Re: [RFC 3/7] drm/xe/svm: Clear CPU_AUTORESET_ACTIVE on first GPU fault

["Yadav, Arvind" <arvind.yadav@intel.com>]

On 21-02-2026 04:03, Matthew Brost wrote:
> On Fri, Feb 20, 2026 at 12:12:32PM -0800, Matthew Brost wrote:
>> On Thu, Feb 19, 2026 at 02:43:08PM +0530, Arvind Yadav wrote:
>>> Clear XE_VMA_CPU_AUTORESET_ACTIVE before installing GPU PTEs for CPU
>>> address mirror VMAs.
>>>
>>> This marks the one-way transition from CPU-only to GPU-touched so munmap
>>> handling can switch from the MADVISE autoreset notifier to the existing
>>> SVM notifier.
>>>
>>> Cc: Matthew Brost <matthew.brost@intel.com>
>>> Cc: Thomas Hellström <thomas.hellstrom@linux.intel.com>
>>> Cc: Himal Prasad Ghimiray <himal.prasad.ghimiray@intel.com>
>>> Signed-off-by: Arvind Yadav <arvind.yadav@intel.com>
>>> ---
>>>   drivers/gpu/drm/xe/xe_svm.c | 10 ++++++++++
>>>   drivers/gpu/drm/xe/xe_vm.h  | 11 +++++++++++
>>>   2 files changed, 21 insertions(+)
>>>
>>> diff --git a/drivers/gpu/drm/xe/xe_svm.c b/drivers/gpu/drm/xe/xe_svm.c
>>> index cda3bf7e2418..b9dbbb245779 100644
>>> --- a/drivers/gpu/drm/xe/xe_svm.c
>>> +++ b/drivers/gpu/drm/xe/xe_svm.c
>>> @@ -1209,6 +1209,9 @@ static int __xe_svm_handle_pagefault(struct xe_vm *vm, struct xe_vma *vma,
>>>   	lockdep_assert_held_write(&vm->lock);
>>>   	xe_assert(vm->xe, xe_vma_is_cpu_addr_mirror(vma));
>>>   
>>> +	/* Invariant: CPU_AUTORESET_ACTIVE cleared before reaching here. */
>>> +	WARN_ON_ONCE(xe_vma_has_cpu_autoreset_active(vma));
>>> +
>>>   	xe_gt_stats_incr(gt, XE_GT_STATS_ID_SVM_PAGEFAULT_COUNT, 1);
>>>   
>>>   retry:
>>> @@ -1360,6 +1363,13 @@ int xe_svm_handle_pagefault(struct xe_vm *vm, struct xe_vma *vma,
>>>   			    bool atomic)
>>>   {
>>>   	int need_vram, ret;
>>> +
>>> +	lockdep_assert_held_write(&vm->lock);
>>> +
>>> +	/* Transition CPU-only -> GPU-touched before installing PTEs. */
>>> +	if (xe_vma_has_cpu_autoreset_active(vma))
>>> +		xe_vma_gpu_touch(vma);
>> I don’t think this will work going forward. I plan on making the fault
>> handler run under vm->lock in read mode [1], and VMA state will only be
>> allowed to be modified under the lockdep constraints in [2], which are
>> vm->lock in write mode or vm->lock in read mode plus the garbage
>> collector lock. Maybe this is fine for now, and we can rework it once
>> [1] and [2] land—most likely by taking the garbage collector mutex
>> introduced in [1] before touching this VMA’s flags.
>>
>> Another issue is what happens if we don’t want to taint the VMA unless
>> we actually fault in a range. It is valid to not find a range to fault
>> in if this is a prefetch, as that fault just gets suppressed on the
>> device. So at a minimum, this needs to be moved to where the function
>> returns zero (i.e., by the out label).
>>
>> [1] https://gitlab.freedesktop.org/mbrost/xe-kernel-driver-svn-perf-6-15-2025/-/commit/08fa2b95800583e804a91caf477f9c30b3440a33#33e7d2d9323cd529c8d587d9d3801e353439d783_181_179
>> [2] https://gitlab.freedesktop.org/mbrost/xe-kernel-driver-svn-perf-6-15-2025/-/commit/08fa2b95800583e804a91caf477f9c30b3440a33#71bf077daec46f3ebd785235e8ecc786681aff99_1112_1128
>>

Good catch on both. I will move the gpu_touch to after 
__xe_svm_handle_pagefault() returns 0 so we donot taint on suppressed 
prefetch faults. For the future locking rework, happy to revisit — 
likely taking the GC mutex there as you mentioned.

>>> +
>>>   retry:
>>>   	need_vram = xe_vma_need_vram_for_atomic(vm->xe, vma, atomic);
>>>   	if (need_vram < 0)
>>> diff --git a/drivers/gpu/drm/xe/xe_vm.h b/drivers/gpu/drm/xe/xe_vm.h
>>> index 7bf400f068ce..3dc549550c91 100644
>>> --- a/drivers/gpu/drm/xe/xe_vm.h
>>> +++ b/drivers/gpu/drm/xe/xe_vm.h
>>> @@ -423,4 +423,15 @@ static inline struct drm_exec *xe_vm_validation_exec(struct xe_vm *vm)
>>>   	((READ_ONCE(tile_present) & ~READ_ONCE(tile_invalidated)) & BIT((tile)->id))
>>>   
>>>   void xe_vma_mem_attr_copy(struct xe_vma_mem_attr *to, struct xe_vma_mem_attr *from);
>>> +
>>> +/**
>>> + * xe_vma_gpu_touch() - Mark VMA as GPU-touched
>>> + * @vma: VMA to mark
>>> + *
>>> + * Clear XE_VMA_CPU_AUTORESET_ACTIVE. Must be done before first GPU PTE install.
>>> + */
>>> +static inline void xe_vma_gpu_touch(struct xe_vma *vma)
>>> +{
>>> +	vma->gpuva.flags &= ~XE_VMA_CPU_AUTORESET_ACTIVE;
>> Thinking out loud — not strictly related to your series, but I think we
>> should route all accesses to vma->gpuva.flags through helpers with
>> lockdep annotations to prove we aren’t violating the rules I mentioned
>> above (right now this would just require vm->lock in write mode).
>>
> Actually I looked at this and exec IOCTLs can set some of the
> vma->gpuva.flags in vm->lock read mode only, so we basically don't have
> any rules for vma->gpuva.flags and it all happens to work by chance. So
> at minimum here let's define some clear rules /w lockdep for
> XE_VMA_CPU_AUTORESET_ACTIVE, if that needs to be moved out
> vma->gpuva.flags for now that would be my preference as we shouldn't
> make the usage of vma->gpuva.flags worse.


I will move it out of vma->gpuva.flags into a dedicated bool 
cpu_autoreset_active in struct xe_vma with explicit lockdep comment 
(write: vm->lock write; read: vm->lock read). 
XE_VMA_CPU_AUTORESET_ACTIVE is kept as a pipeline-only bit in 
op->map.vma_flags to carry state through MAP/REMAP ops, not stored in 
vma->gpuva.flags.

>
> Matt
>
>> Perhaps when I post [1] and [2], I can clean all of that up, or if you
>> want to transition all access to vma->gpuva.flags to helpers (e.g.,
>> xe_vma_write_flags(vma, mask), xe_vma_read_flags(vma, mask),
>> xe_vma_clear_flags(vma, mask)), I wouldn’t complain.


We can add the xe_vma_write_flags/read_flags/clear_flags wrappers too if 
you want that done in this series or prefer to do it alongside your 
locking rework.


Thanks,
Arvind

>>
>> Matt
>>
>>> +}
>>>   #endif
>>> -- 
>>> 2.43.0
>>>