Skip to contentSkip to conversation

drm/gem: Internally test import_attach for imported objects

PATCH Superseded Thomas Zimmermann 4 participants 12 messages Apr 15, 2025 – Apr 16, 2025

Subsystems (3)

v1 of 3 Superseded Latest is v3 →

lore
Share & tools
lore

Thomas Zimmermann

author maintainer
#0 lore raw

PATCH 1/1 drm/gem: Internally test import_attach for imported objects

Test struct drm_gem_object.import_attach to detect imported objects during cleanup. At that point, the imported DMA buffer might have already been released and the dma_buf field is NULL. The object's free callback then does a cleanup as for native objects.

Happens for calls to drm_mode_destroy_dumb_ioctl() that eventually clear the dma_buf field in drm_gem_object_exported_dma_buf_free().

Fixes b57aa47d39e9 ("drm/gem: Test for imported GEM buffers with helper")
Cc Thomas Zimmermann <[email protected]>
Cc Anusha Srivatsa <[email protected]>
Cc Christian König <[email protected]>
Cc Maarten Lankhorst <[email protected]>
Cc Maxime Ripard <[email protected]>
Cc David Airlie <[email protected]>
Cc Simona Vetter <[email protected]>
Cc Sumit Semwal <[email protected]>
Cc "Christian König" <[email protected]>
---
 include/drm/drm_gem.h | 8 +++++++-
 1 file changed, 7 insertions(+), 1 deletion(-)
include/drm/drm_gem.h +7 -1
index 9b71f7a9f3f8..f09b8afcf86d 100644
@@ -589,7 +589,13 @@ static inline bool drm_gem_object_is_shared_for_memory_stats(struct drm_gem_obje
 static inline bool drm_gem_is_imported(const struct drm_gem_object *obj)
 {
 	/* The dma-buf's priv field points to the original GEM object. */
-	return obj->dma_buf && (obj->dma_buf->priv != obj);
+	return (obj->dma_buf && (obj->dma_buf->priv != obj)) ||
+	       /*
+		* TODO: During object release, the dma-buf might already
+		*       be gone. For now keep testing import_attach, but
+		*       this should be removed at some point.
+		*/
+	       obj->import_attach;
 }
 
 #ifdef CONFIG_LOCKDEP
Signature · 1 line
2.49.0
#1 lore raw

Am 15.04.25 um 11:20 schrieb Thomas Zimmermann:

Quoted 4 lines

Test struct drm_gem_object.import_attach to detect imported objects during cleanup. At that point, the imported DMA buffer might have already been released and the dma_buf field is NULL. The object's free callback then does a cleanup as for native objects.

I don't think that this is a good idea.

The DMA-buf is separately reference counted through the import attachment.

Quoted 2 lines

Happens for calls to drm_mode_destroy_dumb_ioctl() that eventually clear the dma_buf field in drm_gem_object_exported_dma_buf_free().

That is for exported DMA-buf and should *NEVER* be used for imported ones.

Regards,
Christian.

Quoted 41 lines
Fixes b57aa47d39e9 ("drm/gem: Test for imported GEM buffers with helper")
Cc Thomas Zimmermann <[email protected]>
Cc Anusha Srivatsa <[email protected]>
Cc Christian König <[email protected]>
Cc Maarten Lankhorst <[email protected]>
Cc Maxime Ripard <[email protected]>
Cc David Airlie <[email protected]>
Cc Simona Vetter <[email protected]>
Cc Sumit Semwal <[email protected]>
Cc "Christian König" <[email protected]>
---
 include/drm/drm_gem.h | 8 +++++++-
 1 file changed, 7 insertions(+), 1 deletion(-)
include/drm/drm_gem.h +7 -1
index 9b71f7a9f3f8..f09b8afcf86d 100644
@@ -589,7 +589,13 @@ static inline bool drm_gem_object_is_shared_for_memory_stats(struct drm_gem_obje
 static inline bool drm_gem_is_imported(const struct drm_gem_object *obj)
 {
 	/* The dma-buf's priv field points to the original GEM object. */
-	return obj->dma_buf && (obj->dma_buf->priv != obj);
+	return (obj->dma_buf && (obj->dma_buf->priv != obj)) ||
+	       /*
+		* TODO: During object release, the dma-buf might already
+		*       be gone. For now keep testing import_attach, but
+		*       this should be removed at some point.
+		*/
+	       obj->import_attach;
 }
 
 #ifdef CONFIG_LOCKDEP
author maintainer
#2 lore raw

Hi

Am 15.04.25 um 11:39 schrieb Christian König:

Quoted 8 lines

Am 15.04.25 um 11:20 schrieb Thomas Zimmermann:

Quoted 4 lines

Test struct drm_gem_object.import_attach to detect imported objects during cleanup. At that point, the imported DMA buffer might have already been released and the dma_buf field is NULL. The object's free callback then does a cleanup as for native objects.

