From: Sultan Alsawaf <sultan@xxxxxxxxxxxxxxx>
The retire and active callbacks can run simultaneously, allowing
intel_context_pin() and intel_context_unpin() to run at the same time,
trashing the ring and page tables. In 5.4, this was more noticeable
because intel_ring_unpin() would set ring->vaddr to NULL and cause a
clean NULL-pointer-dereference panic, but in newer kernels the
use-after-free goes unnoticed.
The NULL-pointer-dereference looks like this:
BUG: unable to handle page fault for address: 0000000000003448
RIP: 0010:gen8_emit_flush_render+0x163/0x190
Call Trace:
execlists_request_alloc+0x25/0x40
__i915_request_create+0x1f4/0x2c0
i915_request_create+0x71/0xc0
i915_gem_do_execbuffer+0xb98/0x1a80
? preempt_count_add+0x68/0xa0
? _raw_spin_lock+0x13/0x30
? _raw_spin_unlock+0x16/0x30
i915_gem_execbuffer2_ioctl+0x1de/0x3c0
? i915_gem_busy_ioctl+0x7f/0x1d0
? i915_gem_execbuffer_ioctl+0x2d0/0x2d0
drm_ioctl_kernel+0xb2/0x100
drm_ioctl+0x209/0x360
? i915_gem_execbuffer_ioctl+0x2d0/0x2d0
ksys_ioctl+0x87/0xc0
__x64_sys_ioctl+0x16/0x20
do_syscall_64+0x4e/0x150
entry_SYSCALL_64_after_hwframe+0x44/0xa9
Protect the retire callback with ref->mutex to complement the active
callback and fix the corruption.
Fixes: 12c255b5dad1 ("drm/i915: Provide an i915_active.acquire callback")
Cc: <stable@xxxxxxxxxxxxxxx>
Signed-off-by: Sultan Alsawaf <sultan@xxxxxxxxxxxxxxx>
---
drivers/gpu/drm/i915/i915_active.c | 2 ++
1 file changed, 2 insertions(+)
diff --git a/drivers/gpu/drm/i915/i915_active.c b/drivers/gpu/drm/i915/i915_active.c
index c4048628188a..0478bcf061b5 100644
--- a/drivers/gpu/drm/i915/i915_active.c
+++ b/drivers/gpu/drm/i915/i915_active.c
@@ -148,8 +148,10 @@ __active_retire(struct i915_active *ref)
spin_unlock_irqrestore(&ref->tree_lock, flags);
/* After the final retire, the entire struct may be freed */
+ mutex_lock(&ref->mutex);
if (ref->retire)
ref->retire(ref);
+ mutex_unlock(&ref->mutex);
/* ... except if you wait on it, you must manage your own references! */
wake_up_var(ref);
--
2.26.0
