Skip to contentSkip to conversation

drm/msm: de-struct_mutex-ification

PATCH In review Rob Clark 4 participants 31 messages Oct 23, 2020 – Oct 29, 2020

Subsystems (5)
lore
Share & tools
lore

Rob Clark

author
#0 lore raw

v4 cover drm/msm: de-struct_mutex-ification

From: Rob Clark <[email protected]>

This doesn't remove *all* the struct_mutex, but it covers the worst of it, ie. shrinker/madvise/free/retire. The submit path still uses struct_mutex, but it still needs *something* serialize a portion of the submit path, and lock_stat mostly just shows the lock contention there being with other submits. And there are a few other bits of struct_mutex usage in less critical paths (debugfs, etc). But this seems like a reasonable step in the right direction.

v2: teach lockdep about shrinker locking patters (danvet) and

    convert to obj->resv locking (danvet)
v3: fix get_vaddr locking for legacy userspace (relocs), devcoredump,
    and rd/hangrd
v4: couple minor review comments (krh), fix deadlock with imported
    dma-buf's (ie. from v4l2, etc)

Rob Clark (23):
drm/msm: Fix a couple incorrect usages of get_vaddr_active() drm/msm/gem: Add obj->lock wrappers
drm/msm/gem: Rename internal get_iova_locked helper drm/msm/gem: Move prototypes to msm_gem.h drm/msm/gem: Add some _locked() helpers drm/msm/gem: Move locking in shrinker path drm/msm/submit: Move copy_from_user ahead of locking bos drm/msm: Do rpm get sooner in the submit path drm/msm/gem: Switch over to obj->resv for locking drm/msm: Use correct drm_gem_object_put() in fail case drm/msm: Drop chatty trace
drm/msm: Move update_fences()
drm/msm: Add priv->mm_lock to protect active/inactive lists drm/msm: Document and rename preempt_lock
drm/msm: Protect ring->submits with it's own lock drm/msm: Refcount submits
drm/msm: Remove obj->gpu
drm/msm: Drop struct_mutex from the retire path drm/msm: Drop struct_mutex in free_object() path drm/msm: Remove msm_gem_free_work
drm/msm: Drop struct_mutex in madvise path drm/msm: Drop struct_mutex in shrinker path drm/msm: Don't implicit-sync if only a single ring

 drivers/gpu/drm/msm/adreno/a5xx_gpu.c     |   6 +-
 drivers/gpu/drm/msm/adreno/a5xx_preempt.c |  12 +-
 drivers/gpu/drm/msm/adreno/a6xx_gpu.c     |   6 +-
 drivers/gpu/drm/msm/disp/mdp4/mdp4_crtc.c |   1 +
 drivers/gpu/drm/msm/disp/mdp5/mdp5_crtc.c |   1 +
 drivers/gpu/drm/msm/dsi/dsi_host.c        |   1 +
 drivers/gpu/drm/msm/msm_debugfs.c         |   7 +
 drivers/gpu/drm/msm/msm_drv.c             |  21 +-
 drivers/gpu/drm/msm/msm_drv.h             |  73 +-----
 drivers/gpu/drm/msm/msm_fbdev.c           |   1 +
 drivers/gpu/drm/msm/msm_gem.c             | 271 +++++++++++-----------
 drivers/gpu/drm/msm/msm_gem.h             | 133 +++++++++--
 drivers/gpu/drm/msm/msm_gem_shrinker.c    |  81 ++-----
 drivers/gpu/drm/msm/msm_gem_submit.c      | 164 ++++++++-----
 drivers/gpu/drm/msm/msm_gpu.c             | 110 +++++----
 drivers/gpu/drm/msm/msm_gpu.h             |   5 +-
 drivers/gpu/drm/msm/msm_rd.c              |   2 +-
 drivers/gpu/drm/msm/msm_ringbuffer.c      |   3 +-
 drivers/gpu/drm/msm/msm_ringbuffer.h      |  13 +-
 19 files changed, 506 insertions(+), 405 deletions(-)
Signature · 1 line
2.26.2
author
#1 lore raw

PATCH v4 1/23 drm/msm: Fix a couple incorrect usages of get_vaddr_active()

From: Rob Clark <[email protected]>

The microcode bo's should never be madvise(WONTNEED), so these should not be using msm_gem_get_vaddr_active().

---
 drivers/gpu/drm/msm/adreno/a5xx_gpu.c | 2 +-
 drivers/gpu/drm/msm/adreno/a6xx_gpu.c | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)
drivers/gpu/drm/msm/adreno/a5xx_gpu.c +1 -1
index c941c8138f25..2180650a03bc 100644
@@ -426,7 +426,7 @@ static int a5xx_preempt_start(struct msm_gpu *gpu)
 static void a5xx_ucode_check_version(struct a5xx_gpu *a5xx_gpu,
 		struct drm_gem_object *obj)
 {
-	u32 *buf = msm_gem_get_vaddr_active(obj);
+	u32 *buf = msm_gem_get_vaddr(obj);
 
 	if (IS_ERR(buf))
 		return;
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +1 -1
index 8915882e4444..16eaaf0804ca 100644
@@ -522,7 +522,7 @@ static int a6xx_cp_init(struct msm_gpu *gpu)
 static void a6xx_ucode_check_version(struct a6xx_gpu *a6xx_gpu,
 		struct drm_gem_object *obj)
 {
-	u32 *buf = msm_gem_get_vaddr_active(obj);
+	u32 *buf = msm_gem_get_vaddr(obj);
 
 	if (IS_ERR(buf))
 		return;
Signature · 6 lines
2.26.2

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
author
#2 lore raw

PATCH v4 2/23 drm/msm/gem: Add obj->lock wrappers

In reply to Rob Clark

From: Rob Clark <[email protected]>

This will make it easier to transition over to obj->resv locking for everything that is per-bo locking.

---
 drivers/gpu/drm/msm/msm_gem.c | 99 ++++++++++++++++-------------------
 drivers/gpu/drm/msm/msm_gem.h | 28 ++++++++++
 2 files changed, 74 insertions(+), 53 deletions(-)
drivers/gpu/drm/msm/msm_gem.c +46 -53
index 14e14caf90f9..afef9c6b1a1c 100644
@@ -178,15 +178,15 @@ struct page **msm_gem_get_pages(struct drm_gem_object *obj)
 	struct msm_gem_object *msm_obj = to_msm_bo(obj);
 	struct page **p;
 
-	mutex_lock(&msm_obj->lock);
+	msm_gem_lock(obj);
 
 	if (WARN_ON(msm_obj->madv != MSM_MADV_WILLNEED)) {
-		mutex_unlock(&msm_obj->lock);
+		msm_gem_unlock(obj);
 		return ERR_PTR(-EBUSY);
 	}
 
 	p = get_pages(obj);
-	mutex_unlock(&msm_obj->lock);
+	msm_gem_unlock(obj);
 	return p;
 }
 
@@ -252,14 +252,14 @@ vm_fault_t msm_gem_fault(struct vm_fault *vmf)
 	 * vm_ops.open/drm_gem_mmap_obj and close get and put
 	 * a reference on obj. So, we dont need to hold one here.
 	 */
-	err = mutex_lock_interruptible(&msm_obj->lock);
+	err = msm_gem_lock_interruptible(obj);
 	if (err) {
 		ret = VM_FAULT_NOPAGE;
 		goto out;
 	}
 
 	if (WARN_ON(msm_obj->madv != MSM_MADV_WILLNEED)) {
-		mutex_unlock(&msm_obj->lock);
+		msm_gem_unlock(obj);
 		return VM_FAULT_SIGBUS;
 	}
 
@@ -280,7 +280,7 @@ vm_fault_t msm_gem_fault(struct vm_fault *vmf)
 
 	ret = vmf_insert_mixed(vma, vmf->address, __pfn_to_pfn_t(pfn, PFN_DEV));
 out_unlock:
-	mutex_unlock(&msm_obj->lock);
+	msm_gem_unlock(obj);
 out:
 	return ret;
 }
@@ -289,10 +289,9 @@ vm_fault_t msm_gem_fault(struct vm_fault *vmf)
 static uint64_t mmap_offset(struct drm_gem_object *obj)
 {
 	struct drm_device *dev = obj->dev;
-	struct msm_gem_object *msm_obj = to_msm_bo(obj);
 	int ret;
 
-	WARN_ON(!mutex_is_locked(&msm_obj->lock));
+	WARN_ON(!msm_gem_is_locked(obj));
 
 	/* Make it mmapable */
 	ret = drm_gem_create_mmap_offset(obj);
@@ -308,11 +307,10 @@ static uint64_t mmap_offset(struct drm_gem_object *obj)
 uint64_t msm_gem_mmap_offset(struct drm_gem_object *obj)
 {
 	uint64_t offset;
-	struct msm_gem_object *msm_obj = to_msm_bo(obj);
 
-	mutex_lock(&msm_obj->lock);
+	msm_gem_lock(obj);
 	offset = mmap_offset(obj);
-	mutex_unlock(&msm_obj->lock);
+	msm_gem_unlock(obj);
 	return offset;
 }
 
@@ -322,7 +320,7 @@ static struct msm_gem_vma *add_vma(struct drm_gem_object *obj,
 	struct msm_gem_object *msm_obj = to_msm_bo(obj);
 	struct msm_gem_vma *vma;
 
-	WARN_ON(!mutex_is_locked(&msm_obj->lock));
+	WARN_ON(!msm_gem_is_locked(obj));
 
 	vma = kzalloc(sizeof(*vma), GFP_KERNEL);
 	if (!vma)
@@ -341,7 +339,7 @@ static struct msm_gem_vma *lookup_vma(struct drm_gem_object *obj,
 	struct msm_gem_object *msm_obj = to_msm_bo(obj);
 	struct msm_gem_vma *vma;
 
-	WARN_ON(!mutex_is_locked(&msm_obj->lock));
+	WARN_ON(!msm_gem_is_locked(obj));
 
 	list_for_each_entry(vma, &msm_obj->vmas, list) {
 		if (vma->aspace == aspace)
@@ -360,14 +358,14 @@ static void del_vma(struct msm_gem_vma *vma)
 	kfree(vma);
 }
 
-/* Called with msm_obj->lock locked */
+/* Called with msm_obj locked */
 static void
 put_iova(struct drm_gem_object *obj)
 {
 	struct msm_gem_object *msm_obj = to_msm_bo(obj);
 	struct msm_gem_vma *vma, *tmp;
 
-	WARN_ON(!mutex_is_locked(&msm_obj->lock));
+	WARN_ON(!msm_gem_is_locked(obj));
 
 	list_for_each_entry_safe(vma, tmp, &msm_obj->vmas, list) {
 		if (vma->aspace) {
@@ -382,11 +380,10 @@ static int msm_gem_get_iova_locked(struct drm_gem_object *obj,
 		struct msm_gem_address_space *aspace, uint64_t *iova,
 		u64 range_start, u64 range_end)
 {
-	struct msm_gem_object *msm_obj = to_msm_bo(obj);
 	struct msm_gem_vma *vma;
 	int ret = 0;
 
-	WARN_ON(!mutex_is_locked(&msm_obj->lock));
+	WARN_ON(!msm_gem_is_locked(obj));
 
 	vma = lookup_vma(obj, aspace);
 
@@ -421,7 +418,7 @@ static int msm_gem_pin_iova(struct drm_gem_object *obj,
 	if (msm_obj->flags & MSM_BO_MAP_PRIV)
 		prot |= IOMMU_PRIV;
 
-	WARN_ON(!mutex_is_locked(&msm_obj->lock));
+	WARN_ON(!msm_gem_is_locked(obj));
 
 	if (WARN_ON(msm_obj->madv != MSM_MADV_WILLNEED))
 		return -EBUSY;
@@ -446,11 +443,10 @@ int msm_gem_get_and_pin_iova_range(struct drm_gem_object *obj,
 		struct msm_gem_address_space *aspace, uint64_t *iova,
 		u64 range_start, u64 range_end)
 {
-	struct msm_gem_object *msm_obj = to_msm_bo(obj);
 	u64 local;
 	int ret;
 
-	mutex_lock(&msm_obj->lock);
+	msm_gem_lock(obj);
 
 	ret = msm_gem_get_iova_locked(obj, aspace, &local,
 		range_start, range_end);
@@ -461,7 +457,7 @@ int msm_gem_get_and_pin_iova_range(struct drm_gem_object *obj,
 	if (!ret)
 		*iova = local;
 
-	mutex_unlock(&msm_obj->lock);
+	msm_gem_unlock(obj);
 	return ret;
 }
 
@@ -479,12 +475,11 @@ int msm_gem_get_and_pin_iova(struct drm_gem_object *obj,
 int msm_gem_get_iova(struct drm_gem_object *obj,
 		struct msm_gem_address_space *aspace, uint64_t *iova)
 {
-	struct msm_gem_object *msm_obj = to_msm_bo(obj);
 	int ret;
 
-	mutex_lock(&msm_obj->lock);
+	msm_gem_lock(obj);
 	ret = msm_gem_get_iova_locked(obj, aspace, iova, 0, U64_MAX);
-	mutex_unlock(&msm_obj->lock);
+	msm_gem_unlock(obj);
 
 	return ret;
 }
@@ -495,12 +490,11 @@ int msm_gem_get_iova(struct drm_gem_object *obj,
 uint64_t msm_gem_iova(struct drm_gem_object *obj,
 		struct msm_gem_address_space *aspace)
 {
-	struct msm_gem_object *msm_obj = to_msm_bo(obj);
 	struct msm_gem_vma *vma;
 
-	mutex_lock(&msm_obj->lock);
+	msm_gem_lock(obj);
 	vma = lookup_vma(obj, aspace);
-	mutex_unlock(&msm_obj->lock);
+	msm_gem_unlock(obj);
 	WARN_ON(!vma);
 
 	return vma ? vma->iova : 0;
@@ -514,16 +508,15 @@ uint64_t msm_gem_iova(struct drm_gem_object *obj,
 void msm_gem_unpin_iova(struct drm_gem_object *obj,
 		struct msm_gem_address_space *aspace)
 {
-	struct msm_gem_object *msm_obj = to_msm_bo(obj);
 	struct msm_gem_vma *vma;
 
-	mutex_lock(&msm_obj->lock);
+	msm_gem_lock(obj);
 	vma = lookup_vma(obj, aspace);
 
 	if (!WARN_ON(!vma))
 		msm_gem_unmap_vma(aspace, vma);
 
-	mutex_unlock(&msm_obj->lock);
+	msm_gem_unlock(obj);
 }
 
 int msm_gem_dumb_create(struct drm_file *file, struct drm_device *dev,
@@ -564,20 +557,20 @@ static void *get_vaddr(struct drm_gem_object *obj, unsigned madv)
 	if (obj->import_attach)
 		return ERR_PTR(-ENODEV);
 
-	mutex_lock(&msm_obj->lock);
+	msm_gem_lock(obj);
 
 	if (WARN_ON(msm_obj->madv > madv)) {
 		DRM_DEV_ERROR(obj->dev->dev, "Invalid madv state: %u vs %u\n",
 			msm_obj->madv, madv);
-		mutex_unlock(&msm_obj->lock);
+		msm_gem_unlock(obj);
 		return ERR_PTR(-EBUSY);
 	}
 
 	/* increment vmap_count *before* vmap() call, so shrinker can
-	 * check vmap_count (is_vunmapable()) outside of msm_obj->lock.
+	 * check vmap_count (is_vunmapable()) outside of msm_obj lock.
 	 * This guarantees that we won't try to msm_gem_vunmap() this
 	 * same object from within the vmap() call (while we already
-	 * hold msm_obj->lock)
+	 * hold msm_obj lock)
 	 */
 	msm_obj->vmap_count++;
 
@@ -595,12 +588,12 @@ static void *get_vaddr(struct drm_gem_object *obj, unsigned madv)
 		}
 	}
 
-	mutex_unlock(&msm_obj->lock);
+	msm_gem_unlock(obj);
 	return msm_obj->vaddr;
 
 fail:
 	msm_obj->vmap_count--;
-	mutex_unlock(&msm_obj->lock);
+	msm_gem_unlock(obj);
 	return ERR_PTR(ret);
 }
 
@@ -624,10 +617,10 @@ void msm_gem_put_vaddr(struct drm_gem_object *obj)
 {
 	struct msm_gem_object *msm_obj = to_msm_bo(obj);
 
-	mutex_lock(&msm_obj->lock);
+	msm_gem_lock(obj);
 	WARN_ON(msm_obj->vmap_count < 1);
 	msm_obj->vmap_count--;
-	mutex_unlock(&msm_obj->lock);
+	msm_gem_unlock(obj);
 }
 