I don't think that this is a good idea.

The DMA-buf is separately reference counted through the import attachment.

I understand that it's not ideal, but testing for import_attach to be set is what we currently do throughout drivers. Putting this behind an interface is already a step forward.

Quoted 4 lines
Quoted 3 lines

Happens for calls to drm_mode_destroy_dumb_ioctl() that eventually clear the dma_buf field in drm_gem_object_exported_dma_buf_free().

That is for exported DMA-buf and should *NEVER* be used for imported ones.

Did you look at the discussion at the Closes tag? Where else could dma-buf be cleared?

Best regards
Thomas

Quoted 44 lines

Regards,
Christian.

Quoted 40 lines
Fixes b57aa47d39e9 ("drm/gem: Test for imported GEM buffers with helper")
Cc Thomas Zimmermann <[email protected]>
Cc Anusha Srivatsa <[email protected]>
Cc Christian König <[email protected]>
Cc Maarten Lankhorst <[email protected]>
Cc Maxime Ripard <[email protected]>
Cc David Airlie <[email protected]>
Cc Simona Vetter <[email protected]>
Cc Sumit Semwal <[email protected]>
Cc "Christian König" <[email protected]>
---
  include/drm/drm_gem.h | 8 +++++++-
  1 file changed, 7 insertions(+), 1 deletion(-)
include/drm/drm_gem.h +7 -1
index 9b71f7a9f3f8..f09b8afcf86d 100644
@@ -589,7 +589,13 @@ static inline bool drm_gem_object_is_shared_for_memory_stats(struct drm_gem_obje
  static inline bool drm_gem_is_imported(const struct drm_gem_object *obj)
  {
  	/* The dma-buf's priv field points to the original GEM object. */
-	return obj->dma_buf && (obj->dma_buf->priv != obj);
+	return (obj->dma_buf && (obj->dma_buf->priv != obj)) ||
+	       /*
+		* TODO: During object release, the dma-buf might already
+		*       be gone. For now keep testing import_attach, but
+		*       this should be removed at some point.
+		*/
+	       obj->import_attach;
  }
 
  #ifdef CONFIG_LOCKDEP
Signature · 7 lines
--
Thomas Zimmermann
Graphics Driver Developer
SUSE Software Solutions Germany GmbH
Frankenstrasse 146, 90461 Nuernberg, Germany
GF: Ivo Totev, Andrew Myers, Andrew McDonald, Boudien Moerman
HRB 36809 (AG Nuernberg)
#3 lore raw

Am 15.04.25 um 12:45 schrieb Thomas Zimmermann:

Quoted 12 lines

Hi

Am 15.04.25 um 11:39 schrieb Christian König:

Quoted 8 lines

Am 15.04.25 um 11:20 schrieb Thomas Zimmermann:

Quoted 4 lines

Test struct drm_gem_object.import_attach to detect imported objects during cleanup. At that point, the imported DMA buffer might have already been released and the dma_buf field is NULL. The object's free callback then does a cleanup as for native objects.

I don't think that this is a good idea.

The DMA-buf is separately reference counted through the import attachment.

Quoted 8 lines

I understand that it's not ideal, but testing for import_attach to be set is what we currently do throughout drivers. Putting this behind an interface is already a step forward.

Quoted 4 lines
Quoted 3 lines

Happens for calls to drm_mode_destroy_dumb_ioctl() that eventually clear the dma_buf field in drm_gem_object_exported_dma_buf_free().

That is for exported DMA-buf and should *NEVER* be used for imported ones.

Did you look at the discussion at the Closes tag? Where else could dma-buf be cleared?

Yeah, I've seen that. The solution is just completely wrong.

See for the export case obj->dma_buf points to the exported DMA-buf and causes a circle dependency when not set to NULL when the last handle is released.

But for the import case obj->dma_buf is actually not that relevant. Instead obj->import_attach->dma_buf should be used.

Regards,
Christian.

Quoted 48 lines

Best regards
Thomas

Quoted 43 lines

Regards,
Christian.

