On 24.04.23 14:35, Janusz Krzysztofik wrote:
Visible glitches have been observed when running graphics applications on
Linux under Xen hypervisor. Those observations have been confirmed with
failures from kms_pwrite_crc Intel GPU test that verifies data coherency
of DRM frame buffer objects using hardware CRC checksums calculated by
display controllers, exposed to userspace via debugfs. Affected
processing paths have then been identified with new test variants that
mmap the objects using different methods and caching modes.
When running as a Xen PV guest, Linux uses Xen provided PAT configuration
which is different from its native one. In particular, Xen specific PTE
encoding of write-combining caching, likely used by graphics applications,
differs from the Linux default one found among statically defined minimal
set of supported modes. Since Xen defines PTE encoding of the WC mode as
_PAGE_PAT, it no longer belongs to the minimal set, depends on correct
handling of _PAGE_PAT bit, and can be mismatched with write-back caching.
When a user calls mmap() for a DRM buffer object, DRM device specific
.mmap file operation, called from mmap_region(), takes care of setting PTE
encoding bits in a vm_page_prot field of an associated virtual memory area
structure. Unfortunately, _PAGE_PAT bit is not preserved when the vma's
.vm_flags are then applied to .vm_page_prot via vm_set_page_prot(). Bits
to be preserved are determined with _PAGE_CHG_MASK symbol that doesn't
cover _PAGE_PAT. As a consequence, WB caching is requested instead of WC
when running under Xen (also, WP is silently changed to WT, and UC
downgraded to UC_MINUS). When running on bare metal, WC is not affected,
but WP and WT extra modes are unintentionally replaced with WC and UC,
respectively.
WP and WT modes, encoded with _PAGE_PAT bit set, were introduced by commit
281d4078bec3 ("x86: Make page cache mode a real type"). Care was taken
to extend _PAGE_CACHE_MASK symbol with that additional bit, but that
symbol has never been used for identification of bits preserved when
applying page protection flags. Support for all cache modes under Xen,
including the problematic WC mode, was then introduced by commit
47591df50512 ("xen: Support Xen pv-domains using PAT").
Extend bitmask used by pgprot_modify() for selecting bits to be preserved
with _PAGE_PAT bit. However, since that bit can be reused as _PAGE_PSE,
and the _PAGE_CHG_MASK symbol, primarly used by pte_modify(), is likely
intentionally defined with that bit not set, keep that symbol unchanged.
Hmm, I wonder whether pte_mkhuge() shouldn't just set _PAGE_PSE, but use
pgprot_4k_2_large() before doing so.
OTOH a use case like in remove_migration_pte(), where pte_mkhuge() is
directly followed by a call of arch_make_huge_pte(), which in turn is
calling pte_mkhuge() again, would set _always_ the PAT bit.
When running as a Xen PV guest this doesn't matter at all, as large or
huge pages aren't supported there. So clearly something for the MM
maintainers. :-)
Juergen
P.S.: Janusz, nice catch! The QubesOS folks who reported the problem
originally will test your patch under Xen soon.
Attachment:
OpenPGP_0xB0DE9DD628BF132F.asc
Description: OpenPGP public key
Attachment:
OpenPGP_signature
Description: OpenPGP digital signature
