From: Eric Huang jinhuieric.huang@amd.com
[ Upstream commit 93aa919ca05bec544b17ee9a1bfe394ce6c94bd8 ]
When it only allocates vram without va, which is 0, and a SVM range allocated stays in this range, the vram allocation returns failure. It should be skipped for this case from SVM usage check.
Signed-off-by: Eric Huang jinhuieric.huang@amd.com Reviewed-by: Harish Kasiviswanathan Harish.Kasiviswanathan@amd.com Signed-off-by: Alex Deucher alexander.deucher@amd.com Signed-off-by: Sasha Levin sashal@kernel.org ---
LLM Generated explanations, may be completely bogus:
YES
This is a small, targeted bug fix that prevents erroneous -EADDRINUSE failures when userspace allocates a VRAM buffer without providing a VA (i.e., `va_addr == 0`). The change is confined to the KFD ioctl path and poses minimal regression risk while fixing a real user-visible issue.
What changed - In `drivers/gpu/drm/amd/amdkfd/kfd_chardev.c:1045`, inside `kfd_ioctl_alloc_memory_of_gpu`, the SVM overlap check was amended to skip a special case: - New guard added at `drivers/gpu/drm/amd/amdkfd/kfd_chardev.c:1071`: - `if (!(!args->va_addr && (flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM)) && interval_tree_iter_first(...)) { ... return -EADDRINUSE; }` - Practically, this means the SVM interval-tree overlap check is bypassed only when: - `args->va_addr == 0` (no VA requested), and - `flags` includes `KFD_IOC_ALLOC_MEM_FLAGS_VRAM`. - Previously, the overlap check was unconditional, which could falsely report “Address already allocated by SVM” when VA is 0 (see the surrounding context at `drivers/gpu/drm/amd/amdkfd/kfd_chardev.c:1064-1079`).
Why it’s a bug fix - The commit message accurately describes a failure mode: when allocating VRAM-only without a VA (VA=0) and there exists an SVM range that falls in that [0, size) range, the ioctl incorrectly returns `-EADDRINUSE`. For VRAM-only allocations without a VA, SVM address- range conflicts are irrelevant and should not block allocation. - The code change corrects this by skipping the SVM overlap check for that specific case, avoiding a false-positive error.
Safety and scope - Minimal, localized change: It adds a single conditional guard and comment in one function. No ABI or architectural changes. - Confined to AMD KFD user memory allocation path; does not touch core MM, scheduler, or unrelated GPU subsystems. - Consistency with mapping rules: mapping requires a non-zero VA. In `kfd_mem_attach` (called during mapping), mapping with `mem->va == 0` is rejected (`drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c:858-930`, check at “if (!va) { ... return -EINVAL; }”). This ensures that skipping the SVM check for VA=0 can’t accidentally permit an overlapping SVM GPU-VA mapping later: mapping at VA=0 is inherently invalid and denied. Thus the change strictly avoids a spurious allocation-time error without enabling unsafe mappings. - Flags behavior matches UAPI: `KFD_IOC_ALLOC_MEM_FLAGS_VRAM` is intended for VRAM allocations (`include/uapi/linux/kfd_ioctl.h:407`). VRAM-only allocations with VA=0 are valid for certain use cases (e.g., export or CPU-visible VRAM on large BAR), and should not be blocked by SVM interval checks.
Stable backport criteria - Fixes a real bug affecting users (spurious -EADDRINUSE on valid VRAM- only allocations). - Change is small and contained, with clear intent and low regression risk. - No new features or architectural shifts. - Touches only driver code in a single path (`kfd_ioctl_alloc_memory_of_gpu`), no widespread side effects.
Conclusion - This is a clear, minimal bug fix that prevents erroneous allocation failures and aligns with the mapping semantics already enforced elsewhere. It is suitable for stable backport.
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 7 ++++++- 1 file changed, 6 insertions(+), 1 deletion(-)
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c index 79ed3be63d0dd..43115a3744694 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_chardev.c @@ -1070,7 +1070,12 @@ static int kfd_ioctl_alloc_memory_of_gpu(struct file *filep, svm_range_list_lock_and_flush_work(&p->svms, current->mm); mutex_lock(&p->svms.lock); mmap_write_unlock(current->mm); - if (interval_tree_iter_first(&p->svms.objects, + + /* Skip a special case that allocates VRAM without VA, + * VA will be invalid of 0. + */ + if (!(!args->va_addr && (flags & KFD_IOC_ALLOC_MEM_FLAGS_VRAM)) && + interval_tree_iter_first(&p->svms.objects, args->va_addr >> PAGE_SHIFT, (args->va_addr + args->size - 1) >> PAGE_SHIFT)) { pr_err("Address: 0x%llx already allocated by SVM\n",