kernel/git/stable/linux.git/fs/btrfs, branch linux-5.3.y

btrfs: record all roots for rename exchange on a subvol

2019-12-17T19:07:35+00:00

commit 3e1740993e43116b3bc71b0aad1e6872f6ccf341 upstream. Testing with the new fsstress support for subvolumes uncovered a pretty bad problem with rename exchange on subvolumes. We're modifying two different subvolumes, but we only start the transaction on one of them, so the other one is not added to the dirty root list. This is caught by btrfs_cow_block() with a warning because the root has not been updated, however if we do not modify this root again we'll end up pointing at an invalid root because the root item is never updated. Fix this by making sure we add the destination root to the trans list, the same as we do with normal renames. This fixes the corruption. Fixes: cdd1fedf8261 ("btrfs: add support for RENAME_EXCHANGE and RENAME_WHITEOUT") CC: stable@vger.kernel.org # 4.9+ Reviewed-by: Filipe Manana Signed-off-by: Josef Bacik Signed-off-by: David Sterba Signed-off-by: Greg Kroah-Hartman

Btrfs: send, skip backreference walking for extents with many references

2019-12-17T19:07:34+00:00

commit fd0ddbe2509568b00df364156f47561e9f469f15 upstream. Backreference walking, which is used by send to figure if it can issue clone operations instead of write operations, can be very slow and use too much memory when extents have many references. This change simply skips backreference walking when an extent has more than 64 references, in which case we fallback to a write operation instead of a clone operation. This limit is conservative and in practice I observed no signicant slowdown with up to 100 references and still low memory usage up to that limit. This is a temporary workaround until there are speedups in the backref walking code, and as such it does not attempt to add extra interfaces or knobs to tweak the threshold. Reported-by: Atemu Link: https://lore.kernel.org/linux-btrfs/CAE4GHgkvqVADtS4AzcQJxo0Q1jKQgKaW3JGp3SGdoinVo=C9eQ@mail.gmail.com/T/#me55dc0987f9cc2acaa54372ce0492c65782be3fa CC: stable@vger.kernel.org # 4.4+ Reviewed-by: Qu Wenruo Signed-off-by: Filipe Manana Signed-off-by: David Sterba Signed-off-by: Greg Kroah-Hartman

btrfs: Remove btrfs_bio::flags member

2019-12-17T19:07:33+00:00

commit 34b127aecd4fe8e6a3903e10f204a7b7ffddca22 upstream. The last user of btrfs_bio::flags was removed in commit 326e1dbb5736 ("block: remove management of bi_remaining when restoring original bi_end_io"), remove it. (Tagged for stable as the structure is heavily used and space savings are desirable.) CC: stable@vger.kernel.org # 4.4+ Signed-off-by: Qu Wenruo Reviewed-by: David Sterba Signed-off-by: David Sterba Signed-off-by: Greg Kroah-Hartman

btrfs: Avoid getting stuck during cyclic writebacks

2019-12-17T19:07:32+00:00

commit f7bddf1e27d18fbc7d3e3056ba449cfbe4e20b0a upstream. During a cyclic writeback, extent_write_cache_pages() uses done_index to update the writeback_index after the current run is over. However, instead of current index + 1, it gets to to the current index itself. Unfortunately, this, combined with returning on EOF instead of looping back, can lead to the following pathlogical behavior. 1. There is a single file which has accumulated enough dirty pages to trigger balance_dirty_pages() and the writer appending to the file with a series of short writes. 2. balance_dirty_pages kicks in, wakes up background writeback and sleeps. 3. Writeback kicks in and the cursor is on the last page of the dirty file. Writeback is started or skipped if already in progress. As it's EOF, extent_write_cache_pages() returns and the cursor is set to done_index which is pointing to the last page. 4. Writeback is done. Nothing happens till balance_dirty_pages finishes, at which point we go back to #1. This can almost completely stall out writing back of the file and keep the system over dirty threshold for a long time which can mess up the whole system. We encountered this issue in production with a package handling application which can reliably reproduce the issue when running under tight memory limits. Reading the comment in the error handling section, this seems to be to avoid accidentally skipping a page in case the write attempt on the page doesn't succeed. However, this concern seems bogus. On each page, the code either: * Skips and moves onto the next page. * Fails issue and sets done_index to index + 1. * Successfully issues and continue to the next page if budget allows and not EOF. IOW, as long as it's not EOF and there's budget, the code never retries writing back the same page. Only when a page happens to be the last page of a particular run, we end up retrying the page, which can't possibly guarantee anything data integrity related. Besides, cyclic writes are only used for non-syncing writebacks meaning that there's no data integrity implication to begin with. Fix it by always setting done_index past the current page being processed. Note that this problem exists in other writepages too. CC: stable@vger.kernel.org # 4.19+ Signed-off-by: Tejun Heo Reviewed-by: David Sterba Signed-off-by: David Sterba Signed-off-by: Greg Kroah-Hartman

