[PATCH AUTOSEL 6.6 01/75] drm/vc4: don't check if plane->state->fb == state->fb

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



From: Maíra Canal <mcanal@xxxxxxxxxx>

[ Upstream commit 5ee0d47dcf33efd8950b347dcf4d20bab12a3fa9 ]

Currently, when using non-blocking commits, we can see the following
kernel warning:

[  110.908514] ------------[ cut here ]------------
[  110.908529] refcount_t: underflow; use-after-free.
[  110.908620] WARNING: CPU: 0 PID: 1866 at lib/refcount.c:87 refcount_dec_not_one+0xb8/0xc0
[  110.908664] Modules linked in: rfcomm snd_seq_dummy snd_hrtimer snd_seq snd_seq_device cmac algif_hash aes_arm64 aes_generic algif_skcipher af_alg bnep hid_logitech_hidpp vc4 brcmfmac hci_uart btbcm brcmutil bluetooth snd_soc_hdmi_codec cfg80211 cec drm_display_helper drm_dma_helper drm_kms_helper snd_soc_core snd_compress snd_pcm_dmaengine fb_sys_fops sysimgblt syscopyarea sysfillrect raspberrypi_hwmon ecdh_generic ecc rfkill libaes i2c_bcm2835 binfmt_misc joydev snd_bcm2835(C) bcm2835_codec(C) bcm2835_isp(C) v4l2_mem2mem videobuf2_dma_contig snd_pcm bcm2835_v4l2(C) raspberrypi_gpiomem bcm2835_mmal_vchiq(C) videobuf2_v4l2 snd_timer videobuf2_vmalloc videobuf2_memops videobuf2_common snd videodev vc_sm_cma(C) mc hid_logitech_dj uio_pdrv_genirq uio i2c_dev drm fuse dm_mod drm_panel_orientation_quirks backlight ip_tables x_tables ipv6
[  110.909086] CPU: 0 PID: 1866 Comm: kodi.bin Tainted: G         C         6.1.66-v8+ #32
[  110.909104] Hardware name: Raspberry Pi 3 Model B Rev 1.2 (DT)
[  110.909114] pstate: 60000005 (nZCv daif -PAN -UAO -TCO -DIT -SSBS BTYPE=--)
[  110.909132] pc : refcount_dec_not_one+0xb8/0xc0
[  110.909152] lr : refcount_dec_not_one+0xb4/0xc0
[  110.909170] sp : ffffffc00913b9c0
[  110.909177] x29: ffffffc00913b9c0 x28: 000000556969bbb0 x27: 000000556990df60
[  110.909205] x26: 0000000000000002 x25: 0000000000000004 x24: ffffff8004448480
[  110.909230] x23: ffffff800570b500 x22: ffffff802e03a7bc x21: ffffffecfca68c78
[  110.909257] x20: ffffff8002b42000 x19: ffffff802e03a600 x18: 0000000000000000
[  110.909283] x17: 0000000000000011 x16: ffffffffffffffff x15: 0000000000000004
[  110.909308] x14: 0000000000000fff x13: ffffffed577e47e0 x12: 0000000000000003
[  110.909333] x11: 0000000000000000 x10: 0000000000000027 x9 : c912d0d083728c00
[  110.909359] x8 : c912d0d083728c00 x7 : 65646e75203a745f x6 : 746e756f63666572
[  110.909384] x5 : ffffffed579f62ee x4 : ffffffed579eb01e x3 : 0000000000000000
[  110.909409] x2 : 0000000000000000 x1 : ffffffc00913b750 x0 : 0000000000000001
[  110.909434] Call trace:
[  110.909441]  refcount_dec_not_one+0xb8/0xc0
[  110.909461]  vc4_bo_dec_usecnt+0x4c/0x1b0 [vc4]
[  110.909903]  vc4_cleanup_fb+0x44/0x50 [vc4]
[  110.910315]  drm_atomic_helper_cleanup_planes+0x88/0xa4 [drm_kms_helper]
[  110.910669]  vc4_atomic_commit_tail+0x390/0x9dc [vc4]
[  110.911079]  commit_tail+0xb0/0x164 [drm_kms_helper]
[  110.911397]  drm_atomic_helper_commit+0x1d0/0x1f0 [drm_kms_helper]
[  110.911716]  drm_atomic_commit+0xb0/0xdc [drm]
[  110.912569]  drm_mode_atomic_ioctl+0x348/0x4b8 [drm]
[  110.913330]  drm_ioctl_kernel+0xec/0x15c [drm]
[  110.914091]  drm_ioctl+0x24c/0x3b0 [drm]
[  110.914850]  __arm64_sys_ioctl+0x9c/0xd4
[  110.914873]  invoke_syscall+0x4c/0x114
[  110.914897]  el0_svc_common+0xd0/0x118
[  110.914917]  do_el0_svc+0x38/0xd0
[  110.914936]  el0_svc+0x30/0x8c
[  110.914958]  el0t_64_sync_handler+0x84/0xf0
[  110.914979]  el0t_64_sync+0x18c/0x190
[  110.914996] ---[ end trace 0000000000000000 ]---

This happens because, although `prepare_fb` and `cleanup_fb` are
perfectly balanced, we cannot guarantee consistency in the check
plane->state->fb == state->fb. This means that sometimes we can increase
the refcount in `prepare_fb` and don't decrease it in `cleanup_fb`. The
opposite can also be true.

In fact, the struct drm_plane .state shouldn't be accessed directly
but instead, the `drm_atomic_get_new_plane_state()` helper function should
be used. So, we could stick to this check, but using
`drm_atomic_get_new_plane_state()`. But actually, this check is not really
needed. We can increase and decrease the refcount symmetrically without
problems.

This is going to make the code more simple and consistent.

Signed-off-by: Maíra Canal <mcanal@xxxxxxxxxx>
Acked-by: Maxime Ripard <mripard@xxxxxxxxxx>
Link: https://patchwork.freedesktop.org/patch/msgid/20240105175908.242000-1-mcanal@xxxxxxxxxx
Signed-off-by: Sasha Levin <sashal@xxxxxxxxxx>
---
 drivers/gpu/drm/vc4/vc4_plane.c | 5 +----
 1 file changed, 1 insertion(+), 4 deletions(-)

diff --git a/drivers/gpu/drm/vc4/vc4_plane.c b/drivers/gpu/drm/vc4/vc4_plane.c
index 00e713faecd5a..5948e34f7f813 100644
--- a/drivers/gpu/drm/vc4/vc4_plane.c
+++ b/drivers/gpu/drm/vc4/vc4_plane.c
@@ -1505,9 +1505,6 @@ static int vc4_prepare_fb(struct drm_plane *plane,
 
 	drm_gem_plane_helper_prepare_fb(plane, state);
 
-	if (plane->state->fb == state->fb)
-		return 0;
-
 	return vc4_bo_inc_usecnt(bo);
 }
 
@@ -1516,7 +1513,7 @@ static void vc4_cleanup_fb(struct drm_plane *plane,
 {
 	struct vc4_bo *bo;
 
-	if (plane->state->fb == state->fb || !state->fb)
+	if (!state->fb)
 		return;
 
 	bo = to_vc4_bo(&drm_fb_dma_get_gem_obj(state->fb, 0)->base);
-- 
2.43.0





[Index of Archives]     [Linux Kernel]     [Kernel Development Newbies]     [Linux USB Devel]     [Video for Linux]     [Linux Audio Users]     [Yosemite Hiking]     [Linux Kernel]     [Linux SCSI]

  Powered by Linux