[PATCH v4 0/8] drm/xe/mmio_gem: fix fault handler and destroy path

[Matthew Auld <matthew.auld@intel.com>]

Picking up the series from Ilia:
https://lore.kernel.org/intel-xe/20260723161832.137153-1-ilia.levi@intel.com/

This series fixes several issues in xe_mmio_gem, introduced by
1ffcf8b8ae8a ("drm/xe: Support for mmap-ing mmio regions"):
split VMA handling, WB/UC aliasing of the dummy page, a rb-tree
leak on destroy, dummy page accumulation, and use-after-free / MMIO
access after destroy. It also converts the existing PCI barrier mmap
to build on xe_mmio_gem now that the infrastructure is in place.

v4:
- One extra patch for VM_EXEC and DONTCOPY.
- Tweak the commit message in "use write-back mapping for dummy page" wrt x86.
v3: (Matt Auld)
- New patch using the xe_mmio_gem infra for PCI barrier mmap
- New patch for rejecting VMA split
- Using dma_resv lock for synchronization
- SIGBUS instead of dummy page on access after destroy
- Some changes in patch splitting (e.g. the vm_pgoff handling now lives
  in the destroy-flow patch alongside the zap that needs it)
v2:
- New patch 1/5: fix dummy page WB/UC aliasing (Sashiko)
- Patch 2/5: no longer modifies xe_mmio_gem_vm_fault_dummy_page() (handled by 1/5)
- Patch 3/5: unchanged
- Patch 4/5: compute pfn inside scoped_guard
- Patch 5/5: adapt to xe_mmio_gem_vm_fault_dummy_page() signature change, fix "objecthas" typo

-- 
2.55.0