Btrfs: fix negative subv_writers counter and data space leak after buffered write

2019-12-17T19:07:32+00:00

commit a0e248bb502d5165b3314ac3819e888fdcdf7d9f upstream. When doing a buffered write it's possible to leave the subv_writers counter of the root, used for synchronization between buffered nocow writers and snapshotting. This happens in an exceptional case like the following: 1) We fail to allocate data space for the write, since there's not enough available data space nor enough unallocated space for allocating a new data block group; 2) Because of that failure, we try to go to NOCOW mode, which succeeds and therefore we set the local variable 'only_release_metadata' to true and set the root's sub_writers counter to 1 through the call to btrfs_start_write_no_snapshotting() made by check_can_nocow(); 3) The call to btrfs_copy_from_user() returns zero, which is very unlikely to happen but not impossible; 4) No pages are copied because btrfs_copy_from_user() returned zero; 5) We call btrfs_end_write_no_snapshotting() which decrements the root's subv_writers counter to 0; 6) We don't set 'only_release_metadata' back to 'false' because we do it only if 'copied', the value returned by btrfs_copy_from_user(), is greater than zero; 7) On the next iteration of the while loop, which processes the same page range, we are now able to allocate data space for the write (we got enough data space released in the meanwhile); 8) After this if we fail at btrfs_delalloc_reserve_metadata(), because now there isn't enough free metadata space, or in some other place further below (prepare_pages(), lock_and_cleanup_extent_if_need(), btrfs_dirty_pages()), we break out of the while loop with 'only_release_metadata' having a value of 'true'; 9) Because 'only_release_metadata' is 'true' we end up decrementing the root's subv_writers counter to -1 (through a call to btrfs_end_write_no_snapshotting()), and we also end up not releasing the data space previously reserved through btrfs_check_data_free_space(). As a consequence the mechanism for synchronizing NOCOW buffered writes with snapshotting gets broken. Fix this by always setting 'only_release_metadata' to false at the start of each iteration. Fixes: 8257b2dc3c1a ("Btrfs: introduce btrfs_{start, end}_nocow_write() for each subvolume") Fixes: 7ee9e4405f26 ("Btrfs: check if we can nocow if we don't have data space") CC: stable@vger.kernel.org # 4.4+ Reviewed-by: Josef Bacik Signed-off-by: Filipe Manana Reviewed-by: David Sterba Signed-off-by: David Sterba Signed-off-by: Greg Kroah-Hartman

Btrfs: fix metadata space leak on fixup worker failure to set range as delalloc

2019-12-17T19:07:31+00:00

commit 536870071dbc4278264f59c9a2f5f447e584d139 upstream. In the fixup worker, if we fail to mark the range as delalloc in the io tree, we must release the previously reserved metadata, as well as update the outstanding extents counter for the inode, otherwise we leak metadata space. In pratice we can't return an error from btrfs_set_extent_delalloc(), which is just a wrapper around __set_extent_bit(), as for most errors __set_extent_bit() does a BUG_ON() (or panics which hits a BUG_ON() as well) and returning an -EEXIST error doesn't happen in this case since the exclusive bits parameter always has a value of 0 through this code path. Nevertheless, just fix the error handling in the fixup worker, in case one day __set_extent_bit() can return an error to this code path. Fixes: f3038ee3a3f101 ("btrfs: Handle btrfs_set_extent_delalloc failure in fixup worker") CC: stable@vger.kernel.org # 4.19+ Reviewed-by: Nikolay Borisov Signed-off-by: Filipe Manana Reviewed-by: David Sterba Signed-off-by: David Sterba Signed-off-by: Greg Kroah-Hartman

