From 9d274c19a71b3a276949933859610721a453946b Mon Sep 17 00:00:00 2001 From: Omar Sandoval Date: Fri, 24 May 2024 13:58:11 -0700 Subject: [PATCH 1/3] btrfs: fix crash on racing fsync and size-extending write into prealloc We have been seeing crashes on duplicate keys in btrfs_set_item_key_safe(): BTRFS critical (device vdb): slot 4 key (450 108 8192) new key (450 108 8192) ------------[ cut here ]------------ kernel BUG at fs/btrfs/ctree.c:2620! invalid opcode: 0000 [#1] PREEMPT SMP PTI CPU: 0 PID: 3139 Comm: xfs_io Kdump: loaded Not tainted 6.9.0 #6 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.16.3-2.fc40 04/01/2014 RIP: 0010:btrfs_set_item_key_safe+0x11f/0x290 [btrfs] With the following stack trace: #0 btrfs_set_item_key_safe (fs/btrfs/ctree.c:2620:4) #1 btrfs_drop_extents (fs/btrfs/file.c:411:4) #2 log_one_extent (fs/btrfs/tree-log.c:4732:9) #3 btrfs_log_changed_extents (fs/btrfs/tree-log.c:4955:9) #4 btrfs_log_inode (fs/btrfs/tree-log.c:6626:9) #5 btrfs_log_inode_parent (fs/btrfs/tree-log.c:7070:8) #6 btrfs_log_dentry_safe (fs/btrfs/tree-log.c:7171:8) #7 btrfs_sync_file (fs/btrfs/file.c:1933:8) #8 vfs_fsync_range (fs/sync.c:188:9) #9 vfs_fsync (fs/sync.c:202:9) #10 do_fsync (fs/sync.c:212:9) #11 __do_sys_fdatasync (fs/sync.c:225:9) #12 __se_sys_fdatasync (fs/sync.c:223:1) #13 __x64_sys_fdatasync (fs/sync.c:223:1) #14 do_syscall_x64 (arch/x86/entry/common.c:52:14) #15 do_syscall_64 (arch/x86/entry/common.c:83:7) #16 entry_SYSCALL_64+0xaf/0x14c (arch/x86/entry/entry_64.S:121) So we're logging a changed extent from fsync, which is splitting an extent in the log tree. But this split part already exists in the tree, triggering the BUG(). This is the state of the log tree at the time of the crash, dumped with drgn (https://github.com/osandov/drgn/blob/main/contrib/btrfs_tree.py) to get more details than btrfs_print_leaf() gives us: >>> print_extent_buffer(prog.crashed_thread().stack_trace()[0]["eb"]) leaf 33439744 level 0 items 72 generation 9 owner 18446744073709551610 leaf 33439744 flags 0x100000000000000 fs uuid e5bd3946-400c-4223-8923-190ef1f18677 chunk uuid d58cb17e-6d02-494a-829a-18b7d8a399da item 0 key (450 INODE_ITEM 0) itemoff 16123 itemsize 160 generation 7 transid 9 size 8192 nbytes 8473563889606862198 block group 0 mode 100600 links 1 uid 0 gid 0 rdev 0 sequence 204 flags 0x10(PREALLOC) atime 1716417703.220000000 (2024-05-22 15:41:43) ctime 1716417704.983333333 (2024-05-22 15:41:44) mtime 1716417704.983333333 (2024-05-22 15:41:44) otime 17592186044416.000000000 (559444-03-08 01:40:16) item 1 key (450 INODE_REF 256) itemoff 16110 itemsize 13 index 195 namelen 3 name: 193 item 2 key (450 XATTR_ITEM 1640047104) itemoff 16073 itemsize 37 location key (0 UNKNOWN.0 0) type XATTR transid 7 data_len 1 name_len 6 name: user.a data a item 3 key (450 EXTENT_DATA 0) itemoff 16020 itemsize 53 generation 9 type 1 (regular) extent data disk byte 303144960 nr 12288 extent data offset 0 nr 4096 ram 12288 extent compression 0 (none) item 4 key (450 EXTENT_DATA 4096) itemoff 15967 itemsize 53 generation 9 type 2 (prealloc) prealloc data disk byte 303144960 nr 12288 prealloc data offset 4096 nr 8192 item 5 key (450 EXTENT_DATA 8192) itemoff 15914 itemsize 53 generation 9 type 2 (prealloc) prealloc data disk byte 303144960 nr 12288 prealloc data offset 8192 nr 4096 ... So the real problem happened earlier: notice that items 4 (4k-12k) and 5 (8k-12k) overlap. Both are prealloc extents. Item 4 straddles i_size and item 5 starts at