem_vunmap_locked (obj);put_pages (obj); -put_iova_vmas (obj); -msm_gem_unlock (obj); +put_iova_vmas_locked (obj); }drm_gem_object_release (obj);drivers/gpu/drm/msm/msm_gem.h +0
-1@@ -253,7 +253,6 @@ static inline bool is_purgeable(struct msm_gem_object *msm_obj)static inline bool is_vunmapable (struct msm_gem_object *msm_obj) { -GEM_WARN_ON (!msm_gem_is_locked (&msm_obj->base));return (msm_obj->vmap_count ==0 ) && msm_obj->vaddr; }drivers/gpu/drm/msm/msm_gem_shrinker.c +1
-0@@ -175,6 +175,7 @@ static const int vmap_shrink_limit = 15;static bool vmap_shrink (struct msm_gem_object *msm_obj) { +GEM_WARN_ON (!msm_gem_is_locked (&msm_obj->base));if (!is_vunmapable (msm_obj))return false;
Quoted 34 lines · Rob Clark
Quoted 25 lines · Stephen Boyd
Quoting Rob Clark (2022-06-13 13:50:32)
Quoted 12 lines
drivers/gpu/drm/msm/msm_gem.h +3
-1@@ -229,7 +229,19 @@ msm_gem_unlock(struct drm_gem_object *obj)static inline bool msm_gem_is_locked (struct drm_gem_object *obj) { -return dma_resv_is_locked (obj->resv); +/* +* Destroying the object is a special case.. msm_gem_free_object() +* calls many things that WARN_ON if the obj lock is not held. But drivers/gpu/drm/msm/msm_gem.h msm_gem_unlock(struct drm_gem_object *obj) · line 244
Quoted 10 lines
drivers/gpu/drm/msm/msm_gem.h +10
-0+* acquiring the obj lock in msm_gem_free_object() can cause a +* locking order inversion between reservation_ww_class_mutex and +* fs_reclaim. +* +* This deadlock is not actually possible, because no one should +* be already holding the lock when msm_gem_free_object() is called. +* Unfortunately lockdep is not aware of this detail. So when the +* refcount drops to zero, we pretend it is already locked. +*/ +return dma_resv_is_locked (obj->resv) || (kref_read (&obj->refcount) ==0 );Quoted 7 lines
Instead of modifying this function can we push down the fact that this function is being called from the free path and skip checking this condition in that case? Or add some "_locked/free_path" wrappers that skip the lock assertion? That would make it clearer to understand while reading the code that it is locked when it is asserted to be locked, and that we don't care when we're freeing because all references to the object are gone.
Quoted 3 lines
that was my earlier attempt, and I wasn't too happy with the result. And then I realized if refcount==0 then by definition we aren't racing with anyone else ;-)
I think that's not entirely correct, at least not for fairly reasonable shrinker designs:
If the shrinker trylocks for throwing stuff out it might race with a concurrent finalization. Depends a bit upon your design, but usually that's possible.
There won't be a problem since you'll serialize on a lock eventually. But if you drop all your locking debug checks like this then it's very hard to figure this out :-)
Ofc you can adjust your refcounting scheme to make this impossible, but then there's also not really any need to call any functions which might need some locking, since by that time all that stuff must have been cleaned up already (or the refcount could not have dropped to zero). Like if any iova mapping holds a reference you never have these problems.
Long story short, this kind of design freaks me out big time. Especially when it involves both a cross-driver refcount like the gem_bo one and a cross driver lock ...
