On kernels built with support for transparent huge pages, different CPUs
can access the PMD concurrently due to e.g. fast GUP or page_vma_mapped_walk
and they must take care to use READ_ONCE to avoid value tearing or caching
of stale values by the compiler. Unfortunately, these functions call into
our pgtable macros, which don't use READ_ONCE, and compiler caching has
been observed to cause the following crash during ext4 writeback:

PC is at check_pte+0x20/0x170
LR is at page_vma_mapped_walk+0x2e0/0x540
[...]
Process doio (pid: 2463, stack limit = 0xffff00000f2e8000)
Call trace:
[<ffff000008233328>] check_pte+0x20/0x170
[<ffff000008233758>] page_vma_mapped_walk+0x2e0/0x540
[<ffff000008234adc>] page_mkclean_one+0xac/0x278
[<ffff000008234d98>] rmap_walk_file+0xf0/0x238
[<ffff000008236e74>] rmap_walk+0x64/0xa0
[<ffff0000082370c8>] page_mkclean+0x90/0xa8
[<ffff0000081f3c64>] clear_page_dirty_for_io+0x84/0x2a8
[<ffff00000832f984>] mpage_submit_page+0x34/0x98
[<ffff00000832fb4c>] mpage_process_page_bufs+0x164/0x170
[<ffff00000832fc8c>] mpage_prepare_extent_to_map+0x134/0x2b8
[<ffff00000833530c>] ext4_writepages+0x484/0xe30
[<ffff0000081f6ab4>] do_writepages+0x44/0xe8
[<ffff0000081e5bd4>] __filemap_fdatawrite_range+0xbc/0x110
[<ffff0000081e5e68>] file_write_and_wait_range+0x48/0xd8
[<ffff000008324310>] ext4_sync_file+0x80/0x4b8
[<ffff0000082bd434>] vfs_fsync_range+0x64/0xc0
[<ffff0000082332b4>] SyS_msync+0x194/0x1e8

This is because page_vma_mapped_walk loads the PMD twice before calling
pte_offset_map: the first time without READ_ONCE (where it gets all zeroes
due to a concurrent pmdp_invalidate) and the second time with READ_ONCE
(where it sees a valid table pointer due to a concurrent pmd_populate).
However, the compiler inlines everything and caches the first value in
a register, which is subsequently used in pte_offset_phys which returns
a junk pointer that is later dereferenced when attempting to access the
relevant pte.

This patch fixes the issue by using READ_ONCE in pte_offset_phys to ensure
that a stale value is not used. Whilst this is a point fix for a known
failure (and simple to backport), a full fix moving all of our page table
accessors over to {READ,WRITE}_ONCE and consistently using READ_ONCE in
page_vma_mapped_walk is in the works for a future kernel release.

Cc: Jon Masters <jcm@redhat.com>
Cc: Timur Tabi <timur@codeaurora.org>
Cc: <stable@vger.kernel.org>
Fixes: f27176cf ("mm: convert page_mkclean_one() to use page_vma_mapped_walk()")
Tested-by: Richard Ruigrok <rruigrok@codeaurora.org>
Signed-off-by: Will Deacon <will.deacon@arm.com>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>