| Message ID | 1508238917-14553-1-git-send-email-will.deacon@arm.com |
|---|---|
| State | Accepted |
| Commit | a7b100953aa33a5bbdc3e5e7f2241b9c0704606e |
| Headers | show |
| Series | mm: page_vma_mapped: ensure pmd is loaded with READ_ONCE outside of lock | expand |
On Tue, Oct 17, 2017 at 12:15:17PM +0100, Will Deacon wrote: > [Commit a7b100953aa33a5bbdc3e5e7f2241b9c0704606e upstream] > > Loading the pmd without holding the pmd_lock exposes us to races with > concurrent updaters of the page tables but, worse still, it also allows > the compiler to cache the pmd value in a register and reuse it later on, > even if we've performed a READ_ONCE in between and seen a more recent > value. > > In the case of page_vma_mapped_walk, this leads to the following crash > when the pmd loaded for the initial pmd_trans_huge check is all zeroes > and a subsequent valid table entry is loaded by check_pmd. We then > proceed into map_pte, but the compiler re-uses the zero entry inside > pte_offset_map, resulting in a junk pointer being installed in > pvmw->pte: > > 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: > check_pte+0x20/0x170 > page_vma_mapped_walk+0x2e0/0x540 > page_mkclean_one+0xac/0x278 > rmap_walk_file+0xf0/0x238 > rmap_walk+0x64/0xa0 > page_mkclean+0x90/0xa8 > clear_page_dirty_for_io+0x84/0x2a8 > mpage_submit_page+0x34/0x98 > mpage_process_page_bufs+0x164/0x170 > mpage_prepare_extent_to_map+0x134/0x2b8 > ext4_writepages+0x484/0xe30 > do_writepages+0x44/0xe8 > __filemap_fdatawrite_range+0xbc/0x110 > file_write_and_wait_range+0x48/0xd8 > ext4_sync_file+0x80/0x4b8 > vfs_fsync_range+0x64/0xc0 > SyS_msync+0x194/0x1e8 > > This patch fixes the problem by ensuring that READ_ONCE is used before > the initial checks on the pmd, and this value is subsequently used when > checking whether or not the pmd is present. pmd_check is removed and > the pmd_present check is inlined directly. > > Link: http://lkml.kernel.org/r/1507222630-5839-1-git-send-email-will.deacon@arm.com > Fixes: f27176cfc363 ("mm: convert page_mkclean_one() to use page_vma_mapped_walk()") > Cc: <stable@vger.kernel.org> # 4.13 > [will: backport to 4.13.y] > Signed-off-by: Will Deacon <will.deacon@arm.com> Thanks for the backport! greg k-h
diff --git a/mm/page_vma_mapped.c b/mm/page_vma_mapped.c index 8ec6ba230bb9..6b9311631aa1 100644 --- a/mm/page_vma_mapped.c +++ b/mm/page_vma_mapped.c @@ -6,17 +6,6 @@ #include "internal.h" -static inline bool check_pmd(struct page_vma_mapped_walk *pvmw) -{ - pmd_t pmde; - /* - * Make sure we don't re-load pmd between present and !trans_huge check. - * We need a consistent view. - */ - pmde = READ_ONCE(*pvmw->pmd); - return pmd_present(pmde) && !pmd_trans_huge(pmde); -} - static inline bool not_found(struct page_vma_mapped_walk *pvmw) { page_vma_mapped_walk_done(pvmw); @@ -106,6 +95,7 @@ bool page_vma_mapped_walk(struct page_vma_mapped_walk *pvmw) pgd_t *pgd; p4d_t *p4d; pud_t *pud; + pmd_t pmde; /* The only possible pmd mapping has been handled on last iteration */ if (pvmw->pmd && !pvmw->pte) @@ -138,7 +128,13 @@ bool page_vma_mapped_walk(struct page_vma_mapped_walk *pvmw) if (!pud_present(*pud)) return false; pvmw->pmd = pmd_offset(pud, pvmw->address); - if (pmd_trans_huge(*pvmw->pmd)) { + /* + * Make sure the pmd value isn't cached in a register by the + * compiler and used as a stale value after we've observed a + * subsequent update. + */ + pmde = READ_ONCE(*pvmw->pmd); + if (pmd_trans_huge(pmde)) { pvmw->ptl = pmd_lock(mm, pvmw->pmd); if (!pmd_present(*pvmw->pmd)) return not_found(pvmw); @@ -153,9 +149,8 @@ bool page_vma_mapped_walk(struct page_vma_mapped_walk *pvmw) spin_unlock(pvmw->ptl); pvmw->ptl = NULL; } - } else { - if (!check_pmd(pvmw)) - return false; + } else if (!pmd_present(pmde)) { + return false; } if (!map_pte(pvmw)) goto next_pte;
[Commit a7b100953aa33a5bbdc3e5e7f2241b9c0704606e upstream] Loading the pmd without holding the pmd_lock exposes us to races with concurrent updaters of the page tables but, worse still, it also allows the compiler to cache the pmd value in a register and reuse it later on, even if we've performed a READ_ONCE in between and seen a more recent value. In the case of page_vma_mapped_walk, this leads to the following crash when the pmd loaded for the initial pmd_trans_huge check is all zeroes and a subsequent valid table entry is loaded by check_pmd. We then proceed into map_pte, but the compiler re-uses the zero entry inside pte_offset_map, resulting in a junk pointer being installed in pvmw->pte: 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: check_pte+0x20/0x170 page_vma_mapped_walk+0x2e0/0x540 page_mkclean_one+0xac/0x278 rmap_walk_file+0xf0/0x238 rmap_walk+0x64/0xa0 page_mkclean+0x90/0xa8 clear_page_dirty_for_io+0x84/0x2a8 mpage_submit_page+0x34/0x98 mpage_process_page_bufs+0x164/0x170 mpage_prepare_extent_to_map+0x134/0x2b8 ext4_writepages+0x484/0xe30 do_writepages+0x44/0xe8 __filemap_fdatawrite_range+0xbc/0x110 file_write_and_wait_range+0x48/0xd8 ext4_sync_file+0x80/0x4b8 vfs_fsync_range+0x64/0xc0 SyS_msync+0x194/0x1e8 This patch fixes the problem by ensuring that READ_ONCE is used before the initial checks on the pmd, and this value is subsequently used when checking whether or not the pmd is present. pmd_check is removed and the pmd_present check is inlined directly. Link: http://lkml.kernel.org/r/1507222630-5839-1-git-send-email-will.deacon@arm.com Fixes: f27176cfc363 ("mm: convert page_mkclean_one() to use page_vma_mapped_walk()") Cc: <stable@vger.kernel.org> # 4.13 [will: backport to 4.13.y] Signed-off-by: Will Deacon <will.deacon@arm.com> --- mm/page_vma_mapped.c | 25 ++++++++++--------------- 1 file changed, 10 insertions(+), 15 deletions(-) -- 2.1.4