[AMD Official Use Only - AMD Internal Distribution Only]

> -----Original Message-----
> From: Koenig, Christian <[email protected]>
> Sent: Tuesday, March 24, 2026 9:34 PM
> To: SHANMUGAM, SRINIVASAN <[email protected]>;
> Deucher, Alexander <[email protected]>
> Cc: [email protected]
> Subject: Re: [PATCH v2] drm/amdgpu: Fix NULL bo_va dereference in VA clear
> path
>
> On 3/24/26 16:35, SHANMUGAM, SRINIVASAN wrote:
> > [Public]
> >
> > Hi Christian,
> >
> >> -----Original Message-----
> >> From: Koenig, Christian <[email protected]>
> >> Sent: Tuesday, March 24, 2026 7:43 PM
> >> To: SHANMUGAM, SRINIVASAN <[email protected]>;
> Deucher,
> >> Alexander <[email protected]>
> >> Cc: [email protected]
> >> Subject: Re: [PATCH v2] drm/amdgpu: Fix NULL bo_va dereference in VA
> >> clear path
> >>
> >> On 3/24/26 14:57, Srinivasan Shanmugam wrote:
> >>> amdgpu_gem_va_ioctl() can call amdgpu_gem_va_update_vm() with bo_va
> >>> == NULL for AMDGPU_VA_OP_CLEAR.
> >>>
> >>> CLEAR operates on a VM address range and is not associated with a
> >>> specific BO. In this case, the update helper should perform only
> >>> VM-level updates and must not access BO-specific fields.
> >>>
> >>> Currently, bo_va may be dereferenced in the MAP/REPLACE handling
> >>> paths without explicitly guarding against NULL, which can lead to a
> >>> NULL pointer dereference when CLEAR is processed.
> >>>
> >>> Fix this by making amdgpu_gem_va_update_vm() explicitly handle bo_va
> >>> == NULL:
> >>> - Guard BO-specific accesses with bo_va checks
> >>> - Warn if MAP/REPLACE ever reaches the helper with NULL bo_va
> >>> - Keep VM update path unchanged for CLEAR
> >>>
> >>> This keeps CLEAR on the common update path while ensuring safe
> >>> handling of NULL bo_va.
> >>>
> >>> Crash signature:
> >>> [  325.716062] [IGT] amd_bo: executing [  325.779102]
> >>>
> >>
> ============================================================
> >> ======
> >>> [  325.786483] BUG: KASAN: null-ptr-deref in
> >>> amdgpu_gem_va_ioctl+0x380/0x1130 [amdgpu] [  325.795105] Write of
> >>> size
> >>> 4 at addr 0000000000000000 by task amd_bo/7893 [  325.801997] [
> >>> 325.803595] CPU: 12 UID: 0 PID: 7893 Comm: amd_bo Not tainted
> >>> 6.19.0-1314135.2.zuul.928a0cbbebc74c4f8d5a99a4d0a7ca55 #1
> >>> PREEMPT(voluntary) [  325.803602] Hardware name: TYAN B8021G88V2HR-
> >> 2T/S8021GM2NR-2T, BIOS V1.03.B10 04/01/2019 [  325.803606] Call Trace:
> >>> [  325.803609]  <TASK>
> >>> [  325.803612]  dump_stack_lvl+0x64/0x80 [  325.803623]
> >>> kasan_report+0xb8/0xf0 [  325.803631]  ?
> >>> amdgpu_gem_va_ioctl+0x380/0x1130 [amdgpu] [  325.804427]
> >>> kasan_check_range+0x105/0x1b0 [  325.804432]
> >>> amdgpu_gem_va_ioctl+0x380/0x1130 [amdgpu] [  325.805229]  ?
> >>> __pfx_amdgpu_gem_create_ioctl+0x10/0x10 [amdgpu] [  325.806022]  ?
> >>> __pfx_amdgpu_gem_va_ioctl+0x10/0x10 [amdgpu] [  325.806815]  ?
> >>> __pfx___drm_dev_dbg+0x10/0x10 [drm] [  325.806894]  ?
> >>> __pfx_amdgpu_gem_va_ioctl+0x10/0x10 [amdgpu] [  325.807686]
> >>> drm_ioctl_kernel+0x13d/0x2b0 [drm] [  325.807767]  ?
> >>> __pfx_file_has_perm+0x10/0x10 [  325.807777]  ?
> >>> __pfx_drm_ioctl_kernel+0x10/0x10 [drm] [  325.807857]
> >>> drm_ioctl+0x4be/0xae0 [drm] [  325.807936]  ?
> >>> __pfx_amdgpu_gem_va_ioctl+0x10/0x10 [amdgpu] [  325.808728]  ?
> >>> __pfx_sock_write_iter+0x10/0x10 [  325.808737]  ?
> >>> __pfx_drm_ioctl+0x10/0x10 [drm] [  325.808816]  ?
> >>> ioctl_has_perm.constprop.0.isra.0+0x2ad/0x490
> >>> [  325.808823]  ? __pfx_ioctl_has_perm.constprop.0.isra.0+0x10/0x10
> >>> [  325.808827]  ? _raw_spin_lock_irqsave+0x86/0xd0 [  325.808835]  ?
> >>> __pfx__raw_spin_lock_irqsave+0x10/0x10
> >>> [  325.808841]  amdgpu_drm_ioctl+0xce/0x180 [amdgpu] [  325.809622]
> >>> __x64_sys_ioctl+0x139/0x1c0 [  325.809630]  do_syscall_64+0x64/0x880
> >>> [ 325.809638]  entry_SYSCALL_64_after_hwframe+0x76/0x7e
> >>> [  325.809645] RIP: 0033:0x7f205fd12e1d [  325.809650] Code: 04 25
> >>> 28
> >>> 00 00 00 48 89 45 c8 31 c0 48 8d 45 10 c7 45 b0 10 00 00 00 48 89 45
> >>> b8 48 8d 45 d0 48 89 45 c0 b8 10 00 00 00 0f 05 <89> c2 3d 00 f0 ff
> >>> ff
> >>> 77 1a 48 8b 45 c8 64 48 2b 04 25 28 00 00 00 [  325.809654] RSP:
> >>> 002b:00007ffe9032b510 EFLAGS: 00000246 ORIG_RAX: 0000000000000010
> [
> >>> 325.809660] RAX: ffffffffffffffda RBX: 0000000000000000 RCX:
> >>> 00007f205fd12e1d [  325.809663] RDX: 00007ffe9032b5b0 RSI:
> >>> 00000000c0406448 RDI: 0000000000000006 [  325.809665] RBP:
> >>> 00007ffe9032b560 R08: 0000000100000000 R09: 000000000000000e [
> >>> 325.809668] R10: 0000000000000000 R11: 0000000000000246 R12:
> >>> 00000000c0406448 [  325.809670] R13: 0000000000000006 R14:
> >>> 0000000000001000 R15: 0000000000000001 [  325.809675]  </TASK> [
> >>> 325.809678]
> >>>
> >>
> ============================================================
> >> ======
> >>>
> >>> Fixes: dc54d3d1744d ("drm/amdgpu: implement AMDGPU_VA_OP_CLEAR
> v2")
> >>> Cc: Christian König <[email protected]>
> >>> Cc: Alex Deucher <[email protected]>
> >>> Signed-off-by: Srinivasan Shanmugam <[email protected]>
> >>> ---
> >>>  drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c | 11 +++++++++--
> >>>  1 file changed, 9 insertions(+), 2 deletions(-)
> >>>
> >>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c
> >>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c
> >>> index b0ba2bdaf43a..145cb222d5cf 100644
> >>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c
> >>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c
> >>> @@ -759,9 +759,15 @@ amdgpu_gem_va_update_vm(struct amdgpu_device
> >> *adev,
> >>>     if (r)
> >>>             goto error;
> >>>
> >>> -   /* For MAP/REPLACE we also need to update the BO mappings. */
> >>> +    /* For MAP/REPLACE we also need to update the BO mappings.
> >>> +     * CLEAR operates on the VM address range only and can come in with
> >>> +     * bo_va == NULL.
> >>> +     */
> >>>     if (operation == AMDGPU_VA_OP_MAP ||
> >>>         operation == AMDGPU_VA_OP_REPLACE) {
> >>> +           if (WARN_ON_ONCE(!bo_va))
> >>> +                   goto error;
> >>> +
> >>>             r = amdgpu_vm_bo_update(adev, bo_va, false);
> >>>             if (r)
> >>>                     goto error;
> >>> @@ -772,7 +778,8 @@ amdgpu_gem_va_update_vm(struct amdgpu_device
> >> *adev,
> >>>     if (r)
> >>>             goto error;
> >>>
> >>> -   if ((operation == AMDGPU_VA_OP_MAP ||
> >>> +   if (bo_va &&
> >>> +       (operation == AMDGPU_VA_OP_MAP ||
> >>>          operation == AMDGPU_VA_OP_REPLACE) &&
> >>
> >> Something else must be broken here. We already check operation ==
> >> AMDGPU_VA_OP_MAP or AMDGPU_VA_OP_REPLACE.
> >>
> >> That should be enough to Ensure that bo_va isn't NULL.
> >
> > Thanks for the review and the clarification. please correct me if I am 
> > mistaken.
> >
> > - For MAP/REPLACE, we are working with a real buffer (BO)
>
> Not quite for PRT the BO is NULL.

