-
Filipe Manana authored
If a transaction aborts it can cause a memory leak of the pages array of a block group's io_ctl structure. The following steps explain how that can happen: 1) Transaction N is committing, currently in state TRANS_STATE_UNBLOCKED and it's about to start writing out dirty extent buffers; 2) Transaction N + 1 already started and another task, task A, just called btrfs_commit_transaction() on it; 3) Block group B was dirtied (extents allocated from it) by transaction N + 1, so when task A calls btrfs_start_dirty_block_groups(), at the very beginning of the transaction commit, it starts writeback for the block group's space cache by calling btrfs_write_out_cache(), which allocates the pages array for the block group's io_ctl with a call to io_ctl_init(). Block group A is added to the io_list of transaction N + 1 by btrfs_start_dirty_block_groups(); 4) While transaction N's commit is writing out the extent buffers, it gets an IO error and aborts transaction N, also setting the file system to RO mode; 5) Task A has already returned from btrfs_start_dirty_block_groups(), is at btrfs_commit_transaction() and has set transaction N + 1 state to TRANS_STATE_COMMIT_START. Immediately after that it checks that the filesystem was turned to RO mode, due to transaction N's abort, and jumps to the "cleanup_transaction" label. After that we end up at btrfs_cleanup_one_transaction() which calls btrfs_cleanup_dirty_bgs(). That helper finds block group B in the transaction's io_list but it never releases the pages array of the block group's io_ctl, resulting in a memory leak. In fact at the point when we are at btrfs_cleanup_dirty_bgs(), the pages array points to pages that were already released by us at __btrfs_write_out_cache() through the call to io_ctl_drop_pages(). We end up freeing the pages array only after waiting for the ordered extent to complete through btrfs_wait_cache_io(), which calls io_ctl_free() to do that. But in the transaction abort case we don't wait for the space cache's ordered extent to complete through a call to btrfs_wait_cache_io(), so that's why we end up with a memory leak - we wait for the ordered extent to complete indirectly by shutting down the work queues and waiting for any jobs in them to complete before returning from close_ctree(). We can solve the leak simply by freeing the pages array right after releasing the pages (with the call to io_ctl_drop_pages()) at __btrfs_write_out_cache(), since we will never use it anymore after that and the pages array points to already released pages at that point, which is currently not a problem since no one will use it after that, but not a good practice anyway since it can easily lead to use-after-free issues. So fix this by freeing the pages array right after releasing the pages at __btrfs_write_out_cache(). This issue can often be reproduced with test case generic/475 from fstests and kmemleak can detect it and reports it with the following trace: unreferenced object 0xffff9bbf009fa600 (size 512): comm "fsstress", pid 38807, jiffies 4298504428 (age 22.028s) hex dump (first 32 bytes): 00 a0 7c 4d 3d ed ff ff 40 a0 7c 4d 3d ed ff ff ..|M=...@.|M=... 80 a0 7c 4d 3d ed ff ff c0 a0 7c 4d 3d ed ff ff ..|M=.....|M=... backtrace: [<00000000f4b5cfe2>] __kmalloc+0x1a8/0x3e0 [<0000000028665e7f>] io_ctl_init+0xa7/0x120 [btrfs] [<00000000a1f95b2d>] __btrfs_write_out_cache+0x86/0x4a0 [btrfs] [<00000000207ea1b0>] btrfs_write_out_cache+0x7f/0xf0 [btrfs] [<00000000af21f534>] btrfs_start_dirty_block_groups+0x27b/0x580 [btrfs] [<00000000c3c23d44>] btrfs_commit_transaction+0xa6f/0xe70 [btrfs] [<000000009588930c>] create_subvol+0x581/0x9a0 [btrfs] [<000000009ef2fd7f>] btrfs_mksubvol+0x3fb/0x4a0 [btrfs] [<00000000474e5187>] __btrfs_ioctl_snap_create+0x119/0x1a0 [btrfs] [<00000000708ee349>] btrfs_ioctl_snap_create_v2+0xb0/0xf0 [btrfs] [<00000000ea60106f>] btrfs_ioctl+0x12c/0x3130 [btrfs] [<000000005c923d6d>] __x64_sys_ioctl+0x83/0xb0 [<0000000043ace2c9>] do_syscall_64+0x33/0x80 [<00000000904efbce>] entry_SYSCALL_64_after_hwframe+0x44/0xa9 CC: stable@vger.kernel.org # 4.9+ Reviewed-by: Josef Bacik <josef@toxicpanda.com> Signed-off-by: Filipe Manana <fdmanana@suse.com> Signed-off-by: David Sterba <dsterba@suse.com>
bbc37d6e