i_size. Here is the state of the filesystem tree at the time of the crash: >>> root = prog.crashed_thread().stack_trace()[2]["inode"].root >>> ret, nodes, slots = btrfs_search_slot(root, BtrfsKey(450, 0, 0)) >>> print_extent_buffer(nodes[0]) leaf 30425088 level 0 items 184 generation 9 owner 5 leaf 30425088 flags 0x100000000000000 fs uuid e5bd3946-400c-4223-8923-190ef1f18677 chunk uuid d58cb17e-6d02-494a-829a-18b7d8a399da ... item 179 key (450 INODE_ITEM 0) itemoff 4907 itemsize 160 generation 7 transid 7 size 4096 nbytes 12288 block group 0 mode 100600 links 1 uid 0 gid 0 rdev 0 sequence 6 flags 0x10(PREALLOC) atime 1716417703.220000000 (2024-05-22 15:41:43) ctime 1716417703.220000000 (2024-05-22 15:41:43) mtime 1716417703.220000000 (2024-05-22 15:41:43) otime 1716417703.220000000 (2024-05-22 15:41:43) item 180 key (450 INODE_REF 256) itemoff 4894 itemsize 13 index 195 namelen 3 name: 193 item 181 key (450 XATTR_ITEM 1640047104) itemoff 4857 itemsize 37 location key (0 UNKNOWN.0 0) type XATTR transid 7 data_len 1 name_len 6 name: user.a data a item 182 key (450 EXTENT_DATA 0) itemoff 4804 itemsize 53 generation 9 type 1 (regular) extent data disk byte 303144960 nr 12288 extent data offset 0 nr 8192 ram 12288 extent compression 0 (none) item 183 key (450 EXTENT_DATA 8192) itemoff 4751 itemsize 53 generation 9 type 2 (prealloc) prealloc data disk byte 303144960 nr 12288 prealloc data offset 8192 nr 4096 Item 5 in the log tree corresponds to item 183 in the filesystem tree, but nothing matches item 4. Furthermore, item 183 is the last item in the leaf. btrfs_log_prealloc_extents() is responsible for logging prealloc extents beyond i_size. It first truncates any previously logged prealloc extents that start beyond i_size. Then, it walks the filesystem tree and copies the prealloc extent items to the log tree. If it hits the end of a leaf, then it calls btrfs_next_leaf(), which unlocks the tree and does another search. However, while the filesystem tree is unlocked, an ordered extent completion may modify the tree. In particular, it may insert an extent item that overlaps with an extent item that was already copied to the log tree. This may manifest in several ways depending on the exact scenario, including an EEXIST error that is silently translated to a full sync, overlapping items in the log tree, or this crash. This particular crash is triggered by the following sequence of events: - Initially, the file has i_size=4k, a regular extent from 0-4k, and a prealloc extent beyond i_size from 4k-12k. The prealloc extent item is the last item in its B-tree leaf. - The file is fsync'd, which copies its inode item and both extent items to the log tree. - An xattr is set on the file, which sets the BTRFS_INODE_COPY_EVERYTHING flag. - The range 4k-8k in the file is written using direct I/O. i_size is extended to 8k, but the ordered extent is still in flight. - The file is fsync'd. Since BTRFS_INODE_COPY_EVERYTHING is set, this calls copy_inode_items_to_log(), which calls btrfs_log_prealloc_extents(). - btrfs_log_prealloc_extents() finds the 4k-12k prealloc extent in the filesystem tree. Since it starts before i_size, it skips it. Since it is the last item in its B-tree leaf, it calls btrfs_next_leaf(). - btrfs_next_leaf() unlocks the path. - The ordered extent completion runs, which converts the 4k-8k part of the prealloc extent to written and inserts the remaining prealloc part from 8k-12k. - btrfs_next_leaf() does a search and finds the