 /* Update madvise status, returns true if not purged, else
@@ -637,7 +630,7 @@ int msm_gem_madvise(struct drm_gem_object *obj, unsigned madv)
 {
 	struct msm_gem_object *msm_obj = to_msm_bo(obj);
 
-	mutex_lock(&msm_obj->lock);
+	msm_gem_lock(obj);
 
 	WARN_ON(!mutex_is_locked(&obj->dev->struct_mutex));
 
@@ -646,7 +639,7 @@ int msm_gem_madvise(struct drm_gem_object *obj, unsigned madv)
 
 	madv = msm_obj->madv;
 
-	mutex_unlock(&msm_obj->lock);
+	msm_gem_unlock(obj);
 
 	return (madv != __MSM_MADV_PURGED);
 }
@@ -683,14 +676,14 @@ void msm_gem_purge(struct drm_gem_object *obj, enum msm_gem_lock subclass)
 	invalidate_mapping_pages(file_inode(obj->filp)->i_mapping,
 			0, (loff_t)-1);
 
-	mutex_unlock(&msm_obj->lock);
+	msm_gem_unlock(obj);
 }
 
 static void msm_gem_vunmap_locked(struct drm_gem_object *obj)
 {
 	struct msm_gem_object *msm_obj = to_msm_bo(obj);
 
-	WARN_ON(!mutex_is_locked(&msm_obj->lock));
+	WARN_ON(!msm_gem_is_locked(obj));
 
 	if (!msm_obj->vaddr || WARN_ON(!is_vunmapable(msm_obj)))
 		return;
@@ -705,7 +698,7 @@ void msm_gem_vunmap(struct drm_gem_object *obj, enum msm_gem_lock subclass)
 
 	mutex_lock_nested(&msm_obj->lock, subclass);
 	msm_gem_vunmap_locked(obj);
-	mutex_unlock(&msm_obj->lock);
+	msm_gem_unlock(obj);
 }
 
 /* must be called before _move_to_active().. */
@@ -816,7 +809,7 @@ void msm_gem_describe(struct drm_gem_object *obj, struct seq_file *m)
 	uint64_t off = drm_vma_node_start(&obj->vma_node);
 	const char *madv;
 
-	mutex_lock(&msm_obj->lock);
+	msm_gem_lock(obj);
 
 	switch (msm_obj->madv) {
 	case __MSM_MADV_PURGED:
@@ -884,7 +877,7 @@ void msm_gem_describe(struct drm_gem_object *obj, struct seq_file *m)
 		describe_fence(fence, "Exclusive", m);
 	rcu_read_unlock();
 
-	mutex_unlock(&msm_obj->lock);
+	msm_gem_unlock(obj);
 }
 
 void msm_gem_describe_objects(struct list_head *list, struct seq_file *m)
@@ -929,7 +922,7 @@ static void free_object(struct msm_gem_object *msm_obj)
 
 	list_del(&msm_obj->mm_list);
 
-	mutex_lock(&msm_obj->lock);
+	msm_gem_lock(obj);
 
 	put_iova(obj);
 
@@ -950,7 +943,7 @@ static void free_object(struct msm_gem_object *msm_obj)
 
 	drm_gem_object_release(obj);
 
-	mutex_unlock(&msm_obj->lock);
+	msm_gem_unlock(obj);
 	kfree(msm_obj);
 }
 
@@ -1070,10 +1063,10 @@ static struct drm_gem_object *_msm_gem_new(struct drm_device *dev,
 		struct msm_gem_vma *vma;
 		struct page **pages;
 
-		mutex_lock(&msm_obj->lock);
+		msm_gem_lock(obj);
 
 		vma = add_vma(obj, NULL);
-		mutex_unlock(&msm_obj->lock);
+		msm_gem_unlock(obj);
 		if (IS_ERR(vma)) {
 			ret = PTR_ERR(vma);
 			goto fail;
@@ -1157,22 +1150,22 @@ struct drm_gem_object *msm_gem_import(struct drm_device *dev,
 	npages = size / PAGE_SIZE;
 
 	msm_obj = to_msm_bo(obj);
-	mutex_lock(&msm_obj->lock);
+	msm_gem_lock(obj);
 	msm_obj->sgt = sgt;
 	msm_obj->pages = kvmalloc_array(npages, sizeof(struct page *), GFP_KERNEL);
 	if (!msm_obj->pages) {
-		mutex_unlock(&msm_obj->lock);
+		msm_gem_unlock(obj);
 		ret = -ENOMEM;
 		goto fail;
 	}
 
 	ret = drm_prime_sg_to_page_addr_arrays(sgt, msm_obj->pages, NULL, npages);
 	if (ret) {
-		mutex_unlock(&msm_obj->lock);
+		msm_gem_unlock(obj);
 		goto fail;
 	}
 
-	mutex_unlock(&msm_obj->lock);
+	msm_gem_unlock(obj);
 
 	mutex_lock(&dev->struct_mutex);
 	list_add_tail(&msm_obj->mm_list, &priv->inactive_list);
drivers/gpu/drm/msm/msm_gem.h +28 -0
index a1bf741b9b89..f6482154e8bb 100644
@@ -93,6 +93,34 @@ struct msm_gem_object {
 };
 #define to_msm_bo(x) container_of(x, struct msm_gem_object, base)
 
+static inline void
+msm_gem_lock(struct drm_gem_object *obj)
+{
+	struct msm_gem_object *msm_obj = to_msm_bo(obj);
+	mutex_lock(&msm_obj->lock);
+}
+
+static inline int
+msm_gem_lock_interruptible(struct drm_gem_object *obj)
+{
+	struct msm_gem_object *msm_obj = to_msm_bo(obj);
+	return mutex_lock_interruptible(&msm_obj->lock);
+}
+
+static inline void
+msm_gem_unlock(struct drm_gem_object *obj)
+{
+	struct msm_gem_object *msm_obj = to_msm_bo(obj);
+	mutex_unlock(&msm_obj->lock);
+}
+
+static inline bool
+msm_gem_is_locked(struct drm_gem_object *obj)
+{
+	struct msm_gem_object *msm_obj = to_msm_bo(obj);
+	return mutex_is_locked(&msm_obj->lock);
+}
+
 static inline bool is_active(struct msm_gem_object *msm_obj)
 {
 	return atomic_read(&msm_obj->active_count);
Signature · 6 lines
2.26.2

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
author
#3 lore raw

PATCH v4 3/23 drm/msm/gem: Rename internal get_iova_locked helper

In reply to Rob Clark

From: Rob Clark <[email protected]>

We'll need to introduce a _locked() version of msm_gem_get_iova(), so we need to make that name available.

---
 drivers/gpu/drm/msm/msm_gem.c | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)
drivers/gpu/drm/msm/msm_gem.c +3 -3
index afef9c6b1a1c..dec89fe79025 100644
@@ -376,7 +376,7 @@ put_iova(struct drm_gem_object *obj)
 	}
 }
 
-static int msm_gem_get_iova_locked(struct drm_gem_object *obj,
+static int get_iova_locked(struct drm_gem_object *obj,
 		struct msm_gem_address_space *aspace, uint64_t *iova,
 		u64 range_start, u64 range_end)
 {
@@ -448,7 +448,7 @@ int msm_gem_get_and_pin_iova_range(struct drm_gem_object *obj,
 
 	msm_gem_lock(obj);
 
-	ret = msm_gem_get_iova_locked(obj, aspace, &local,
+	ret = get_iova_locked(obj, aspace, &local,
 		range_start, range_end);
 
 	if (!ret)
@@ -478,7 +478,7 @@ int msm_gem_get_iova(struct drm_gem_object *obj,
 	int ret;
 
 	msm_gem_lock(obj);
-	ret = msm_gem_get_iova_locked(obj, aspace, iova, 0, U64_MAX);
+	ret = get_iova_locked(obj, aspace, iova, 0, U64_MAX);
 	msm_gem_unlock(obj);
 
 	return ret;
Signature · 6 lines
2.26.2

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
author
#4 lore raw

PATCH v4 4/23 drm/msm/gem: Move prototypes to msm_gem.h

In reply to Rob Clark

From: Rob Clark <[email protected]>

---
 drivers/gpu/drm/msm/disp/mdp4/mdp4_crtc.c |  1 +
 drivers/gpu/drm/msm/disp/mdp5/mdp5_crtc.c |  1 +
 drivers/gpu/drm/msm/dsi/dsi_host.c        |  1 +
 drivers/gpu/drm/msm/msm_drv.h             | 54 ----------------------
 drivers/gpu/drm/msm/msm_fbdev.c           |  1 +
 drivers/gpu/drm/msm/msm_gem.h             | 56 +++++++++++++++++++++++
 6 files changed, 60 insertions(+), 54 deletions(-)
drivers/gpu/drm/msm/disp/mdp4/mdp4_crtc.c +1 -0
index a0253297bc76..b65b2329cc8d 100644
@@ -11,6 +11,7 @@
 #include <drm/drm_vblank.h>
 
 #include "mdp4_kms.h"
+#include "msm_gem.h"
 
 struct mdp4_crtc {
 	struct drm_crtc base;
drivers/gpu/drm/msm/disp/mdp5/mdp5_crtc.c +1 -0
index c39dad151bb6..81fbd52ad7e7 100644
@@ -15,6 +15,7 @@
 #include <drm/drm_vblank.h>
 
 #include "mdp5_kms.h"
+#include "msm_gem.h"
 
 #define CURSOR_WIDTH	64
 #define CURSOR_HEIGHT	64
drivers/gpu/drm/msm/dsi/dsi_host.c +1 -0
index b17ac6c27554..5e7cdc11c764 100644
@@ -26,6 +26,7 @@
 #include "sfpb.xml.h"
 #include "dsi_cfg.h"
 #include "msm_kms.h"
+#include "msm_gem.h"
 
 #define DSI_RESET_TOGGLE_DELAY_MS 20
 
drivers/gpu/drm/msm/msm_drv.h +0 -54

74 lines not on this page · raw message

drivers/gpu/drm/msm/msm_fbdev.c +1 -0
index 47235f8c5922..678dba1725a6 100644
@@ -9,6 +9,7 @@
 #include <drm/drm_fourcc.h>
 
 #include "msm_drv.h"
+#include "msm_gem.h"
 #include "msm_kms.h"
 
 extern int msm_gem_mmap_obj(struct drm_gem_object *obj,
drivers/gpu/drm/msm/msm_gem.h +56 -0

62 lines not on this page · raw message

Signature · 6 lines
2.26.2

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
author
#5 lore raw

PATCH v4 5/23 drm/msm/gem: Add some _locked() helpers

In reply to Rob Clark

From: Rob Clark <[email protected]>

When we cut-over to using dma_resv_lock/etc instead of msm_obj->lock, we'll need these for the submit path (where resv->lock is already held).

---
 drivers/gpu/drm/msm/msm_gem.c | 89 +++++++++++++++++++++++++++--------
 drivers/gpu/drm/msm/msm_gem.h |  6 +++
 2 files changed, 75 insertions(+), 20 deletions(-)
drivers/gpu/drm/msm/msm_gem.c +69 -20
index dec89fe79025..e0d8d739b068 100644
@@ -435,18 +435,14 @@ static int msm_gem_pin_iova(struct drm_gem_object *obj,
 			msm_obj->sgt, obj->size >> PAGE_SHIFT);
 }
 
-/*
- * get iova and pin it. Should have a matching put
- * limits iova to specified range (in pages)
- */
-int msm_gem_get_and_pin_iova_range(struct drm_gem_object *obj,
+static int get_and_pin_iova_range_locked(struct drm_gem_object *obj,
 		struct msm_gem_address_space *aspace, uint64_t *iova,
 		u64 range_start, u64 range_end)
 {
 	u64 local;
 	int ret;
 
-	msm_gem_lock(obj);
+	WARN_ON(!msm_gem_is_locked(obj));
 
 	ret = get_iova_locked(obj, aspace, &local,
 		range_start, range_end);
@@ -457,10 +453,32 @@ int msm_gem_get_and_pin_iova_range(struct drm_gem_object *obj,
 	if (!ret)
 		*iova = local;
 
+	return ret;
+}
+
+/*
+ * get iova and pin it. Should have a matching put
+ * limits iova to specified range (in pages)
+ */
+int msm_gem_get_and_pin_iova_range(struct drm_gem_object *obj,
+		struct msm_gem_address_space *aspace, uint64_t *iova,
+		u64 range_start, u64 range_end)
+{
+	int ret;
+
+	msm_gem_lock(obj);
+	ret = get_and_pin_iova_range_locked(obj, aspace, iova, range_start, range_end);
 	msm_gem_unlock(obj);
+
 	return ret;
 }
 
+int msm_gem_get_and_pin_iova_locked(struct drm_gem_object *obj,
+		struct msm_gem_address_space *aspace, uint64_t *iova)
+{
+	return get_and_pin_iova_range_locked(obj, aspace, iova, 0, U64_MAX);
+}
+
 /* get iova and pin it. Should have a matching put */
 int msm_gem_get_and_pin_iova(struct drm_gem_object *obj,
 		struct msm_gem_address_space *aspace, uint64_t *iova)
@@ -501,21 +519,31 @@ uint64_t msm_gem_iova(struct drm_gem_object *obj,
 }
 
 /*
- * Unpin a iova by updating the reference counts. The memory isn't actually
- * purged until something else (shrinker, mm_notifier, destroy, etc) decides
- * to get rid of it
+ * Locked variant of msm_gem_unpin_iova()
  */
-void msm_gem_unpin_iova(struct drm_gem_object *obj,
+void msm_gem_unpin_iova_locked(struct drm_gem_object *obj,
 		struct msm_gem_address_space *aspace)
 {
 	struct msm_gem_vma *vma;
 
-	msm_gem_lock(obj);
+	WARN_ON(!msm_gem_is_locked(obj));
+
 	vma = lookup_vma(obj, aspace);
 
 	if (!WARN_ON(!vma))
 		msm_gem_unmap_vma(aspace, vma);
+}
 
+/*
+ * Unpin a iova by updating the reference counts. The memory isn't actually
+ * purged until something else (shrinker, mm_notifier, destroy, etc) decides
+ * to get rid of it
+ */
+void msm_gem_unpin_iova(struct drm_gem_object *obj,
+		struct msm_gem_address_space *aspace)
+{
+	msm_gem_lock(obj);
+	msm_gem_unpin_iova_locked(obj, aspace);
 	msm_gem_unlock(obj);
 }
 
@@ -554,15 +582,14 @@ static void *get_vaddr(struct drm_gem_object *obj, unsigned madv)
 	struct msm_gem_object *msm_obj = to_msm_bo(obj);
 	int ret = 0;
 
+	WARN_ON(!msm_gem_is_locked(obj));
+
 	if (obj->import_attach)
 		return ERR_PTR(-ENODEV);
 
-	msm_gem_lock(obj);
-
 	if (WARN_ON(msm_obj->madv > madv)) {
 		DRM_DEV_ERROR(obj->dev->dev, "Invalid madv state: %u vs %u\n",
 			msm_obj->madv, madv);
-		msm_gem_unlock(obj);
 		return ERR_PTR(-EBUSY);
 	}
 
@@ -588,20 +615,29 @@ static void *get_vaddr(struct drm_gem_object *obj, unsigned madv)
 		}
 	}
 
-	msm_gem_unlock(obj);
 	return msm_obj->vaddr;
 
 fail:
 	msm_obj->vmap_count--;
-	msm_gem_unlock(obj);
 	return ERR_PTR(ret);
 }
 
-void *msm_gem_get_vaddr(struct drm_gem_object *obj)
+void *msm_gem_get_vaddr_locked(struct drm_gem_object *obj)
 {
 	return get_vaddr(obj, MSM_MADV_WILLNEED);
 }
 
+void *msm_gem_get_vaddr(struct drm_gem_object *obj)
+{
+	void *ret;
+
+	msm_gem_lock(obj);
+	ret = msm_gem_get_vaddr_locked(obj);
+	msm_gem_unlock(obj);
+
+	return ret;
+}
+
 /*
  * Don't use this!  It is for the very special case of dumping
  * submits from GPU hangs or faults, were the bo may already
@@ -610,16 +646,29 @@ void *msm_gem_get_vaddr(struct drm_gem_object *obj)
  */
 void *msm_gem_get_vaddr_active(struct drm_gem_object *obj)
 {
-	return get_vaddr(obj, __MSM_MADV_PURGED);
+	void *ret;
+
+	msm_gem_lock(obj);
+	ret = get_vaddr(obj, __MSM_MADV_PURGED);
+	msm_gem_unlock(obj);
+
+	return ret;
 }
 
-void msm_gem_put_vaddr(struct drm_gem_object *obj)
+void msm_gem_put_vaddr_locked(struct drm_gem_object *obj)
 {
 	struct msm_gem_object *msm_obj = to_msm_bo(obj);
 
-	msm_gem_lock(obj);
+	WARN_ON(!msm_gem_is_locked(obj));
 	WARN_ON(msm_obj->vmap_count < 1);
+
 	msm_obj->vmap_count--;
+}
+
+void msm_gem_put_vaddr(struct drm_gem_object *obj)
+{
+	msm_gem_lock(obj);
+	msm_gem_put_vaddr_locked(obj);
 	msm_gem_unlock(obj);
 }
 
drivers/gpu/drm/msm/msm_gem.h +6 -0
index fbad08badf43..d55d5401a2d2 100644
@@ -103,10 +103,14 @@ int msm_gem_get_iova(struct drm_gem_object *obj,
 int msm_gem_get_and_pin_iova_range(struct drm_gem_object *obj,
 		struct msm_gem_address_space *aspace, uint64_t *iova,
 		u64 range_start, u64 range_end);
+int msm_gem_get_and_pin_iova_locked(struct drm_gem_object *obj,
+		struct msm_gem_address_space *aspace, uint64_t *iova);
 int msm_gem_get_and_pin_iova(struct drm_gem_object *obj,
 		struct msm_gem_address_space *aspace, uint64_t *iova);
 uint64_t msm_gem_iova(struct drm_gem_object *obj,
 		struct msm_gem_address_space *aspace);
+void msm_gem_unpin_iova_locked(struct drm_gem_object *obj,
+		struct msm_gem_address_space *aspace);
 void msm_gem_unpin_iova(struct drm_gem_object *obj,
 		struct msm_gem_address_space *aspace);
 struct page **msm_gem_get_pages(struct drm_gem_object *obj);
@@ -115,8 +119,10 @@ int msm_gem_dumb_create(struct drm_file *file, struct drm_device *dev,
 		struct drm_mode_create_dumb *args);
 int msm_gem_dumb_map_offset(struct drm_file *file, struct drm_device *dev,
 		uint32_t handle, uint64_t *offset);
+void *msm_gem_get_vaddr_locked(struct drm_gem_object *obj);
 void *msm_gem_get_vaddr(struct drm_gem_object *obj);
 void *msm_gem_get_vaddr_active(struct drm_gem_object *obj);
+void msm_gem_put_vaddr_locked(struct drm_gem_object *obj);
 void msm_gem_put_vaddr(struct drm_gem_object *obj);
 int msm_gem_madvise(struct drm_gem_object *obj, unsigned madv);
 int msm_gem_sync_object(struct drm_gem_object *obj,
Signature · 6 lines
2.26.2

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
author
#6 lore raw

PATCH v4 6/23 drm/msm/gem: Move locking in shrinker path

In reply to Rob Clark

From: Rob Clark <[email protected]>

Move grabbing the bo lock into shrinker, with a msm_gem_trylock() to skip over bo's that are already locked. This gets rid of the nested lock classes.

---
 drivers/gpu/drm/msm/msm_gem.c          | 24 +++++----------------
 drivers/gpu/drm/msm/msm_gem.h          | 29 ++++++++++----------------
 drivers/gpu/drm/msm/msm_gem_shrinker.c | 27 +++++++++++++++++-------
 3 files changed, 35 insertions(+), 45 deletions(-)
drivers/gpu/drm/msm/msm_gem.c +5 -19
index e0d8d739b068..1195847714ba 100644
@@ -17,8 +17,6 @@
 #include "msm_gpu.h"
 #include "msm_mmu.h"
 
-static void msm_gem_vunmap_locked(struct drm_gem_object *obj);
-
 
 static dma_addr_t physaddr(struct drm_gem_object *obj)
 {
@@ -693,20 +691,19 @@ int msm_gem_madvise(struct drm_gem_object *obj, unsigned madv)
 	return (madv != __MSM_MADV_PURGED);
 }
 
-void msm_gem_purge(struct drm_gem_object *obj, enum msm_gem_lock subclass)
+void msm_gem_purge(struct drm_gem_object *obj)
 {
 	struct drm_device *dev = obj->dev;
 	struct msm_gem_object *msm_obj = to_msm_bo(obj);
 
 	WARN_ON(!mutex_is_locked(&dev->struct_mutex));
+	WARN_ON(!msm_gem_is_locked(obj));
 	WARN_ON(!is_purgeable(msm_obj));
 	WARN_ON(obj->import_attach);
 
-	mutex_lock_nested(&msm_obj->lock, subclass);
-
 	put_iova(obj);
 
-	msm_gem_vunmap_locked(obj);
+	msm_gem_vunmap(obj);
 
 	put_pages(obj);
 
@@ -724,11 +721,9 @@ void msm_gem_purge(struct drm_gem_object *obj, enum msm_gem_lock subclass)
 
 	invalidate_mapping_pages(file_inode(obj->filp)->i_mapping,
 			0, (loff_t)-1);
-
-	msm_gem_unlock(obj);
 }
 
-static void msm_gem_vunmap_locked(struct drm_gem_object *obj)
+void msm_gem_vunmap(struct drm_gem_object *obj)
 {
 	struct msm_gem_object *msm_obj = to_msm_bo(obj);
 
@@ -741,15 +736,6 @@ static void msm_gem_vunmap_locked(struct drm_gem_object *obj)
 	msm_obj->vaddr = NULL;
 }
 
-void msm_gem_vunmap(struct drm_gem_object *obj, enum msm_gem_lock subclass)
-{
-	struct msm_gem_object *msm_obj = to_msm_bo(obj);
-
-	mutex_lock_nested(&msm_obj->lock, subclass);
-	msm_gem_vunmap_locked(obj);
-	msm_gem_unlock(obj);
-}
-
 /* must be called before _move_to_active().. */
 int msm_gem_sync_object(struct drm_gem_object *obj,
 		struct msm_fence_context *fctx, bool exclusive)
@@ -986,7 +972,7 @@ static void free_object(struct msm_gem_object *msm_obj)
 
 		drm_prime_gem_destroy(obj, msm_obj->sgt);
 	} else {
-		msm_gem_vunmap_locked(obj);
+		msm_gem_vunmap(obj);
 		put_pages(obj);
 	}
 
drivers/gpu/drm/msm/msm_gem.h +11 -18
index d55d5401a2d2..766ce278c74a 100644
@@ -162,6 +162,13 @@ msm_gem_lock(struct drm_gem_object *obj)
 	mutex_lock(&msm_obj->lock);
 }
 
+static inline bool __must_check
+msm_gem_trylock(struct drm_gem_object *obj)
+{
+	struct msm_gem_object *msm_obj = to_msm_bo(obj);
+	return mutex_trylock(&msm_obj->lock) == 1;
+}
+
 static inline int
 msm_gem_lock_interruptible(struct drm_gem_object *obj)
 {
@@ -190,6 +197,7 @@ static inline bool is_active(struct msm_gem_object *msm_obj)
 
 static inline bool is_purgeable(struct msm_gem_object *msm_obj)
 {
+	WARN_ON(!msm_gem_is_locked(&msm_obj->base));
 	WARN_ON(!mutex_is_locked(&msm_obj->base.dev->struct_mutex));
 	return (msm_obj->madv == MSM_MADV_DONTNEED) && msm_obj->sgt &&
 			!msm_obj->base.dma_buf && !msm_obj->base.import_attach;
@@ -197,27 +205,12 @@ static inline bool is_purgeable(struct msm_gem_object *msm_obj)
 
 static inline bool is_vunmapable(struct msm_gem_object *msm_obj)
 {
+	WARN_ON(!msm_gem_is_locked(&msm_obj->base));
 	return (msm_obj->vmap_count == 0) && msm_obj->vaddr;
 }
 
-/* The shrinker can be triggered while we hold objA->lock, and need
- * to grab objB->lock to purge it.  Lockdep just sees these as a single
- * class of lock, so we use subclasses to teach it the difference.
- *
- * OBJ_LOCK_NORMAL is implicit (ie. normal mutex_lock() call), and
- * OBJ_LOCK_SHRINKER is used by shrinker.
- *
- * It is *essential* that we never go down paths that could trigger the
- * shrinker for a purgable object.  This is ensured by checking that
- * msm_obj->madv == MSM_MADV_WILLNEED.
- */
-enum msm_gem_lock {
-	OBJ_LOCK_NORMAL,
-	OBJ_LOCK_SHRINKER,
-};
-
-void msm_gem_purge(struct drm_gem_object *obj, enum msm_gem_lock subclass);
-void msm_gem_vunmap(struct drm_gem_object *obj, enum msm_gem_lock subclass);
+void msm_gem_purge(struct drm_gem_object *obj);
+void msm_gem_vunmap(struct drm_gem_object *obj);
 void msm_gem_free_work(struct work_struct *work);
 
 /* Created per submit-ioctl, to track bo's and cmdstream bufs, etc,
drivers/gpu/drm/msm/msm_gem_shrinker.c +19 -8
index 482576d7a39a..2dc0ffa925b4 100644
@@ -52,8 +52,11 @@ msm_gem_shrinker_count(struct shrinker *shrinker, struct shrink_control *sc)
 		return 0;
 
 	list_for_each_entry(msm_obj, &priv->inactive_list, mm_list) {
+		if (!msm_gem_trylock(&msm_obj->base))
+			continue;
 		if (is_purgeable(msm_obj))
 			count += msm_obj->base.size >> PAGE_SHIFT;
+		msm_gem_unlock(&msm_obj->base);
 	}
 
 	if (unlock)
@@ -78,10 +81,13 @@ msm_gem_shrinker_scan(struct shrinker *shrinker, struct shrink_control *sc)
 	list_for_each_entry(msm_obj, &priv->inactive_list, mm_list) {
 		if (freed >= sc->nr_to_scan)
 			break;
+		if (!msm_gem_trylock(&msm_obj->base))
+			continue;
 		if (is_purgeable(msm_obj)) {
-			msm_gem_purge(&msm_obj->base, OBJ_LOCK_SHRINKER);
+			msm_gem_purge(&msm_obj->base);
 			freed += msm_obj->base.size >> PAGE_SHIFT;
 		}
+		msm_gem_unlock(&msm_obj->base);
 	}
 
 	if (unlock)
@@ -107,15 +113,20 @@ msm_gem_shrinker_vmap(struct notifier_block *nb, unsigned long event, void *ptr)
 		return NOTIFY_DONE;
 
 	list_for_each_entry(msm_obj, &priv->inactive_list, mm_list) {
+		if (!msm_gem_trylock(&msm_obj->base))
+			continue;
 		if (is_vunmapable(msm_obj)) {
-			msm_gem_vunmap(&msm_obj->base, OBJ_LOCK_SHRINKER);
-			/* since we don't know any better, lets bail after a few
-			 * and if necessary the shrinker will be invoked again.
-			 * Seems better than unmapping *everything*
-			 */
-			if (++unmapped >= 15)
-				break;
+			msm_gem_vunmap(&msm_obj->base);
+			unmapped++;
 		}
+		msm_gem_unlock(&msm_obj->base);
+
+		/* since we don't know any better, lets bail after a few
+		 * and if necessary the shrinker will be invoked again.
+		 * Seems better than unmapping *everything*
+		 */
+		if (++unmapped >= 15)
+			break;
 	}
 
 	if (unlock)
Signature · 6 lines
2.26.2

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
author
#7 lore raw

PATCH v4 7/23 drm/msm/submit: Move copy_from_user ahead of locking bos

In reply to Rob Clark

From: Rob Clark <[email protected]>

We cannot switch to using obj->resv for locking without first moving all the copy_from_user() ahead of submit_lock_objects(). Otherwise in the mm fault path we aquire mm->mmap_sem before obj lock, but in the submit path the order is reversed.

---
 drivers/gpu/drm/msm/msm_gem.h        |   3 +
 drivers/gpu/drm/msm/msm_gem_submit.c | 127 +++++++++++++++++----------
 2 files changed, 82 insertions(+), 48 deletions(-)
drivers/gpu/drm/msm/msm_gem.h +3 -0
index 766ce278c74a..1f4e5d871a23 100644
@@ -240,7 +240,10 @@ struct msm_gem_submit {
 		uint32_t type;
 		uint32_t size;  /* in dwords */
 		uint64_t iova;
+		uint32_t offset;/* in dwords */
 		uint32_t idx;   /* cmdstream buffer idx in bos[] */
+		uint32_t nr_relocs;
+		struct drm_msm_gem_submit_reloc *relocs;
 	} *cmd;  /* array of size nr_cmds */
 	struct {
 		uint32_t flags;
drivers/gpu/drm/msm/msm_gem_submit.c +79 -48

187 lines not on this page · raw message

Signature · 6 lines
2.26.2

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
author
#9 lore raw

PATCH v4 8/23 drm/msm: Do rpm get sooner in the submit path

In reply to Rob Clark

From: Rob Clark <[email protected]>

Unfortunately, due to an dev_pm_opp locking interaction with mm->mmap_sem, we need to do pm get before aquiring obj locks, otherwise we can have anger lockdep with the chain:

opp_table_lock --> &mm->mmap_sem --> reservation_ww_class_mutex

For an explicit fencing userspace, the impact should be minimal as we do all the fence waits before this point. It could result in some needless resumes in error cases, etc.

---
 drivers/gpu/drm/msm/msm_gem_submit.c | 15 +++++++++++++--
 1 file changed, 13 insertions(+), 2 deletions(-)
drivers/gpu/drm/msm/msm_gem_submit.c +13 -2
index b6c258c89290..aa3c7af54079 100644
@@ -750,11 +750,20 @@ int msm_ioctl_gem_submit(struct drm_device *dev, void *data,
 
 	ret = submit_lookup_objects(submit, args, file);
 	if (ret)
-		goto out;
+		goto out_pre_pm;
 
 	ret = submit_lookup_cmds(submit, args, file);
 	if (ret)
-		goto out;
+		goto out_pre_pm;
+
+	/*
+	 * Thanks to dev_pm_opp opp_table_lock interactions with mm->mmap_sem
+	 * in the resume path, we need to to rpm get before we lock objs.
+	 * Which unfortunately might involve powering up the GPU sooner than
+	 * is necessary.  But at least in the explicit fencing case, we will
+	 * have already done all the fence waiting.
+	 */
+	pm_runtime_get_sync(&gpu->pdev->dev);
 
 	/* copy_*_user while holding a ww ticket upsets lockdep */
 	ww_acquire_init(&submit->ticket, &reservation_ww_class);
@@ -831,6 +840,8 @@ int msm_ioctl_gem_submit(struct drm_device *dev, void *data,
 
 
 out:
+	pm_runtime_put(&gpu->pdev->dev);
+out_pre_pm:
 	submit_cleanup(submit);
 	if (has_ww_ticket)
 		ww_acquire_fini(&submit->ticket);
Signature · 6 lines
2.26.2

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
author
#10 lore raw

PATCH v4 9/23 drm/msm/gem: Switch over to obj->resv for locking

In reply to Rob Clark

From: Rob Clark <[email protected]>

This also converts the special msm_gem_get_vaddr_active() to expect the lock to already be held. There are two call-sites for this, one already has the lock held, so it is more straightforward to just open-code the locking for the other caller.

---
 drivers/gpu/drm/msm/msm_gem.c        | 17 +++++++----------
 drivers/gpu/drm/msm/msm_gem.h        | 16 +++++-----------
 drivers/gpu/drm/msm/msm_gem_submit.c |  8 ++++----
 drivers/gpu/drm/msm/msm_gpu.c        | 14 ++++++++++++--
 drivers/gpu/drm/msm/msm_rd.c         |  2 +-
 5 files changed, 29 insertions(+), 28 deletions(-)
drivers/gpu/drm/msm/msm_gem.c +7 -10
index 1195847714ba..17afa627ea3d 100644
@@ -644,13 +644,7 @@ void *msm_gem_get_vaddr(struct drm_gem_object *obj)
  */
 void *msm_gem_get_vaddr_active(struct drm_gem_object *obj)
 {
-	void *ret;
-
-	msm_gem_lock(obj);
-	ret = get_vaddr(obj, __MSM_MADV_PURGED);
-	msm_gem_unlock(obj);
-
-	return ret;
+	return get_vaddr(obj, __MSM_MADV_PURGED);
 }
 
 void msm_gem_put_vaddr_locked(struct drm_gem_object *obj)
@@ -970,15 +964,20 @@ static void free_object(struct msm_gem_object *msm_obj)
 		if (msm_obj->pages)
 			kvfree(msm_obj->pages);
 
+		/* 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);
+		msm_gem_unlock(obj);
 	}
 
 	drm_gem_object_release(obj);
 
-	msm_gem_unlock(obj);
 	kfree(msm_obj);
 }
 
@@ -1050,8 +1049,6 @@ static int msm_gem_new_impl(struct drm_device *dev,
 	if (!msm_obj)
 		return -ENOMEM;
 
-	mutex_init(&msm_obj->lock);
-
 	msm_obj->flags = flags;
 	msm_obj->madv = MSM_MADV_WILLNEED;
 
drivers/gpu/drm/msm/msm_gem.h +5 -11
index 1f4e5d871a23..f0608d96ef03 100644
@@ -85,7 +85,6 @@ struct msm_gem_object {
 	 * an IOMMU.  Also used for stolen/splashscreen buffer.
 	 */
 	struct drm_mm_node *vram_node;
-	struct mutex lock; /* Protects resources associated with bo */
 
 	char name[32]; /* Identifier to print for the debugfs files */
 
@@ -158,36 +157,31 @@ void msm_gem_describe_objects(struct list_head *list, struct seq_file *m);
 static inline void
 msm_gem_lock(struct drm_gem_object *obj)
 {
-	struct msm_gem_object *msm_obj = to_msm_bo(obj);
-	mutex_lock(&msm_obj->lock);
+	dma_resv_lock(obj->resv, NULL);
 }
 
 static inline bool __must_check
 msm_gem_trylock(struct drm_gem_object *obj)
 {
-	struct msm_gem_object *msm_obj = to_msm_bo(obj);
-	return mutex_trylock(&msm_obj->lock) == 1;
+	return dma_resv_trylock(obj->resv);
 }
 
 static inline int
 msm_gem_lock_interruptible(struct drm_gem_object *obj)
 {
-	struct msm_gem_object *msm_obj = to_msm_bo(obj);
-	return mutex_lock_interruptible(&msm_obj->lock);
+	return dma_resv_lock_interruptible(obj->resv, NULL);
 }
 
 static inline void
 msm_gem_unlock(struct drm_gem_object *obj)
 {
-	struct msm_gem_object *msm_obj = to_msm_bo(obj);
-	mutex_unlock(&msm_obj->lock);
+	dma_resv_unlock(obj->resv);
 }
 
 static inline bool
 msm_gem_is_locked(struct drm_gem_object *obj)
 {
-	struct msm_gem_object *msm_obj = to_msm_bo(obj);
-	return mutex_is_locked(&msm_obj->lock);
+	return dma_resv_is_locked(obj->resv);
 }
 
 static inline bool is_active(struct msm_gem_object *msm_obj)
drivers/gpu/drm/msm/msm_gem_submit.c +4 -4
index aa3c7af54079..044e9bee70a2 100644
@@ -221,7 +221,7 @@ static void submit_unlock_unpin_bo(struct msm_gem_submit *submit,
 	struct msm_gem_object *msm_obj = submit->bos[i].obj;
 
 	if (submit->bos[i].flags & BO_PINNED)
-		msm_gem_unpin_iova(&msm_obj->base, submit->aspace);
+		msm_gem_unpin_iova_locked(&msm_obj->base, submit->aspace);
 
 	if (submit->bos[i].flags & BO_LOCKED)
 		dma_resv_unlock(msm_obj->base.resv);
@@ -324,7 +324,7 @@ static int submit_pin_objects(struct msm_gem_submit *submit)
 		uint64_t iova;
 
 		/* if locking succeeded, pin bo: */
-		ret = msm_gem_get_and_pin_iova(&msm_obj->base,
+		ret = msm_gem_get_and_pin_iova_locked(&msm_obj->base,
 				submit->aspace, &iova);
 
 		if (ret)
@@ -383,7 +383,7 @@ static int submit_reloc(struct msm_gem_submit *submit, struct msm_gem_object *ob
 	/* For now, just map the entire thing.  Eventually we probably
 	 * to do it page-by-page, w/ kmap() if not vmap()d..
 	 */
-	ptr = msm_gem_get_vaddr(&obj->base);
+	ptr = msm_gem_get_vaddr_locked(&obj->base);
 
 	if (IS_ERR(ptr)) {
 		ret = PTR_ERR(ptr);
@@ -434,7 +434,7 @@ static int submit_reloc(struct msm_gem_submit *submit, struct msm_gem_object *ob
 	}
 
 out:
-	msm_gem_put_vaddr(&obj->base);
+	msm_gem_put_vaddr_locked(&obj->base);
 
 	return ret;
 }
drivers/gpu/drm/msm/msm_gpu.c +12 -2
index 55d16489d0f3..015f6b884e2e 100644
@@ -326,7 +326,9 @@ static void msm_gpu_crashstate_get_bo(struct msm_gpu_state *state,
 		if (!state_bo->data)
 			goto out;
 
+		msm_gem_lock(&obj->base);
 		ptr = msm_gem_get_vaddr_active(&obj->base);
+		msm_gem_unlock(&obj->base);
 		if (IS_ERR(ptr)) {
 			kvfree(state_bo->data);
 			state_bo->data = NULL;
@@ -470,14 +472,22 @@ static void recover_worker(struct work_struct *work)
 			put_task_struct(task);
 		}
 
+		/* msm_rd_dump_submit() needs bo locked to dump: */
+		for (i = 0; i < submit->nr_bos; i++)
+			msm_gem_lock(&submit->bos[i].obj->base);
+
 		if (comm && cmd) {
 			DRM_DEV_ERROR(dev->dev, "%s: offending task: %s (%s)\n",
 				gpu->name, comm, cmd);
 
 			msm_rd_dump_submit(priv->hangrd, submit,
 				"offending task: %s (%s)", comm, cmd);
-		} else
+		} else {
 			msm_rd_dump_submit(priv->hangrd, submit, NULL);
+		}
+
+		for (i = 0; i < submit->nr_bos; i++)
+			msm_gem_unlock(&submit->bos[i].obj->base);
 	}
 
 	/* Record the crash state */
@@ -784,7 +794,7 @@ void msm_gpu_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit)
 
 		/* submit takes a reference to the bo and iova until retired: */
 		drm_gem_object_get(&msm_obj->base);
-		msm_gem_get_and_pin_iova(&msm_obj->base, submit->aspace, &iova);
+		msm_gem_get_and_pin_iova_locked(&msm_obj->base, submit->aspace, &iova);
 
 		if (submit->bos[i].flags & MSM_SUBMIT_BO_WRITE)
 			dma_resv_add_excl_fence(drm_obj->resv, submit->fence);
drivers/gpu/drm/msm/msm_rd.c +1 -1
index fea30e7aa9e8..659e5cc4b40a 100644
@@ -333,7 +333,7 @@ static void snapshot_buf(struct msm_rd_state *rd,
 
 	rd_write_section(rd, RD_BUFFER_CONTENTS, buf, size);
 
-	msm_gem_put_vaddr(&obj->base);
+	msm_gem_put_vaddr_locked(&obj->base);
 }
 
 /* called under struct_mutex */
Signature · 1 line
2.26.2
author
#11 lore raw

PATCH v4 10/23 drm/msm: Use correct drm_gem_object_put() in fail case

In reply to Rob Clark

From: Rob Clark <[email protected]>

We only want to use the _unlocked() variant in the unlocked case.

---
 drivers/gpu/drm/msm/msm_gem.c | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)
drivers/gpu/drm/msm/msm_gem.c +5 -1
index 17afa627ea3d..992cda7e4995 100644
@@ -1140,7 +1140,11 @@ static struct drm_gem_object *_msm_gem_new(struct drm_device *dev,
 	return obj;
 
 fail:
-	drm_gem_object_put(obj);
+	if (struct_mutex_locked) {
+		drm_gem_object_put_locked(obj);
+	} else {
+		drm_gem_object_put(obj);
+	}
 	return ERR_PTR(ret);
 }
 
Signature · 6 lines
2.26.2

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
author
#12 lore raw

PATCH v4 11/23 drm/msm: Drop chatty trace

In reply to Rob Clark

From: Rob Clark <[email protected]>

It is somewhat redundant with the gpu tracepoints, and anyways not too useful to justify spamming the log when debug traces are enabled.

---
 drivers/gpu/drm/msm/msm_gpu.c | 1 -
 1 file changed, 1 deletion(-)
drivers/gpu/drm/msm/msm_gpu.c +0 -1
index 015f6b884e2e..ed6645aa0ae5 100644
@@ -545,7 +545,6 @@ static void recover_worker(struct work_struct *work)
 
 static void hangcheck_timer_reset(struct msm_gpu *gpu)
 {
-	DBG("%s", gpu->name);
 	mod_timer(&gpu->hangcheck_timer,
 			round_jiffies_up(jiffies + DRM_MSM_HANGCHECK_JIFFIES));
 }
Signature · 1 line
2.26.2
author
#13 lore raw

PATCH v4 12/23 drm/msm: Move update_fences()

In reply to Rob Clark

From: Rob Clark <[email protected]>

Small cleanup, update_fences() is used in the hangcheck path, but also in the normal retire path.

---
 drivers/gpu/drm/msm/msm_gpu.c | 28 ++++++++++++++--------------
 1 file changed, 14 insertions(+), 14 deletions(-)
drivers/gpu/drm/msm/msm_gpu.c +14 -14
index ed6645aa0ae5..1667d8066897 100644
@@ -265,6 +265,20 @@ int msm_gpu_hw_init(struct msm_gpu *gpu)
 	return ret;
 }
 
+static void update_fences(struct msm_gpu *gpu, struct msm_ringbuffer *ring,
+		uint32_t fence)
+{
+	struct msm_gem_submit *submit;
+
+	list_for_each_entry(submit, &ring->submits, node) {
+		if (submit->seqno > fence)
+			break;
+
+		msm_update_fence(submit->ring->fctx,
+			submit->fence->seqno);
+	}
+}
+
 #ifdef CONFIG_DEV_COREDUMP
 static ssize_t msm_gpu_devcoredump_read(char *buffer, loff_t offset,
 		size_t count, void *data, size_t datalen)
@@ -413,20 +427,6 @@ static void msm_gpu_crashstate_capture(struct msm_gpu *gpu,
  * Hangcheck detection for locked gpu:
  */
 
-static void update_fences(struct msm_gpu *gpu, struct msm_ringbuffer *ring,
-		uint32_t fence)
-{
-	struct msm_gem_submit *submit;
-
-	list_for_each_entry(submit, &ring->submits, node) {
-		if (submit->seqno > fence)
-			break;
-
-		msm_update_fence(submit->ring->fctx,
-			submit->fence->seqno);
-	}
-}
-
 static struct msm_gem_submit *
 find_submit(struct msm_ringbuffer *ring, uint32_t fence)
 {
Signature · 6 lines
2.26.2

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
author
#14 lore raw

PATCH v4 13/23 drm/msm: Add priv->mm_lock to protect active/inactive lists

In reply to Rob Clark

From: Rob Clark <[email protected]>

Rather than relying on the big dev->struct_mutex hammer, introduce a more specific lock for protecting the bo lists.

---
 drivers/gpu/drm/msm/msm_debugfs.c      |  7 +++++++
 drivers/gpu/drm/msm/msm_drv.c          |  7 +++++++
 drivers/gpu/drm/msm/msm_drv.h          | 13 +++++++++++-
 drivers/gpu/drm/msm/msm_gem.c          | 28 +++++++++++++++-----------
 drivers/gpu/drm/msm/msm_gem_shrinker.c | 12 +++++++++++
 drivers/gpu/drm/msm/msm_gpu.h          |  5 ++++-
 6 files changed, 58 insertions(+), 14 deletions(-)
drivers/gpu/drm/msm/msm_debugfs.c +7 -0
index ee2e270f464c..64afbed89821 100644
@@ -112,6 +112,11 @@ static int msm_gem_show(struct drm_device *dev, struct seq_file *m)
 {
 	struct msm_drm_private *priv = dev->dev_private;
 	struct msm_gpu *gpu = priv->gpu;
+	int ret;
+
+	ret = mutex_lock_interruptible(&priv->mm_lock);
+	if (ret)
+		return ret;
 
 	if (gpu) {
 		seq_printf(m, "Active Objects (%s):\n", gpu->name);
@@ -121,6 +126,8 @@ static int msm_gem_show(struct drm_device *dev, struct seq_file *m)
 	seq_printf(m, "Inactive Objects:\n");
 	msm_gem_describe_objects(&priv->inactive_list, m);
 
+	mutex_unlock(&priv->mm_lock);
+
 	return 0;
 }
 
drivers/gpu/drm/msm/msm_drv.c +7 -0
index 49685571dc0e..81cb2cecc829 100644
@@ -7,6 +7,7 @@
 
 #include <linux/dma-mapping.h>
 #include <linux/kthread.h>
+#include <linux/sched/mm.h>
 #include <linux/uaccess.h>
 #include <uapi/linux/sched/types.h>
 
@@ -441,6 +442,12 @@ static int msm_drm_init(struct device *dev, struct drm_driver *drv)
 	init_llist_head(&priv->free_list);
 
 	INIT_LIST_HEAD(&priv->inactive_list);
+	mutex_init(&priv->mm_lock);
+
+	/* Teach lockdep about lock ordering wrt. shrinker: */
+	fs_reclaim_acquire(GFP_KERNEL);
+	might_lock(&priv->mm_lock);
+	fs_reclaim_release(GFP_KERNEL);
 
 	drm_mode_config_init(ddev);
 
drivers/gpu/drm/msm/msm_drv.h +12 -1
index 79ee7d05b363..a17dadd38685 100644
@@ -174,8 +174,19 @@ struct msm_drm_private {
 	struct msm_rd_state *hangrd;   /* debugfs to dump hanging submits */
 	struct msm_perf_state *perf;
 
-	/* list of GEM objects: */
+	/*
+	 * List of inactive GEM objects.  Every bo is either in the inactive_list
+	 * or gpu->active_list (for the gpu it is active on[1])
+	 *
+	 * These lists are protected by mm_lock.  If struct_mutex is involved, it
+	 * should be aquired prior to mm_lock.  One should *not* hold mm_lock in
+	 * get_pages()/vmap()/etc paths, as they can trigger the shrinker.
+	 *
+	 * [1] if someone ever added support for the old 2d cores, there could be
+	 *     more than one gpu object
+	 */
 	struct list_head inactive_list;
+	struct mutex mm_lock;
 
 	/* worker for delayed free of objects: */
 	struct work_struct free_work;
drivers/gpu/drm/msm/msm_gem.c +16 -12
index 992cda7e4995..81ed03322a74 100644
@@ -768,13 +768,17 @@ int msm_gem_sync_object(struct drm_gem_object *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);
-	WARN_ON(!mutex_is_locked(&obj->dev->struct_mutex));
+	struct msm_drm_private *priv = obj->dev->dev_private;
+
+	might_sleep();
 	WARN_ON(msm_obj->madv != MSM_MADV_WILLNEED);
 
 	if (!atomic_fetch_inc(&msm_obj->active_count)) {
+		mutex_lock(&priv->mm_lock);
 		msm_obj->gpu = gpu;
 		list_del_init(&msm_obj->mm_list);
 		list_add_tail(&msm_obj->mm_list, &gpu->active_list);
+		mutex_unlock(&priv->mm_lock);
 	}
 }
 
@@ -783,12 +787,14 @@ void msm_gem_active_put(struct drm_gem_object *obj)
 	struct msm_gem_object *msm_obj = to_msm_bo(obj);
 	struct msm_drm_private *priv = obj->dev->dev_private;
 
-	WARN_ON(!mutex_is_locked(&obj->dev->struct_mutex));
+	might_sleep();
 
 	if (!atomic_dec_return(&msm_obj->active_count)) {
+		mutex_lock(&priv->mm_lock);
 		msm_obj->gpu = NULL;
 		list_del_init(&msm_obj->mm_list);
 		list_add_tail(&msm_obj->mm_list, &priv->inactive_list);
+		mutex_unlock(&priv->mm_lock);
 	}
 }
 
@@ -943,13 +949,16 @@ static void free_object(struct msm_gem_object *msm_obj)
 {
 	struct drm_gem_object *obj = &msm_obj->base;
 	struct drm_device *dev = obj->dev;
+	struct msm_drm_private *priv = dev->dev_private;
 
 	WARN_ON(!mutex_is_locked(&dev->struct_mutex));
 
 	/* object should not be on active list: */
 	WARN_ON(is_active(msm_obj));
 
+	mutex_lock(&priv->mm_lock);
 	list_del(&msm_obj->mm_list);
+	mutex_unlock(&priv->mm_lock);
 
 	msm_gem_lock(obj);
 
@@ -1128,14 +1137,9 @@ static struct drm_gem_object *_msm_gem_new(struct drm_device *dev,
 		mapping_set_gfp_mask(obj->filp->f_mapping, GFP_HIGHUSER);
 	}
 
-	if (struct_mutex_locked) {
-		WARN_ON(!mutex_is_locked(&dev->struct_mutex));
-		list_add_tail(&msm_obj->mm_list, &priv->inactive_list);
-	} else {
-		mutex_lock(&dev->struct_mutex);
-		list_add_tail(&msm_obj->mm_list, &priv->inactive_list);
-		mutex_unlock(&dev->struct_mutex);
-	}
+	mutex_lock(&priv->mm_lock);
+	list_add_tail(&msm_obj->mm_list, &priv->inactive_list);
+	mutex_unlock(&priv->mm_lock);
 
 	return obj;
 
@@ -1203,9 +1207,9 @@ struct drm_gem_object *msm_gem_import(struct drm_device *dev,
 
 	msm_gem_unlock(obj);
 
-	mutex_lock(&dev->struct_mutex);
+	mutex_lock(&priv->mm_lock);
 	list_add_tail(&msm_obj->mm_list, &priv->inactive_list);
-	mutex_unlock(&dev->struct_mutex);
+	mutex_unlock(&priv->mm_lock);
 
 	return obj;
 
drivers/gpu/drm/msm/msm_gem_shrinker.c +12 -0
index 2dc0ffa925b4..6be073b8ca08 100644
@@ -51,6 +51,8 @@ msm_gem_shrinker_count(struct shrinker *shrinker, struct shrink_control *sc)
 	if (!msm_gem_shrinker_lock(dev, &unlock))
 		return 0;
 
+	mutex_lock(&priv->mm_lock);
+
 	list_for_each_entry(msm_obj, &priv->inactive_list, mm_list) {
 		if (!msm_gem_trylock(&msm_obj->base))
 			continue;
@@ -59,6 +61,8 @@ msm_gem_shrinker_count(struct shrinker *shrinker, struct shrink_control *sc)
 		msm_gem_unlock(&msm_obj->base);
 	}
 
+	mutex_unlock(&priv->mm_lock);
+
 	if (unlock)
 		mutex_unlock(&dev->struct_mutex);
 
@@ -78,6 +82,8 @@ msm_gem_shrinker_scan(struct shrinker *shrinker, struct shrink_control *sc)
 	if (!msm_gem_shrinker_lock(dev, &unlock))
 		return SHRINK_STOP;
 
+	mutex_lock(&priv->mm_lock);
+
 	list_for_each_entry(msm_obj, &priv->inactive_list, mm_list) {
 		if (freed >= sc->nr_to_scan)
 			break;
@@ -90,6 +96,8 @@ msm_gem_shrinker_scan(struct shrinker *shrinker, struct shrink_control *sc)
 		msm_gem_unlock(&msm_obj->base);
 	}
 
+	mutex_unlock(&priv->mm_lock);
+
 	if (unlock)
 		mutex_unlock(&dev->struct_mutex);
 
@@ -112,6 +120,8 @@ msm_gem_shrinker_vmap(struct notifier_block *nb, unsigned long event, void *ptr)
 	if (!msm_gem_shrinker_lock(dev, &unlock))
 		return NOTIFY_DONE;
 
+	mutex_lock(&priv->mm_lock);
+
 	list_for_each_entry(msm_obj, &priv->inactive_list, mm_list) {
 		if (!msm_gem_trylock(&msm_obj->base))
 			continue;
@@ -129,6 +139,8 @@ msm_gem_shrinker_vmap(struct notifier_block *nb, unsigned long event, void *ptr)
 			break;
 	}
 
+	mutex_unlock(&priv->mm_lock);
+
 	if (unlock)
 		mutex_unlock(&dev->struct_mutex);
 
drivers/gpu/drm/msm/msm_gpu.h +4 -1
index 6c9e1fdc1a76..1806e87600c0 100644
@@ -94,7 +94,10 @@ struct msm_gpu {
 	struct msm_ringbuffer *rb[MSM_GPU_MAX_RINGS];
 	int nr_rings;
 
-	/* list of GEM active objects: */
+	/*
+	 * List of GEM active objects on this gpu.  Protected by
+	 * msm_drm_private::mm_lock
+	 */
 	struct list_head active_list;
 
 	/* does gpu need hw_init? */
Signature · 6 lines
2.26.2

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
author
#15 lore raw

PATCH v4 14/23 drm/msm: Document and rename preempt_lock

In reply to Rob Clark

From: Rob Clark <[email protected]>

Before adding another lock, give ring->lock a more descriptive name.

---
 drivers/gpu/drm/msm/adreno/a5xx_gpu.c     |  4 ++--
 drivers/gpu/drm/msm/adreno/a5xx_preempt.c | 12 ++++++------
 drivers/gpu/drm/msm/adreno/a6xx_gpu.c     |  4 ++--
 drivers/gpu/drm/msm/msm_ringbuffer.c      |  2 +-
 drivers/gpu/drm/msm/msm_ringbuffer.h      |  7 ++++++-
 5 files changed, 17 insertions(+), 12 deletions(-)
drivers/gpu/drm/msm/adreno/a5xx_gpu.c +2 -2
index 2180650a03bc..a1b9419a59c9 100644
@@ -36,7 +36,7 @@ void a5xx_flush(struct msm_gpu *gpu, struct msm_ringbuffer *ring,
 		OUT_RING(ring, upper_32_bits(shadowptr(a5xx_gpu, ring)));
 	}
 
-	spin_lock_irqsave(&ring->lock, flags);
+	spin_lock_irqsave(&ring->preempt_lock, flags);
 
 	/* Copy the shadow to the actual register */
 	ring->cur = ring->next;
@@ -44,7 +44,7 @@ void a5xx_flush(struct msm_gpu *gpu, struct msm_ringbuffer *ring,
 	/* Make sure to wrap wptr if we need to */
 	wptr = get_wptr(ring);
 
-	spin_unlock_irqrestore(&ring->lock, flags);
+	spin_unlock_irqrestore(&ring->preempt_lock, flags);
 
 	/* Make sure everything is posted before making a decision */
 	mb();
drivers/gpu/drm/msm/adreno/a5xx_preempt.c +6 -6
index 7e04509c4e1f..183de1139eeb 100644
@@ -45,9 +45,9 @@ static inline void update_wptr(struct msm_gpu *gpu, struct msm_ringbuffer *ring)
 	if (!ring)
 		return;
 
-	spin_lock_irqsave(&ring->lock, flags);
+	spin_lock_irqsave(&ring->preempt_lock, flags);
 	wptr = get_wptr(ring);
-	spin_unlock_irqrestore(&ring->lock, flags);
+	spin_unlock_irqrestore(&ring->preempt_lock, flags);
 
 	gpu_write(gpu, REG_A5XX_CP_RB_WPTR, wptr);
 }
@@ -62,9 +62,9 @@ static struct msm_ringbuffer *get_next_ring(struct msm_gpu *gpu)
 		bool empty;
 		struct msm_ringbuffer *ring = gpu->rb[i];
 
-		spin_lock_irqsave(&ring->lock, flags);
+		spin_lock_irqsave(&ring->preempt_lock, flags);
 		empty = (get_wptr(ring) == ring->memptrs->rptr);
-		spin_unlock_irqrestore(&ring->lock, flags);
+		spin_unlock_irqrestore(&ring->preempt_lock, flags);
 
 		if (!empty)
 			return ring;
@@ -132,9 +132,9 @@ void a5xx_preempt_trigger(struct msm_gpu *gpu)
 	}
 
 	/* Make sure the wptr doesn't update while we're in motion */
-	spin_lock_irqsave(&ring->lock, flags);
+	spin_lock_irqsave(&ring->preempt_lock, flags);
 	a5xx_gpu->preempt[ring->id]->wptr = get_wptr(ring);
-	spin_unlock_irqrestore(&ring->lock, flags);
+	spin_unlock_irqrestore(&ring->preempt_lock, flags);
 
 	/* Set the address of the incoming preemption record */
 	gpu_write64(gpu, REG_A5XX_CP_CONTEXT_SWITCH_RESTORE_ADDR_LO,
drivers/gpu/drm/msm/adreno/a6xx_gpu.c +2 -2
index 16eaaf0804ca..eb44e0dbef34 100644
@@ -65,7 +65,7 @@ static void a6xx_flush(struct msm_gpu *gpu, struct msm_ringbuffer *ring)
 		OUT_RING(ring, upper_32_bits(shadowptr(a6xx_gpu, ring)));
 	}
 
-	spin_lock_irqsave(&ring->lock, flags);
+	spin_lock_irqsave(&ring->preempt_lock, flags);
 
 	/* Copy the shadow to the actual register */
 	ring->cur = ring->next;
@@ -73,7 +73,7 @@ static void a6xx_flush(struct msm_gpu *gpu, struct msm_ringbuffer *ring)
 	/* Make sure to wrap wptr if we need to */
 	wptr = get_wptr(ring);
 
-	spin_unlock_irqrestore(&ring->lock, flags);
+	spin_unlock_irqrestore(&ring->preempt_lock, flags);
 
 	/* Make sure everything is posted before making a decision */
 	mb();
drivers/gpu/drm/msm/msm_ringbuffer.c +1 -1
index 935bf9b1d941..1b6958e908dc 100644
@@ -46,7 +46,7 @@ struct msm_ringbuffer *msm_ringbuffer_new(struct msm_gpu *gpu, int id,
 	ring->memptrs_iova = memptrs_iova;
 
 	INIT_LIST_HEAD(&ring->submits);
-	spin_lock_init(&ring->lock);
+	spin_lock_init(&ring->preempt_lock);
 
 	snprintf(name, sizeof(name), "gpu-ring-%d", ring->id);
 
drivers/gpu/drm/msm/msm_ringbuffer.h +6 -1
index 0987d6bf848c..4956d1bc5d0e 100644
@@ -46,7 +46,12 @@ struct msm_ringbuffer {
 	struct msm_rbmemptrs *memptrs;
 	uint64_t memptrs_iova;
 	struct msm_fence_context *fctx;
-	spinlock_t lock;
+
+	/*
+	 * preempt_lock protects preemption and serializes wptr updates against
+	 * preemption.  Can be aquired from irq context.
+	 */
+	spinlock_t preempt_lock;
 };
 
 struct msm_ringbuffer *msm_ringbuffer_new(struct msm_gpu *gpu, int id,
Signature · 6 lines
2.26.2

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
author
#16 lore raw

PATCH v4 15/23 drm/msm: Protect ring->submits with it's own lock

In reply to Rob Clark

From: Rob Clark <[email protected]>

One less place to rely on dev->struct_mutex.

---
 drivers/gpu/drm/msm/msm_gem_submit.c |  2 ++
 drivers/gpu/drm/msm/msm_gpu.c        | 37 ++++++++++++++++++++++------
 drivers/gpu/drm/msm/msm_ringbuffer.c |  1 +
 drivers/gpu/drm/msm/msm_ringbuffer.h |  6 +++++
 4 files changed, 39 insertions(+), 7 deletions(-)
drivers/gpu/drm/msm/msm_gem_submit.c +2 -0
index 044e9bee70a2..24ce4c65429d 100644
@@ -65,7 +65,9 @@ void msm_gem_submit_free(struct msm_gem_submit *submit)
 	unsigned i;
 
 	dma_fence_put(submit->fence);
+	spin_lock(&submit->ring->submit_lock);
 	list_del(&submit->node);
+	spin_unlock(&submit->ring->submit_lock);
 	put_pid(submit->pid);
 	msm_submitqueue_put(submit->queue);
 
drivers/gpu/drm/msm/msm_gpu.c +30 -7
index 1667d8066897..1d6f3dc3fe78 100644
@@ -270,6 +270,7 @@ static void update_fences(struct msm_gpu *gpu, struct msm_ringbuffer *ring,
 {
 	struct msm_gem_submit *submit;
 
+	spin_lock(&ring->submit_lock);
 	list_for_each_entry(submit, &ring->submits, node) {
 		if (submit->seqno > fence)
 			break;
@@ -277,6 +278,7 @@ static void update_fences(struct msm_gpu *gpu, struct msm_ringbuffer *ring,
 		msm_update_fence(submit->ring->fctx,
 			submit->fence->seqno);
 	}
+	spin_unlock(&ring->submit_lock);
 }
 
 #ifdef CONFIG_DEV_COREDUMP
@@ -432,11 +434,14 @@ find_submit(struct msm_ringbuffer *ring, uint32_t fence)
 {
 	struct msm_gem_submit *submit;
 
-	WARN_ON(!mutex_is_locked(&ring->gpu->dev->struct_mutex));
-
-	list_for_each_entry(submit, &ring->submits, node)
-		if (submit->seqno == fence)
+	spin_lock(&ring->submit_lock);
+	list_for_each_entry(submit, &ring->submits, node) {
+		if (submit->seqno == fence) {
+			spin_unlock(&ring->submit_lock);
 			return submit;
+		}
+	}
+	spin_unlock(&ring->submit_lock);
 
 	return NULL;
 }
@@ -533,8 +538,10 @@ static void recover_worker(struct work_struct *work)
 		for (i = 0; i < gpu->nr_rings; i++) {
 			struct msm_ringbuffer *ring = gpu->rb[i];
 
+			spin_lock(&ring->submit_lock);
 			list_for_each_entry(submit, &ring->submits, node)
 				gpu->funcs->submit(gpu, submit);
+			spin_unlock(&ring->submit_lock);
 		}
 	}
 
@@ -721,7 +728,6 @@ static void retire_submit(struct msm_gpu *gpu, struct msm_ringbuffer *ring,
 static void retire_submits(struct msm_gpu *gpu)
 {
 	struct drm_device *dev = gpu->dev;
-	struct msm_gem_submit *submit, *tmp;
 	int i;
 
 	WARN_ON(!mutex_is_locked(&dev->struct_mutex));
@@ -730,9 +736,24 @@ static void retire_submits(struct msm_gpu *gpu)
 	for (i = 0; i < gpu->nr_rings; i++) {
 		struct msm_ringbuffer *ring = gpu->rb[i];
 
-		list_for_each_entry_safe(submit, tmp, &ring->submits, node) {
-			if (dma_fence_is_signaled(submit->fence))
+		while (true) {
+			struct msm_gem_submit *submit = NULL;
+
+			spin_lock(&ring->submit_lock);
+			submit = list_first_entry_or_null(&ring->submits,
+					struct msm_gem_submit, node);
+			spin_unlock(&ring->submit_lock);
+
+			/*
+			 * If no submit, we are done.  If submit->fence hasn't
+			 * been signalled, then later submits are not signalled
+			 * either, so we are also done.
+			 */
+			if (submit && dma_fence_is_signaled(submit->fence)) {
 				retire_submit(gpu, ring, submit);
+			} else {
+				break;
+			}
 		}
 	}
 }
@@ -775,7 +796,9 @@ void msm_gpu_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit)
 
 	submit->seqno = ++ring->seqno;
 
+	spin_lock(&ring->submit_lock);
 	list_add_tail(&submit->node, &ring->submits);
+	spin_unlock(&ring->submit_lock);
 
 	msm_rd_dump_submit(priv->rd, submit, NULL);
 
drivers/gpu/drm/msm/msm_ringbuffer.c +1 -0
index 1b6958e908dc..4d2a2a4abef8 100644
@@ -46,6 +46,7 @@ struct msm_ringbuffer *msm_ringbuffer_new(struct msm_gpu *gpu, int id,
 	ring->memptrs_iova = memptrs_iova;
 
 	INIT_LIST_HEAD(&ring->submits);
+	spin_lock_init(&ring->submit_lock);
 	spin_lock_init(&ring->preempt_lock);
 
 	snprintf(name, sizeof(name), "gpu-ring-%d", ring->id);
drivers/gpu/drm/msm/msm_ringbuffer.h +6 -0
index 4956d1bc5d0e..fe55d4a1aa16 100644
@@ -39,7 +39,13 @@ struct msm_ringbuffer {
 	int id;
 	struct drm_gem_object *bo;
 	uint32_t *start, *end, *cur, *next;
+
+	/*
+	 * List of in-flight submits on this ring.  Protected by submit_lock.
+	 */
 	struct list_head submits;
+	spinlock_t submit_lock;
+
 	uint64_t iova;
 	uint32_t seqno;
 	uint32_t hangcheck_fence;
Signature · 6 lines
2.26.2

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
author
#17 lore raw

PATCH v4 16/23 drm/msm: Refcount submits

In reply to Rob Clark

From: Rob Clark <[email protected]>

Before we remove dev->struct_mutex from the retire path, we have to deal with the situation of a submit retiring before the submit ioctl returns.

