Hi Ray,
On Wed, 21 Oct 2026 10:58:47 +0800 Huang Rui wrote:
> + * Invalidate the GPU view of @range over the notifier window: zap the
> + * affected PTEs and clear the gpu_mapped flag. The range is not removed
> + * here; structural removal if needed happens in the event-end phase.
I think this can leave stale PTEs behind when only part of a range is
invalidated. Take a one-page notifier event on a larger system-memory
range mapped with 4K PTEs. This zaps that one PTE but clears gpu_mapped
for the whole range, and then the event-end callback unmaps DMA for the
whole range. The other PTEs are still valid at that point, but their DMA
mappings are gone. A later notifier won't zap them either, because
gpu_mapped is already false.
Maybe the zap should cover the full range when its DMA mapping is
going away?
> + if (crosses_boundary) {
Another case is GPU_RO going from 1 to 0 over a whole mapped
GPUSVM range. This branch isn't taken, so the PTEs are zapped but the
old DMA_TO_DEVICE mapping stays around. The next get_pages() sees valid
pages and reuses it, so the writable GPU PTE goes in without an
HMM_PFN_REQ_WRITE fault or a DMA_BIDIRECTIONAL remap. If it's a VM_WRITE
MAP_PRIVATE mapping whose CPU PTE still needs COW, I think it can miss
the write fault. So could the old DMA mapping be dropped before the
writable PTE goes in?
I couldn't tell from the cover which drm_gpusvm base this is on, so I
used 671b7825dbfe for both of these. I haven't reproduced either one on
a GPU, though.
Thanks,
Aveline Noir