Skip to contentSkip to conversation

drm: etnaviv: Add lockdep annotations for context lock

RFC PATCH In review Guido Günther 3 participants 10 messages Oct 29, 2020 – Oct 30, 2020 1 message not in the archive

Subsystems (2)
lore
Share & tools
lore
#2 lore raw

RFC PATCH 1/2 drm: etnaviv: Add lockdep annotations for context lock

In reply to Guido Günther

etnaviv_iommu_find_iova has it so etnaviv_iommu_insert_exact and lockdep_assert_held should have it as well.

---
 drivers/gpu/drm/etnaviv/etnaviv_mmu.c | 4 ++++
 1 file changed, 4 insertions(+)
drivers/gpu/drm/etnaviv/etnaviv_mmu.c +4 -0
index 3607d348c298..cd599ac04663 100644
@@ -131,6 +131,8 @@ static void etnaviv_iommu_remove_mapping(struct etnaviv_iommu_context *context,
 {
 	struct etnaviv_gem_object *etnaviv_obj = mapping->object;
 
+	lockdep_assert_held(&context->lock);
+
 	etnaviv_iommu_unmap(context, mapping->vram_node.start,
 			    etnaviv_obj->sgt, etnaviv_obj->base.size);
 	drm_mm_remove_node(&mapping->vram_node);
@@ -223,6 +225,8 @@ static int etnaviv_iommu_find_iova(struct etnaviv_iommu_context *context,
 static int etnaviv_iommu_insert_exact(struct etnaviv_iommu_context *context,
 		   struct drm_mm_node *node, size_t size, u64 va)
 {
+	lockdep_assert_held(&context->lock);
+
 	return drm_mm_insert_node_in_range(&context->mm, node, size, 0, 0, va,
 					   va + size, DRM_MM_INSERT_LOWEST);
 }
Signature · 6 lines
2.28.0

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

cover drm: etnaviv: Unmap gems on gem_close

In reply to a message not in the archive

This is meant as a RFC since i'm not sure if this is the right way to fix the problem:

So far the unmap from gpu address space only happened when dropping the last ref in gem_free_object_unlocked, however that is skipped if there's still multiple handles to the same GEM object.

Since userspace (here mesa) in the case of softpin hands back the memory region to the pool of available GPU virtual memory closing the handle via DRM_IOCTL_GEM_CLOSE this can lead to etnaviv_iommu_insert_exact failing later since userspace thinks the vaddr is available while the kernel thinks it isn't making the submit fail like

[E] submit failed: -14 (No space left on device) (etna_cmd_stream_flush:244)

Fix this by unmapping the memory via the .gem_close_object callback.

The patch is against 5.9 and will need to be redone for drm-misc-next due to the conversion to GEM object functions but i'm happy to do that it looks like the right approach.

I can trigger the problem when plugging/unplugging a DP screen driven by DCSS while DSI is driven by mxsfb. It preferably happens with 4k since this allocates bigger chunks.

I also folded in a commit checking for the context->lock in etnaviv_iommu_insert_exact and etnaviv_iommu_remove_mapping too to make it match etnaviv_iommu_find_iova.

To: Lucas Stach <[email protected]>,Russell King <[email protected]>,Christian Gmeiner <[email protected]>,David Airlie <[email protected]>,Daniel Vetter <[email protected]>,[email protected], [email protected]

Guido Günther (2):
drm: etnaviv: Add lockdep annotations for context lock
drm: etnaviv: Unmap gems on gem_close

 drivers/gpu/drm/etnaviv/etnaviv_drv.c |  1 +
 drivers/gpu/drm/etnaviv/etnaviv_drv.h |  1 +
 drivers/gpu/drm/etnaviv/etnaviv_gem.c | 32 +++++++++++++++++++++++++++
 drivers/gpu/drm/etnaviv/etnaviv_mmu.c |  4 ++++
 4 files changed, 38 insertions(+)
Signature · 6 lines
2.28.0

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

RFC PATCH 2/2 drm: etnaviv: Unmap gems on gem_close

So far the unmap from gpu address space only happened when dropping the last ref in gem_free_object_unlocked, however that is skipped if there's still multiple handles to the same GEM object.

Since userspace (here mesa) in the case of softpin hands back the memory region to the pool of available GPU virtual memory closing the handle via DRM_IOCTL_GEM_CLOSE this can lead to etnaviv_iommu_insert_exact failing later since userspace thinks the vaddr is available while the kernel thinks it isn't making the submit fail like

[E] submit failed: -14 (No space left on device) (etna_cmd_stream_flush:244)

Fix this by unmapping the memory via the .gem_close_object callback.

---
 drivers/gpu/drm/etnaviv/etnaviv_drv.c |  1 +
 drivers/gpu/drm/etnaviv/etnaviv_drv.h |  1 +
 drivers/gpu/drm/etnaviv/etnaviv_gem.c | 32 +++++++++++++++++++++++++++
 3 files changed, 34 insertions(+)
drivers/gpu/drm/etnaviv/etnaviv_drv.c +1 -0
index a9a3afaef9a1..fdcc6405068c 100644
@@ -491,6 +491,7 @@ static struct drm_driver etnaviv_drm_driver = {
 	.open               = etnaviv_open,
 	.postclose           = etnaviv_postclose,
 	.gem_free_object_unlocked = etnaviv_gem_free_object,
+	.gem_close_object   = etnaviv_gem_close_object,
 	.gem_vm_ops         = &vm_ops,
 	.prime_handle_to_fd = drm_gem_prime_handle_to_fd,
 	.prime_fd_to_handle = drm_gem_prime_fd_to_handle,
drivers/gpu/drm/etnaviv/etnaviv_drv.h +1 -0
index 4d8dc9236e5f..2226a9af0d63 100644
@@ -65,6 +65,7 @@ int etnaviv_gem_cpu_prep(struct drm_gem_object *obj, u32 op,
 		struct drm_etnaviv_timespec *timeout);
 int etnaviv_gem_cpu_fini(struct drm_gem_object *obj);
 void etnaviv_gem_free_object(struct drm_gem_object *obj);
+void etnaviv_gem_close_object(struct drm_gem_object *obj, struct drm_file *file);
 int etnaviv_gem_new_handle(struct drm_device *dev, struct drm_file *file,
 		u32 size, u32 flags, u32 *handle);
 int etnaviv_gem_new_userptr(struct drm_device *dev, struct drm_file *file,
drivers/gpu/drm/etnaviv/etnaviv_gem.c +32 -0
index f06e19e7be04..5aec4a23c77e 100644
@@ -515,6 +515,38 @@ static const struct etnaviv_gem_ops etnaviv_gem_shmem_ops = {
 	.mmap = etnaviv_gem_mmap_obj,
 };
 