The standard way to fix this is to trylock dma_resv on cleanup and push to a worker if you can't get it. This is what ttm and i915 does. Might be good to lift that into drm_gem.c helpers and just use it. -Daniel
Quoted 162 lines
Quoted 26 lines
Here's a totally untested patch to show the idea. The comment about pretending the lock is held can be put in
msm_gem_free_object()to clarify why it's OK to call the locked variants of the functions.---8<---
drivers/gpu/drm/msm/msm_gem.c +1
-3@@ -346,13 +346,11 @@ static void del_vma(struct msm_gem_vma *vma)* mapping. */ static void -put_iova_spaces (struct drm_gem_object *obj,bool close) +put_iova_spaces_locked (struct drm_gem_object *obj,bool close) {struct msm_gem_object *msm_obj =to_msm_bo (obj);struct msm_gem_vma *vma; -GEM_WARN_ON (!msm_gem_is_locked (obj)); -list_for_each_entry (vma, &msm_obj->vmas, list) {if (vma->aspace) {msm_gem_purge_vma (vma->aspace, vma);drivers/gpu/drm/msm/msm_gem.c put_iova_spaces(struct drm_gem_object *obj, bool close) · line 367
Quoted 9 lines
drivers/gpu/drm/msm/msm_gem.c +5
-1@@ -362,18 +360,28 @@ put_iova_spaces(struct drm_gem_object *obj, bool close) } } -/* Called with msm_obj locked */ +static void put_iova_spaces (struct drm_gem_object *obj,bool close) +{ +GEM_WARN_ON (!msm_gem_is_locked (obj)); +put_iova_spaces_locked (obj, close); +}they are both
_lockedpaths ;-)But in general I think the parallel code paths make things more confusing about what is the right thing to call. And I would like to put more WARN_ON(!locked()) in the
gem_vmacode, to make it clear that the obj lock is protecting things there as well.. which, with this approach would turn into parallel code paths multiple levels deepBR,
-RQuoted 115 lines
drivers/gpu/drm/msm/msm_gem.c +23
-20+ +/* Called with msm_obj locked or on the free path */ static void -put_iova_vmas (struct drm_gem_object *obj) +put_iova_vmas_locked (struct drm_gem_object *obj) {struct msm_gem_object *msm_obj =to_msm_bo (obj);struct msm_gem_vma *vma, *tmp; -GEM_WARN_ON (!msm_gem_is_locked (obj)); - -list_for_each_entry_safe (vma, tmp, &msm_obj->vmas, list) { +list_for_each_entry_safe (vma, tmp, &msm_obj->vmas, list)del_vma (vma); - } +} + +static void +put_iova_vmas (struct drm_gem_object *obj) +{ +GEM_WARN_ON (!msm_gem_is_locked (obj)); +put_iova_vmas_locked (obj); } static struct msm_gem_vma *get_vma_locked(struct drm_gem_object *obj,@@ -795,12 +803,10 @@ void msm_gem_evict(struct drm_gem_object *obj)update_inactive (msm_obj); } -void msm_gem_vunmap (struct drm_gem_object *obj) +static void msm_gem_vunmap_locked (struct drm_gem_object *obj) {struct msm_gem_object *msm_obj =to_msm_bo (obj); -GEM_WARN_ON (!msm_gem_is_locked (obj)); -if (!msm_obj->vaddr ||GEM_WARN_ON (!is_vunmapable (msm_obj)))return ;@@ -808,6 +814,12 @@ void msm_gem_vunmap(struct drm_gem_object *obj) msm_obj->vaddr =NULL ; } +void msm_gem_vunmap (struct drm_gem_object *obj) +{ +GEM_WARN_ON (!msm_gem_is_locked (obj)); +msm_gem_vunmap_locked (obj); +} +void msm_gem_active_get (struct drm_gem_object *obj,struct msm_gpu *gpu) {struct msm_gem_object *msm_obj =to_msm_bo (obj);@@ -1021,12 +1033,11 @@ void msm_gem_free_object(struct drm_gem_object *obj)list_del (&msm_obj->mm_list);mutex_unlock (&priv->mm_lock); -msm_gem_lock (obj); -/* object should not be on active list: */ GEM_WARN_ON (is_active (msm_obj)); -put_iova_spaces (obj, true); +put_iova_spaces_locked (obj, true); +if (obj->import_attach) {GEM_WARN_ON (msm_obj->vaddr);@@ -1036,19 +1047,13 @@ void msm_gem_free_object(struct drm_gem_object *obj)*/ kvfree (msm_obj->pages); -put_iova_vmas (obj); - -/* dma_buf_detach() grabs resv lock, so we need to unlock -* prior to drm_prime_gem_destroy -*/ -msm_gem_unlock (obj); +put_iova_vmas_locked (obj);drm_prime_gem_destroy (obj, msm_obj->sgt); }else { -msm_gem_vunmap (obj); +msm_gem_vunmap_locked (obj);put_pages (obj); -put_iova_vmas (obj); -msm_gem_unlock (obj); +put_iova_vmas_locked (obj); }drm_gem_object_release (obj);drivers/gpu/drm/msm/msm_gem.h +0
