All of lore.kernel.org
 help / color / mirror / Atom feed
* [PATCH] drm/v3d: Don't free the core's sg_table when PRIME import fails
@ 2026-11-20 11:58 Roy
  2026-11-20 13:21 ` Maíra Canal
  0 siblings, 1 reply; 2+ messages in thread
From: Roy @ 2026-11-20 11:58 UTC (permalink / raw)
  To: mwen, mcanal; +Cc: dri-devel, linux-kernel, Roy, stable

When v3d_bo_create_finish() fails in v3d_prime_import_sg_table(), for
example with -ENOSPC because the V3D address space is full, the object
is freed with drm_gem_shmem_free(). At that point obj->import_attach is
not set yet, since drm_gem_prime_import_dev() only sets it after the
driver callback returns. drm_gem_shmem_release() therefore takes the
non-imported path and treats shmem->sgt as its own: it unmaps it with
the V3D device, calls sg_free_table() and kfree()s it.

That sg_table belongs to drm_gem_prime_import_dev(), which then unmaps
it again through dma_buf_unmap_attachment_unlocked() on its fail_unmap
path. The exporter's unmap runs on the freed table, and the kernel
crashes:

  Unable to handle kernel NULL pointer dereference at virtual address 000000000000001c
  pc : dma_direct_unmap_sg+0x84/0x260
  lr : dma_unmap_sg_attrs+0x60/0x148
  Call trace:
   dma_direct_unmap_sg+0x84/0x260 (P)
   dma_unmap_sg_attrs+0x60/0x148
   drm_gem_unmap_dma_buf+0x38/0x68 [drm]
   dma_buf_unmap_attachment+0x3c/0xa0
   dma_buf_unmap_attachment_unlocked+0x58/0x98
   drm_gem_prime_import_dev+0x1a0/0x1e0 [drm]
   drm_gem_prime_fd_to_handle+0x1a8/0x280 [drm]
   drm_prime_fd_to_handle_ioctl+0x40/0x58 [drm]

Users have hit this on Raspberry Pi 5 with Xorg importing vc4 buffers
into v3d.

Clear shmem->sgt before freeing the object, so that only the PRIME core
unmaps it.

Tested on a Raspberry Pi 5 running 6.18.39 with a small program that
exports a vc4 dumb buffer, fills the V3D address space with BOs until
creation fails with -ENOSPC, then imports the dma-buf into v3d. Without
this patch the import oopses as above. With it, the import fails
cleanly with -ENOSPC (60 runs), BO stats stay at zero, and imports
succeed again once BOs are freed. The fix was tested as an out-of-tree
v3d module built against 6.18.39; this tree was not built for arm64.

The bug was found and the fix written with the help of an AI coding
assistant (Claude); I reproduced and tested it on hardware and reviewed
the change.

Fixes: 40609d4820b2 ("drm/v3d: Use the new shmem helpers to reduce driver boilerplate.")
Cc: stable@vger.kernel.org
Link: https://github.com/raspberrypi/linux/issues/7504
Signed-off-by: Roy <rongh7426@gmail.com>
---
 drivers/gpu/drm/v3d/v3d_bo.c | 7 +++++++
 1 file changed, 7 insertions(+)

diff --git a/drivers/gpu/drm/v3d/v3d_bo.c b/drivers/gpu/drm/v3d/v3d_bo.c
index db95a2f..d0b672d 100644
--- a/drivers/gpu/drm/v3d/v3d_bo.c
+++ b/drivers/gpu/drm/v3d/v3d_bo.c
@@ -185,6 +185,13 @@ v3d_prime_import_sg_table(struct drm_device *dev,
 
 	ret = v3d_bo_create_finish(obj);
 	if (ret) {
+		/*
+		 * The sg_table still belongs to drm_gem_prime_import_dev(),
+		 * which unmaps it when we return an error. import_attach is
+		 * not set yet, so drm_gem_shmem_free() would otherwise unmap
+		 * and free it as if it were ours.
+		 */
+		to_v3d_bo(obj)->base.sgt = NULL;
 		drm_gem_shmem_free(&to_v3d_bo(obj)->base);
 		return ERR_PTR(ret);
 	}
-- 
2.51.1.windows.1


^ permalink raw reply related	[flat|nested] 2+ messages in thread

end of thread, other threads:[~2026-11-20 13:21 UTC | newest]

Thread overview: 2+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2026-11-20 11:58 [PATCH] drm/v3d: Don't free the core's sg_table when PRIME import fails Roy
2026-11-20 13:21 ` Maíra Canal

This is an external index of several public inboxes,
see mirroring instructions on how to clone and mirror
all data and code used by this external index.