+void etnaviv_gem_close_object(struct drm_gem_object *obj, struct drm_file *unused)
+{
+	struct etnaviv_gem_object *etnaviv_obj = to_etnaviv_bo(obj);
+	struct etnaviv_vram_mapping *mapping, *tmp;
+
+	/* Handle this via etnaviv_gem_free_object */
+	if (obj->handle_count == 1)
+		return;
+
+	WARN_ON(is_active(etnaviv_obj));
+
+	/*
+	 * userspace wants to release the handle so we need to remove
+	 * the mapping from the gpu's virtual address space to stay
+	 * in sync.
+	 */
+	list_for_each_entry_safe(mapping, tmp, &etnaviv_obj->vram_list,
+				 obj_node) {
+		struct etnaviv_iommu_context *context = mapping->context;
+
+		WARN_ON(mapping->use);
+
+		if (context) {
+			etnaviv_iommu_unmap_gem(context, mapping);
+			etnaviv_iommu_context_put(context);
+		}
+
+		list_del(&mapping->obj_node);
+		kfree(mapping);
+	}
+}
+
 void etnaviv_gem_free_object(struct drm_gem_object *obj)
 {
 	struct etnaviv_gem_object *etnaviv_obj = to_etnaviv_bo(obj);
Signature · 6 lines
2.28.0

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

Re: [RFC PATCH 2/2] drm: etnaviv: Unmap gems on gem_close

Hi Guido,

Am Donnerstag, den 29.10.2020, 15:20 +0100 schrieb Guido Günther:

Quoted 68 lines

So far the unmap from gpu address space only happened when dropping the last ref in gem_free_object_unlocked, however that is skipped if there's still multiple handles to the same GEM object.

Since userspace (here mesa) in the case of softpin hands back the memory region to the pool of available GPU virtual memory closing the handle via DRM_IOCTL_GEM_CLOSE this can lead to etnaviv_iommu_insert_exact failing later since userspace thinks the vaddr is available while the kernel thinks it isn't making the submit fail like

[E] submit failed: -14 (No space left on device) (etna_cmd_stream_flush:244)

Fix this by unmapping the memory via the .gem_close_object callback.

---
 drivers/gpu/drm/etnaviv/etnaviv_drv.c |  1 +
 drivers/gpu/drm/etnaviv/etnaviv_drv.h |  1 +
 drivers/gpu/drm/etnaviv/etnaviv_gem.c | 32 +++++++++++++++++++++++++++
 3 files changed, 34 insertions(+)
drivers/gpu/drm/etnaviv/etnaviv_drv.c +1 -0
index a9a3afaef9a1..fdcc6405068c 100644
@@ -491,6 +491,7 @@ static struct drm_driver etnaviv_drm_driver = {
 	.open               = etnaviv_open,
 	.postclose           = etnaviv_postclose,
 	.gem_free_object_unlocked = etnaviv_gem_free_object,
+	.gem_close_object   = etnaviv_gem_close_object,
 	.gem_vm_ops         = &vm_ops,
 	.prime_handle_to_fd = drm_gem_prime_handle_to_fd,
 	.prime_fd_to_handle = drm_gem_prime_fd_to_handle,
drivers/gpu/drm/etnaviv/etnaviv_drv.h +1 -0
index 4d8dc9236e5f..2226a9af0d63 100644
@@ -65,6 +65,7 @@ int etnaviv_gem_cpu_prep(struct drm_gem_object *obj, u32 op,
 		struct drm_etnaviv_timespec *timeout);
 int etnaviv_gem_cpu_fini(struct drm_gem_object *obj);
 void etnaviv_gem_free_object(struct drm_gem_object *obj);
+void etnaviv_gem_close_object(struct drm_gem_object *obj, struct drm_file *file);
 int etnaviv_gem_new_handle(struct drm_device *dev, struct drm_file *file,
 		u32 size, u32 flags, u32 *handle);
 int etnaviv_gem_new_userptr(struct drm_device *dev, struct drm_file *file,
drivers/gpu/drm/etnaviv/etnaviv_gem.c +15 -0
index f06e19e7be04..5aec4a23c77e 100644
@@ -515,6 +515,38 @@ static const struct etnaviv_gem_ops etnaviv_gem_shmem_ops = {
 	.mmap = etnaviv_gem_mmap_obj,
 };
 
+void etnaviv_gem_close_object(struct drm_gem_object *obj, struct drm_file *unused)
+{
+	struct etnaviv_gem_object *etnaviv_obj = to_etnaviv_bo(obj);
+	struct etnaviv_vram_mapping *mapping, *tmp;
+
+	/* Handle this via etnaviv_gem_free_object */
+	if (obj->handle_count == 1)
+		return;
+
+	WARN_ON(is_active(etnaviv_obj));
+
+	/*
+	 * userspace wants to release the handle so we need to remove
+	 * the mapping from the gpu's virtual address space to stay
+	 * in sync.