Thanks for the clarification.

Ah, got it.

So for PRT, bo_va is valid but bo_va->bo is NULL.

>
> > - So bo_va should always be valid in those cases
> > - bo_va should be NULL only for CLEAR or PRT, where no real buffer is
> > used
>
> No, bo_va is a valid pointer for MAP/REPLACE, but for clear it is NULL.
>
> For PRT bo_va->bo is NULL.

Thanks, that helps clarify.

So:
- CLEAR → bo_va = NULL
- PRT → bo_va is valid, but bo_va->bo = NULL
- MAP/REPLACE → bo_va should always be valid

I was incorrectly treating PRT similar to CLEAR earlier.

>
> >
> > Based on this, it seems the issue may not be in
> > amdgpu_gem_va_update_vm(), but earlier in amdgpu_gem_va_ioctl(), in the part
> where we decide:
> >
> > - which buffer to use (abo)
> > - and where it is mapped (bo_va)
> >
> > This is how I currently understand the flow:
> >
> > 1. First, we check the operation and flags
> >
> >    - If it is NOT CLEAR and NOT PRT,
> >      we take the normal path and get the real buffer (abo) from the
> > handle
>
> Ah, I see. Yeah, that is most likely wrong.
>
> We should use the special PRT bo_va in fpriv for that case, otherwise 
> userqueues
> would have quite a problem.

Got it, thanks.

So for the PRT case, we should use fpriv->prt_va instead of treating it like a 
NULL case.

I will fix the ioctl() path to ensure PRT uses prt_va correctly.

Based on this, I will focus on fixing the bo_va selection in 
amdgpu_gem_va_ioctl(),
so that PRT uses prt_va and bo_va is not incorrectly NULL.

Please let me know if this direction looks correct.

Best regards,
Srini

>
> Thanks,
> Christian.

Reply via email to