-1@@ -253,7 +253,6 @@ static inline bool is_purgeable(struct msm_gem_object *msm_obj)static inline bool is_vunmapable (struct msm_gem_object *msm_obj) { -GEM_WARN_ON (!msm_gem_is_locked (&msm_obj->base));return (msm_obj->vmap_count ==0 ) && msm_obj->vaddr; }drivers/gpu/drm/msm/msm_gem_shrinker.c +1
-0@@ -175,6 +175,7 @@ static const int vmap_shrink_limit = 15;static bool vmap_shrink (struct msm_gem_object *msm_obj) { +GEM_WARN_ON (!msm_gem_is_locked (&msm_obj->base));if (!is_vunmapable (msm_obj))return false;
Signature · 3 lines
Daniel Vetter Software Engineer, Intel Corporation http://blog.ffwll.ch
Quoted 40 lines · Daniel Vetter
Quoted 34 lines · Rob Clark
Quoted 25 lines · Stephen Boyd
Quoting Rob Clark (2022-06-13 13:50:32)
Quoted 12 lines
drivers/gpu/drm/msm/msm_gem.h +3
-1@@ -229,7 +229,19 @@ msm_gem_unlock(struct drm_gem_object *obj)static inline bool msm_gem_is_locked (struct drm_gem_object *obj) { -return dma_resv_is_locked (obj->resv); +/* +* Destroying the object is a special case.. msm_gem_free_object() +* calls many things that WARN_ON if the obj lock is not held. But drivers/gpu/drm/msm/msm_gem.h msm_gem_unlock(struct drm_gem_object *obj) · line 244
Quoted 10 lines
drivers/gpu/drm/msm/msm_gem.h +10
-0+* acquiring the obj lock in msm_gem_free_object() can cause a +* locking order inversion between reservation_ww_class_mutex and +* fs_reclaim. +* +* This deadlock is not actually possible, because no one should +* be already holding the lock when msm_gem_free_object() is called. +* Unfortunately lockdep is not aware of this detail. So when the +* refcount drops to zero, we pretend it is already locked. +*/ +return dma_resv_is_locked (obj->resv) || (kref_read (&obj->refcount) ==0 );Quoted 7 lines
Instead of modifying this function can we push down the fact that this function is being called from the free path and skip checking this condition in that case? Or add some "_locked/free_path" wrappers that skip the lock assertion? That would make it clearer to understand while reading the code that it is locked when it is asserted to be locked, and that we don't care when we're freeing because all references to the object are gone.
Quoted 3 lines
that was my earlier attempt, and I wasn't too happy with the result. And then I realized if refcount==0 then by definition we aren't racing with anyone else ;-)
Quoted 6 lines
I think that's not entirely correct, at least not for fairly reasonable shrinker designs:
If the shrinker trylocks for throwing stuff out it might race with a concurrent finalization. Depends a bit upon your design, but usually that's possible.
Kinda but in fact no. At least not if your shrinker is designed properly.
The shrinker does kref_get_unless_zero() and bails in the case that we've already started finalizing. See:
https://patchwork.freedesktop.org/patch/490160/
Quoted 10 lines
There won't be a problem since you'll serialize on a lock eventually. But if you drop all your locking debug checks like this then it's very hard to figure this out :-)
Ofc you can adjust your refcounting scheme to make this impossible, but then there's also not really any need to call any functions which might need some locking, since by that time all that stuff must have been cleaned up already (or the refcount could not have dropped to zero). Like if any iova mapping holds a reference you never have these problems.