Quoted 39 lines
Fixes b57aa47d39e9 ("drm/gem: Test for imported GEM buffers with helper")
Cc Thomas Zimmermann <[email protected]>
Cc Anusha Srivatsa <[email protected]>
Cc Christian König <[email protected]>
Cc Maarten Lankhorst <[email protected]>
Cc Maxime Ripard <[email protected]>
Cc David Airlie <[email protected]>
Cc Simona Vetter <[email protected]>
Cc Sumit Semwal <[email protected]>
Cc "Christian König" <[email protected]>

---
  include/drm/drm_gem.h | 8 +++++++-

  1 file changed, 7 insertions(+), 1 deletion(-)
include/drm/drm_gem.h +7 -1
index 9b71f7a9f3f8..f09b8afcf86d 100644
@@ -589,7 +589,13 @@ static inline bool drm_gem_object_is_shared_for_memory_stats(struct drm_gem_obje
   static inline bool drm_gem_is_imported(const struct drm_gem_object *obj)
   {
       /* The dma-buf's priv field points to the original GEM object. */
-    return obj->dma_buf && (obj->dma_buf->priv != obj);
+    return (obj->dma_buf && (obj->dma_buf->priv != obj)) ||
+           /*
+        * TODO: During object release, the dma-buf might already
+        *       be gone. For now keep testing import_attach, but
+        *       this should be removed at some point.
+        */
+           obj->import_attach;
   }
     #ifdef CONFIG_LOCKDEP
author maintainer
#4 lore raw

Hi

Am 15.04.25 um 14:19 schrieb Christian König:

Quoted 16 lines

Am 15.04.25 um 12:45 schrieb Thomas Zimmermann:

Quoted 15 lines

Hi

Am 15.04.25 um 11:39 schrieb Christian König:

Quoted 8 lines

Am 15.04.25 um 11:20 schrieb Thomas Zimmermann:

Quoted 4 lines

Test struct drm_gem_object.import_attach to detect imported objects during cleanup. At that point, the imported DMA buffer might have already been released and the dma_buf field is NULL. The object's free callback then does a cleanup as for native objects.

I don't think that this is a good idea.

The DMA-buf is separately reference counted through the import attachment.

I understand that it's not ideal, but testing for import_attach to be set is what we currently do throughout drivers. Putting this behind an interface is already a step forward.

Quoted 2 lines
Quoted 2 lines

Happens for calls to drm_mode_destroy_dumb_ioctl() that eventually clear the dma_buf field in drm_gem_object_exported_dma_buf_free().

Quoted 7 lines
Quoted 2 lines
Quoted 1 line

That is for exported DMA-buf and should *NEVER* be used for imported ones.

Did you look at the discussion at the Closes tag? Where else could dma-buf be cleared?

Yeah, I've seen that. The solution is just completely wrong.

See for the export case obj->dma_buf points to the exported DMA-buf and causes a circle dependency when not set to NULL when the last handle is released.

But for the import case obj->dma_buf is actually not that relevant. Instead obj->import_attach->dma_buf should be used.

So if I understand correctly, the tests in that helper should be done by looking at import_attach->dma_buf.

The long-term goal is to make import_attach optional because its setup requires a DMA-capable device.

Best regards
Thomas

Quoted 49 lines

Regards,
Christian.

Quoted 45 lines

Best regards
Thomas

Quoted 42 lines

Regards,
Christian.

Quoted 39 lines
Fixes b57aa47d39e9 ("drm/gem: Test for imported GEM buffers with helper")
Cc Thomas Zimmermann <[email protected]>
Cc Anusha Srivatsa <[email protected]>
Cc Christian König <[email protected]>
Cc Maarten Lankhorst <[email protected]>
Cc Maxime Ripard <[email protected]>
Cc David Airlie <[email protected]>
Cc Simona Vetter <[email protected]>
Cc Sumit Semwal <[email protected]>
Cc "Christian König" <[email protected]>
---
   include/drm/drm_gem.h | 8 +++++++-
   1 file changed, 7 insertions(+), 1 deletion(-)
include/drm/drm_gem.h +7 -1
index 9b71f7a9f3f8..f09b8afcf86d 100644
@@ -589,7 +589,13 @@ static inline bool drm_gem_object_is_shared_for_memory_stats(struct drm_gem_obje
   static inline bool drm_gem_is_imported(const struct drm_gem_object *obj)
   {
       /* The dma-buf's priv field points to the original GEM object. */
-    return obj->dma_buf && (obj->dma_buf->priv != obj);
+    return (obj->dma_buf && (obj->dma_buf->priv != obj)) ||
+           /*
+        * TODO: During object release, the dma-buf might already
+        *       be gone. For now keep testing import_attach, but
+        *       this should be removed at some point.
+        */
+           obj->import_attach;
   }
     #ifdef CONFIG_LOCKDEP
Signature · 7 lines
--
Thomas Zimmermann
Graphics Driver Developer
SUSE Software Solutions Germany GmbH
Frankenstrasse 146, 90461 Nuernberg, Germany
GF: Ivo Totev, Andrew Myers, Andrew McDonald, Boudien Moerman
HRB 36809 (AG Nuernberg)
Reviewed-by Simona Vetter Reviewed-by Dmitry Osipenko
#5 lore raw

Am 15.04.25 um 14:40 schrieb Thomas Zimmermann:

Quoted 20 lines

Hi

Am 15.04.25 um 14:19 schrieb Christian König:

Quoted 17 lines

Am 15.04.25 um 12:45 schrieb Thomas Zimmermann:

Quoted 16 lines

Hi

Am 15.04.25 um 11:39 schrieb Christian König:

Quoted 8 lines

Am 15.04.25 um 11:20 schrieb Thomas Zimmermann:

Quoted 4 lines

Test struct drm_gem_object.import_attach to detect imported objects during cleanup. At that point, the imported DMA buffer might have already been released and the dma_buf field is NULL. The object's free callback then does a cleanup as for native objects.

I don't think that this is a good idea.

The DMA-buf is separately reference counted through the import attachment.

I understand that it's not ideal, but testing for import_attach to be set is what we currently do throughout drivers. Putting this behind an interface is already a step forward.

Quoted 3 lines
Quoted 2 lines

Happens for calls to drm_mode_destroy_dumb_ioctl() that eventually clear the dma_buf field in drm_gem_object_exported_dma_buf_free().

That is for exported DMA-buf and should *NEVER* be used for imported ones.

Quoted 8 lines
Quoted 6 lines
Quoted 1 line

Did you look at the discussion at the Closes tag? Where else could dma-buf be cleared?

Yeah, I've seen that. The solution is just completely wrong.

See for the export case obj->dma_buf points to the exported DMA-buf and causes a circle dependency when not set to NULL when the last handle is released.

But for the import case obj->dma_buf is actually not that relevant. Instead obj->import_attach->dma_buf should be used.

So if I understand correctly, the tests in that helper should be done by looking at import_attach->dma_buf.

At least in theory yes. IIRC we also set obj->dma_buf to the same value as import_attach->dma_buf for convenient so that code doesn't need to check both.

But it can be that obj->dma_buf is already NULL while the import attachment is still in the process of being cleaned up. So there are a couple of cases where we have to look at obj->import_attach->dma_buf.

Quoted 1 line

The long-term goal is to make import_attach optional because its setup requires a DMA-capable device.

HUI WHAT?

Dmitry and I put quite some effort into being able to create an import_attach without the requirement to have a DMA-capable device.

The last puzzle piece of that landed a month ago in the form of this patch here:

commit b72f66f22c0e39ae6684c43fead774c13db24e73
Author: Christian König <[email protected]> Date:   Tue Feb 11 17:20:53 2025 +0100

    dma-buf: drop caching of sg_tables

    That was purely for the transition from static to dynamic dma-buf handling and can be removed again now.

When you don't create an import attachment the exporter wouldn't know that his buffer is actually used which is usually a quite bad idea.

This is for devices who only want to do a vmap of the buffer, isn't it?

Regards,
Christian.

Quoted 54 lines

Best regards
Thomas

Quoted 49 lines

Regards,
Christian.

Quoted 45 lines

Best regards
Thomas

Quoted 42 lines

Regards,
Christian.

Quoted 39 lines
Fixes b57aa47d39e9 ("drm/gem: Test for imported GEM buffers with helper")
Cc Thomas Zimmermann <[email protected]>
Cc Anusha Srivatsa <[email protected]>
Cc Christian König <[email protected]>
Cc Maarten Lankhorst <[email protected]>
Cc Maxime Ripard <[email protected]>
Cc David Airlie <[email protected]>
Cc Simona Vetter <[email protected]>
Cc Sumit Semwal <[email protected]>
Cc "Christian König" <[email protected]>

---
   include/drm/drm_gem.h | 8 +++++++-

   1 file changed, 7 insertions(+), 1 deletion(-)
include/drm/drm_gem.h +7 -1
index 9b71f7a9f3f8..f09b8afcf86d 100644
@@ -589,7 +589,13 @@ static inline bool drm_gem_object_is_shared_for_memory_stats(struct drm_gem_obje
    static inline bool drm_gem_is_imported(const struct drm_gem_object *obj)
    {
        /* The dma-buf's priv field points to the original GEM object. */
-    return obj->dma_buf && (obj->dma_buf->priv != obj);
+    return (obj->dma_buf && (obj->dma_buf->priv != obj)) ||
+           /*
+        * TODO: During object release, the dma-buf might already
+        *       be gone. For now keep testing import_attach, but
+        *       this should be removed at some point.
+        */
+           obj->import_attach;
    }
      #ifdef CONFIG_LOCKDEP
#6 lore raw
Quoted 26 lines · Christian König

Am 15.04.25 um 14:40 schrieb Thomas Zimmermann:

Quoted 25 lines

Hi

Am 15.04.25 um 14:19 schrieb Christian König:

Quoted 22 lines

Am 15.04.25 um 12:45 schrieb Thomas Zimmermann:

Quoted 10 lines

Hi

Am 15.04.25 um 11:39 schrieb Christian König:

Quoted 7 lines

Am 15.04.25 um 11:20 schrieb Thomas Zimmermann:

Quoted 4 lines

Test struct drm_gem_object.import_attach to detect imported objects during cleanup. At that point, the imported DMA buffer might have already been released and the dma_buf field is NULL. The object's free callback then does a cleanup as for native objects.

I don't think that this is a good idea.

Quoted 7 lines
Quoted 1 line

The DMA-buf is separately reference counted through the import attachment.

I understand that it's not ideal, but testing for import_attach to be set is what we currently do throughout drivers. Putting this behind an interface is already a step forward.

Quoted 3 lines
Quoted 2 lines

Happens for calls to drm_mode_destroy_dumb_ioctl() that eventually clear the dma_buf field in drm_gem_object_exported_dma_buf_free().

That is for exported DMA-buf and should *NEVER* be used for imported ones.

Did you look at the discussion at the Closes tag? Where else could dma-buf be cleared?

Yeah, I've seen that. The solution is just completely wrong.

See for the export case obj->dma_buf points to the exported DMA-buf and causes a circle dependency when not set to NULL when the last handle is released.

Quoted 7 lines
Quoted 3 lines
Quoted 1 line

But for the import case obj->dma_buf is actually not that relevant. Instead obj->import_attach->dma_buf should be used.

So if I understand correctly, the tests in that helper should be done by looking at import_attach->dma_buf.

At least in theory yes. IIRC we also set obj->dma_buf to the same value as import_attach->dma_buf for convenient so that code doesn't need to check both.

Uh, given that we already have a confusion between in the importer and exporter cases I think it'd be better to more strictly separate them than to mix them up even more for convenience. We need more clarity here instead.

Quoted 3 lines

But it can be that obj->dma_buf is already NULL while the import attachment is still in the process of being cleaned up. So there are a couple of cases where we have to look at obj->import_attach->dma_buf.

Yeah this sounds better imo.

Quoted 17 lines
Quoted 1 line

The long-term goal is to make import_attach optional because its setup requires a DMA-capable device.

HUI WHAT?

Dmitry and I put quite some effort into being able to create an import_attach without the requirement to have a DMA-capable device.

The last puzzle piece of that landed a month ago in the form of this patch here:

commit b72f66f22c0e39ae6684c43fead774c13db24e73
Author: Christian König <[email protected]> Date:   Tue Feb 11 17:20:53 2025 +0100

    dma-buf: drop caching of sg_tables

    That was purely for the transition from static to dynamic dma-buf handling and can be removed again now.

Quoted 6 lines

When you don't create an import attachment the exporter wouldn't know that his buffer is actually used which is usually a quite bad idea.

This is im entirely unrelated because ...

Quoted 1 line

This is for devices who only want to do a vmap of the buffer, isn't it?

... it's for the vmap only case, where you might not even have a struct device. Or definitely not a reasonable one, like maybe a faux_bus device or some device on a bus that really doesn't do dma (e.g. spi or i2c), and where hence dma_buf_map_attachment is just something you never ever want to do.

I think we might want to transform obj->import_attach into a union or tagged pointer or something like that, which can cover both cases. And maybe a drm_gem_bo_imported_dma_buf() helper that gives you the dma_buf no matter what if it's imported, or NULL if it's allocated on that drm_device?

Cheers, Sima

Quoted 59 lines

Regards,
Christian.

Quoted 54 lines

Best regards
Thomas

Quoted 49 lines

Regards,
Christian.

Quoted 45 lines

Best regards
Thomas

Quoted 42 lines

Regards,
Christian.

Quoted 39 lines
Fixes b57aa47d39e9 ("drm/gem: Test for imported GEM buffers with helper")
Cc Thomas Zimmermann <[email protected]>
Cc Anusha Srivatsa <[email protected]>
Cc Christian König <[email protected]>
Cc Maarten Lankhorst <[email protected]>
Cc Maxime Ripard <[email protected]>
Cc David Airlie <[email protected]>
Cc Simona Vetter <[email protected]>
Cc Sumit Semwal <[email protected]>
Cc "Christian König" <[email protected]>

---
   include/drm/drm_gem.h | 8 +++++++-

   1 file changed, 7 insertions(+), 1 deletion(-)
include/drm/drm_gem.h +7 -1
index 9b71f7a9f3f8..f09b8afcf86d 100644
@@ -589,7 +589,13 @@ static inline bool drm_gem_object_is_shared_for_memory_stats(struct drm_gem_obje
    static inline bool drm_gem_is_imported(const struct drm_gem_object *obj)
    {
        /* The dma-buf's priv field points to the original GEM object. */
-    return obj->dma_buf && (obj->dma_buf->priv != obj);
+    return (obj->dma_buf && (obj->dma_buf->priv != obj)) ||
+           /*
+        * TODO: During object release, the dma-buf might already
+        *       be gone. For now keep testing import_attach, but
+        *       this should be removed at some point.
+        */
+           obj->import_attach;
    }
      #ifdef CONFIG_LOCKDEP
Signature · 3 lines
Simona Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch
#7 lore raw

Am 15.04.25 um 15:10 schrieb Simona Vetter:

Quoted 6 lines
Quoted 1 line

This is for devices who only want to do a vmap of the buffer, isn't it?

... it's for the vmap only case, where you might not even have a struct device. Or definitely not a reasonable one, like maybe a faux_bus device or some device on a bus that really doesn't do dma (e.g. spi or i2c), and where hence dma_buf_map_attachment is just something you never ever want to do.

Even in that case I would still suggest to at least create an attachment to let the exporter know that somebody is doing something with it's buffer.

That is also important for move notification since you can't do those without an attachment.

BTW: What is keeping a vmap alive after dropping the reservation lock? There is no pinning whatsoever as far as I can see.

Quoted 5 lines

I think we might want to transform obj->import_attach into a union or tagged pointer or something like that, which can cover both cases. And maybe a drm_gem_bo_imported_dma_buf() helper that gives you the dma_buf no matter what if it's imported, or NULL if it's allocated on that drm_device?

Yeah, I had the same idea before as well. Just didn't know if that was something worth looking into.

Regards,
Christian.

Quoted 2 lines

Cheers, Sima

#8 lore raw
Quoted 8 lines · Christian König

Am 15.04.25 um 15:10 schrieb Simona Vetter:

Quoted 6 lines
Quoted 1 line

This is for devices who only want to do a vmap of the buffer, isn't it?

... it's for the vmap only case, where you might not even have a struct device. Or definitely not a reasonable one, like maybe a faux_bus device or some device on a bus that really doesn't do dma (e.g. spi or i2c), and where hence dma_buf_map_attachment is just something you never ever want to do.

Quoted 5 lines

Even in that case I would still suggest to at least create an attachment to let the exporter know that somebody is doing something with it's buffer.

That is also important for move notification since you can't do those without an attachment.

BTW: What is keeping a vmap alive after dropping the reservation lock? There is no pinning whatsoever as far as I can see.

dma_buf_vmap should also pin, or something will go wrong very badly. And that also can tell the exporter what exactly the buffer is used for (kernel cpu access). I really don't think we should mix that up with device access as a dma_buf_attachment.
-Sima

Quoted 14 lines
Quoted 6 lines

I think we might want to transform obj->import_attach into a union or tagged pointer or something like that, which can cover both cases. And maybe a drm_gem_bo_imported_dma_buf() helper that gives you the dma_buf no matter what if it's imported, or NULL if it's allocated on that drm_device?

Yeah, I had the same idea before as well. Just didn't know if that was something worth looking into.

Regards,
Christian.

Quoted 2 lines

Cheers, Sima

Signature · 3 lines
Simona Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch
author maintainer
#9 lore raw
In reply to Christian König

Hi

Am 15.04.25 um 14:52 schrieb Christian König:

Quoted 24 lines

Am 15.04.25 um 14:40 schrieb Thomas Zimmermann:

Quoted 23 lines

Hi

Am 15.04.25 um 14:19 schrieb Christian König:

Quoted 20 lines

Am 15.04.25 um 12:45 schrieb Thomas Zimmermann:

Quoted 10 lines

Hi

Am 15.04.25 um 11:39 schrieb Christian König:

Quoted 7 lines

Am 15.04.25 um 11:20 schrieb Thomas Zimmermann:

Quoted 4 lines

Test struct drm_gem_object.import_attach to detect imported objects during cleanup. At that point, the imported DMA buffer might have already been released and the dma_buf field is NULL. The object's free callback then does a cleanup as for native objects.

I don't think that this is a good idea.

Quoted 7 lines
Quoted 1 line

The DMA-buf is separately reference counted through the import attachment.

I understand that it's not ideal, but testing for import_attach to be set is what we currently do throughout drivers. Putting this behind an interface is already a step forward.

Quoted 3 lines
Quoted 2 lines

Happens for calls to drm_mode_destroy_dumb_ioctl() that eventually clear the dma_buf field in drm_gem_object_exported_dma_buf_free().

That is for exported DMA-buf and should *NEVER* be used for imported ones.

Did you look at the discussion at the Closes tag? Where else could dma-buf be cleared?

Yeah, I've seen that. The solution is just completely wrong.

Quoted 7 lines
Quoted 4 lines
Quoted 3 lines

See for the export case obj->dma_buf points to the exported DMA-buf and causes a circle dependency when not set to NULL when the last handle is released.

But for the import case obj->dma_buf is actually not that relevant. Instead obj->import_attach->dma_buf should be used.

So if I understand correctly, the tests in that helper should be done by looking at import_attach->dma_buf.

At least in theory yes. IIRC we also set obj->dma_buf to the same value as import_attach->dma_buf for convenient so that code doesn't need to check both.

But it can be that obj->dma_buf is already NULL while the import attachment is still in the process of being cleaned up. So there are a couple of cases where we have to look at obj->import_attach->dma_buf.

Ok, that should also solve the problem for now. The point here is to limit exposure of import_attach.

Quoted 17 lines
Quoted 2 lines

The long-term goal is to make import_attach optional because its setup requires a DMA-capable device.

HUI WHAT?

Dmitry and I put quite some effort into being able to create an import_attach without the requirement to have a DMA-capable device.

The last puzzle piece of that landed a month ago in the form of this patch here:

commit b72f66f22c0e39ae6684c43fead774c13db24e73
Author: Christian König <[email protected]> Date:   Tue Feb 11 17:20:53 2025 +0100

    dma-buf: drop caching of sg_tables

    That was purely for the transition from static to dynamic dma-buf handling and can be removed again now.

Quoted 8 lines

When you don't create an import attachment the exporter wouldn't know that his buffer is actually used which is usually a quite bad idea.

This is for devices who only want to do a vmap of the buffer, isn't it?

Yeah, the vmap needs the sgtable, which needs import_attach IIRC. Somewhere in there a DMA device is required. But it's not of high priority as we have workarounds.

Best regards
Thomas

Quoted 55 lines

Regards,
Christian.

Quoted 51 lines

Best regards
Thomas

Quoted 48 lines

Regards,
Christian.

Quoted 45 lines

Best regards
Thomas

Quoted 42 lines

Regards,
Christian.

Quoted 39 lines
Fixes b57aa47d39e9 ("drm/gem: Test for imported GEM buffers with helper")
Cc Thomas Zimmermann <[email protected]>
Cc Anusha Srivatsa <[email protected]>
Cc Christian König <[email protected]>
Cc Maarten Lankhorst <[email protected]>
Cc Maxime Ripard <[email protected]>
Cc David Airlie <[email protected]>
Cc Simona Vetter <[email protected]>
Cc Sumit Semwal <[email protected]>
Cc "Christian König" <[email protected]>
---
    include/drm/drm_gem.h | 8 +++++++-
    1 file changed, 7 insertions(+), 1 deletion(-)
include/drm/drm_gem.h +7 -1
index 9b71f7a9f3f8..f09b8afcf86d 100644
@@ -589,7 +589,13 @@ static inline bool drm_gem_object_is_shared_for_memory_stats(struct drm_gem_obje
    static inline bool drm_gem_is_imported(const struct drm_gem_object *obj)
    {
        /* The dma-buf's priv field points to the original GEM object. */
-    return obj->dma_buf && (obj->dma_buf->priv != obj);
+    return (obj->dma_buf && (obj->dma_buf->priv != obj)) ||
+           /*
+        * TODO: During object release, the dma-buf might already
+        *       be gone. For now keep testing import_attach, but
+        *       this should be removed at some point.
+        */
+           obj->import_attach;
    }
      #ifdef CONFIG_LOCKDEP
Signature · 7 lines
--
Thomas Zimmermann
Graphics Driver Developer
SUSE Software Solutions Germany GmbH
Frankenstrasse 146, 90461 Nuernberg, Germany
GF: Ivo Totev, Andrew Myers, Andrew McDonald, Boudien Moerman
HRB 36809 (AG Nuernberg)
#10 lore raw

Am 15.04.25 um 15:14 schrieb Thomas Zimmermann:

Quoted 20 lines
Quoted 20 lines
Quoted 2 lines

The long-term goal is to make import_attach optional because its setup requires a DMA-capable device.

HUI WHAT?

Dmitry and I put quite some effort into being able to create an import_attach without the requirement to have a DMA-capable device.

The last puzzle piece of that landed a month ago in the form of this patch here:

commit b72f66f22c0e39ae6684c43fead774c13db24e73
Author: Christian König <[email protected]> Date:   Tue Feb 11 17:20:53 2025 +0100

     dma-buf: drop caching of sg_tables
          That was purely for the transition from static to dynamic dma-buf handling and can be removed again now.

Quoted 5 lines
Quoted 3 lines

When you don't create an import attachment the exporter wouldn't know that his buffer is actually used which is usually a quite bad idea.

This is for devices who only want to do a vmap of the buffer, isn't it?

Yeah, the vmap needs the sgtable, which needs import_attach IIRC. Somewhere in there a DMA device is required. But it's not of high priority as we have workarounds.

I've removed the need to create an sgtable just to create an import_attach.

The crux is that exporters sometimes need to distinct between the case when DMA-buf is just used for inter process passing of buffers and inter device passing of buffers. Usually we use the list of attachments for that.

Because of this I though of changing the dma_buf_vmap functions to take an attachment instead of an dma_buf as parameter. That would also resolve the problem is making sure to signal buffer movement through the move notifier.

BTW: Where do we currently pin the buffers to make sure that the pointers returned by dma_buf_vmap() stays valid after dropping the reservation lock?

Regards,
Christian.

Quoted 60 lines

Best regards
Thomas

Quoted 55 lines

Regards,
Christian.

Quoted 51 lines

Best regards
Thomas

Quoted 48 lines

Regards,
Christian.

Quoted 45 lines

Best regards
Thomas

Quoted 42 lines

Regards,
Christian.

Quoted 39 lines
Fixes b57aa47d39e9 ("drm/gem: Test for imported GEM buffers with helper")
Cc Thomas Zimmermann <[email protected]>
Cc Anusha Srivatsa <[email protected]>
Cc Christian König <[email protected]>
Cc Maarten Lankhorst <[email protected]>
Cc Maxime Ripard <[email protected]>
Cc David Airlie <[email protected]>
Cc Simona Vetter <[email protected]>
Cc Sumit Semwal <[email protected]>
Cc "Christian König" <[email protected]>

---
    include/drm/drm_gem.h | 8 +++++++-

    1 file changed, 7 insertions(+), 1 deletion(-)
include/drm/drm_gem.h +7 -1
index 9b71f7a9f3f8..f09b8afcf86d 100644
@@ -589,7 +589,13 @@ static inline bool drm_gem_object_is_shared_for_memory_stats(struct drm_gem_obje
     static inline bool drm_gem_is_imported(const struct drm_gem_object *obj)
     {
         /* The dma-buf's priv field points to the original GEM object. */
-    return obj->dma_buf && (obj->dma_buf->priv != obj);
+    return (obj->dma_buf && (obj->dma_buf->priv != obj)) ||
+           /*
+        * TODO: During object release, the dma-buf might already
+        *       be gone. For now keep testing import_attach, but
+        *       this should be removed at some point.
+        */
+           obj->import_attach;
     }
       #ifdef CONFIG_LOCKDEP
#11 lore raw
In reply to Christian König
Quoted 25 lines · On Tue, 15 Apr 2025 14:19:20 +0200 Christian König <[email protected]>…

Am 15.04.25 um 12:45 schrieb Thomas Zimmermann:

Quoted 24 lines

Hi

Am 15.04.25 um 11:39 schrieb Christian König:

Quoted 10 lines

Am 15.04.25 um 11:20 schrieb Thomas Zimmermann:

Quoted 5 lines

Test struct drm_gem_object.import_attach to detect imported objects during cleanup. At that point, the imported DMA buffer might have already been released and the dma_buf field is NULL. The object's free callback then does a cleanup as for native objects.

I don't think that this is a good idea.

The DMA-buf is separately reference counted through the import attachment.

I understand that it's not ideal, but testing for import_attach to be set is what we currently do throughout drivers. Putting this behind an interface is already a step forward.

Quoted 6 lines
Quoted 4 lines

Happens for calls to drm_mode_destroy_dumb_ioctl() that eventually clear the dma_buf field in drm_gem_object_exported_dma_buf_free().

That is for exported DMA-buf and should *NEVER* be used for imported ones.

Quoted 8 lines
Quoted 2 lines

Did you look at the discussion at the Closes tag? Where else could dma-buf be cleared?

Yeah, I've seen that. The solution is just completely wrong.

See for the export case obj->dma_buf points to the exported DMA-buf and causes a circle dependency when not set to NULL when the last handle is released.

I can confirm this is causing a regression in Panthor, where the driver holds references to GEMs that might have been released by userspace, so it's not just drivers calling drm_mode_destroy_dumb_ioctl(). This leads drm_gem_is_imported() to return inconsistent values depending on when the function is called (before/after the handle creation/destruction). This patch seems to fix the problem, but I can't tell if it's the right thing to do.