new prealloc extent 8k-12k. - btrfs_log_prealloc_extents() copies the 8k-12k prealloc extent into the log tree. Note that it overlaps with the 4k-12k prealloc extent that was copied to the log tree by the first fsync. - fsync calls btrfs_log_changed_extents(), which tries to log the 4k-8k extent that was written. - This tries to drop the range 4k-8k in the log tree, which requires adjusting the start of the 4k-12k prealloc extent in the log tree to 8k. - btrfs_set_item_key_safe() sees that there is already an extent starting at 8k in the log tree and calls BUG(). Fix this by detecting when we're about to insert an overlapping file extent item in the log tree and truncating the part that would overlap. CC: stable@vger.kernel.org # 6.1+ Reviewed-by: Filipe Manana Signed-off-by: Omar Sandoval Signed-off-by: David Sterba --- fs/btrfs/tree-log.c | 17 +++++++++++------ 1 file changed, 11 insertions(+), 6 deletions(-) diff --git a/fs/btrfs/tree-log.c b/fs/btrfs/tree-log.c index 5146387b416b..26a2e5aa08e9 100644 --- a/fs/btrfs/tree-log.c +++ b/fs/btrfs/tree-log.c @@ -4860,18 +4860,23 @@ static int btrfs_log_prealloc_extents(struct btrfs_trans_handle *trans, path->slots[0]++; continue; } - if (!dropped_extents) { - /* - * Avoid logging extent items logged in past fsync calls - * and leading to duplicate keys in the log tree. - */ + /* + * Avoid overlapping items in the log tree. The first time we + * get here, get rid of everything from a past fsync. After + * that, if the current extent starts before the end of the last + * extent we copied, truncate the last one. This can happen if + * an ordered extent completion modifies the subvolume tree + * while btrfs_next_leaf() has the tree unlocked. + */ + if (!dropped_extents || key.offset < truncate_offset) { ret = truncate_inode_items(trans, root->log_root, inode, - truncate_offset, + min(key.offset, truncate_offset), BTRFS_EXTENT_DATA_KEY); if (ret) goto out; dropped_extents = true; } + truncate_offset = btrfs_file_extent_end(path); if (ins_nr == 0) start_slot = slot; ins_nr++; From fb33eb2ef0d88e75564983ef057b44c5b7e4fded Mon Sep 17 00:00:00 2001 From: Filipe Manana Date: Mon, 3 Jun 2024 12:49:08 +0100 Subject: [PATCH 2/3] btrfs: fix leak of qgroup extent records after transaction abort Qgroup extent records are created when delayed ref heads are created and then released after accounting extents at btrfs_qgroup_account_extents(), called during the transaction commit path. If a transaction is aborted we free the qgroup records by calling btrfs_qgroup_destroy_extent_records() at btrfs_destroy_delayed_refs(), unless we don't have delayed references. We are incorrectly assuming that no delayed references means we don't have qgroup extents records. We can currently have no delayed references because we ran them all during a transaction commit and the transaction was aborted after that due to some error in the commit path. So fix this by ensuring we btrfs_qgroup_destroy_extent_records() at btrfs_destroy_delayed_refs() even if we don't have any delayed references. Reported-by: syzbot+0fecc032fa134afd49df@syzkaller.appspotmail.com Link: https://lore.kernel.org/linux-btrfs/0000000000004e7f980619f91835@google.com/ Fixes: 81f7eb00ff5b ("btrfs: destroy qgroup extent records on transaction abort") CC: stable@vger.kernel.org # 6.1+ Reviewed-by: Josef Bacik Reviewed-by: Qu Wenruo Signed-off-by: Filipe Manana Signed-off-by: David Sterba --- fs/btrfs/disk-io.c | 10 +--------- 1 file changed, 1 insertion(+), 