Quoted 7 lines
Long story short, this kind of design freaks me out big time. Especially when it involves both a cross-driver refcount like the
gem_boone and a cross driver lock ...The standard way to fix this is to trylock
dma_resvon cleanup and push to a worker if you can't get it. This is what ttm and i915 does. Might be good to lift that intodrm_gem.chelpers and just use it.
We used to do that (but unconditionally).. and got rid of it because it was causing jank issues (lots of queued up finalizers delaying retire work, or something along those lines, IIRC). I guess back then struct_mutex was also involved, and it might not be as bad if we only took the async path if we didn't win the trylock. But IMO that is totally unnecessary. And I kinda am skeptical about pushing too much locking stuff to drm core.
BR,
-R
Quoted 169 lines
-Daniel
Quoted 150 lines
Quoted 26 lines
Here's a totally untested patch to show the idea. The comment about pretending the lock is held can be put in
msm_gem_free_object()to clarify why it's OK to call the locked variants of the functions.---8<---
drivers/gpu/drm/msm/msm_gem.c +1
-3@@ -346,13 +346,11 @@ static void del_vma(struct msm_gem_vma *vma)* mapping. */ static void -put_iova_spaces (struct drm_gem_object *obj,bool close) +put_iova_spaces_locked (struct drm_gem_object *obj,bool close) {struct msm_gem_object *msm_obj =to_msm_bo (obj);struct msm_gem_vma *vma; -GEM_WARN_ON (!msm_gem_is_locked (obj)); -list_for_each_entry (vma, &msm_obj->vmas, list) {if (vma->aspace) {msm_gem_purge_vma (vma->aspace, vma);drivers/gpu/drm/msm/msm_gem.c put_iova_spaces(struct drm_gem_object *obj, bool close) · line 367
Quoted 9 lines
drivers/gpu/drm/msm/msm_gem.c +5
-1@@ -362,18 +360,28 @@ put_iova_spaces(struct drm_gem_object *obj, bool close) } } -/* Called with msm_obj locked */ +static void put_iova_spaces (struct drm_gem_object *obj,bool close) +{ +GEM_WARN_ON (!msm_gem_is_locked (obj)); +put_iova_spaces_locked (obj, close); +}they are both
_lockedpaths ;-)But in general I think the parallel code paths make things more confusing about what is the right thing to call. And I would like to put more WARN_ON(!locked()) in the
gem_vmacode, to make it clear that the obj lock is protecting things there as well.. which, with this approach would turn into parallel code paths multiple levels deepBR,
-RQuoted 103 lines
drivers/gpu/drm/msm/msm_gem.c +23
-20+ +/* Called with msm_obj locked or on the free path */ static void -put_iova_vmas (struct drm_gem_object *obj) +put_iova_vmas_locked (struct drm_gem_object *obj) {struct msm_gem_object *msm_obj =to_msm_bo (obj);struct msm_gem_vma *vma, *tmp; -GEM_WARN_ON (!msm_gem_is_locked (obj)); - -list_for_each_entry_safe (vma, tmp, &msm_obj->vmas, list) { +list_for_each_entry_safe (vma, tmp, &msm_obj->vmas, list)del_vma (vma); - } +} + +static void +put_iova_vmas (struct drm_gem_object *obj) +{ +GEM_WARN_ON (!msm_gem_is_locked (obj)); +put_iova_vmas_locked (obj); } static struct msm_gem_vma *get_vma_locked(struct drm_gem_object *obj,@@ -795,12 +803,10 @@ void msm_gem_evict(struct drm_gem_object *obj)update_inactive (msm_obj); } -void msm_gem_vunmap (struct drm_gem_object *obj) +static void msm_gem_vunmap_locked (struct drm_gem_object *obj) {struct msm_gem_object *msm_obj =to_msm_bo (obj); -GEM_WARN_ON (!msm_gem_is_locked (obj)); -if (!msm_obj->vaddr ||GEM_WARN_ON (!is_vunmapable (msm_obj)))return ;@@ -808,6 +814,12 @@ void msm_gem_vunmap(struct drm_gem_object *obj) msm_obj->vaddr =NULL ; } +void msm_gem_vunmap (struct drm_gem_object *obj) +{ +GEM_WARN_ON (!msm_gem_is_locked (obj)); +msm_gem_vunmap_locked (obj); +} +void msm_gem_active_get (struct drm_gem_object *obj,struct msm_gpu *gpu) {struct msm_gem_object *msm_obj =to_msm_bo (obj);@@ -1021,12 +1033,11 @@ void msm_gem_free_object(struct drm_gem_object *obj)list_del (&msm_obj->mm_list);mutex_unlock (&priv->mm_lock); -msm_gem_lock (obj); -/* object should not be on active list: */ GEM_WARN_ON (is_active (msm_obj)); -put_iova_spaces (obj, true); +put_iova_spaces_locked (obj, true); +if (obj->import_attach) {GEM_WARN_ON (msm_obj->vaddr);@@ -1036,19 +1047,13 @@ void msm_gem_free_object(struct drm_gem_object *obj)*/ kvfree (msm_obj->pages); -put_iova_vmas (obj); - -/* dma_buf_detach() grabs resv lock, so we need to unlock -* prior to drm_prime_gem_destroy -*/ -msm_gem_unlock (obj); +put_iova_vmas_locked (obj);drm_prime_gem_destroy (obj, msm_obj->sgt); }else { -msm_gem_vunmap (obj); +msm_gem_vunmap_locked (obj);put_pages (obj); -put_iova_vmas (obj); -msm_gem_unlock (obj); +put_iova_vmas_locked (obj); }drm_gem_object_release (obj);drivers/gpu/drm/msm/msm_gem.h +0
-1@@ -253,7 +253,6 @@ static inline bool is_purgeable(struct msm_gem_object *msm_obj)static inline bool is_vunmapable (struct msm_gem_object *msm_obj) { -GEM_WARN_ON (!msm_gem_is_locked (&msm_obj->base));return (msm_obj->vmap_count ==0 ) && msm_obj->vaddr; }Quoted 12 lines
drivers/gpu/drm/msm/msm_gem_shrinker.c static const int vmap_shrink_limit = 15; · line 180
Quoted 12 lines
drivers/gpu/drm/msm/msm_gem_shrinker.c +1
-0@@ -175,6 +175,7 @@ static const int vmap_shrink_limit = 15;static bool vmap_shrink (struct msm_gem_object *msm_obj) { +GEM_WARN_ON (!msm_gem_is_locked (&msm_obj->base));if (!is_vunmapable (msm_obj))return false;--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch
Quoted 48 lines · Rob Clark
Quoted 40 lines · Daniel Vetter
Quoted 34 lines · Rob Clark
Quoted 25 lines · Stephen Boyd
Quoting Rob Clark (2022-06-13 13:50:32)
Quoted 12 lines
drivers/gpu/drm/msm/msm_gem.h +3
-1@@ -229,7 +229,19 @@ msm_gem_unlock(struct drm_gem_object *obj)static inline bool msm_gem_is_locked (struct drm_gem_object *obj) { -return dma_resv_is_locked (obj->resv); +/* +* Destroying the object is a special case.. msm_gem_free_object() +* calls many things that WARN_ON if the obj lock is not held. But drivers/gpu/drm/msm/msm_gem.h msm_gem_unlock(struct drm_gem_object *obj) · line 244
Quoted 10 lines
drivers/gpu/drm/msm/msm_gem.h +10
-0+* acquiring the obj lock in msm_gem_free_object() can cause a +* locking order inversion between reservation_ww_class_mutex and +* fs_reclaim. +* +* This deadlock is not actually possible, because no one should +* be already holding the lock when msm_gem_free_object() is called. +* Unfortunately lockdep is not aware of this detail. So when the +* refcount drops to zero, we pretend it is already locked. +*/ +return dma_resv_is_locked (obj->resv) || (kref_read (&obj->refcount) ==0 );Quoted 7 lines
Instead of modifying this function can we push down the fact that this function is being called from the free path and skip checking this condition in that case? Or add some "_locked/free_path" wrappers that skip the lock assertion? That would make it clearer to understand while reading the code that it is locked when it is asserted to be locked, and that we don't care when we're freeing because all references to the object are gone.