btrfs: use refcount_inc_not_zero in kill_all_nodes

2019-12-17T19:07:30+00:00

commit baf320b9d531f1cfbf64c60dd155ff80a58b3796 upstream. We hit the following warning while running down a different problem [ 6197.175850] ------------[ cut here ]------------ [ 6197.185082] refcount_t: underflow; use-after-free. [ 6197.194704] WARNING: CPU: 47 PID: 966 at lib/refcount.c:190 refcount_sub_and_test_checked+0x53/0x60 [ 6197.521792] Call Trace: [ 6197.526687] __btrfs_release_delayed_node+0x76/0x1c0 [ 6197.536615] btrfs_kill_all_delayed_nodes+0xec/0x130 [ 6197.546532] ? __btrfs_btree_balance_dirty+0x60/0x60 [ 6197.556482] btrfs_clean_one_deleted_snapshot+0x71/0xd0 [ 6197.566910] cleaner_kthread+0xfa/0x120 [ 6197.574573] kthread+0x111/0x130 [ 6197.581022] ? kthread_create_on_node+0x60/0x60 [ 6197.590086] ret_from_fork+0x1f/0x30 [ 6197.597228] ---[ end trace 424bb7ae00509f56 ]--- This is because the free side drops the ref without the lock, and then takes the lock if our refcount is 0. So you can have nodes on the tree that have a refcount of 0. Fix this by zero'ing out that element in our temporary array so we don't try to kill it again. CC: stable@vger.kernel.org # 4.14+ Reviewed-by: Nikolay Borisov Signed-off-by: Josef Bacik Reviewed-by: David Sterba [ add comment ] Signed-off-by: David Sterba Signed-off-by: Greg Kroah-Hartman

btrfs: check page->mapping when loading free space cache

2019-12-17T19:07:29+00:00

commit 3797136b626ad4b6582223660c041efdea8f26b2 upstream. While testing 5.2 we ran into the following panic [52238.017028] BUG: kernel NULL pointer dereference, address: 0000000000000001 [52238.105608] RIP: 0010:drop_buffers+0x3d/0x150 [52238.304051] Call Trace: [52238.308958] try_to_free_buffers+0x15b/0x1b0 [52238.317503] shrink_page_list+0x1164/0x1780 [52238.325877] shrink_inactive_list+0x18f/0x3b0 [52238.334596] shrink_node_memcg+0x23e/0x7d0 [52238.342790] ? do_shrink_slab+0x4f/0x290 [52238.350648] shrink_node+0xce/0x4a0 [52238.357628] balance_pgdat+0x2c7/0x510 [52238.365135] kswapd+0x216/0x3e0 [52238.371425] ? wait_woken+0x80/0x80 [52238.378412] ? balance_pgdat+0x510/0x510 [52238.386265] kthread+0x111/0x130 [52238.392727] ? kthread_create_on_node+0x60/0x60 [52238.401782] ret_from_fork+0x1f/0x30 The page we were trying to drop had a page->private, but had no page->mapping and so called drop_buffers, assuming that we had a buffer_head on the page, and then panic'ed trying to deref 1, which is our page->private for data pages. This is happening because we're truncating the free space cache while we're trying to load the free space cache. This isn't supposed to happen, and I'll fix that in a followup patch. However we still shouldn't allow those sort of mistakes to result in messing with pages that do not belong to us. So add the page->mapping check to verify that we still own this page after dropping and re-acquiring the page lock. This page being unlocked as: btrfs_readpage extent_read_full_page __extent_read_full_page __do_readpage if (!nr) unlock_page <-- nr can be 0 only if submit_extent_page returns an error CC: stable@vger.kernel.org # 4.4+ Reviewed-by: Filipe Manana Reviewed-by: Nikolay Borisov Signed-off-by: Josef Bacik [ add callchain ] Signed-off-by: David Sterba Signed-off-by: Greg Kroah-Hartman

Btrfs: fix log context list corruption after rename exchange operation

2019-11-20T15:49:03+00:00

commit e6c617102c7e4ac1398cb0b98ff1f0727755b520 upstream. During rename exchange we might have successfully log the new name in the source root's log tree, in which case we leave our log context (allocated on stack) in the root's list of log contextes. However we might fail to log the new name in the destination root, in which case we fallback to a transaction commit later and never sync the log of the source root, which causes the source root log context to remain in the list of log contextes. This later causes invalid memory accesses because the context was allocated on stack and after rename exchange finishes the stack gets reused and overwritten for other purposes. The kernel's linked list corruption detector (CONFIG_DEBUG_LIST=y) can detect this and report something like the following: [ 691.489929] ------------[ cut here ]------------ [ 691.489947] list_add corruption. prev->next should be next (ffff88819c944530), but was ffff8881c23f7be4. (prev=ffff8881c23f7a38). [ 691.489967] WARNING: CPU: 2 PID: 28933 at lib/list_debug.c:28 __list_add_valid+0x95/0xe0 (...) [ 691.489998] CPU: 2 PID: 28933 Comm: fsstress Not tainted 5.4.0-rc6-btrfs-next-62 #1 [ 691.490001] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.12.0-0-ga698c8995f-prebuilt.qemu.org 04/01/2014 [ 691.490003] RIP: 0010:__list_add_valid+0x95/0xe0 (...) [ 691.490007] RSP: 0018:ffff8881f0b3faf8 EFLAGS: 00010282 [ 691.490010] RAX: 0000000000000000 RBX: ffff88819c944530 RCX: 0000000000000000 [ 691.490011] RDX: 0000000000000001 RSI: 0000000000000008 RDI: ffffffffa2c497e0 [ 691.490013] RBP: ffff8881f0b3fe68 R08: ffffed103eaa4115 R09: ffffed103eaa4114 [ 691.490015] R10: ffff88819c944000 R11: ffffed103eaa4115 R12: 7fffffffffffffff [ 691.490016] R13: ffff8881b4035610 R14: ffff8881e7b84728 R15: 1ffff1103e167f7b [ 691.490019] FS: 00007f4b25ea2e80(0000) GS:ffff8881f5500000(0000) knlGS:0000000000000000 [ 691.490021] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [ 691.490022] CR2: 00007fffbb2d4eec CR3: 00000001f2a4a004 CR4: 00000000003606e0 [ 691.490025] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000 [ 691.490027] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400 [ 691.490029] Call Trace: [ 691.490058] btrfs_log_inode_parent+0x667/0x2730 [btrfs] [ 691.490083] ? join_transaction+0x24a/0xce0 [btrfs] [ 691.490107] ? btrfs_end_log_trans+0x80/0x80 [btrfs] [ 691.490111] ? dget_parent+0xb8/0x460 [ 691.490116] ? lock_downgrade+0x6b0/0x6b0 [ 691.490121] ? rwlock_bug.part.0+0x90/0x90 [ 691.490127] ? do_raw_spin_unlock+0x142/0x220 [ 691.490151] btrfs_log_dentry_safe+0x65/0x90 [btrfs] [ 691.490172] btrfs_sync_file+0x9f1/0xc00 [btrfs] [ 691.490195] ? btrfs_file_write_iter+0x1800/0x1800 [btrfs] [ 691.490198] ? rcu_read_lock_any_held.part.11+0x20/0x20 [ 691.490204] ? __do_sys_newstat+0x88/0xd0 [ 691.490207] ? cp_new_stat+0x5d0/0x5d0 [ 691.490218] ? do_fsync+0x38/0x60 [ 691.490220] do_fsync+0x38/0x60 [ 691.490224] __x64_sys_fdatasync+0x32/0x40 [ 691.490228] do_syscall_64+0x9f/0x540 [ 691.490233] entry_SYSCALL_64_after_hwframe+0x49/0xbe [ 691.490235] RIP: 0033:0x7f4b253ad5f0 (...) [ 691.490239] RSP: 002b:00007fffbb2d6078 EFLAGS: 00000246 ORIG_RAX: 000000000000004b [ 691.490242] RAX: ffffffffffffffda RBX: 0000000000000003 RCX: 00007f4b253ad5f0 [ 691.490244] RDX: 00007fffbb2d5fe0 RSI: 00007fffbb2d5fe0 RDI: 0000000000000003 [ 691.490245] RBP: 000000000000000d R08: 0000000000000001 R09: 00007fffbb2d608c [ 691.490247] R10: 00000000000002e8 R11: 0000000000000246 R12: 00000000000001f4 [ 691.490248] R13: 0000000051eb851f R14: 00007fffbb2d6120 R15: 00005635a498bda0 This started happening recently when running some test cases from fstests like btrfs/004 for example, because support for rename exchange was added last week to fsstress from fstests. So fix this by deleting the log context for the source root from the list if we have logged the new name in the source root. Reported-by: Su Yue Fixes: d4682ba03ef618 ("Btrfs: sync log after logging new name") CC: stable@vger.kernel.org # 4.19+ Tested-by: Su Yue Signed-off-by: Filipe Manana Signed-off-by: David Sterba Signed-off-by: Greg Kroah-Hartman

btrfs: save i_size to avoid double evaluation of i_size_read in compress_file_range

2019-11-12T18:26:59+00:00

commit d98da49977f67394db492f06c00b1fb1cc090c05 upstream. We hit a regression while rolling out 5.2 internally where we were hitting the following panic kernel BUG at mm/page-writeback.c:2659! RIP: 0010:clear_page_dirty_for_io+0xe6/0x1f0 Call Trace: __process_pages_contig+0x25a/0x350 ? extent_clear_unlock_delalloc+0x43/0x70 submit_compressed_extents+0x359/0x4d0 normal_work_helper+0x15a/0x330 process_one_work+0x1f5/0x3f0 worker_thread+0x2d/0x3d0 ? rescuer_thread+0x340/0x340 kthread+0x111/0x130 ? kthread_create_on_node+0x60/0x60 ret_from_fork+0x1f/0x30 This is happening because the page is not locked when doing clear_page_dirty_for_io. Looking at the core dump it was because our async_extent had a ram_size of 24576 but our async_chunk range only spanned 20480, so we had a whole extra page in our ram_size for our async_extent. This happened because we try not to compress pages outside of our i_size, however a cleanup patch changed us to do actual_end = min_t(u64, i_size_read(inode), end + 1); which is problematic because i_size_read() can evaluate to different values in between checking and assigning. So either an expanding truncate or a fallocate could increase our i_size while we're doing writeout and actual_end would end up being past the range we have locked. I confirmed this was what was happening by installing a debug kernel that had actual_end = min_t(u64, i_size_read(inode), end + 1); if (actual_end > end + 1) { printk(KERN_ERR "KABOOM\n"); actual_end = end + 1; } and installing it onto 500 boxes of the tier that had been seeing the problem regularly. Last night I got my debug message and no panic, confirming what I expected. [ dsterba: the assembly confirms a tiny race window: mov 0x20(%rsp),%rax cmp %rax,0x48(%r15) # read movl $0x0,0x18(%rsp) mov %rax,%r12 mov %r14,%rax cmovbe 0x48(%r15),%r12 # eval Where r15 is inode and 0x48 is offset of i_size. The original fix was to revert 62b37622718c that would do an intermediate assignment and this would also avoid the doulble evaluation but is not future-proof, should the compiler merge the stores and call i_size_read anyway. There's a patch adding READ_ONCE to i_size_read but that's not being applied at the moment and we need to fix the bug. Instead, emulate READ_ONCE by two barrier()s that's what effectively happens. The assembly confirms single evaluation: mov 0x48(%rbp),%rax # read once mov 0x20(%rsp),%rcx mov $0x20,%edx cmp %rax,%rcx cmovbe %rcx,%rax mov %rax,(%rsp) mov %rax,%rcx mov %r14,%rax Where 0x48(%rbp) is inode->i_size stored to %eax. ] Fixes: 62b37622718c ("btrfs: Remove isize local variable in compress_file_range") CC: stable@vger.kernel.org # v5.1+ Reviewed-by: Filipe Manana Signed-off-by: Josef Bacik Reviewed-by: David Sterba [ changelog updated ] Signed-off-by: David Sterba Signed-off-by: Greg Kroah-Hartman