drivers/gpu/drm/etnaviv/etnaviv_gem.c static const struct etnaviv_gem_ops etnaviv_gem_shmem_ops = { · line 542

Quoted 10 lines
drivers/gpu/drm/etnaviv/etnaviv_gem.c +10 -0
+	 */
+	list_for_each_entry_safe(mapping, tmp, &etnaviv_obj->vram_list,
+				 obj_node) {
+		struct etnaviv_iommu_context *context = mapping->context;
+
+		WARN_ON(mapping->use);
+
+		if (context) {
+			etnaviv_iommu_unmap_gem(context, mapping);
+			etnaviv_iommu_context_put(context);

I see the issue you are trying to fix here, but this is not a viable fix. While userspace may close the handle, the GPU may still be processing jobs referening the BO, so we can't just unmap it from the address space.

I think what we need to do here is waiting for the current jobs/fences of the BO when we hit the case where userspace tries to assign a new address to the BO. Basically wait for current jobs -> unamp from the address space -> map at new userspace assigned address.

Regards,
Lucas

drivers/gpu/drm/etnaviv/etnaviv_gem.c static const struct etnaviv_gem_ops etnaviv_gem_shmem_ops = { · line 552

Quoted 10 lines
drivers/gpu/drm/etnaviv/etnaviv_gem.c +7 -0
+		}
+
+		list_del(&mapping->obj_node);
+		kfree(mapping);
+	}
+}
+
 void etnaviv_gem_free_object(struct drm_gem_object *obj)
 {
 	struct etnaviv_gem_object *etnaviv_obj = to_etnaviv_bo(obj);

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

#6 lore raw

Re: [RFC PATCH 2/2] drm: etnaviv: Unmap gems on gem_close

Quoted 87 lines · Lucas Stach

Hi Guido,

Am Donnerstag, den 29.10.2020, 15:20 +0100 schrieb Guido Günther:

Quoted 68 lines

So far the unmap from gpu address space only happened when dropping the last ref in gem_free_object_unlocked, however that is skipped if there's still multiple handles to the same GEM object.

Since userspace (here mesa) in the case of softpin hands back the memory region to the pool of available GPU virtual memory closing the handle via DRM_IOCTL_GEM_CLOSE this can lead to etnaviv_iommu_insert_exact failing later since userspace thinks the vaddr is available while the kernel thinks it isn't making the submit fail like

[E] submit failed: -14 (No space left on device) (etna_cmd_stream_flush:244)

Fix this by unmapping the memory via the .gem_close_object callback.

---
 drivers/gpu/drm/etnaviv/etnaviv_drv.c |  1 +
 drivers/gpu/drm/etnaviv/etnaviv_drv.h |  1 +
 drivers/gpu/drm/etnaviv/etnaviv_gem.c | 32 +++++++++++++++++++++++++++
 3 files changed, 34 insertions(+)
drivers/gpu/drm/etnaviv/etnaviv_drv.c +1 -0
index a9a3afaef9a1..fdcc6405068c 100644
@@ -491,6 +491,7 @@ static struct drm_driver etnaviv_drm_driver = {
 	.open               = etnaviv_open,
 	.postclose           = etnaviv_postclose,
 	.gem_free_object_unlocked = etnaviv_gem_free_object,
+	.gem_close_object   = etnaviv_gem_close_object,
 	.gem_vm_ops         = &vm_ops,
 	.prime_handle_to_fd = drm_gem_prime_handle_to_fd,
 	.prime_fd_to_handle = drm_gem_prime_fd_to_handle,
drivers/gpu/drm/etnaviv/etnaviv_drv.h +1 -0
index 4d8dc9236e5f..2226a9af0d63 100644
@@ -65,6 +65,7 @@ int etnaviv_gem_cpu_prep(struct drm_gem_object *obj, u32 op,
 		struct drm_etnaviv_timespec *timeout);
 int etnaviv_gem_cpu_fini(struct drm_gem_object *obj);
 void etnaviv_gem_free_object(struct drm_gem_object *obj);
+void etnaviv_gem_close_object(struct drm_gem_object *obj, struct drm_file *file);
 int etnaviv_gem_new_handle(struct drm_device *dev, struct drm_file *file,
 		u32 size, u32 flags, u32 *handle);
 int etnaviv_gem_new_userptr(struct drm_device *dev, struct drm_file *file,
drivers/gpu/drm/etnaviv/etnaviv_gem.c +15 -0
index f06e19e7be04..5aec4a23c77e 100644
@@ -515,6 +515,38 @@ static const struct etnaviv_gem_ops etnaviv_gem_shmem_ops = {
 	.mmap = etnaviv_gem_mmap_obj,
 };
 
+void etnaviv_gem_close_object(struct drm_gem_object *obj, struct drm_file *unused)
+{
+	struct etnaviv_gem_object *etnaviv_obj = to_etnaviv_bo(obj);
+	struct etnaviv_vram_mapping *mapping, *tmp;
+
+	/* Handle this via etnaviv_gem_free_object */
+	if (obj->handle_count == 1)
+		return;
+
+	WARN_ON(is_active(etnaviv_obj));
+
+	/*
+	 * userspace wants to release the handle so we need to remove
+	 * the mapping from the gpu's virtual address space to stay
+	 * in sync.

drivers/gpu/drm/etnaviv/etnaviv_gem.c static const struct etnaviv_gem_ops etnaviv_gem_shmem_ops = { · line 542

Quoted 10 lines
drivers/gpu/drm/etnaviv/etnaviv_gem.c +10 -0
+	 */
+	list_for_each_entry_safe(mapping, tmp, &etnaviv_obj->vram_list,
+				 obj_node) {
+		struct etnaviv_iommu_context *context = mapping->context;
+
+		WARN_ON(mapping->use);
+
+		if (context) {
+			etnaviv_iommu_unmap_gem(context, mapping);
+			etnaviv_iommu_context_put(context);

I see the issue you are trying to fix here, but this is not a viable fix. While userspace may close the handle, the GPU may still be processing jobs referening the BO, so we can't just unmap it from the address space.

Quoted 4 lines

I think what we need to do here is waiting for the current jobs/fences of the BO when we hit the case where userspace tries to assign a new address to the BO. Basically wait for current jobs -> unamp from the address space -> map at new userspace assigned address.

Yeah was about to say the same. There's two solutions to this:
- let the kernel manage the VA space. This is what amdgpu does in some cases (but still no relocations)
- pipeline the VA/PTE updates in your driver, because userspace has a somewhat hard time figuring out when a buffer is done. Doing that would either at complexity or stalls when the kernel is doing all the tracking already anyway. Minimal fix is to do what Lucas explained above, but importantly with the kernel solution we have the option to fully pipeline everything and avoid stalls. I think this is what everyone else who lets userspace manage VA does in their kernel side.

-Daniel

Quoted 15 lines

Regards,
Lucas

Quoted 10 lines
drivers/gpu/drm/etnaviv/etnaviv_gem.c +7 -0
+		}
+
+		list_del(&mapping->obj_node);
+		kfree(mapping);
+	}
+}
+
 void etnaviv_gem_free_object(struct drm_gem_object *obj)
 {
 	struct etnaviv_gem_object *etnaviv_obj = to_etnaviv_bo(obj);
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
maintainer
#7 lore raw

Re: [RFC PATCH 2/2] drm: etnaviv: Unmap gems on gem_close

Am Donnerstag, den 29.10.2020, 19:20 +0100 schrieb Daniel Vetter:

Quoted 93 lines · Lucas Stach
Quoted 87 lines

Hi Guido,

Am Donnerstag, den 29.10.2020, 15:20 +0100 schrieb Guido Günther:

Quoted 68 lines

So far the unmap from gpu address space only happened when dropping the last ref in gem_free_object_unlocked, however that is skipped if there's still multiple handles to the same GEM object.

Since userspace (here mesa) in the case of softpin hands back the memory region to the pool of available GPU virtual memory closing the handle via DRM_IOCTL_GEM_CLOSE this can lead to etnaviv_iommu_insert_exact failing later since userspace thinks the vaddr is available while the kernel thinks it isn't making the submit fail like

[E] submit failed: -14 (No space left on device) (etna_cmd_stream_flush:244)

Fix this by unmapping the memory via the .gem_close_object callback.

---
 drivers/gpu/drm/etnaviv/etnaviv_drv.c |  1 +
 drivers/gpu/drm/etnaviv/etnaviv_drv.h |  1 +
 drivers/gpu/drm/etnaviv/etnaviv_gem.c | 32 +++++++++++++++++++++++++++
 3 files changed, 34 insertions(+)
drivers/gpu/drm/etnaviv/etnaviv_drv.c +1 -0
index a9a3afaef9a1..fdcc6405068c 100644
@@ -491,6 +491,7 @@ static struct drm_driver etnaviv_drm_driver = {
 	.open               = etnaviv_open,
 	.postclose           = etnaviv_postclose,
 	.gem_free_object_unlocked = etnaviv_gem_free_object,
+	.gem_close_object   = etnaviv_gem_close_object,
 	.gem_vm_ops         = &vm_ops,
 	.prime_handle_to_fd = drm_gem_prime_handle_to_fd,
 	.prime_fd_to_handle = drm_gem_prime_fd_to_handle,
drivers/gpu/drm/etnaviv/etnaviv_drv.h +1 -0
index 4d8dc9236e5f..2226a9af0d63 100644
@@ -65,6 +65,7 @@ int etnaviv_gem_cpu_prep(struct drm_gem_object *obj, u32 op,
 		struct drm_etnaviv_timespec *timeout);
 int etnaviv_gem_cpu_fini(struct drm_gem_object *obj);
 void etnaviv_gem_free_object(struct drm_gem_object *obj);
+void etnaviv_gem_close_object(struct drm_gem_object *obj, struct drm_file *file);
 int etnaviv_gem_new_handle(struct drm_device *dev, struct drm_file *file,
 		u32 size, u32 flags, u32 *handle);
 int etnaviv_gem_new_userptr(struct drm_device *dev, struct drm_file *file,
drivers/gpu/drm/etnaviv/etnaviv_gem.c +15 -0
index f06e19e7be04..5aec4a23c77e 100644
@@ -515,6 +515,38 @@ static const struct etnaviv_gem_ops etnaviv_gem_shmem_ops = {
 	.mmap = etnaviv_gem_mmap_obj,
 };
 
+void etnaviv_gem_close_object(struct drm_gem_object *obj, struct drm_file *unused)
+{
+	struct etnaviv_gem_object *etnaviv_obj = to_etnaviv_bo(obj);
+	struct etnaviv_vram_mapping *mapping, *tmp;
+
+	/* Handle this via etnaviv_gem_free_object */
+	if (obj->handle_count == 1)
+		return;
+
+	WARN_ON(is_active(etnaviv_obj));
+
+	/*
+	 * userspace wants to release the handle so we need to remove
+	 * the mapping from the gpu's virtual address space to stay
+	 * in sync.

drivers/gpu/drm/etnaviv/etnaviv_gem.c static const struct etnaviv_gem_ops etnaviv_gem_shmem_ops = { · line 542

Quoted 10 lines
drivers/gpu/drm/etnaviv/etnaviv_gem.c +10 -0
+	 */
+	list_for_each_entry_safe(mapping, tmp, &etnaviv_obj->vram_list,
+				 obj_node) {
+		struct etnaviv_iommu_context *context = mapping->context;
+
+		WARN_ON(mapping->use);
+
+		if (context) {
+			etnaviv_iommu_unmap_gem(context, mapping);
+			etnaviv_iommu_context_put(context);

I see the issue you are trying to fix here, but this is not a viable fix. While userspace may close the handle, the GPU may still be processing jobs referening the BO, so we can't just unmap it from the address space.

Quoted 4 lines

I think what we need to do here is waiting for the current jobs/fences of the BO when we hit the case where userspace tries to assign a new address to the BO. Basically wait for current jobs -> unamp from the address space -> map at new userspace assigned address.

Quoted 10 lines

Yeah was about to say the same. There's two solutions to this:
- let the kernel manage the VA space. This is what amdgpu does in some cases (but still no relocations)
- pipeline the VA/PTE updates in your driver, because userspace has a somewhat hard time figuring out when a buffer is done. Doing that would either at complexity or stalls when the kernel is doing all the tracking already anyway. Minimal fix is to do what Lucas explained above, but importantly with the kernel solution we have the option to fully pipeline everything and avoid stalls. I think this is what everyone else who lets userspace manage VA does in their kernel side.

I thought a bit more about this and the issue is only about userspace reusing regions of its _private_ address space, before they are GPU inactive. So the problem of tracking BO active for figuring out when it is safe to close the handle (and thus forget about the last placement of the BO in the AS) goes from "is this BO globally active anywhere?" to "is this BO still active in one of the jobs I submitted?", which is not hard at all for the userspace to figure out.

With this in mind I think we can reasonably declare the kernel behavior of rejecting the submit as okay and just add the tiny bit of userspace fencing needed for the userspace driver to not close the handle until the last submit referencing this BO has finished.

Regards,
Lucas

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

#8 lore raw

Re: [RFC PATCH 2/2] drm: etnaviv: Unmap gems on gem_close

Quoted 113 lines · Lucas Stach

Am Donnerstag, den 29.10.2020, 19:20 +0100 schrieb Daniel Vetter:

Quoted 93 lines · Lucas Stach
Quoted 87 lines

Hi Guido,

Am Donnerstag, den 29.10.2020, 15:20 +0100 schrieb Guido Günther:

Quoted 68 lines

So far the unmap from gpu address space only happened when dropping the last ref in gem_free_object_unlocked, however that is skipped if there's still multiple handles to the same GEM object.

Since userspace (here mesa) in the case of softpin hands back the memory region to the pool of available GPU virtual memory closing the handle via DRM_IOCTL_GEM_CLOSE this can lead to etnaviv_iommu_insert_exact failing later since userspace thinks the vaddr is available while the kernel thinks it isn't making the submit fail like

[E] submit failed: -14 (No space left on device) (etna_cmd_stream_flush:244)

Fix this by unmapping the memory via the .gem_close_object callback.

---
 drivers/gpu/drm/etnaviv/etnaviv_drv.c |  1 +
 drivers/gpu/drm/etnaviv/etnaviv_drv.h |  1 +
 drivers/gpu/drm/etnaviv/etnaviv_gem.c | 32 +++++++++++++++++++++++++++
 3 files changed, 34 insertions(+)
drivers/gpu/drm/etnaviv/etnaviv_drv.c +1 -0
index a9a3afaef9a1..fdcc6405068c 100644
@@ -491,6 +491,7 @@ static struct drm_driver etnaviv_drm_driver = {
 	.open               = etnaviv_open,
 	.postclose           = etnaviv_postclose,
 	.gem_free_object_unlocked = etnaviv_gem_free_object,
+	.gem_close_object   = etnaviv_gem_close_object,
 	.gem_vm_ops         = &vm_ops,
 	.prime_handle_to_fd = drm_gem_prime_handle_to_fd,
 	.prime_fd_to_handle = drm_gem_prime_fd_to_handle,
drivers/gpu/drm/etnaviv/etnaviv_drv.h +1 -0
index 4d8dc9236e5f..2226a9af0d63 100644
@@ -65,6 +65,7 @@ int etnaviv_gem_cpu_prep(struct drm_gem_object *obj, u32 op,
 		struct drm_etnaviv_timespec *timeout);
 int etnaviv_gem_cpu_fini(struct drm_gem_object *obj);
 void etnaviv_gem_free_object(struct drm_gem_object *obj);
+void etnaviv_gem_close_object(struct drm_gem_object *obj, struct drm_file *file);
 int etnaviv_gem_new_handle(struct drm_device *dev, struct drm_file *file,
 		u32 size, u32 flags, u32 *handle);
 int etnaviv_gem_new_userptr(struct drm_device *dev, struct drm_file *file,
drivers/gpu/drm/etnaviv/etnaviv_gem.c +15 -0
index f06e19e7be04..5aec4a23c77e 100644
@@ -515,6 +515,38 @@ static const struct etnaviv_gem_ops etnaviv_gem_shmem_ops = {
 	.mmap = etnaviv_gem_mmap_obj,
 };
 
+void etnaviv_gem_close_object(struct drm_gem_object *obj, struct drm_file *unused)
+{
+	struct etnaviv_gem_object *etnaviv_obj = to_etnaviv_bo(obj);
+	struct etnaviv_vram_mapping *mapping, *tmp;
+
+	/* Handle this via etnaviv_gem_free_object */
+	if (obj->handle_count == 1)
+		return;
+
+	WARN_ON(is_active(etnaviv_obj));
+
+	/*
+	 * userspace wants to release the handle so we need to remove
+	 * the mapping from the gpu's virtual address space to stay
+	 * in sync.

drivers/gpu/drm/etnaviv/etnaviv_gem.c static const struct etnaviv_gem_ops etnaviv_gem_shmem_ops = { · line 542

Quoted 10 lines
drivers/gpu/drm/etnaviv/etnaviv_gem.c +10 -0
+	 */
+	list_for_each_entry_safe(mapping, tmp, &etnaviv_obj->vram_list,
+				 obj_node) {
+		struct etnaviv_iommu_context *context = mapping->context;
+
+		WARN_ON(mapping->use);
+
+		if (context) {
+			etnaviv_iommu_unmap_gem(context, mapping);
+			etnaviv_iommu_context_put(context);

I see the issue you are trying to fix here, but this is not a viable fix. While userspace may close the handle, the GPU may still be processing jobs referening the BO, so we can't just unmap it from the address space.

Quoted 4 lines

I think what we need to do here is waiting for the current jobs/fences of the BO when we hit the case where userspace tries to assign a new address to the BO. Basically wait for current jobs -> unamp from the address space -> map at new userspace assigned address.

Quoted 10 lines

Yeah was about to say the same. There's two solutions to this:
- let the kernel manage the VA space. This is what amdgpu does in some cases (but still no relocations)
- pipeline the VA/PTE updates in your driver, because userspace has a somewhat hard time figuring out when a buffer is done. Doing that would either at complexity or stalls when the kernel is doing all the tracking already anyway. Minimal fix is to do what Lucas explained above, but importantly with the kernel solution we have the option to fully pipeline everything and avoid stalls. I think this is what everyone else who lets userspace manage VA does in their kernel side.

I thought a bit more about this and the issue is only about userspace reusing regions of its _private_ address space, before they are GPU inactive. So the problem of tracking BO active for figuring out when it is safe to close the handle (and thus forget about the last placement of the BO in the AS) goes from "is this BO globally active anywhere?" to "is this BO still active in one of the jobs I submitted?", which is not hard at all for the userspace to figure out.

Quoted 4 lines

With this in mind I think we can reasonably declare the kernel behavior of rejecting the submit as okay and just add the tiny bit of userspace fencing needed for the userspace driver to not close the handle until the last submit referencing this BO has finished.

Hm why don't you have the same issue for shared buffers? Are these handled with relocs?
-Daniel

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
maintainer
#9 lore raw

Re: [RFC PATCH 2/2] drm: etnaviv: Unmap gems on gem_close

Am Freitag, den 30.10.2020, 10:33 +0100 schrieb Daniel Vetter:

Quoted 114 lines · Lucas Stach
Quoted 113 lines

Am Donnerstag, den 29.10.2020, 19:20 +0100 schrieb Daniel Vetter:

Quoted 93 lines · Lucas Stach
Quoted 87 lines

Hi Guido,

Am Donnerstag, den 29.10.2020, 15:20 +0100 schrieb Guido Günther:

Quoted 68 lines

So far the unmap from gpu address space only happened when dropping the last ref in gem_free_object_unlocked, however that is skipped if there's still multiple handles to the same GEM object.

Since userspace (here mesa) in the case of softpin hands back the memory region to the pool of available GPU virtual memory closing the handle via DRM_IOCTL_GEM_CLOSE this can lead to etnaviv_iommu_insert_exact failing later since userspace thinks the vaddr is available while the kernel thinks it isn't making the submit fail like

[E] submit failed: -14 (No space left on device) (etna_cmd_stream_flush:244)

Fix this by unmapping the memory via the .gem_close_object callback.

---
 drivers/gpu/drm/etnaviv/etnaviv_drv.c |  1 +
 drivers/gpu/drm/etnaviv/etnaviv_drv.h |  1 +
 drivers/gpu/drm/etnaviv/etnaviv_gem.c | 32 +++++++++++++++++++++++++++
 3 files changed, 34 insertions(+)
drivers/gpu/drm/etnaviv/etnaviv_drv.c +1 -0
index a9a3afaef9a1..fdcc6405068c 100644
@@ -491,6 +491,7 @@ static struct drm_driver etnaviv_drm_driver = {
 	.open               = etnaviv_open,
 	.postclose           = etnaviv_postclose,
 	.gem_free_object_unlocked = etnaviv_gem_free_object,
+	.gem_close_object   = etnaviv_gem_close_object,
 	.gem_vm_ops         = &vm_ops,
 	.prime_handle_to_fd = drm_gem_prime_handle_to_fd,
 	.prime_fd_to_handle = drm_gem_prime_fd_to_handle,
drivers/gpu/drm/etnaviv/etnaviv_drv.h +1 -0
index 4d8dc9236e5f..2226a9af0d63 100644
@@ -65,6 +65,7 @@ int etnaviv_gem_cpu_prep(struct drm_gem_object *obj, u32 op,
 		struct drm_etnaviv_timespec *timeout);
 int etnaviv_gem_cpu_fini(struct drm_gem_object *obj);
 void etnaviv_gem_free_object(struct drm_gem_object *obj);
+void etnaviv_gem_close_object(struct drm_gem_object *obj, struct drm_file *file);
 int etnaviv_gem_new_handle(struct drm_device *dev, struct drm_file *file,
 		u32 size, u32 flags, u32 *handle);
 int etnaviv_gem_new_userptr(struct drm_device *dev, struct drm_file *file,
drivers/gpu/drm/etnaviv/etnaviv_gem.c +15 -0
index f06e19e7be04..5aec4a23c77e 100644
@@ -515,6 +515,38 @@ static const struct etnaviv_gem_ops etnaviv_gem_shmem_ops = {
 	.mmap = etnaviv_gem_mmap_obj,
 };
 
+void etnaviv_gem_close_object(struct drm_gem_object *obj, struct drm_file *unused)
+{
+	struct etnaviv_gem_object *etnaviv_obj = to_etnaviv_bo(obj);
+	struct etnaviv_vram_mapping *mapping, *tmp;
+
+	/* Handle this via etnaviv_gem_free_object */
+	if (obj->handle_count == 1)
+		return;
+
+	WARN_ON(is_active(etnaviv_obj));
+
+	/*
+	 * userspace wants to release the handle so we need to remove
+	 * the mapping from the gpu's virtual address space to stay
+	 * in sync.

drivers/gpu/drm/etnaviv/etnaviv_gem.c static const struct etnaviv_gem_ops etnaviv_gem_shmem_ops = { · line 542

Quoted 10 lines
drivers/gpu/drm/etnaviv/etnaviv_gem.c +10 -0
+	 */
+	list_for_each_entry_safe(mapping, tmp, &etnaviv_obj->vram_list,
+				 obj_node) {
+		struct etnaviv_iommu_context *context = mapping->context;
+
+		WARN_ON(mapping->use);
+
+		if (context) {
+			etnaviv_iommu_unmap_gem(context, mapping);
+			etnaviv_iommu_context_put(context);

I see the issue you are trying to fix here, but this is not a viable fix. While userspace may close the handle, the GPU may still be processing jobs referening the BO, so we can't just unmap it from the address space.

Quoted 4 lines

I think what we need to do here is waiting for the current jobs/fences of the BO when we hit the case where userspace tries to assign a new address to the BO. Basically wait for current jobs -> unamp from the address space -> map at new userspace assigned address.

Quoted 10 lines

Yeah was about to say the same. There's two solutions to this:
- let the kernel manage the VA space. This is what amdgpu does in some cases (but still no relocations)
- pipeline the VA/PTE updates in your driver, because userspace has a somewhat hard time figuring out when a buffer is done. Doing that would either at complexity or stalls when the kernel is doing all the tracking already anyway. Minimal fix is to do what Lucas explained above, but importantly with the kernel solution we have the option to fully pipeline everything and avoid stalls. I think this is what everyone else who lets userspace manage VA does in their kernel side.

I thought a bit more about this and the issue is only about userspace reusing regions of its _private_ address space, before they are GPU inactive. So the problem of tracking BO active for figuring out when it is safe to close the handle (and thus forget about the last placement of the BO in the AS) goes from "is this BO globally active anywhere?" to "is this BO still active in one of the jobs I submitted?", which is not hard at all for the userspace to figure out.

Quoted 7 lines
Quoted 4 lines

With this in mind I think we can reasonably declare the kernel behavior of rejecting the submit as okay and just add the tiny bit of userspace fencing needed for the userspace driver to not close the handle until the last submit referencing this BO has finished.

Hm why don't you have the same issue for shared buffers? Are these handled with relocs?

What's clashing here is the mapping in the private address space. Shared buffers are mapped separately in the address space of each client. So for the issue of unmapping and reusing the spot in the address space we only care about the buffer being active for that single client, we don't need to care about the buffer potentially still being active with another client, as this won't hurt us here.

Regards,
Lucas

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

#10 lore raw

Re: [RFC PATCH 2/2] drm: etnaviv: Unmap gems on gem_close

In reply to Lucas Stach

Hi,

Quoted 105 lines · Lucas Stach

Am Donnerstag, den 29.10.2020, 19:20 +0100 schrieb Daniel Vetter:

Quoted 93 lines · Lucas Stach
Quoted 87 lines

Hi Guido,

Am Donnerstag, den 29.10.2020, 15:20 +0100 schrieb Guido Günther:

Quoted 68 lines

So far the unmap from gpu address space only happened when dropping the last ref in gem_free_object_unlocked, however that is skipped if there's still multiple handles to the same GEM object.

Since userspace (here mesa) in the case of softpin hands back the memory region to the pool of available GPU virtual memory closing the handle via DRM_IOCTL_GEM_CLOSE this can lead to etnaviv_iommu_insert_exact failing later since userspace thinks the vaddr is available while the kernel thinks it isn't making the submit fail like

[E] submit failed: -14 (No space left on device) (etna_cmd_stream_flush:244)

Fix this by unmapping the memory via the .gem_close_object callback.

---
 drivers/gpu/drm/etnaviv/etnaviv_drv.c |  1 +
 drivers/gpu/drm/etnaviv/etnaviv_drv.h |  1 +
 drivers/gpu/drm/etnaviv/etnaviv_gem.c | 32 +++++++++++++++++++++++++++
 3 files changed, 34 insertions(+)
drivers/gpu/drm/etnaviv/etnaviv_drv.c +1 -0
index a9a3afaef9a1..fdcc6405068c 100644
@@ -491,6 +491,7 @@ static struct drm_driver etnaviv_drm_driver = {
 	.open               = etnaviv_open,
 	.postclose           = etnaviv_postclose,
 	.gem_free_object_unlocked = etnaviv_gem_free_object,
+	.gem_close_object   = etnaviv_gem_close_object,
 	.gem_vm_ops         = &vm_ops,
 	.prime_handle_to_fd = drm_gem_prime_handle_to_fd,
 	.prime_fd_to_handle = drm_gem_prime_fd_to_handle,
drivers/gpu/drm/etnaviv/etnaviv_drv.h +1 -0
index 4d8dc9236e5f..2226a9af0d63 100644
@@ -65,6 +65,7 @@ int etnaviv_gem_cpu_prep(struct drm_gem_object *obj, u32 op,
 		struct drm_etnaviv_timespec *timeout);
 int etnaviv_gem_cpu_fini(struct drm_gem_object *obj);
 void etnaviv_gem_free_object(struct drm_gem_object *obj);
+void etnaviv_gem_close_object(struct drm_gem_object *obj, struct drm_file *file);
 int etnaviv_gem_new_handle(struct drm_device *dev, struct drm_file *file,
 		u32 size, u32 flags, u32 *handle);
 int etnaviv_gem_new_userptr(struct drm_device *dev, struct drm_file *file,
drivers/gpu/drm/etnaviv/etnaviv_gem.c +15 -0
index f06e19e7be04..5aec4a23c77e 100644
@@ -515,6 +515,38 @@ static const struct etnaviv_gem_ops etnaviv_gem_shmem_ops = {
 	.mmap = etnaviv_gem_mmap_obj,
 };
 