Quoted 3 lines
that was my earlier attempt, and I wasn't too happy with the result. And then I realized if refcount==0 then by definition we aren't racing with anyone else ;-)
Quoted 6 lines
I think that's not entirely correct, at least not for fairly reasonable shrinker designs:
If the shrinker trylocks for throwing stuff out it might race with a concurrent finalization. Depends a bit upon your design, but usually that's possible.
Quoted 6 lines
Kinda but in fact no. At least not if your shrinker is designed properly.
The shrinker does
kref_get_unless_zero()and bails in the case that we've already started finalizing. See:
Oh you have the order differently than what I'd have typed. If you do dma_resv_trylock under the lru lock then the kref_get_unless_zero isn't needed.
Ofc if you do it like you do then you need your locking check trickery.
Quoted 18 lines
Quoted 10 lines
There won't be a problem since you'll serialize on a lock eventually. But if you drop all your locking debug checks like this then it's very hard to figure this out :-)
Ofc you can adjust your refcounting scheme to make this impossible, but then there's also not really any need to call any functions which might need some locking, since by that time all that stuff must have been cleaned up already (or the refcount could not have dropped to zero). Like if any iova mapping holds a reference you never have these problems.
Quoted 7 lines
Long story short, this kind of design freaks me out big time. Especially when it involves both a cross-driver refcount like the
gem_boone and a cross driver lock ...The standard way to fix this is to trylock
dma_resvon cleanup and push to a worker if you can't get it. This is what ttm and i915 does. Might be good to lift that intodrm_gem.chelpers and just use it.
Quoted 7 lines
We used to do that (but unconditionally).. and got rid of it because it was causing jank issues (lots of queued up finalizers delaying retire work, or something along those lines, IIRC). I guess back then
struct_mutexwas also involved, and it might not be as bad if we only took the async path if we didn't win the trylock. But IMO that is totally unnecessary. And I kinda am skeptical about pushing too much locking stuff to drm core.
Yeah with dev->struct_mutex and unconditionally it's going to be terrible. It really should't be that bad.
Pulling back into the big picture, I really don't like drivers to spin their own world for this stuff. And with the ttm drivers (and the i915-gem one or so) doing one thing, I think msm should do the same. Unless there's a reason why that's really stupid, and then we should probably switch ttm over to that too?
If ever driver uses dma_resv differently in the cleanup paths (which are really the tricky ones) then cross driver code reading becomes an exercise in pitfall counting and leg regeneration :-(
Also I really don't care about which bikeshed we settle on, as least as they're all the same.
-Daniel
Signature · 3 lines
Daniel Vetter Software Engineer, Intel Corporation http://blog.ffwll.ch
Quoted 55 lines · Daniel Vetter
Quoted 53 lines · Rob Clark
Quoted 40 lines · Daniel Vetter
Quoted 34 lines · Rob Clark
Quoted 25 lines · Stephen Boyd
Quoting Rob Clark (2022-06-13 13:50:32)
Quoted 12 lines
drivers/gpu/drm/msm/msm_gem.h +3
-1@@ -229,7 +229,19 @@ msm_gem_unlock(struct drm_gem_object *obj)static inline bool msm_gem_is_locked (struct drm_gem_object *obj) { -return dma_resv_is_locked (obj->resv); +/* +* Destroying the object is a special case.. msm_gem_free_object() +* calls many things that WARN_ON if the obj lock is not held. But drivers/gpu/drm/msm/msm_gem.h msm_gem_unlock(struct drm_gem_object *obj) · line 244
Quoted 10 lines
drivers/gpu/drm/msm/msm_gem.h +10
-0+* acquiring the obj lock in msm_gem_free_object() can cause a +* locking order inversion between reservation_ww_class_mutex and +* fs_reclaim. +* +* This deadlock is not actually possible, because no one should +* be already holding the lock when msm_gem_free_object() is called. +* Unfortunately lockdep is not aware of this detail. So when the +* refcount drops to zero, we pretend it is already locked. +*/ +return dma_resv_is_locked (obj->resv) || (kref_read (&obj->refcount) ==0 );Quoted 7 lines
Instead of modifying this function can we push down the fact that this function is being called from the free path and skip checking this condition in that case? Or add some "_locked/free_path" wrappers that skip the lock assertion? That would make it clearer to understand while reading the code that it is locked when it is asserted to be locked, and that we don't care when we're freeing because all references to the object are gone.
