On Thu, Nov 12, 2026 at 02:15:41PM +0000, [email protected] wrote:
> Thank you for your contribution! Sashiko AI review found 2 potential
> issues to consider:
> - [High] drm/amdgpu: unmapped split endpoints in
>   amdgpu_vm_bo_clear_mappings()
> - [Low] drm/amdgpu: missing Fixes trailer in commit message
> --
> 
> commit 193793bc04d0356a7faf15b1ecf5a2a8fdd585ce
> Author: Thadeu Lima de Souza Cascardo <[email protected]>
> 
> > drm/amdgpu: clear the entire mapping when partially unmapping
> >
> > When we partially clear a BO mapping, we only add the partially unmapped
> > region to the freed list. That will update the page table to clear that
> > mapping and leave the other endpoints mapped.
> >
> > However, we move the endpoints to the invalids list. When later closing
> > that BO, it will simply free the mapping and not clear the GPU page
> > tables.
> >
> > Drop the optimization and clear the entire mapping from the page tables.
> > As the endpoints are already in the invalids list, when that BO mapping
> > needs to be accessed, the page tables will be updated.
> >
> > Though that may clear, then write page table entries and even lead to
> > new page table allocations in certain situations, those would already
> > happen in the clear path.
> >
> > When measuring a given workload, it was noticed an extra allocation in
> > less than 10% of the cases, but no noticeable extra latency. In any
> > case, correctness is more important and this fixes the scenario
> > described.
> >
> > Signed-off-by: Thadeu Lima de Souza Cascardo <[email protected]>
> 
> [Severity: Low]
> Since this fixes a page table leak on partial unmap, should it include a
> Fixes: tag for 387f49e54672 ("drm/amdgpu: fix clear_all and replace
> handling in the VM (v2)") or dc54d3d1744d ("drm/amdgpu: implement
> AMDGPU_VA_OP_CLEAR v2")?
> 

Okay, I have just addressed this in a different message.

> > diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c 
> > b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
> > index a99c2ef14be96..808b77083a5e4 100644
> > --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
> > +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
> > @@ -2108,12 +2108,6 @@ int amdgpu_vm_bo_clear_mappings(struct amdgpu_device 
> > *adev,
> >     list_for_each_entry_safe(tmp, next, &removed, list) {
> >             amdgpu_vm_it_remove(tmp, &vm->va);
> >             list_del(&tmp->list);
> > -
> > -           if (tmp->start < saddr)
> > -               tmp->start = saddr;
> > -           if (tmp->last > eaddr)
> > -               tmp->last = eaddr;
> > -
> >             tmp->bo_va = NULL;
> >             list_add(&tmp->list, &vm->freed);
> >             trace_amdgpu_vm_bo_unmap(NULL, tmp);
> 
> [Severity: High]
> Does queuing the unclamped mapping onto vm->freed leave the surviving
> before and after endpoints unmapped in the hardware page tables?
> 
> When amdgpu_vm_clear_freed() processes vm->freed, it now zeroes the PTEs
> for the entire original [tmp->start, tmp->last] range, wiping the PTEs of
> the surviving before and after sub-ranges.  However, two paths fail to
> re-populate those endpoints from bo_va->invalids:
> 
> 1. In amdgpu_vm_bo_clear_mappings(), amdgpu_vm_bo_needs_update() is only
> called when amdgpu_vm_is_bo_always_valid(vm, bo) is true:
> 
> drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c:amdgpu_vm_bo_clear_mappings() {
>     ...
>       if (!list_empty(&before->list)) {
>               struct amdgpu_bo *bo = before->bo_va->base.bo;
>     ...
>               if (amdgpu_vm_is_bo_always_valid(vm, bo) &&
>                   !before->bo_va->base.moved)
>                       amdgpu_vm_bo_needs_update(&before->bo_va->base);
>       }
>     ...
> }
> 
> For normal individual BOs, before->bo_va and after->bo_va stay on
> vm->individual.idle instead of moving to vm->individual.needs_update, so
> amdgpu_vm_handle_moved() and amdgpu_userq_bo_validate() skip them and
> never restore their cleared PTEs unless the BO is explicitly passed in a
> legacy CS bo_list.
> 

Which is exactly what I would expect, so all seem fine here.

> 2. On immediate updates without AMDGPU_VM_DELAY_UPDATE,
> amdgpu_gem_va_update_vm() runs amdgpu_vm_clear_freed() and only calls
> amdgpu_vm_bo_update() on the ioctl target bo_va for MAP or REPLACE:
> 
> drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c:amdgpu_gem_va_update_vm() {
>     ...
>       r = amdgpu_vm_clear_freed(adev, vm, &vm->last_update);
>       if (r)
>               goto error;
> 
>       /* For MAP/REPLACE we also need to update the BO mappings. */
>       if (operation == AMDGPU_VA_OP_MAP ||
>           operation == AMDGPU_VA_OP_REPLACE) {
>               r = amdgpu_vm_bo_update(adev, bo_va, false);
>               if (r)
>                       goto error;
>       }
>     ...
> }
> 
> When userspace partially unmaps via AMDGPU_VA_OP_CLEAR (where bo_va is
> NULL) or splits a mapping belonging to another bo_va via
> AMDGPU_VA_OP_REPLACE, the surviving endpoints are cleared by
> amdgpu_vm_clear_freed() and never re-populated, causing GPU page faults
> when accessed (for instance on usermode queues synchronizing via
> vm_timeline_syncobj_out).
> 

Christian, do you want me to address this case here? It is non trivial as
we don't have the single bo_va as in the case of AMDGPU_VA_OP_MAP or
AMDGPU_VA_OP_REPLACE. Any tips on what to use to test this?

Thanks.
Cascardo.

> -- 
> Sashiko AI review · 
> https://sashiko.dev/#/patchset/20261111-amdgpu_partial_mappings_fix-v1-1-02c5dfa20...@igalia.com?part=1

Reply via email to