To deal with this, ring->submits will hold a reference to the submit, which is dropped when the submit is retired. And the submit ioctl path holds it's own ref, which it drops when it is done with the submit.

Also, add to submit list *after* getting/pinning bo's, to prevent badness in case the completed fence is corrupted, and retire_worker mistakenly believes the submit is done too early.

---
 drivers/gpu/drm/msm/msm_drv.h        |  1 -
 drivers/gpu/drm/msm/msm_gem.h        | 13 +++++++++++++
 drivers/gpu/drm/msm/msm_gem_submit.c | 11 +++++------
 drivers/gpu/drm/msm/msm_gpu.c        | 21 ++++++++++++++++-----
 4 files changed, 34 insertions(+), 12 deletions(-)
drivers/gpu/drm/msm/msm_drv.h +0 -1
index a17dadd38685..2ef5cff19883 100644
@@ -277,7 +277,6 @@ void msm_unregister_mmu(struct drm_device *dev, struct msm_mmu *mmu);
 
 bool msm_use_mmu(struct drm_device *dev);
 
-void msm_gem_submit_free(struct msm_gem_submit *submit);
 int msm_ioctl_gem_submit(struct drm_device *dev, void *data,
 		struct drm_file *file);
 
drivers/gpu/drm/msm/msm_gem.h +13 -0
index f0608d96ef03..2f289c436ddd 100644
@@ -213,6 +213,7 @@ void msm_gem_free_work(struct work_struct *work);
  * lasts for the duration of the submit-ioctl.
  */
 struct msm_gem_submit {
+	struct kref ref;
 	struct drm_device *dev;
 	struct msm_gpu *gpu;
 	struct msm_gem_address_space *aspace;
@@ -249,6 +250,18 @@ struct msm_gem_submit {
 	} bos[];
 };
 
+void __msm_gem_submit_destroy(struct kref *kref);
+
+static inline void msm_gem_submit_get(struct msm_gem_submit *submit)
+{
+	kref_get(&submit->ref);
+}
+
+static inline void msm_gem_submit_put(struct msm_gem_submit *submit)
+{
+	kref_put(&submit->ref, __msm_gem_submit_destroy);
+}
+
 /* helper to determine of a buffer in submit should be dumped, used for both
  * devcoredump and debugfs cmdstream dumping:
  */
drivers/gpu/drm/msm/msm_gem_submit.c +5 -6
index 24ce4c65429d..d04c349d8112 100644
@@ -42,6 +42,7 @@ static struct msm_gem_submit *submit_create(struct drm_device *dev,
 	if (!submit)
 		return NULL;
 
+	kref_init(&submit->ref);
 	submit->dev = dev;
 	submit->aspace = queue->ctx->aspace;
 	submit->gpu = gpu;
@@ -60,14 +61,13 @@ static struct msm_gem_submit *submit_create(struct drm_device *dev,
 	return submit;
 }
 
-void msm_gem_submit_free(struct msm_gem_submit *submit)
+void __msm_gem_submit_destroy(struct kref *kref)
 {
+	struct msm_gem_submit *submit =
+			container_of(kref, struct msm_gem_submit, ref);
 	unsigned i;
 
 	dma_fence_put(submit->fence);
-	spin_lock(&submit->ring->submit_lock);
-	list_del(&submit->node);
-	spin_unlock(&submit->ring->submit_lock);
 	put_pid(submit->pid);
 	msm_submitqueue_put(submit->queue);
 
@@ -847,8 +847,7 @@ int msm_ioctl_gem_submit(struct drm_device *dev, void *data,
 	submit_cleanup(submit);
 	if (has_ww_ticket)
 		ww_acquire_fini(&submit->ticket);
-	if (ret)
-		msm_gem_submit_free(submit);
+	msm_gem_submit_put(submit);
 out_unlock:
 	if (ret && (out_fence_fd >= 0))
 		put_unused_fd(out_fence_fd);
drivers/gpu/drm/msm/msm_gpu.c +16 -5
index 1d6f3dc3fe78..bcd9b4fa98b2 100644
@@ -722,7 +722,12 @@ static void retire_submit(struct msm_gpu *gpu, struct msm_ringbuffer *ring,
 
 	pm_runtime_mark_last_busy(&gpu->pdev->dev);
 	pm_runtime_put_autosuspend(&gpu->pdev->dev);
-	msm_gem_submit_free(submit);
+
+	spin_lock(&ring->submit_lock);
+	list_del(&submit->node);
+	spin_unlock(&ring->submit_lock);
+
+	msm_gem_submit_put(submit);
 }
 
 static void retire_submits(struct msm_gpu *gpu)
@@ -796,10 +801,6 @@ void msm_gpu_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit)
 
 	submit->seqno = ++ring->seqno;
 
-	spin_lock(&ring->submit_lock);
-	list_add_tail(&submit->node, &ring->submits);
-	spin_unlock(&ring->submit_lock);
-
 	msm_rd_dump_submit(priv->rd, submit, NULL);
 
 	update_sw_cntrs(gpu);
@@ -826,6 +827,16 @@ void msm_gpu_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit)
 		msm_gem_active_get(drm_obj, gpu);
 	}
 
+	/*
+	 * ring->submits holds a ref to the submit, to deal with the case
+	 * that a submit completes before msm_ioctl_gem_submit() returns.
+	 */
+	msm_gem_submit_get(submit);
+
+	spin_lock(&ring->submit_lock);
+	list_add_tail(&submit->node, &ring->submits);
+	spin_unlock(&ring->submit_lock);
+
 	gpu->funcs->submit(gpu, submit);
 	priv->lastctx = submit->queue->ctx;
 
Signature · 6 lines
2.26.2

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
author
#18 lore raw

PATCH v4 17/23 drm/msm: Remove obj->gpu

In reply to Rob Clark

From: Rob Clark <[email protected]>

It cannot be atomically updated with obj->active_count, and the only purpose is a useless WARN_ON() (which becomes a buggy WARN_ON() once retire_submits() is not serialized with incoming submits via struct_mutex)

---
 drivers/gpu/drm/msm/msm_gem.c | 2 --
 drivers/gpu/drm/msm/msm_gem.h | 1 -
 drivers/gpu/drm/msm/msm_gpu.c | 5 -----
 3 files changed, 8 deletions(-)
drivers/gpu/drm/msm/msm_gem.c +0 -2
index 81ed03322a74..b9de19462109 100644
@@ -775,7 +775,6 @@ void msm_gem_active_get(struct drm_gem_object *obj, struct msm_gpu *gpu)
 
 	if (!atomic_fetch_inc(&msm_obj->active_count)) {
 		mutex_lock(&priv->mm_lock);
-		msm_obj->gpu = gpu;
 		list_del_init(&msm_obj->mm_list);
 		list_add_tail(&msm_obj->mm_list, &gpu->active_list);
 		mutex_unlock(&priv->mm_lock);
@@ -791,7 +790,6 @@ void msm_gem_active_put(struct drm_gem_object *obj)
 
 	if (!atomic_dec_return(&msm_obj->active_count)) {
 		mutex_lock(&priv->mm_lock);
-		msm_obj->gpu = NULL;
 		list_del_init(&msm_obj->mm_list);
 		list_add_tail(&msm_obj->mm_list, &priv->inactive_list);
 		mutex_unlock(&priv->mm_lock);
drivers/gpu/drm/msm/msm_gem.h +0 -1
index 2f289c436ddd..f4e73c6f07bf 100644
@@ -64,7 +64,6 @@ struct msm_gem_object {
 	 *
 	 */
 	struct list_head mm_list;
-	struct msm_gpu *gpu;     /* non-null if active */
 
 	/* Transiently in the process of submit ioctl, objects associated
 	 * with the submit are on submit->bo_list.. this only lasts for
drivers/gpu/drm/msm/msm_gpu.c +0 -5
index bcd9b4fa98b2..d0f625112a97 100644
@@ -810,11 +810,6 @@ void msm_gpu_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit)
 		struct drm_gem_object *drm_obj = &msm_obj->base;
 		uint64_t iova;
 
-		/* can't happen yet.. but when we add 2d support we'll have
-		 * to deal w/ cross-ring synchronization:
-		 */
-		WARN_ON(is_active(msm_obj) && (msm_obj->gpu != gpu));
-
 		/* submit takes a reference to the bo and iova until retired: */
 		drm_gem_object_get(&msm_obj->base);
 		msm_gem_get_and_pin_iova_locked(&msm_obj->base, submit->aspace, &iova);
Signature · 6 lines
2.26.2

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
author
#19 lore raw

PATCH v4 18/23 drm/msm: Drop struct_mutex from the retire path

In reply to Rob Clark

From: Rob Clark <[email protected]>

Now that we are not relying on dev->struct_mutex to protect the ring->submits lists, drop the struct_mutex lock.

---
 drivers/gpu/drm/msm/msm_gpu.c | 8 +-------
 1 file changed, 1 insertion(+), 7 deletions(-)
drivers/gpu/drm/msm/msm_gpu.c +1 -7
index d0f625112a97..30ba3beaad0a 100644
@@ -717,7 +717,7 @@ static void retire_submit(struct msm_gpu *gpu, struct msm_ringbuffer *ring,
 
 		msm_gem_active_put(&msm_obj->base);
 		msm_gem_unpin_iova(&msm_obj->base, submit->aspace);
-		drm_gem_object_put_locked(&msm_obj->base);
+		drm_gem_object_put(&msm_obj->base);
 	}
 
 	pm_runtime_mark_last_busy(&gpu->pdev->dev);
@@ -732,11 +732,8 @@ static void retire_submit(struct msm_gpu *gpu, struct msm_ringbuffer *ring,
 
 static void retire_submits(struct msm_gpu *gpu)
 {
-	struct drm_device *dev = gpu->dev;
 	int i;
 
-	WARN_ON(!mutex_is_locked(&dev->struct_mutex));
-
 	/* Retire the commits starting with highest priority */
 	for (i = 0; i < gpu->nr_rings; i++) {
 		struct msm_ringbuffer *ring = gpu->rb[i];
@@ -766,15 +763,12 @@ static void retire_submits(struct msm_gpu *gpu)
 static void retire_worker(struct work_struct *work)
 {
 	struct msm_gpu *gpu = container_of(work, struct msm_gpu, retire_work);
-	struct drm_device *dev = gpu->dev;
 	int i;
 
 	for (i = 0; i < gpu->nr_rings; i++)
 		update_fences(gpu, gpu->rb[i], gpu->rb[i]->memptrs->fence);
 
-	mutex_lock(&dev->struct_mutex);
 	retire_submits(gpu);
-	mutex_unlock(&dev->struct_mutex);
 }
 
 /* call from irq handler to schedule work to retire bo's */
Signature · 6 lines
2.26.2

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
author
#20 lore raw

PATCH v4 19/23 drm/msm: Drop struct_mutex in free_object() path

In reply to Rob Clark

From: Rob Clark <[email protected]>

Now that active_list/inactive_list is protected by mm_lock, we no longer need dev->struct_mutex in the free_object() path.

---
 drivers/gpu/drm/msm/msm_gem.c | 8 --------
 1 file changed, 8 deletions(-)
drivers/gpu/drm/msm/msm_gem.c +0 -8
index b9de19462109..5d555750943e 100644
@@ -949,8 +949,6 @@ static void free_object(struct msm_gem_object *msm_obj)
 	struct drm_device *dev = obj->dev;
 	struct msm_drm_private *priv = dev->dev_private;
 
-	WARN_ON(!mutex_is_locked(&dev->struct_mutex));
-
 	/* object should not be on active list: */
 	WARN_ON(is_active(msm_obj));
 
@@ -992,20 +990,14 @@ void msm_gem_free_work(struct work_struct *work)
 {
 	struct msm_drm_private *priv =
 		container_of(work, struct msm_drm_private, free_work);
-	struct drm_device *dev = priv->dev;
 	struct llist_node *freed;
 	struct msm_gem_object *msm_obj, *next;
 
 	while ((freed = llist_del_all(&priv->free_list))) {
-
-		mutex_lock(&dev->struct_mutex);
-
 		llist_for_each_entry_safe(msm_obj, next,
 					  freed, freed)
 			free_object(msm_obj);
 
-		mutex_unlock(&dev->struct_mutex);
-
 		if (need_resched())
 			break;
 	}
Signature · 6 lines
2.26.2

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
author
#21 lore raw

PATCH v4 20/23 drm/msm: Remove msm_gem_free_work

In reply to Rob Clark

From: Rob Clark <[email protected]>

Now that we don't need struct_mutex in the free path, we can get rid of the asynchronous free altogether.

---
 drivers/gpu/drm/msm/msm_drv.c |  3 ---
 drivers/gpu/drm/msm/msm_drv.h |  5 -----
 drivers/gpu/drm/msm/msm_gem.c | 27 ---------------------------
 drivers/gpu/drm/msm/msm_gem.h |  1 -
 4 files changed, 36 deletions(-)
drivers/gpu/drm/msm/msm_drv.c +0 -3
index 81cb2cecc829..49e6daf30b42 100644
@@ -438,9 +438,6 @@ static int msm_drm_init(struct device *dev, struct drm_driver *drv)
 
 	priv->wq = alloc_ordered_workqueue("msm", 0);
 
-	INIT_WORK(&priv->free_work, msm_gem_free_work);
-	init_llist_head(&priv->free_list);
-
 	INIT_LIST_HEAD(&priv->inactive_list);
 	mutex_init(&priv->mm_lock);
 
drivers/gpu/drm/msm/msm_drv.h +0 -5
index 2ef5cff19883..af296712eae8 100644
@@ -188,10 +188,6 @@ struct msm_drm_private {
 	struct list_head inactive_list;
 	struct mutex mm_lock;
 
-	/* worker for delayed free of objects: */
-	struct work_struct free_work;
-	struct llist_head free_list;
-
 	struct workqueue_struct *wq;
 
 	unsigned int num_planes;
@@ -291,7 +287,6 @@ struct drm_gem_object *msm_gem_prime_import_sg_table(struct drm_device *dev,
 		struct dma_buf_attachment *attach, struct sg_table *sg);
 int msm_gem_prime_pin(struct drm_gem_object *obj);
 void msm_gem_prime_unpin(struct drm_gem_object *obj);
-void msm_gem_free_work(struct work_struct *work);
 
 int msm_framebuffer_prepare(struct drm_framebuffer *fb,
 		struct msm_gem_address_space *aspace);
drivers/gpu/drm/msm/msm_gem.c +0 -27
index 5d555750943e..00b4788366c5 100644
@@ -939,16 +939,6 @@ void msm_gem_free_object(struct drm_gem_object *obj)
 	struct drm_device *dev = obj->dev;
 	struct msm_drm_private *priv = dev->dev_private;
 
-	if (llist_add(&msm_obj->freed, &priv->free_list))
-		queue_work(priv->wq, &priv->free_work);
-}
-
-static void free_object(struct msm_gem_object *msm_obj)
-{
-	struct drm_gem_object *obj = &msm_obj->base;
-	struct drm_device *dev = obj->dev;
-	struct msm_drm_private *priv = dev->dev_private;
-
 	/* object should not be on active list: */
 	WARN_ON(is_active(msm_obj));
 
@@ -986,23 +976,6 @@ static void free_object(struct msm_gem_object *msm_obj)
 	kfree(msm_obj);
 }
 
-void msm_gem_free_work(struct work_struct *work)
-{
-	struct msm_drm_private *priv =
-		container_of(work, struct msm_drm_private, free_work);
-	struct llist_node *freed;
-	struct msm_gem_object *msm_obj, *next;
-
-	while ((freed = llist_del_all(&priv->free_list))) {
-		llist_for_each_entry_safe(msm_obj, next,
-					  freed, freed)
-			free_object(msm_obj);
-
-		if (need_resched())
-			break;
-	}
-}
-
 /* convenience method to construct a GEM buffer object, and userspace handle */
 int msm_gem_new_handle(struct drm_device *dev, struct drm_file *file,
 		uint32_t size, uint32_t flags, uint32_t *handle,
drivers/gpu/drm/msm/msm_gem.h +0 -1
index f4e73c6f07bf..ffa2130ee97d 100644
@@ -204,7 +204,6 @@ static inline bool is_vunmapable(struct msm_gem_object *msm_obj)
 
 void msm_gem_purge(struct drm_gem_object *obj);
 void msm_gem_vunmap(struct drm_gem_object *obj);
-void msm_gem_free_work(struct work_struct *work);
 
 /* Created per submit-ioctl, to track bo's and cmdstream bufs, etc,
  * associated with the cmdstream submission for synchronization (and
Signature · 6 lines
2.26.2

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
author
#24 lore raw

PATCH v4 23/23 drm/msm: Don't implicit-sync if only a single ring

In reply to Rob Clark

From: Rob Clark <[email protected]>

If there is only a single ring (no-preemption), everything is FIFO order and there is no need to implicit-sync.

Mesa should probably just always use MSM_SUBMIT_NO_IMPLICIT, as behavior is undefined when fences are not used to synchronize buffer usage across contexts (which is the only case where multiple different priority rings could come into play).

---
 drivers/gpu/drm/msm/msm_gem_submit.c | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)
drivers/gpu/drm/msm/msm_gem_submit.c +4 -3
index d04c349d8112..b6babc7f9bb8 100644
@@ -283,7 +283,7 @@ static int submit_lock_objects(struct msm_gem_submit *submit)
 	return ret;
 }
 
-static int submit_fence_sync(struct msm_gem_submit *submit, bool no_implicit)
+static int submit_fence_sync(struct msm_gem_submit *submit, bool implicit_sync)
 {
 	int i, ret = 0;
 
@@ -303,7 +303,7 @@ static int submit_fence_sync(struct msm_gem_submit *submit, bool no_implicit)
 				return ret;
 		}
 
-		if (no_implicit)
+		if (!implicit_sync)
 			continue;
 
 		ret = msm_gem_sync_object(&msm_obj->base, submit->ring->fctx,
@@ -774,7 +774,8 @@ int msm_ioctl_gem_submit(struct drm_device *dev, void *data,
 	if (ret)
 		goto out;
 
-	ret = submit_fence_sync(submit, !!(args->flags & MSM_SUBMIT_NO_IMPLICIT));
+	ret = submit_fence_sync(submit, (gpu->nr_rings > 1) &&
+			!(args->flags & MSM_SUBMIT_NO_IMPLICIT));
 	if (ret)
 		goto out;
 
Signature · 6 lines
2.26.2

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
author
#23 lore raw

PATCH v4 22/23 drm/msm: Drop struct_mutex in shrinker path

In reply to Rob Clark

From: Rob Clark <[email protected]>

Now that the inactive_list is protected by mm_lock, and everything else on per-obj basis is protected by obj->lock, we no longer depend on struct_mutex.

---
 drivers/gpu/drm/msm/msm_gem.c          |  1 -
 drivers/gpu/drm/msm/msm_gem_shrinker.c | 54 --------------------------
 2 files changed, 55 deletions(-)
drivers/gpu/drm/msm/msm_gem.c +0 -1
index 7f24a58c0390..a9a834bb7794 100644
@@ -688,7 +688,6 @@ void msm_gem_purge(struct drm_gem_object *obj)
 	struct drm_device *dev = obj->dev;
 	struct msm_gem_object *msm_obj = to_msm_bo(obj);
 
-	WARN_ON(!mutex_is_locked(&dev->struct_mutex));
 	WARN_ON(!is_purgeable(msm_obj));
 	WARN_ON(obj->import_attach);
 
drivers/gpu/drm/msm/msm_gem_shrinker.c +0 -54

101 lines not on this page · raw message

Signature · 6 lines
2.26.2

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
author
#22 lore raw

PATCH v4 21/23 drm/msm: Drop struct_mutex in madvise path

In reply to Rob Clark

From: Rob Clark <[email protected]>

The obj->lock is sufficient for what we need.

This *does* have the implication that userspace can try to shoot themselves in the foot by racing madvise(DONTNEED) with submit. But the result will be about the same if they did madvise(DONTNEED) before the submit ioctl, ie. they might not get want they want if they race with shrinker. But iova fault handling is robust enough, and userspace is only shooting it's own foot.

---
 drivers/gpu/drm/msm/msm_drv.c | 11 ++---------
 drivers/gpu/drm/msm/msm_gem.c |  4 +---
 drivers/gpu/drm/msm/msm_gem.h |  2 --
 3 files changed, 3 insertions(+), 14 deletions(-)
drivers/gpu/drm/msm/msm_drv.c +2 -9
index 49e6daf30b42..f2d58fe25497 100644
@@ -912,14 +912,9 @@ static int msm_ioctl_gem_madvise(struct drm_device *dev, void *data,
 		return -EINVAL;
 	}
 
-	ret = mutex_lock_interruptible(&dev->struct_mutex);
-	if (ret)
-		return ret;
-
 	obj = drm_gem_object_lookup(file, args->handle);
 	if (!obj) {
-		ret = -ENOENT;
-		goto unlock;
+		return -ENOENT;
 	}
 
 	ret = msm_gem_madvise(obj, args->madv);
@@ -928,10 +923,8 @@ static int msm_ioctl_gem_madvise(struct drm_device *dev, void *data,
 		ret = 0;
 	}
 
-	drm_gem_object_put_locked(obj);
+	drm_gem_object_put(obj);
 
-unlock:
-	mutex_unlock(&dev->struct_mutex);
 	return ret;
 }
 
drivers/gpu/drm/msm/msm_gem.c +1 -3
index 00b4788366c5..7f24a58c0390 100644
@@ -673,8 +673,6 @@ int msm_gem_madvise(struct drm_gem_object *obj, unsigned madv)
 
 	msm_gem_lock(obj);
 
-	WARN_ON(!mutex_is_locked(&obj->dev->struct_mutex));
-
 	if (msm_obj->madv != __MSM_MADV_PURGED)
 		msm_obj->madv = madv;
 
@@ -691,7 +689,6 @@ void msm_gem_purge(struct drm_gem_object *obj)
 	struct msm_gem_object *msm_obj = to_msm_bo(obj);
 
 	WARN_ON(!mutex_is_locked(&dev->struct_mutex));
-	WARN_ON(!msm_gem_is_locked(obj));
 	WARN_ON(!is_purgeable(msm_obj));
 	WARN_ON(obj->import_attach);
 
@@ -771,6 +768,7 @@ void msm_gem_active_get(struct drm_gem_object *obj, struct msm_gpu *gpu)
 	struct msm_drm_private *priv = obj->dev->dev_private;
 
 	might_sleep();
+	WARN_ON(!msm_gem_is_locked(obj));
 	WARN_ON(msm_obj->madv != MSM_MADV_WILLNEED);
 
 	if (!atomic_fetch_inc(&msm_obj->active_count)) {
drivers/gpu/drm/msm/msm_gem.h +0 -2
index ffa2130ee97d..d79e7019cc88 100644
@@ -190,8 +190,6 @@ static inline bool is_active(struct msm_gem_object *msm_obj)
 
 static inline bool is_purgeable(struct msm_gem_object *msm_obj)
 {
-	WARN_ON(!msm_gem_is_locked(&msm_obj->base));
-	WARN_ON(!mutex_is_locked(&msm_obj->base.dev->struct_mutex));
 	return (msm_obj->madv == MSM_MADV_DONTNEED) && msm_obj->sgt &&
 			!msm_obj->base.dma_buf && !msm_obj->base.import_attach;
 }
Signature · 6 lines
2.26.2

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
#25 lore raw

Re: [PATCH v4 23/23] drm/msm: Don't implicit-sync if only a single ring

In reply to Rob Clark
Quoted 9 lines · Rob Clark

From: Rob Clark <[email protected]>

If there is only a single ring (no-preemption), everything is FIFO order and there is no need to implicit-sync.

Mesa should probably just always use MSM_SUBMIT_NO_IMPLICIT, as behavior is undefined when fences are not used to synchronize buffer usage across contexts (which is the only case where multiple different priority rings could come into play).

Really, doesn't this break cross-device implicit sync? Okay, you may not have many peripherals that rely on implicit sync on devices where Adreno is usually found, but it seems rather heavy-handed.

Wouldn't it be better to only ignore fences from your own ring context in the implicit sync, like we do in the common DRM scheduler (drm_sched_dependency_optimized)?

Regards,
Lucas

Quoted 29 lines
---
 drivers/gpu/drm/msm/msm_gem_submit.c | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)
drivers/gpu/drm/msm/msm_gem_submit.c +2 -2
index d04c349d8112..b6babc7f9bb8 100644
@@ -283,7 +283,7 @@ static int submit_lock_objects(struct msm_gem_submit *submit)
 	return ret;
 }
 
-static int submit_fence_sync(struct msm_gem_submit *submit, bool no_implicit)
+static int submit_fence_sync(struct msm_gem_submit *submit, bool implicit_sync)
 {
 	int i, ret = 0;
 
@@ -303,7 +303,7 @@ static int submit_fence_sync(struct msm_gem_submit *submit, bool no_implicit)
 				return ret;
 		}
 
-		if (no_implicit)
+		if (!implicit_sync)
 			continue;
 
 		ret = msm_gem_sync_object(&msm_obj->base, submit->ring->fctx,

drivers/gpu/drm/msm/msm_gem_submit.c int msm_ioctl_gem_submit(struct drm_device *dev, void *data, · line 781

Quoted 9 lines
drivers/gpu/drm/msm/msm_gem_submit.c +2 -1
@@ -774,7 +774,8 @@ int msm_ioctl_gem_submit(struct drm_device *dev, void *data,
 	if (ret)
 		goto out;
 
-	ret = submit_fence_sync(submit, !!(args->flags & MSM_SUBMIT_NO_IMPLICIT));
+	ret = submit_fence_sync(submit, (gpu->nr_rings > 1) &&
+			!(args->flags & MSM_SUBMIT_NO_IMPLICIT));
 	if (ret)
 		goto out;
 

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

#8 lore raw

Re: [PATCH v4 07/23] drm/msm/submit: Move copy_from_user ahead of locking bos

In reply to Rob Clark
I love your patch! Perhaps something to improve:

Hi Rob,

I love your patch! Perhaps something to improve:

[auto build test WARNING on drm-exynos/exynos-drm-next]
[also build test WARNING on linus/master next-20201023]
[cannot apply to drm-intel/for-linux-next tegra-drm/drm/tegra/for-next drm-tip/drm-tip drm/drm-next v5.9] [If your patch is applied to the wrong git tree, kindly drop us a note.
And when submitting patch, we suggest to use '--base' as documented in
https://git-scm.com/docs/git-format-patch]

url:    https://github.com/0day-ci/linux/commits/Rob-Clark/drm-msm-de-struct_mutex-ification/20201024-005320
base:   https://git.kernel.org/pub/scm/linux/kernel/git/daeinki/drm-exynos.git exynos-drm-next
config: arm64-randconfig-r002-20201022 (attached as .config)
compiler: clang version 12.0.0 (https://github.com/llvm/llvm-project 147b9497e79a98a8614b2b5eb4ba653b44f6b6f0)

reproduce (this is a W=1 build):
wget https://raw.githubusercontent.com/intel/lkp-tests/master/sbin/make.cross -O ~/bin/make.cross chmod +x ~/bin/make.cross
# install arm64 cross compiling tool for clang build
# apt-get install binutils-aarch64-linux-gnu
# https://github.com/0day-ci/linux/commit/3b6b5d8e737a640fc5a98f7e08cfff14fc7ca89b
git remote add linux-review https://github.com/0day-ci/linux
git fetch --no-tags linux-review Rob-Clark/drm-msm-de-struct_mutex-ification/20201024-005320 git checkout 3b6b5d8e737a640fc5a98f7e08cfff14fc7ca89b
# save the attached .config to linux build tree
COMPILER_INSTALL_PATH=$HOME/0day COMPILER=clang make.cross ARCH=arm64

If you fix the issue, kindly add following tag as appropriate

All warnings (new ones prefixed by >>):

>> drivers/gpu/drm/msm/msm_gem_submit.c:202:10: warning: result of comparison of constant 18446744073709551615 with expression of type 'unsigned int' is always false [-Wtautological-constant-out-of-range-compare]
                   if (sz == SIZE_MAX) {
                       ~~ ^  ~~~~~~~~

1 warning generated.

vim +202 drivers/gpu/drm/msm/msm_gem_submit.c

157
158 static int submit_lookup_cmds(struct msm_gem_submit *submit,
159 struct drm_msm_gem_submit *args, struct drm_file *file)
160 {
161 unsigned i, sz;
162 int ret = 0;
163
164 for (i = 0; i < args->nr_cmds; i++) {
165 struct drm_msm_gem_submit_cmd submit_cmd;
166 void __user *userptr =
167 u64_to_user_ptr(args->cmds + (i * sizeof(submit_cmd)));
168
169 ret = copy_from_user(&submit_cmd, userptr, sizeof(submit_cmd));
170 if (ret) {
171 ret = -EFAULT;
172 goto out;
173 }
174
175 /* validate input from userspace: */
176 switch (submit_cmd.type) {
177 case MSM_SUBMIT_CMD_BUF:
178 case MSM_SUBMIT_CMD_IB_TARGET_BUF:
179 case MSM_SUBMIT_CMD_CTX_RESTORE_BUF:
180 break;
181 default:
182 DRM_ERROR("invalid type: %08x\n", submit_cmd.type);
183 return -EINVAL;
184 }
185
186 if (submit_cmd.size % 4) {
187 DRM_ERROR("non-aligned cmdstream buffer size: %u\n",
188 submit_cmd.size);
189 ret = -EINVAL;
190 goto out;
191 }
192
193 submit->cmd[i].type = submit_cmd.type;
194 submit->cmd[i].size = submit_cmd.size / 4;
195 submit->cmd[i].offset = submit_cmd.submit_offset / 4;
196 submit->cmd[i].idx = submit_cmd.submit_idx;
197 submit->cmd[i].nr_relocs = submit_cmd.nr_relocs;
198
199 sz = array_size(submit_cmd.nr_relocs,
200 sizeof(struct drm_msm_gem_submit_reloc));
201 /* check for overflow: */

 > 202			if (sz == SIZE_MAX) {

203 ret = -ENOMEM;
204 goto out;
205 }
206 submit->cmd[i].relocs = kmalloc(sz, GFP_KERNEL);
207 ret = copy_from_user(submit->cmd[i].relocs, userptr, sz);
208 if (ret) {
209 ret = -EFAULT;
210 goto out;
211 }
212 }
213
214 out:
215 return ret;
216 }
217

---
0-DAY CI Kernel Test Service, Intel Corporation
https://lists.01.org/hyperkitty/list/kbuild-all(a)lists.01.org

author
#26 lore raw

Re: [PATCH v4 23/23] drm/msm: Don't implicit-sync if only a single ring

In reply to Lucas Stach
Quoted 12 lines · Lucas Stach
Quoted 9 lines · Rob Clark

From: Rob Clark <[email protected]>

If there is only a single ring (no-preemption), everything is FIFO order and there is no need to implicit-sync.

Mesa should probably just always use MSM_SUBMIT_NO_IMPLICIT, as behavior is undefined when fences are not used to synchronize buffer usage across contexts (which is the only case where multiple different priority rings could come into play).

Quoted 7 lines

Really, doesn't this break cross-device implicit sync? Okay, you may not have many peripherals that rely on implicit sync on devices where Adreno is usually found, but it seems rather heavy-handed.

Wouldn't it be better to only ignore fences from your own ring context in the implicit sync, like we do in the common DRM scheduler (drm_sched_dependency_optimized)?

we already do this.. as was discussed on an earlier iteration of this patchset

But I'm not aware of any other non-gpu related implicit sync use-case (even on imx devices where display is decoupled from gpu).. I'll revert the patch if someone comes up with one, but otherwise lets let the implicit sync baggage die

BR,
-R

Quoted 43 lines

Regards,
Lucas

Quoted 38 lines
---
 drivers/gpu/drm/msm/msm_gem_submit.c | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)
drivers/gpu/drm/msm/msm_gem_submit.c +4 -3
index d04c349d8112..b6babc7f9bb8 100644
@@ -283,7 +283,7 @@ static int submit_lock_objects(struct msm_gem_submit *submit)
      return ret;
 }
 
-static int submit_fence_sync(struct msm_gem_submit *submit, bool no_implicit)
+static int submit_fence_sync(struct msm_gem_submit *submit, bool implicit_sync)
 {
      int i, ret = 0;
 
@@ -303,7 +303,7 @@ static int submit_fence_sync(struct msm_gem_submit *submit, bool no_implicit)
                              return ret;
              }
 
-             if (no_implicit)
+             if (!implicit_sync)
                      continue;
 
              ret = msm_gem_sync_object(&msm_obj->base, submit->ring->fctx,
@@ -774,7 +774,8 @@ int msm_ioctl_gem_submit(struct drm_device *dev, void *data,
      if (ret)
              goto out;
 
-     ret = submit_fence_sync(submit, !!(args->flags & MSM_SUBMIT_NO_IMPLICIT));
+     ret = submit_fence_sync(submit, (gpu->nr_rings > 1) &&
+                     !(args->flags & MSM_SUBMIT_NO_IMPLICIT));
      if (ret)
              goto out;
 
#27 lore raw

Re: [PATCH v4 23/23] drm/msm: Don't implicit-sync if only a single ring

Quoted 21 lines · Rob Clark
Quoted 12 lines · Lucas Stach
Quoted 9 lines · Rob Clark

From: Rob Clark <[email protected]>

If there is only a single ring (no-preemption), everything is FIFO order and there is no need to implicit-sync.