Quoted 3 lines
that was my earlier attempt, and I wasn't too happy with the result. And then I realized if refcount==0 then by definition we aren't racing with anyone else ;-)
Quoted 6 lines
I think that's not entirely correct, at least not for fairly reasonable shrinker designs:
If the shrinker trylocks for throwing stuff out it might race with a concurrent finalization. Depends a bit upon your design, but usually that's possible.
Kinda but in fact no. At least not if your shrinker is designed properly.
The shrinker does
kref_get_unless_zero()and bails in the case that we've already started finalizing. See:
Quoted 7 lines
Quoted 1 line
Oh you have the order differently than what I'd have typed. If you do
dma_resv_trylockunder the lru lock then thekref_get_unless_zeroisn't needed.Ofc if you do it like you do then you need your locking check trickery.
Just as a side note, if I didn't sprinkle around so liberally WARN_ON(!obj_is_locked(obj)) so much, we also wouldn't need this trickery. (But I'm a fan of those, both to remember where which locks are expected to be held, and to be shouty if $future_me screws it up)
Quoted 36 lines
Quoted 19 lines
Quoted 11 lines
There won't be a problem since you'll serialize on a lock eventually. But if you drop all your locking debug checks like this then it's very hard to figure this out :-)
Ofc you can adjust your refcounting scheme to make this impossible, but then there's also not really any need to call any functions which might need some locking, since by that time all that stuff must have been cleaned up already (or the refcount could not have dropped to zero). Like if any iova mapping holds a reference you never have these problems.
Quoted 7 lines
Long story short, this kind of design freaks me out big time. Especially when it involves both a cross-driver refcount like the
gem_boone and a cross driver lock ...The standard way to fix this is to trylock
dma_resvon cleanup and push to a worker if you can't get it. This is what ttm and i915 does. Might be good to lift that intodrm_gem.chelpers and just use it.Quoted 7 lines
We used to do that (but unconditionally).. and got rid of it because it was causing jank issues (lots of queued up finalizers delaying retire work, or something along those lines, IIRC). I guess back then
struct_mutexwas also involved, and it might not be as bad if we only took the async path if we didn't win the trylock. But IMO that is totally unnecessary. And I kinda am skeptical about pushing too much locking stuff to drm core.Yeah with
dev->struct_mutexand unconditionally it's going to be terrible. It really should't be that bad.Pulling back into the big picture, I really don't like drivers to spin their own world for this stuff. And with the ttm drivers (and the i915-gem one or so) doing one thing, I think msm should do the same. Unless there's a reason why that's really stupid, and then we should probably switch ttm over to that too?
Quoted 6 lines
If ever driver uses
dma_resvdifferently in the cleanup paths (which are really the tricky ones) then cross driver code reading becomes an exercise in pitfall counting and leg regeneration :-(Also I really don't care about which bikeshed we settle on, as least as they're all the same.
Start by bikeshedding my RFC for lru/shrinker helpers? At least then we could pull the tricky bit about parallel references to objects being finalized into a helper.
(I'll send a non-RFC version with a few fixes, hopefully maybe even today.. I'm just putting finishing touches on larger series that it is a part of)
BR,
-R
Quoted 5 lines
-Daniel
--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch