Re: BUG: KCSAN: data-race in btrfs_sync_log [btrfs] / btrfs_update_inode [btrfs]

From: David Sterba
Date: Fri Oct 06 2023 - 12:33:39 EST


On Thu, Sep 21, 2023 at 10:15:49PM +0200, Mirsad Todorovac wrote:
> Hi,
>
> On the vanilla 6.6-rc2 torvalds kernel KCSAN reported this data-race:
>
> [ 2690.990793] ==================================================================
> [ 2690.991470] BUG: KCSAN: data-race in btrfs_sync_log [btrfs] / btrfs_update_inode [btrfs]
>
> [ 2690.992804] write to 0xffff88811b57faf8 of 4 bytes by task 40555 on cpu 20:
> [ 2690.992815] btrfs_sync_log (/home/marvin/linux/kernel/torvalds2/fs/btrfs/tree-log.c:2964) btrfs
> [ 2690.993484] btrfs_sync_file (/home/marvin/linux/kernel/torvalds2/fs/btrfs/file.c:1954) btrfs
> [ 2690.994149] vfs_fsync_range (/home/marvin/linux/kernel/torvalds2/fs/sync.c:188)
> [ 2690.994161] __x64_sys_fsync (/home/marvin/linux/kernel/torvalds2/./include/linux/file.h:45 /home/marvin/linux/kernel/torvalds2/fs/sync.c:213 /home/marvin/linux/kernel/torvalds2/fs/sync.c:220 /home/marvin/linux/kernel/torvalds2/fs/sync.c:218 /home/marvin/linux/kernel/torvalds2/fs/sync.c:218)
> [ 2690.994172] do_syscall_64 (/home/marvin/linux/kernel/torvalds2/arch/x86/entry/common.c:50 /home/marvin/linux/kernel/torvalds2/arch/x86/entry/common.c:80)
> [ 2690.994186] entry_SYSCALL_64_after_hwframe (/home/marvin/linux/kernel/torvalds2/arch/x86/entry/entry_64.S:120)
>
> [ 2690.994203] read to 0xffff88811b57faf8 of 4 bytes by task 5338 on cpu 21:
> [ 2690.994214] btrfs_update_inode (/home/marvin/linux/kernel/torvalds2/fs/btrfs/transaction.h:175 /home/marvin/linux/kernel/torvalds2/fs/btrfs/inode.c:4016) btrfs
> [ 2690.994877] btrfs_finish_one_ordered (/home/marvin/linux/kernel/torvalds2/fs/btrfs/inode.c:4028 /home/marvin/linux/kernel/torvalds2/fs/btrfs/inode.c:3139) btrfs
> [ 2690.995541] btrfs_finish_ordered_io (/home/marvin/linux/kernel/torvalds2/fs/btrfs/inode.c:3230) btrfs
> [ 2690.996205] finish_ordered_fn (/home/marvin/linux/kernel/torvalds2/fs/btrfs/ordered-data.c:304) btrfs
> [ 2690.996871] btrfs_work_helper (/home/marvin/linux/kernel/torvalds2/fs/btrfs/async-thread.c:314) btrfs
> [ 2690.997539] process_one_work (/home/marvin/linux/kernel/torvalds2/kernel/workqueue.c:2630)
> [ 2690.997551] worker_thread (/home/marvin/linux/kernel/torvalds2/kernel/workqueue.c:2697 /home/marvin/linux/kernel/torvalds2/kernel/workqueue.c:2784)
> [ 2690.997562] kthread (/home/marvin/linux/kernel/torvalds2/kernel/kthread.c:388)
> [ 2690.997571] ret_from_fork (/home/marvin/linux/kernel/torvalds2/arch/x86/kernel/process.c:147)
> [ 2690.997583] ret_from_fork_asm (/home/marvin/linux/kernel/torvalds2/arch/x86/entry/entry_64.S:312)
>
> [ 2690.997598] value changed: 0x00000004 -> 0x00000005
>
> [ 2690.997613] Reported by Kernel Concurrency Sanitizer on:
> [ 2690.997621] CPU: 21 PID: 5338 Comm: kworker/u65:7 Tainted: G L 6.6.0-rc2-kcsan-00143-gb5cbe7c00aa0 #41
> [ 2690.997633] Hardware name: ASRock X670E PG Lightning/X670E PG Lightning, BIOS 1.21 04/26/2023
> [ 2690.997640] Workqueue: btrfs-endio-write btrfs_work_helper [btrfs]
> [ 2690.998311] ==================================================================
>
> fs/btrfs/tree-log.c
> -------------------------------------
> 2948 /*
> 2949 * We _must_ update under the root->log_mutex in order to make sure we
> 2950 * have a consistent view of the log root we are trying to commit at
> 2951 * this moment.
> 2952 *
> 2953 * We _must_ copy this into a local copy, because we are not holding the
> 2954 * log_root_tree->log_mutex yet. This is important because when we
> 2955 * commit the log_root_tree we must have a consistent view of the
> 2956 * log_root_tree when we update the super block to point at the
> 2957 * log_root_tree bytenr. If we update the log_root_tree here we'll race
> 2958 * with the commit and possibly point at the new block which we may not
> 2959 * have written out.
> 2960 */
> 2961 btrfs_set_root_node(&log->root_item, log->node);
> 2962 memcpy(&new_root_item, &log->root_item, sizeof(new_root_item));
> 2963
> 2964 → root->log_transid++;
> 2965 log->log_transid = root->log_transid;
> 2966 root->log_start_pid = 0;
> 2967 /*
> 2968 * IO has been started, blocks of the log tree have WRITTEN flag set
> 2969 * in their headers. new modifications of the log will be written to
> 2970 * new positions. so it's safe to allow log writers to go in.
> 2971 */
> 2972 mutex_unlock(&root->log_mutex);
>
> fs/btrfs/transaction.h
> ----------------------------------
> 170 static inline void btrfs_set_inode_last_trans(struct btrfs_trans_handle *trans,
> 171 struct btrfs_inode *inode)
> 172 {
> 173 spin_lock(&inode->lock);
> 174 inode->last_trans = trans->transaction->transid;
> 175 inode->last_sub_trans = inode->root->log_transid;
> 176 inode->last_log_commit = inode->last_sub_trans - 1;
> 177 spin_unlock(&inode->lock);
> 178 }
>
> I am not certain whether the reader and writer side contend for the same lock, but it
> seems that on the safe side would be putting the reader into READ_ONCE() to get a consistent
> value?:

Filipe send a series adding the READ_ONCE/WRITE_ONCE annotations for the
log_transid (and more):
https://lore.kernel.org/linux-btrfs/cover.1696415673.git.fdmanana@xxxxxxxx/

This will appear in linux-next soon.