+void etnaviv_gem_close_object(struct drm_gem_object *obj, struct drm_file *unused)
+{
+	struct etnaviv_gem_object *etnaviv_obj = to_etnaviv_bo(obj);
+	struct etnaviv_vram_mapping *mapping, *tmp;
+
+	/* Handle this via etnaviv_gem_free_object */
+	if (obj->handle_count == 1)
+		return;
+
+	WARN_ON(is_active(etnaviv_obj));
+
+	/*
+	 * userspace wants to release the handle so we need to remove
+	 * the mapping from the gpu's virtual address space to stay
+	 * in sync.

drivers/gpu/drm/etnaviv/etnaviv_gem.c static const struct etnaviv_gem_ops etnaviv_gem_shmem_ops = { · line 542

Quoted 10 lines
drivers/gpu/drm/etnaviv/etnaviv_gem.c +10 -0
+	 */
+	list_for_each_entry_safe(mapping, tmp, &etnaviv_obj->vram_list,
+				 obj_node) {
+		struct etnaviv_iommu_context *context = mapping->context;
+
+		WARN_ON(mapping->use);
+
+		if (context) {
+			etnaviv_iommu_unmap_gem(context, mapping);
+			etnaviv_iommu_context_put(context);

I see the issue you are trying to fix here, but this is not a viable fix. While userspace may close the handle, the GPU may still be processing jobs referening the BO, so we can't just unmap it from the address space.

Quoted 4 lines

I think what we need to do here is waiting for the current jobs/fences of the BO when we hit the case where userspace tries to assign a new address to the BO. Basically wait for current jobs -> unamp from the address space -> map at new userspace assigned address.

Quoted 10 lines

Yeah was about to say the same. There's two solutions to this:
- let the kernel manage the VA space. This is what amdgpu does in some cases (but still no relocations)
- pipeline the VA/PTE updates in your driver, because userspace has a somewhat hard time figuring out when a buffer is done. Doing that would either at complexity or stalls when the kernel is doing all the tracking already anyway. Minimal fix is to do what Lucas explained above, but importantly with the kernel solution we have the option to fully pipeline everything and avoid stalls. I think this is what everyone else who lets userspace manage VA does in their kernel side.

Quoted 7 lines

I thought a bit more about this and the issue is only about userspace reusing regions of its _private_ address space, before they are GPU inactive. So the problem of tracking BO active for figuring out when it is safe to close the handle (and thus forget about the last placement of the BO in the AS) goes from "is this BO globally active anywhere?" to "is this BO still active in one of the jobs I submitted?", which is not hard at all for the userspace to figure out.

It's about private address space, yes.

Quoted 4 lines

With this in mind I think we can reasonably declare the kernel behavior of rejecting the submit as okay and just add the tiny bit of userspace fencing needed for the userspace driver to not close the handle until the last submit referencing this BO has finished.

Sounds reasonable, i'll have a look.
Cheers,
-- Guido

Quoted 4 lines

Regards,
Lucas

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

maintainer
#3 lore raw
In reply to Guido Günther

Am Donnerstag, den 29.10.2020, 15:20 +0100 schrieb Guido Günther:

Quoted 2 lines

etnaviv_iommu_find_iova has it so etnaviv_iommu_insert_exact and lockdep_assert_held should have it as well.

This sounds reasonable to me. I've added this patch to my etnaviv/next branch.

Regards,
Lucas

Quoted 19 lines
---
 drivers/gpu/drm/etnaviv/etnaviv_mmu.c | 4 ++++
 1 file changed, 4 insertions(+)
drivers/gpu/drm/etnaviv/etnaviv_mmu.c +2 -0
index 3607d348c298..cd599ac04663 100644
@@ -131,6 +131,8 @@ static void etnaviv_iommu_remove_mapping(struct etnaviv_iommu_context *context,
 {
 	struct etnaviv_gem_object *etnaviv_obj = mapping->object;
 
+	lockdep_assert_held(&context->lock);
+
 	etnaviv_iommu_unmap(context, mapping->vram_node.start,
 			    etnaviv_obj->sgt, etnaviv_obj->base.size);
 	drm_mm_remove_node(&mapping->vram_node);

drivers/gpu/drm/etnaviv/etnaviv_mmu.c static int etnaviv_iommu_find_iova(struct etnaviv_iommu_context *context, · line 232

Quoted 8 lines
drivers/gpu/drm/etnaviv/etnaviv_mmu.c +2 -0
@@ -223,6 +225,8 @@ static int etnaviv_iommu_find_iova(struct etnaviv_iommu_context *context,
 static int etnaviv_iommu_insert_exact(struct etnaviv_iommu_context *context,
 		   struct drm_mm_node *node, size_t size, u64 va)
 {
+	lockdep_assert_held(&context->lock);
+
 	return drm_mm_insert_node_in_range(&context->mm, node, size, 0, 0, va,
 					   va + size, DRM_MM_INSERT_LOWEST);
 }

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