9 deletions(-) diff --git a/fs/btrfs/disk-io.c b/fs/btrfs/disk-io.c index a91a8056758a..242ada7e47b4 100644 --- a/fs/btrfs/disk-io.c +++ b/fs/btrfs/disk-io.c @@ -4538,18 +4538,10 @@ static void btrfs_destroy_delayed_refs(struct btrfs_transaction *trans, struct btrfs_fs_info *fs_info) { struct rb_node *node; - struct btrfs_delayed_ref_root *delayed_refs; + struct btrfs_delayed_ref_root *delayed_refs = &trans->delayed_refs; struct btrfs_delayed_ref_node *ref; - delayed_refs = &trans->delayed_refs; - spin_lock(&delayed_refs->lock); - if (atomic_read(&delayed_refs->num_entries) == 0) { - spin_unlock(&delayed_refs->lock); - btrfs_debug(fs_info, "delayed_refs has NO entry"); - return; - } - while ((node = rb_first_cached(&delayed_refs->href_root)) != NULL) { struct btrfs_delayed_ref_head *head; struct rb_node *n; From f3a5367c679d31473d3fbb391675055b4792c309 Mon Sep 17 00:00:00 2001 From: Qu Wenruo Date: Thu, 6 Jun 2024 11:01:51 +0930 Subject: [PATCH 3/3] btrfs: protect folio::private when attaching extent buffer folios MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit [BUG] Since v6.8 there are rare kernel crashes reported by various people, the common factor is bad page status error messages like this: BUG: Bad page state in process kswapd0 pfn:d6e840 page: refcount:0 mapcount:0 mapping:000000007512f4f2 index:0x2796c2c7c pfn:0xd6e840 aops:btree_aops ino:1 flags: 0x17ffffe0000008(uptodate|node=0|zone=2|lastcpupid=0x3fffff) page_type: 0xffffffff() raw: 0017ffffe0000008 dead000000000100 dead000000000122 ffff88826d0be4c0 raw: 00000002796c2c7c 0000000000000000 00000000ffffffff 0000000000000000 page dumped because: non-NULL mapping [CAUSE] Commit 09e6cef19c9f ("btrfs: refactor alloc_extent_buffer() to allocate-then-attach method") changes the sequence when allocating a new extent buffer. Previously we always called grab_extent_buffer() under mapping->i_private_lock, to ensure the safety on modification on folio::private (which is a pointer to extent buffer for regular sectorsize). This can lead to the following race: Thread A is trying to allocate an extent buffer at bytenr X, with 4 4K pages, meanwhile thread B is trying to release the page at X + 4K (the second page of the extent buffer at X). Thread A | Thread B -----------------------------------+------------------------------------- | btree_release_folio() | | This is for the page at X + 4K, | | Not page X. | | alloc_extent_buffer() | |- release_extent_buffer() |- filemap_add_folio() for the | | |- atomic_dec_and_test(eb->refs) | page at bytenr X (the first | | | | page). | | | | Which returned -EEXIST. | | | | | | | |- filemap_lock_folio() | | | | Returned the first page locked. | | | | | | | |- grab_extent_buffer() | | | | |- atomic_inc_not_zero() | | | | | Returned false | | | | |- folio_detach_private() | | |- folio_detach_private() for X | |- folio_test_private() | | |- folio_test_private() | Returned true | | | Returned true |- folio_put() | |- folio_put() Now there are two puts on the same folio at folio X, leading to refcount underflow of the folio X, and eventually causing the BUG_ON() on the page->mapping. The condition is not that easy to hit: - The release must be triggered for the middle page of an eb If the release is on the same first page of an eb, page lock would kick in and prevent the race. - folio_detach_private() has a very small race window It's only between folio_test_private() and folio_clear_private(). That's exactly when mapping->i_private_lock is used to prevent such race, and commit 09e6cef19c9f ("btrfs: refactor alloc_extent_buffer() to allocate-then-attach method") screwed that up. At that time, I thought the page lock would kick in as filemap_release_folio() also requires the page to be locked, but forgot the filemap_release_folio() only locks one page, not all pages of an extent buffer. [FIX] Move all the code requiring i_private_lock into attach_eb_folio_to_filemap(), so that everything is done with proper lock protection. Furthermore to prevent future problems, add an extra lockdep_assert_locked() to ensure we're holding the proper lock. To reproducer that is able to hit the race (takes a few minutes with instrumented code inserting delays to alloc_extent_buffer()): #!/bin/sh drop_caches () { while(true); do echo 3 > /proc/sys/vm/drop_caches echo 1 > /proc/sys/vm/compact_memory done } run_tar () { while(true); do for x in `seq 1 80` ; do tar cf /dev/zero /mnt > /dev/null & done wait done } mkfs.btrfs -f -d single -m single /dev/vda mount -o noatime /dev/vda /mnt # create 200,000 files, 1K each ./simoop -n 200000 -E -f 1k /mnt drop_caches & (run_tar) Reported-by: Linus Torvalds Link: https://lore.kernel.org/linux-btrfs/CAHk-=wgt362nGfScVOOii8cgKn2LVVHeOvOA7OBwg1OwbuJQcw@mail.gmail.com/ Reported-by: Mikhail Gavrilov Link: https://lore.kernel.org/lkml/CABXGCsPktcHQOvKTbPaTwegMExije=Gpgci5NW=hqORo-s7diA@mail.gmail.com/ Reported-by: Toralf Förster Link: https://lore.kernel.org/linux-btrfs/e8b3311c-9a75-4903-907f-fc0f7a3fe423@gmx.de/ Reported-by: syzbot+f80b066392366b4af85e@syzkaller.appspotmail.com Fixes: 09e6cef19c9f ("btrfs: refactor alloc_extent_buffer() to allocate-then-attach method") CC: stable@vger.kernel.org # 6.8+ CC: Chris Mason Reviewed-by: Filipe Manana Reviewed-by: Josef Bacik Signed-off-by: Qu Wenruo Reviewed-by: David Sterba Signed-off-by: David Sterba --- fs/btrfs/extent_io.c | 60 +++++++++++++++++++++++--------------------- 1 file changed, 31 insertions(+), 29 deletions(-) diff --git a/fs/btrfs/extent_io.c b/fs/btrfs/extent_io.c index 597387e9f040..f688fab55251 100644 --- a/fs/btrfs/extent_io.c +++ b/fs/btrfs/extent_io.c @@ -3689,6 +3689,8 @@ static struct extent_buffer *grab_extent_buffer( struct folio *folio = page_folio(page); struct extent_buffer *exists; + lockdep_assert_held(&page->mapping->i_private_lock); + /* * For subpage case, we completely rely on radix tree to ensure we * don't try to insert two ebs for the same bytenr. So here we always @@ -3756,13 +3758,14 @@ static int check_eb_alignment(struct btrfs_fs_info *fs_info, u64 start) * The caller needs to free the existing folios and retry using the same order. */ static int attach_eb_folio_to_filemap(struct extent_buffer *eb, int i, + struct btrfs_subpage *prealloc, struct extent_buffer **found_eb_ret) { struct btrfs_fs_info *fs_info = eb->fs_info; struct address_space *mapping = fs_info->btree_inode->i_mapping; const unsigned long index = eb->start >> PAGE_SHIFT; - struct folio *existing_folio; + struct folio *existing_folio = NULL; int ret; ASSERT(found_eb_ret); @@ -3774,12 +3777,14 @@ static int attach_eb_folio_to_filemap(struct extent_buffer *eb, int i, ret = filemap_add_folio(mapping, eb->folios[i], index + i, GFP_NOFS | __GFP_NOFAIL); if (!ret) - return 0; + goto finish; existing_folio = filemap_lock_folio(mapping, index + i); /* The page cache only exists for a very short time, just retry. */ - if (IS_ERR(existing_folio)) + if (IS_ERR(existing_folio)) { + existing_folio = NULL; goto retry; + } /* For now, we should only have single-page folios for btree inode. */ ASSERT(folio_nr_pages(existing_folio) == 1); @@ -3790,14 +3795,13 @@ static int attach_eb_folio_to_filemap(struct extent_buffer *eb, int i, return -EAGAIN; } - if (fs_info->nodesize < PAGE_SIZE) { - /* - * We're going to reuse the existing page, can drop our page - * and subpage structure now. - */ +finish: + spin_lock(&mapping->i_private_lock); + if (existing_folio && fs_info->nodesize < PAGE_SIZE) { + /* We're going to reuse the existing page, can drop our folio now. */ __free_page(folio_page(eb->folios[i], 0)); eb->folios[i] = existing_folio; - } else { + } else if (existing_folio) { struct extent_buffer *existing_eb; existing_eb = grab_extent_buffer(fs_info, @@ -3805,6 +3809,7 @@ static int attach_eb_folio_to_filemap(struct extent_buffer *eb, int i, if (existing_eb) { /* The extent buffer still exists, we can use it directly. */ *found_eb_ret = existing_eb; + spin_unlock(&mapping->i_private_lock); folio_unlock(existing_folio); folio_put(existing_folio); return 1; @@ -3813,6 +3818,22 @@ static int attach_eb_folio_to_filemap(struct extent_buffer *eb, int i, __free_page(folio_page(eb->folios[i], 0)); eb->folios[i] = existing_folio; } + eb->folio_size = folio_size(eb->folios[i]); + eb->folio_shift = folio_shift(eb->folios[i]); + /* Should not fail, as we have preallocated the memory. */ + ret = attach_extent_buffer_folio(eb, eb->folios[i], prealloc); + ASSERT(!ret); + /* + * To inform we have an extra eb under allocation, so that + * detach_extent_buffer_page() won't release the folio private when the + * eb hasn't been inserted into radix tree yet. + * + * The ref will be decreased when the eb releases the page, in + * detach_extent_buffer_page(). Thus needs no special handling in the + * error path. + */ + btrfs_folio_inc_eb_refs(fs_info, eb->folios[i]); + spin_unlock(&mapping->i_private_lock); return 0; } @@ -3824,7 +3845,6 @@ struct extent_buffer *alloc_extent_buffer(struct btrfs_fs_info *fs_info, int attached = 0; struct extent_buffer *eb; struct extent_buffer *existing_eb = NULL; - struct address_space *mapping = fs_info->btree_inode->i_mapping; struct btrfs_subpage *prealloc = NULL; u64 lockdep_owner = owner_root; bool page_contig = true; @@ -3890,7 +3910,7 @@ struct extent_buffer *alloc_extent_buffer(struct btrfs_fs_info *fs_info, for (int i = 0; i < num_folios; i++) { struct folio *folio; - ret = attach_eb_folio_to_filemap(eb, i, &existing_eb); + ret = attach_eb_folio_to_filemap(eb, i, prealloc, &existing_eb); if (ret > 0) { ASSERT(existing_eb); goto out; @@ -3927,24 +3947,6 @@ struct extent_buffer *alloc_extent_buffer(struct btrfs_fs_info *fs_info, * and free the allocated page. */ folio = eb->folios[i]; - eb->folio_size = folio_size(folio); - eb->folio_shift = folio_shift(folio); - spin_lock(&mapping->i_private_lock); - /* Should not fail, as we have preallocated the memory */ - ret = attach_extent_buffer_folio(eb, folio, prealloc); - ASSERT(!ret); - /* - * To inform we have extra eb under allocation, so that - * detach_extent_buffer_page() won't release the folio private - * when the eb hasn't yet been inserted into radix tree. - * - * The ref will be decreased when the eb released the page, in - * detach_extent_buffer_page(). - * Thus needs no special handling in error path. - */ - btrfs_folio_inc_eb_refs(fs_info, folio); - spin_unlock(&mapping->i_private_lock); - WARN_ON(btrfs_folio_test_dirty(fs_info, folio, eb->start, eb->len)); /*