@@ -1599,7 +1599,30 @@ static bool try_to_unmap_one(struct page *page, struct vm_area_struct *vma,
/* MADV_FREE page check */
if (!PageSwapBacked(page)) {
- if (!PageDirty(page)) {
+ int ref_count, map_count;
+
+ /*
+ * Synchronize with gup_pte_range():
+ * - clear PTE; barrier; read refcount
+ * - inc refcount; barrier; read PTE
+ */
+ smp_mb();
+
+ ref_count = page_count(page);
+ map_count = page_mapcount(page);
+
+ /*
+ * Order reads for page refcount and dirty flag
+ * (see comments in __remove_mapping()).
+ */
+ smp_rmb();
+
+ /*
+ * The only page refs must be one from isolation
+ * plus the rmap(s) (dropped by discard:).
+ */
+ if (ref_count == 1 + map_count &&
+ !PageDirty(page)) {
/* Invalidate as we cleared the pte */
mmu_notifier_invalidate_range(mm,
address, address + PAGE_SIZE);
@@ -1729,7 +1729,7 @@ static unsigned int shrink_page_list(struct list_head *page_list,
mapping = page_mapping(page);
}
} else if (unlikely(PageTransHuge(page))) {
- /* Split file THP */
+ /* Split file/lazyfree THP */
if (split_huge_page_to_list(page, page_list))
goto keep_locked;
}
Problem: ======= Userspace might read the zero-page instead of actual data from a direct IO read on a block device if the buffers have been called madvise(MADV_FREE) on earlier (this is discussed below) due to a race between page reclaim on MADV_FREE and blkdev direct IO read. - Race condition: ============== During page reclaim, the MADV_FREE page check in try_to_unmap_one() checks if the page is not dirty, then discards its rmap PTE(s) (vs. remap back if the page is dirty). However, after try_to_unmap_one() returns to shrink_page_list(), it might keep the page _anyway_ if page_ref_freeze() fails (it expects exactly _one_ page reference, from the isolation for page reclaim). Well, blkdev_direct_IO() gets references for all pages, and on READ operations it only sets them dirty _later_. So, if MADV_FREE'd pages (i.e., not dirty) are used as buffers for direct IO read from block devices, and page reclaim happens during __blkdev_direct_IO[_simple]() exactly AFTER bio_iov_iter_get_pages() returns, but BEFORE the pages are set dirty, the situation happens. The direct IO read eventually completes. Now, when userspace reads the buffers, the PTE is no longer there and the page fault handler do_anonymous_page() services that with the zero-page, NOT the data! A synthetic reproducer is provided. - Page faults: =========== If page reclaim happens BEFORE bio_iov_iter_get_pages() the issue doesn't happen, because that faults-in all pages as writeable, so do_anonymous_page() sets up a new page/rmap/PTE, and that is used by direct IO. The userspace reads don't fault as the PTE is there (thus zero-page is not used/setup). But if page reclaim happens AFTER it / BEFORE setting pages dirty, the PTE is no longer there; the subsequent page faults can't help: The data-read from the block device probably won't generate faults due to DMA (no MMU) but even in the case it wouldn't use DMA, that happens on different virtual addresses (not user-mapped addresses) because `struct bio_vec` stores `struct page` to figure addresses out (which are different from user-mapped addresses) for the read. Thus userspace reads (to user-mapped addresses) still fault, then do_anonymous_page() gets another `struct page` that would address/ map to other memory than the `struct page` used by `struct bio_vec` for the read. (The original `struct page` is not available, since it wasn't freed, as page_ref_freeze() failed due to more page refs. And even if it were available, its data cannot be trusted anymore.) Solution: ======== One solution is to check for the expected page reference count in try_to_unmap_one(). There should be one reference from the isolation (that is also checked in shrink_page_list() with page_ref_freeze()) plus one or more references from page mapping(s) (put in discard: label). Further references mean that rmap/PTE cannot be unmapped/nuked. (Note: there might be more than one reference from mapping due to fork()/clone() without CLONE_VM, which use the same `struct page` for references, until the copy-on-write page gets copied.) So, additional page references (e.g., from direct IO read) now prevent the rmap/PTE from being unmapped/dropped; similarly to the page is not freed per shrink_page_list()/page_ref_freeze()). - Races and Barriers: ================== The new check in try_to_unmap_one() should be safe in races with bio_iov_iter_get_pages() in get_user_pages() fast and slow paths, as it's done under the PTE lock. The fast path doesn't take the lock, but it checks if the PTE has changed and if so, it drops the reference and leaves the page for the slow path (which does take that lock). The fast path requires synchronization w/ full memory barrier: it writes the page reference count first then it reads the PTE later, while try_to_unmap() writes PTE first then it reads page refcount. And a second barrier is needed, as the page dirty flag should not be read before the page reference count (as in __remove_mapping()). (This can be a load memory barrier only; no writes are involved.) Call stack/comments: - try_to_unmap_one() - page_vma_mapped_walk() - map_pte() # see pte_offset_map_lock(): pte_offset_map() spin_lock() - ptep_get_and_clear() # write PTE - smp_mb() # (new barrier) GUP fast path - page_ref_count() # (new check) read refcount - page_vma_mapped_walk_done() # see pte_unmap_unlock(): pte_unmap() spin_unlock() - bio_iov_iter_get_pages() - __bio_iov_iter_get_pages() - iov_iter_get_pages() - get_user_pages_fast() - internal_get_user_pages_fast() # fast path - lockless_pages_from_mm() - gup_{pgd,p4d,pud,pmd,pte}_range() ptep = pte_offset_map() # not _lock() pte = ptep_get_lockless(ptep) page = pte_page(pte) try_grab_compound_head(page) # inc refcount # (RMW/barrier # on success) if (pte_val(pte) != pte_val(*ptep)) # read PTE put_compound_head(page) # dec refcount # go slow path # slow path - __gup_longterm_unlocked() - get_user_pages_unlocked() - __get_user_pages_locked() - __get_user_pages() - follow_{page,p4d,pud,pmd}_mask() - follow_page_pte() ptep = pte_offset_map_lock() pte = *ptep page = vm_normal_page(pte) try_grab_page(page) # inc refcount pte_unmap_unlock() - Huge Pages: ========== Regarding transparent hugepages, that logic shouldn't change, as MADV_FREE (aka lazyfree) pages are PageAnon() && !PageSwapBacked() (madvise_free_pte_range() -> mark_page_lazyfree() -> lru_lazyfree_fn()) thus should reach shrink_page_list() -> split_huge_page_to_list() before try_to_unmap[_one](), so it deals with normal pages only. (And in case unlikely/TTU_SPLIT_HUGE_PMD/split_huge_pmd_address() happens, which should not or be rare, the page refcount should be greater than mapcount: the head page is referenced by tail pages. That also prevents checking the head `page` then incorrectly call page_remove_rmap(subpage) for a tail page, that isn't even in the shrink_page_list()'s page_list (an effect of split huge pmd/pmvw), as it might happen today in this unlikely scenario.) MADV_FREE'd buffers: =================== So, back to the "if MADV_FREE pages are used as buffers" note. The case is arguable, and subject to multiple interpretations. The madvise(2) manual page on the MADV_FREE advice value says: 1) 'After a successful MADV_FREE ... data will be lost when the kernel frees the pages.' 2) 'the free operation will be canceled if the caller writes into the page' / 'subsequent writes ... will succeed and then [the] kernel cannot free those dirtied pages' 3) 'If there is no subsequent write, the kernel can free the pages at any time.' Thoughts, questions, considerations... respectively: 1) Since the kernel didn't actually free the page (page_ref_freeze() failed), should the data not have been lost? (on userspace read.) 2) Should writes performed by the direct IO read be able to cancel the free operation? - Should the direct IO read be considered as 'the caller' too, as it's been requested by 'the caller'? - Should the bio technique to dirty pages on return to userspace (bio_check_pages_dirty() is called/used by __blkdev_direct_IO()) be considered in another/special way here? 3) Should an upcoming write from a previously requested direct IO read be considered as a subsequent write, so the kernel should not free the pages? (as it's known at the time of page reclaim.) At last: Technically, the last point would seem a reasonable consideration and balance, as the madvise(2) manual page apparently (and fairly) seem to assume that 'writes' are memory access from the userspace process (not explicitly considering writes from the kernel or its corner cases; again, fairly).. plus the kernel fix implementation for the corner case of the largely 'non-atomic write' encompassed by a direct IO read operation, is relatively simple; and it helps. Reproducer: ========== @ test.c (simplified, but works) #define _GNU_SOURCE #include <fcntl.h> #include <stdio.h> #include <unistd.h> #include <sys/mman.h> int main() { int fd, i; char *buf; fd = open(DEV, O_RDONLY | O_DIRECT); buf = mmap(NULL, BUF_SIZE, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0); for (i = 0; i < BUF_SIZE; i += PAGE_SIZE) buf[i] = 1; // init to non-zero madvise(buf, BUF_SIZE, MADV_FREE); read(fd, buf, BUF_SIZE); for (i = 0; i < BUF_SIZE; i += PAGE_SIZE) printf("%p: 0x%x\n", &buf[i], buf[i]); return 0; } @ block/fops.c (formerly fs/block_dev.c) +#include <linux/swap.h> ... ... __blkdev_direct_IO[_simple](...) { ... + if (!strcmp(current->comm, "good")) + shrink_all_memory(ULONG_MAX); + ret = bio_iov_iter_get_pages(...); + + if (!strcmp(current->comm, "bad")) + shrink_all_memory(ULONG_MAX); ... } @ shell # NUM_PAGES=4 # PAGE_SIZE=$(getconf PAGE_SIZE) # yes | dd of=test.img bs=${PAGE_SIZE} count=${NUM_PAGES} # DEV=$(losetup -f --show test.img) # gcc -DDEV=\"$DEV\" \ -DBUF_SIZE=$((PAGE_SIZE * NUM_PAGES)) \ -DPAGE_SIZE=${PAGE_SIZE} \ test.c -o test # od -tx1 $DEV 0000000 79 0a 79 0a 79 0a 79 0a 79 0a 79 0a 79 0a 79 0a * 0040000 # mv test good # ./good 0x7f7c10418000: 0x79 0x7f7c10419000: 0x79 0x7f7c1041a000: 0x79 0x7f7c1041b000: 0x79 # mv good bad # ./bad 0x7fa1b8050000: 0x0 0x7fa1b8051000: 0x0 0x7fa1b8052000: 0x0 0x7fa1b8053000: 0x0 Note: the issue is consistent on v5.17-rc3, but it's intermittent with the support of MADV_FREE on v4.5 (60%-70% error; needs swap). [wrap do_direct_IO() in do_blockdev_direct_IO() @ fs/direct-io.c]. - v5.17-rc3: # for i in {1..1000}; do ./good; done \ | cut -d: -f2 | sort | uniq -c 4000 0x79 # mv good bad # for i in {1..1000}; do ./bad; done \ | cut -d: -f2 | sort | uniq -c 4000 0x0 # free | grep Swap Swap: 0 0 0 - v4.5: # for i in {1..1000}; do ./good; done \ | cut -d: -f2 | sort | uniq -c 4000 0x79 # mv good bad # for i in {1..1000}; do ./bad; done \ | cut -d: -f2 | sort | uniq -c 2702 0x0 1298 0x79 # swapoff -av swapoff /swap # for i in {1..1000}; do ./bad; done \ | cut -d: -f2 | sort | uniq -c 4000 0x79 Ceph/TCMalloc: ============= For documentation purposes, the use case driving the analysis/fix is Ceph on Ubuntu 18.04, as the TCMalloc library there still uses MADV_FREE to release unused memory to the system from the mmap'ed page heap (might be committed back/used again; it's not munmap'ed.) - PageHeap::DecommitSpan() -> TCMalloc_SystemRelease() -> madvise() - PageHeap::CommitSpan() -> TCMalloc_SystemCommit() -> do nothing. Note: TCMalloc switched back to MADV_DONTNEED a few commits after the release in Ubuntu 18.04 (google-perftools/gperftools 2.5), so the issue just 'disappeared' on Ceph on later Ubuntu releases but is still present in the kernel, and can be hit by other use cases. The observed issue seems to be the old Ceph bug #22464 [1], where checksum mismatches are observed (and instrumentation with buffer dumps shows zero-pages read from mmap'ed/MADV_FREE'd page ranges). The issue in Ceph was reasonably deemed a kernel bug (comment #50) and mostly worked around with a retry mechanism, but other parts of Ceph could still hit that (rocksdb). Anyway, it's less likely to be hit again as TCMalloc switched out of MADV_FREE by default. (Some kernel versions/reports from the Ceph bug, and relation with the MADV_FREE introduction/changes; TCMalloc versions not checked.) - 4.4 good - 4.5 (madv_free: introduction) - 4.9 bad - 4.10 good? maybe a swapless system - 4.12 (madv_free: no longer free instantly on swapless systems) - 4.13 bad [1] https://tracker.ceph.com/issues/22464 Thanks: ====== Several people contributed to analysis/discussions/tests/reproducers in the first stages when drilling down on ceph/tcmalloc/linux kernel: - Dan Hill <daniel.hill@canonical.com> - Dan Streetman <dan.streetman@canonical.com> - Dongdong Tao <dongdong.tao@canonical.com> - Gavin Guo <gavin.guo@canonical.com> - Gerald Yang <gerald.yang@canonical.com> - Heitor Alves de Siqueira <halves@canonical.com> - Ioanna Alifieraki <ioanna-maria.alifieraki@canonical.com> - Jay Vosburgh <jay.vosburgh@canonical.com> - Matthew Ruffell <matthew.ruffell@canonical.com> - Ponnuvel Palaniyappan <ponnuvel.palaniyappan@canonical.com> Reviews, suggestions, corrections, comments: - Minchan Kim <minchan@kernel.org> - Yu Zhao <yuzhao@google.com> - Huang, Ying <ying.huang@intel.com> - John Hubbard <jhubbard@nvidia.com> - Christoph Hellwig <hch@infradead.org> Fixes: 854e9ed09ded ("mm: support madvise(MADV_FREE)") Reviewed-by: "Huang, Ying" <ying.huang@intel.com> Signed-off-by: Mauricio Faria de Oliveira <mfo@canonical.com> --- v4: - fixed Fixes: tag to first support, after tests on v4.5; updated commit message/reproducer section with results. - tested on v5.17-rc3 and v4.5. - shorten comment line; remove inner braces in the check. (Thanks: Christoph Hellwig <hch@infradead.org>) - clarify comment about __remove_mapping() (Thanks: Yu Zhao <yuzhao@google.com>) v3: - add full memory barrier to sync against GUP fast path; update comments. (Thanks: Yu Zhao <yuzhao@google.com>) - add Reviewed-by: "Huang, Ying" <ying.huang@intel.com> (no significant changes from v2). - add Fixes: tag. - minor changes/corrections to the commit message. - tested on v5.17-rc2. v2: - check refcount against mapcount rather than a static 2. (Thanks: Minchan Kim <minchan@kernel.org>) mm/rmap.c | 25 ++++++++++++++++++++++++- mm/vmscan.c | 2 +- 2 files changed, 25 insertions(+), 2 deletions(-)