On 6/16/2026 10:14 AM, Huang, Honglei wrote:
On 6/15/2026 7:21 PM, Christian König wrote:
On 6/12/26 17:07, Huang, Honglei wrote:
On 6/12/2026 8:21 PM, Christian König wrote:
...
On the SVM and BO overlap issue: if planning to mirror xe is acceptable?
I'm not sure yet but I don't think so.
We have a mixture of engines which can do page faults (3D, SDMA) and
engines which can't (multimedia).
We most likely will need a way to completely separate the mapping from
each other.
Maybe lower address range for SVM mappings and higher one for BO
mappings? But IDK if that will be sufficient for ROCm.
Yeah, I agreed with you, and it seems like xe use a place holder in bo
tree to avoid the BO and SVM overlap.
I have remove the previous overlap check, cause it is a defensive check
so zero impact on functions. Will use the check method you raised when
it is done in amdgpu vm.
Regards,
Honglei
Hi Christian,
thanks again for the thorough review. I fixed the code locally, and the
series is ready and addresses comments from the last round.
1. amdgpu_svm.h: dropped the AMDGPU_SVM_KMEM_CACHE_* macros; SVM ranges use
plain kzalloc()/kfree().
2. amdgpu_svm.h: dropped the flush_tlb callback; amdgpu_svm_flush_tlb() is
called directly.
3. amdgpu_svm_attr.h: added kerneldoc for the interval tree and list.
4. amdgpu_svm_attr.c: dropped the attr slab cache; uses kzalloc()/kfree().
5. amdgpu_svm_attr.c: squashed the types header into its implementation
patch
6. amdgpu_svm_attr.c: renamed the local "flags" to "unsupported_vm_flags".
7. amdgpu_svm_attr.c: dropped amdgpu_svm_attr_validate_range_vma() from the
fault path; hmm_range_fault() covers it. A lighter sync check is
kept only at the
set/get attr ioctl entry for fix regression.
8. amdgpu_svm_range.h: AMDGPU_SVM_RANGE_DEBUG removed, replaced by a
tracepoint.
9. amdgpu_svm_range.h: dropped the UNMAP_WORK macro.
10. amdgpu_svm_range.c: the fast range valid check is now lock free.
11. amdgpu_svm_range.c: zap_ptes always drains the fence regardless of the
update return value.
12. per IP PTE flags: moved the switch out of the SVM core into
gmc_v9_0.c / gmc_v12_0.c behind a new gmc_funcs->get_svm_pte_flags()
callback.
13. eviction_lock: drm_gpusvm framwork's limitation blocks using the
eviction_lock
Kept as is.
14. SVM/BO overlap: dropped the scan-based amdgpu_svm_attr_check_vm_bo()
helper.
15. This series has been rebased on top of the amdgpu vm fix.
Please let me know if there are any remaining review comments. And the
follow up features xnack-off support, multi GPU, all build directly on
this code, and they are much easier to develop incrementally once there
is a stable baseline to build on rather than a moving out of tree base.
So please point out anything else you'd like addressed.
Regards,
Honglei
Regards,
Christian.
And maybe need some modificaitons in amdgpu vm, may I know the next
step/design?
Regards,
Honglei
