Message ID | 20210226162200.20548-1-vbabka@suse.cz |
---|---|
State | New |
Headers | show |
Series | [4.9] mm, thp: make do_huge_pmd_wp_page() lock page for testing mapcount | expand |
On Fri, Feb 26, 2021 at 05:22:00PM +0100, Vlastimil Babka wrote: >Jann reported [1] a race between __split_huge_pmd_locked() and >page_trans_huge_map_swapcount() which can result in a page to be reused >instead of COWed. This was later assigned CVE-2020-29368. > >This was fixed by commit c444eb564fb1 ("mm: thp: make the THP mapcount atomic >against __split_huge_pmd_locked()") by doing the split under the page lock, >while all users of page_trans_huge_map_swapcount() were already also under page >lock. The fix was backported also to 4.9 stable series. > >When testing the backport on a 4.12 based kernel, Nicolai noticed the POC from >[1] still reproduces after backporting c444eb564fb1 and identified a missing >page lock in do_huge_pmd_wp_page() around the call to >page_trans_huge_mapcount(). The page lock was only added in ba3c4ce6def4 ("mm, >THP, swap: make reuse_swap_page() works for THP swapped out") in 4.14. The >commit also wrapped page_trans_huge_mapcount() into >page_trans_huge_map_swapcount() for the purposes of COW decisions. > >I have verified that 4.9.y indeed also reproduces with the POC. Backporting >ba3c4ce6def4 alone however is not possible as it's part of a larger effort of >optimizing THP swapping, which would be risky to backport fully. > >Therefore this 4.9-stable-only patch just wraps page_trans_huge_mapcount() >in page_trans_huge_mapcount() under page lock the same way as ba3c4ce6def4 >does, but without the page_trans_huge_map_swapcount() part. Other callers >of page_trans_huge_mapcount() are all under page lock already. I have verified >the POC no longer reproduces afterwards. > >[1] https://bugs.chromium.org/p/project-zero/issues/detail?id=2045 > >Reported-by: Nicolai Stange <nstange@suse.de> >Signed-off-by: Vlastimil Babka <vbabka@suse.cz> Queued up, thanks!
diff --git a/mm/huge_memory.c b/mm/huge_memory.c index 05ca01ef97f7..14cd0ef33b62 100644 --- a/mm/huge_memory.c +++ b/mm/huge_memory.c @@ -1022,6 +1022,19 @@ int do_huge_pmd_wp_page(struct fault_env *fe, pmd_t orig_pmd) * We can only reuse the page if nobody else maps the huge page or it's * part. */ + if (!trylock_page(page)) { + get_page(page); + spin_unlock(fe->ptl); + lock_page(page); + spin_lock(fe->ptl); + if (unlikely(!pmd_same(*fe->pmd, orig_pmd))) { + unlock_page(page); + put_page(page); + goto out_unlock; + } + put_page(page); + } + if (page_trans_huge_mapcount(page, NULL) == 1) { pmd_t entry; entry = pmd_mkyoung(orig_pmd); @@ -1029,8 +1042,10 @@ int do_huge_pmd_wp_page(struct fault_env *fe, pmd_t orig_pmd) if (pmdp_set_access_flags(vma, haddr, fe->pmd, entry, 1)) update_mmu_cache_pmd(vma, fe->address, fe->pmd); ret |= VM_FAULT_WRITE; + unlock_page(page); goto out_unlock; } + unlock_page(page); get_page(page); spin_unlock(fe->ptl); alloc:
Jann reported [1] a race between __split_huge_pmd_locked() and page_trans_huge_map_swapcount() which can result in a page to be reused instead of COWed. This was later assigned CVE-2020-29368. This was fixed by commit c444eb564fb1 ("mm: thp: make the THP mapcount atomic against __split_huge_pmd_locked()") by doing the split under the page lock, while all users of page_trans_huge_map_swapcount() were already also under page lock. The fix was backported also to 4.9 stable series. When testing the backport on a 4.12 based kernel, Nicolai noticed the POC from [1] still reproduces after backporting c444eb564fb1 and identified a missing page lock in do_huge_pmd_wp_page() around the call to page_trans_huge_mapcount(). The page lock was only added in ba3c4ce6def4 ("mm, THP, swap: make reuse_swap_page() works for THP swapped out") in 4.14. The commit also wrapped page_trans_huge_mapcount() into page_trans_huge_map_swapcount() for the purposes of COW decisions. I have verified that 4.9.y indeed also reproduces with the POC. Backporting ba3c4ce6def4 alone however is not possible as it's part of a larger effort of optimizing THP swapping, which would be risky to backport fully. Therefore this 4.9-stable-only patch just wraps page_trans_huge_mapcount() in page_trans_huge_mapcount() under page lock the same way as ba3c4ce6def4 does, but without the page_trans_huge_map_swapcount() part. Other callers of page_trans_huge_mapcount() are all under page lock already. I have verified the POC no longer reproduces afterwards. [1] https://bugs.chromium.org/p/project-zero/issues/detail?id=2045 Reported-by: Nicolai Stange <nstange@suse.de> Signed-off-by: Vlastimil Babka <vbabka@suse.cz> --- mm/huge_memory.c | 15 +++++++++++++++ 1 file changed, 15 insertions(+)