Mesa should probably just always use MSM_SUBMIT_NO_IMPLICIT, as behavior is undefined when fences are not used to synchronize buffer usage across contexts (which is the only case where multiple different priority rings could come into play).

Quoted 7 lines

Really, doesn't this break cross-device implicit sync? Okay, you may not have many peripherals that rely on implicit sync on devices where Adreno is usually found, but it seems rather heavy-handed.

Wouldn't it be better to only ignore fences from your own ring context in the implicit sync, like we do in the common DRM scheduler (drm_sched_dependency_optimized)?

Quoted 6 lines

we already do this.. as was discussed on an earlier iteration of this patchset

But I'm not aware of any other non-gpu related implicit sync use-case (even on imx devices where display is decoupled from gpu).. I'll revert the patch if someone comes up with one, but otherwise lets let the implicit sync baggage die

The thing is, dma_resv won't die, even if implicit sync is dead. We're using internally for activity tracking and memory management. If you don't set these, then we can't share generic code with msm, and I think everyone inventing their own memory management is a bit a mistake.

Now you only kill the implicit write sync stuff here, but I'm not sure that's worth much since you still install all the read fences for consistency. And if userspace doesn't want to be synced, they can set the flag and do this on their own: I think you should be able to achieve exactly the same thing in mesa.

Aside: If you're worried about overhead, you can do O(1) submit if you manage your ppgtt like amdgpu does.
-Daniel

Quoted 53 lines

BR,
-R

Quoted 36 lines

Regards,
Lucas

Quoted 32 lines
---
 drivers/gpu/drm/msm/msm_gem_submit.c | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)
drivers/gpu/drm/msm/msm_gem_submit.c +2 -2
index d04c349d8112..b6babc7f9bb8 100644
@@ -283,7 +283,7 @@ static int submit_lock_objects(struct msm_gem_submit *submit)
      return ret;
 }
 
-static int submit_fence_sync(struct msm_gem_submit *submit, bool no_implicit)
+static int submit_fence_sync(struct msm_gem_submit *submit, bool implicit_sync)
 {
      int i, ret = 0;
 
@@ -303,7 +303,7 @@ static int submit_fence_sync(struct msm_gem_submit *submit, bool no_implicit)
                              return ret;
              }
 
-             if (no_implicit)
+             if (!implicit_sync)
                      continue;
 
              ret = msm_gem_sync_object(&msm_obj->base, submit->ring->fctx,
@@ -774,7 +774,8 @@ int msm_ioctl_gem_submit(struct drm_device *dev, void *data,
      if (ret)
              goto out;
 
Quoted 7 lines

drivers/gpu/drm/msm/msm_gem_submit.c int msm_ioctl_gem_submit(struct drm_device *dev, void *data, · line 781

Quoted 6 lines
drivers/gpu/drm/msm/msm_gem_submit.c +2 -1
-     ret = submit_fence_sync(submit, !!(args->flags & MSM_SUBMIT_NO_IMPLICIT));
+     ret = submit_fence_sync(submit, (gpu->nr_rings > 1) &&
+                     !(args->flags & MSM_SUBMIT_NO_IMPLICIT));
      if (ret)
              goto out;
 

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

Signature · 7 lines
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch
_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
#28 lore raw

Re: [PATCH v4 23/23] drm/msm: Don't implicit-sync if only a single ring

Quoted 35 lines · Daniel Vetter
Quoted 21 lines · Rob Clark
Quoted 12 lines · Lucas Stach
Quoted 9 lines · Rob Clark

From: Rob Clark <[email protected]>

If there is only a single ring (no-preemption), everything is FIFO order and there is no need to implicit-sync.

Mesa should probably just always use MSM_SUBMIT_NO_IMPLICIT, as behavior is undefined when fences are not used to synchronize buffer usage across contexts (which is the only case where multiple different priority rings could come into play).

Quoted 7 lines

Really, doesn't this break cross-device implicit sync? Okay, you may not have many peripherals that rely on implicit sync on devices where Adreno is usually found, but it seems rather heavy-handed.

Wouldn't it be better to only ignore fences from your own ring context in the implicit sync, like we do in the common DRM scheduler (drm_sched_dependency_optimized)?

Quoted 6 lines

we already do this.. as was discussed on an earlier iteration of this patchset

But I'm not aware of any other non-gpu related implicit sync use-case (even on imx devices where display is decoupled from gpu).. I'll revert the patch if someone comes up with one, but otherwise lets let the implicit sync baggage die

The thing is, dma_resv won't die, even if implicit sync is dead. We're using internally for activity tracking and memory management. If you don't set these, then we can't share generic code with msm, and I think everyone inventing their own memory management is a bit a mistake.

Quoted 8 lines

Now you only kill the implicit write sync stuff here, but I'm not sure that's worth much since you still install all the read fences for consistency. And if userspace doesn't want to be synced, they can set the flag and do this on their own: I think you should be able to achieve exactly the same thing in mesa.

Aside: If you're worried about overhead, you can do O(1) submit if you manage your ppgtt like amdgpu does.

So just remember a use-case which is maybe a bit yucky, but it is actually possible to implement race-free. If you have implicit sync.

There's screen-capture tool in mplayer and obs which capture your compositor by running getfb2 in a loop. It works, and after some initial screaming I realized it does actually work race-free. If you have implicit sync.

I really don't think you can sunset this, as much as you want to. And sunsetting it inconsistently is probably the worst. -Daniel

Quoted 60 lines

-Daniel

Quoted 49 lines

BR,
-R

Quoted 36 lines

Regards,
Lucas

Quoted 32 lines
---
 drivers/gpu/drm/msm/msm_gem_submit.c | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)
drivers/gpu/drm/msm/msm_gem_submit.c +2 -2
index d04c349d8112..b6babc7f9bb8 100644
@@ -283,7 +283,7 @@ static int submit_lock_objects(struct msm_gem_submit *submit)
      return ret;
 }
 
-static int submit_fence_sync(struct msm_gem_submit *submit, bool no_implicit)
+static int submit_fence_sync(struct msm_gem_submit *submit, bool implicit_sync)
 {
      int i, ret = 0;
 
@@ -303,7 +303,7 @@ static int submit_fence_sync(struct msm_gem_submit *submit, bool no_implicit)
                              return ret;
              }
 
-             if (no_implicit)
+             if (!implicit_sync)
                      continue;
 
              ret = msm_gem_sync_object(&msm_obj->base, submit->ring->fctx,
@@ -774,7 +774,8 @@ int msm_ioctl_gem_submit(struct drm_device *dev, void *data,
      if (ret)
              goto out;
 
Quoted 7 lines

drivers/gpu/drm/msm/msm_gem_submit.c int msm_ioctl_gem_submit(struct drm_device *dev, void *data, · line 781

Quoted 6 lines
drivers/gpu/drm/msm/msm_gem_submit.c +2 -1
-     ret = submit_fence_sync(submit, !!(args->flags & MSM_SUBMIT_NO_IMPLICIT));
+     ret = submit_fence_sync(submit, (gpu->nr_rings > 1) &&
+                     !(args->flags & MSM_SUBMIT_NO_IMPLICIT));
      if (ret)
              goto out;
 
Quoted 4 lines

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch

Signature · 3 lines
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch
author
#29 lore raw

Re: [PATCH v4 23/23] drm/msm: Don't implicit-sync if only a single ring

Quoted 49 lines · Daniel Vetter
Quoted 35 lines · Daniel Vetter
Quoted 21 lines · Rob Clark
Quoted 12 lines · Lucas Stach
Quoted 9 lines · Rob Clark

From: Rob Clark <[email protected]>

If there is only a single ring (no-preemption), everything is FIFO order and there is no need to implicit-sync.

Mesa should probably just always use MSM_SUBMIT_NO_IMPLICIT, as behavior is undefined when fences are not used to synchronize buffer usage across contexts (which is the only case where multiple different priority rings could come into play).

Quoted 7 lines

Really, doesn't this break cross-device implicit sync? Okay, you may not have many peripherals that rely on implicit sync on devices where Adreno is usually found, but it seems rather heavy-handed.

Wouldn't it be better to only ignore fences from your own ring context in the implicit sync, like we do in the common DRM scheduler (drm_sched_dependency_optimized)?

Quoted 6 lines

we already do this.. as was discussed on an earlier iteration of this patchset

But I'm not aware of any other non-gpu related implicit sync use-case (even on imx devices where display is decoupled from gpu).. I'll revert the patch if someone comes up with one, but otherwise lets let the implicit sync baggage die

The thing is, dma_resv won't die, even if implicit sync is dead. We're using internally for activity tracking and memory management. If you don't set these, then we can't share generic code with msm, and I think everyone inventing their own memory management is a bit a mistake.

Quoted 8 lines

Now you only kill the implicit write sync stuff here, but I'm not sure that's worth much since you still install all the read fences for consistency. And if userspace doesn't want to be synced, they can set the flag and do this on their own: I think you should be able to achieve exactly the same thing in mesa.

Aside: If you're worried about overhead, you can do O(1) submit if you manage your ppgtt like amdgpu does.

So just remember a use-case which is maybe a bit yucky, but it is actually possible to implement race-free. If you have implicit sync.

Quoted 7 lines

There's screen-capture tool in mplayer and obs which capture your compositor by running getfb2 in a loop. It works, and after some initial screaming I realized it does actually work race-free. If you have implicit sync.

I really don't think you can sunset this, as much as you want to. And sunsetting it inconsistently is probably the worst.

For the case where you only have a single ring, as long as it is importing the fb in to egl to read it (which it would need to do to get a linear view), this would still all work

(but I may drop this patch because it is just a micro-optimization and seems to cause more confusion)

BR,
-R

Quoted 69 lines

-Daniel

Quoted 56 lines

-Daniel

Quoted 49 lines

BR,
-R

Quoted 36 lines

Regards,
Lucas

Quoted 32 lines
---
 drivers/gpu/drm/msm/msm_gem_submit.c | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)
drivers/gpu/drm/msm/msm_gem_submit.c +2 -2
index d04c349d8112..b6babc7f9bb8 100644
@@ -283,7 +283,7 @@ static int submit_lock_objects(struct msm_gem_submit *submit)
      return ret;
 }
 
-static int submit_fence_sync(struct msm_gem_submit *submit, bool no_implicit)
+static int submit_fence_sync(struct msm_gem_submit *submit, bool implicit_sync)
 {
      int i, ret = 0;
 
@@ -303,7 +303,7 @@ static int submit_fence_sync(struct msm_gem_submit *submit, bool no_implicit)
                              return ret;
              }
 
-             if (no_implicit)
+             if (!implicit_sync)
                      continue;
 
              ret = msm_gem_sync_object(&msm_obj->base, submit->ring->fctx,
@@ -774,7 +774,8 @@ int msm_ioctl_gem_submit(struct drm_device *dev, void *data,
      if (ret)
              goto out;
 
Quoted 7 lines

drivers/gpu/drm/msm/msm_gem_submit.c int msm_ioctl_gem_submit(struct drm_device *dev, void *data, · line 781

Quoted 6 lines
drivers/gpu/drm/msm/msm_gem_submit.c +2 -1
-     ret = submit_fence_sync(submit, !!(args->flags & MSM_SUBMIT_NO_IMPLICIT));
+     ret = submit_fence_sync(submit, (gpu->nr_rings > 1) &&
+                     !(args->flags & MSM_SUBMIT_NO_IMPLICIT));
      if (ret)
              goto out;
 
Quoted 4 lines

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

Quoted 4 lines

--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch

--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch

#30 lore raw

Re: [PATCH v4 23/23] drm/msm: Don't implicit-sync if only a single ring

Quoted 55 lines · Rob Clark
Quoted 54 lines · Daniel Vetter
Quoted 35 lines · Daniel Vetter
Quoted 21 lines · Rob Clark
Quoted 12 lines · Lucas Stach
Quoted 9 lines · Rob Clark

From: Rob Clark <[email protected]>

If there is only a single ring (no-preemption), everything is FIFO order and there is no need to implicit-sync.

Mesa should probably just always use MSM_SUBMIT_NO_IMPLICIT, as behavior is undefined when fences are not used to synchronize buffer usage across contexts (which is the only case where multiple different priority rings could come into play).

Quoted 7 lines

Really, doesn't this break cross-device implicit sync? Okay, you may not have many peripherals that rely on implicit sync on devices where Adreno is usually found, but it seems rather heavy-handed.

Wouldn't it be better to only ignore fences from your own ring context in the implicit sync, like we do in the common DRM scheduler (drm_sched_dependency_optimized)?

Quoted 6 lines

we already do this.. as was discussed on an earlier iteration of this patchset

But I'm not aware of any other non-gpu related implicit sync use-case (even on imx devices where display is decoupled from gpu).. I'll revert the patch if someone comes up with one, but otherwise lets let the implicit sync baggage die

The thing is, dma_resv won't die, even if implicit sync is dead. We're using internally for activity tracking and memory management. If you don't set these, then we can't share generic code with msm, and I think everyone inventing their own memory management is a bit a mistake.

Quoted 8 lines

Now you only kill the implicit write sync stuff here, but I'm not sure that's worth much since you still install all the read fences for consistency. And if userspace doesn't want to be synced, they can set the flag and do this on their own: I think you should be able to achieve exactly the same thing in mesa.

Aside: If you're worried about overhead, you can do O(1) submit if you manage your ppgtt like amdgpu does.

So just remember a use-case which is maybe a bit yucky, but it is actually possible to implement race-free. If you have implicit sync.

There's screen-capture tool in mplayer and obs which capture your compositor by running getfb2 in a loop. It works, and after some initial screaming I realized it does actually work race-free. If you have implicit sync.

Quoted 6 lines
Quoted 2 lines

I really don't think you can sunset this, as much as you want to. And sunsetting it inconsistently is probably the worst.

For the case where you only have a single ring, as long as it is importing the fb in to egl to read it (which it would need to do to get a linear view), this would still all work

Hm right we still have the implicit sync of the ringbuffer. At least until you add a submit scheduler to msm ...

Quoted 2 lines

(but I may drop this patch because it is just a micro-optimization and seems to cause more confusion)

Yeah I'd say without numbers to justify it it feels a bit on thin ice :-) -Daniel

Quoted 74 lines

BR,
-R

Quoted 69 lines

-Daniel

Quoted 56 lines

-Daniel

Quoted 49 lines

BR,
-R

Quoted 36 lines

Regards,
Lucas

Quoted 32 lines
---
 drivers/gpu/drm/msm/msm_gem_submit.c | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)
drivers/gpu/drm/msm/msm_gem_submit.c +2 -2
index d04c349d8112..b6babc7f9bb8 100644
@@ -283,7 +283,7 @@ static int submit_lock_objects(struct msm_gem_submit *submit)
      return ret;
 }
 
-static int submit_fence_sync(struct msm_gem_submit *submit, bool no_implicit)
+static int submit_fence_sync(struct msm_gem_submit *submit, bool implicit_sync)
 {
      int i, ret = 0;
 
@@ -303,7 +303,7 @@ static int submit_fence_sync(struct msm_gem_submit *submit, bool no_implicit)
                              return ret;
              }
 
-             if (no_implicit)
+             if (!implicit_sync)
                      continue;
 
              ret = msm_gem_sync_object(&msm_obj->base, submit->ring->fctx,
@@ -774,7 +774,8 @@ int msm_ioctl_gem_submit(struct drm_device *dev, void *data,
      if (ret)
              goto out;
 
Quoted 7 lines

drivers/gpu/drm/msm/msm_gem_submit.c int msm_ioctl_gem_submit(struct drm_device *dev, void *data, · line 781

Quoted 6 lines
drivers/gpu/drm/msm/msm_gem_submit.c +2 -1
-     ret = submit_fence_sync(submit, !!(args->flags & MSM_SUBMIT_NO_IMPLICIT));
+     ret = submit_fence_sync(submit, (gpu->nr_rings > 1) &&
+                     !(args->flags & MSM_SUBMIT_NO_IMPLICIT));
      if (ret)
              goto out;
 
Quoted 4 lines

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel

Quoted 4 lines

--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch

--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch

Signature · 3 lines
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch