[REGRESSION][BISECTED] drm/i915: GPU HANGs on v6.12.20+

[REGRESSION][BISECTED] drm/i915: GPU HANGs on v6.12.20+

[Danyil Demchenko (student)]

TLDR: on kernels v6.12.x i915_gem_mmap_gtt_version() should return 4

Hi,

This patch shipped with v6.12.20 introduces GPU HANGs on Intel UHD Graphics 620 
(8gen igpu)

index 21274aa9bdddc1..c3dabb85796052 100644
--- a/drivers/gpu/drm/i915/gem/i915_gem_mman.c
+++ b/drivers/gpu/drm/i915/gem/i915_gem_mman.c
@@ -164,6 +164,9 @@ static unsigned int tile_row_pages(const struct drm_i915_gem
_object *obj)
  * 4 - Support multiple fault handlers per object depending on object's
  *     backing storage (a.k.a. MMAP_OFFSET).
  *
+ * 5 - Support multiple partial mmaps(mmap part of BO + unmap a offset, multipl
e
+ *     times with different size and offset).
+ *
  * Restrictions:
  *
  *  * snoopable objects cannot be accessed via the GTT. It can cause machine
@@ -191,7 +194,7 @@ static unsigned int tile_row_pages(const struct drm_i915_gem
_object *obj)
  */
 int i915_gem_mmap_gtt_version(void)
 {
-   return 4;
+   return 5;
 }

kernel: i915 0000:00:02.0: [drm] Resetting rcs0 for preemption time out
kernel: i915 0000:00:02.0: [drm] aces[2257] context reset due to GPU hang
kernel: i915 0000:00:02.0: [drm] GPU HANG: ecode 9:1:e757fefe, in aces [2257]
kernel: i915 0000:00:02.0: [drm] Resetting rcs0 for preemption time out
kernel: i915 0000:00:02.0: [drm] aces[2257] context reset due to GPU hang
kernel: i915 0000:00:02.0: [drm] GPU HANG: ecode 9:1:e757fefe, in aces [2257]
kernel: BUG: unable to handle page fault for address: 0000000078280000
kernel: #PF: supervisor read access in kernel mode
kernel: #PF: error_code(0x0000) - not-present page
kernel: PGD 1462e9067 P4D 1462e9067 PUD 0 
kernel: Oops: Oops: 0000 [#1] PREEMPT SMP

GPU hangs if i915_gem_mmap_gtt_version() returns 5. If set to return 4(as it was
 in v6.12.19) no HANG occur. This happens on v6.12.20+ kernels ONLY, v6.6.x, v6.
18.x, v7.1.3 and v6.12.19 are unaffected.

The actual commit introducing this is bfef148f3680e6b9d28e7fca46d9520f80c5e50e, 
but it doesn't seem to affect any other kernel tree.

#regzbot introduced: v6.12.19..v6.12.20

Re: [REGRESSION][BISECTED] drm/i915: GPU HANGs on v6.12.20+

[Greg KH]

On Sun, Jul 12, 2026 at 08:46:08AM +0000, Danyil Demchenko (student) wrote:
> TLDR: on kernels v6.12.x i915_gem_mmap_gtt_version() should return 4
> 
> Hi,
> 
> This patch shipped with v6.12.20 introduces GPU HANGs on Intel UHD Graphics 620 
> (8gen igpu)

What is the git id?  6.12.20 was released well over a year ago, are you
sure that something newer hasn't fixed this already?  What about looking
at 7.1.3?

thanks,

greg k-h

Re: [REGRESSION][BISECTED] drm/i915: GPU HANGs on v6.12.20+

[Danyil Demchenko (student)]

Yes, this issue was fixed in recent versions(or rather it never occurred there). This issue exists in 6.12.95 too.

________________________________________
From: Greg KH <gregkh@linuxfoundation.org>
Sent: Sunday, July 12, 2026 10:48 AM
To: Danyil Demchenko (student)
Cc: jose.souza@intel.com; stable@vger.kernel.org; regressions@lists.linux.dev
Subject: Re: [REGRESSION][BISECTED] drm/i915: GPU HANGs on v6.12.20+

On Sun, Jul 12, 2026 at 08:46:08AM +0000, Danyil Demchenko (student) wrote:
> TLDR: on kernels v6.12.x i915_gem_mmap_gtt_version() should return 4
>
> Hi,
>
> This patch shipped with v6.12.20 introduces GPU HANGs on Intel UHD Graphics 620
> (8gen igpu)

What is the git id?  6.12.20 was released well over a year ago, are you
sure that something newer hasn't fixed this already?  What about looking
at 7.1.3?

thanks,

greg k-h

Re: [REGRESSION][BISECTED] drm/i915: GPU HANGs on v6.12.20+

[Sasha Levin]

> GPU hangs if i915_gem_mmap_gtt_version() returns 5. If set to return 4(as it was
>  in v6.12.19) no HANG occur. This happens on v6.12.20+ kernels ONLY, v6.6.x, v6.
> 18.x, v7.1.3 and v6.12.19 are unaffected.
>
> The actual commit introducing this is bfef148f3680e6b9d28e7fca46d9520f80c5e50e,
> but it doesn't seem to affect any other kernel tree.

Thanks for the report and the bisection. This looks like a missing
prerequisite rather than a bad commit: upstream merged the remap_io_sg()
fix 2e0438f9c3d25e ("drm/i915: ensure segment offset never exceeds
allowed max", v6.14-rc1) before it started advertising GTT mmap version
5, but 6.12.y picked up only the version bump. That leaves userspace
doing multiple partial mmaps against the unfixed remap path, which
matches the wrong-PFN page fault and GPU hangs you're seeing. It also
explains your version matrix: 6.6.x never got the bump, and 6.18.x/7.1.x
have both the bump and the fix in their base.

I've queued 2e0438f9c3d25e for 6.12, which keeps 6.12.y matching
upstream instead of diverging via a revert.

-- 
Thanks,
Sasha