lkml.org 
[lkml]   [2020]   [Feb]   [24]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH] fs/jbd2: fix data races at struct journal_head
On Fri 21-02-20 23:31:11, Qian Cai wrote:
> journal_head::b_transaction and journal_head::b_next_transaction could
> be accessed concurrently as noticed by KCSAN,
>
> LTP: starting fsync04
> /dev/zero: Can't open blockdev
> EXT4-fs (loop0): mounting ext3 file system using the ext4 subsystem
> EXT4-fs (loop0): mounted filesystem with ordered data mode. Opts: (null)
> ==================================================================
> BUG: KCSAN: data-race in __jbd2_journal_refile_buffer [jbd2] / jbd2_write_access_granted [jbd2]
>
> write to 0xffff99f9b1bd0e30 of 8 bytes by task 25721 on cpu 70:
> __jbd2_journal_refile_buffer+0xdd/0x210 [jbd2]
> __jbd2_journal_refile_buffer at fs/jbd2/transaction.c:2569
> jbd2_journal_commit_transaction+0x2d15/0x3f20 [jbd2]
> (inlined by) jbd2_journal_commit_transaction at fs/jbd2/commit.c:1034
> kjournald2+0x13b/0x450 [jbd2]
> kthread+0x1cd/0x1f0
> ret_from_fork+0x27/0x50
>
> read to 0xffff99f9b1bd0e30 of 8 bytes by task 25724 on cpu 68:
> jbd2_write_access_granted+0x1b2/0x250 [jbd2]
> jbd2_write_access_granted at fs/jbd2/transaction.c:1155
> jbd2_journal_get_write_access+0x2c/0x60 [jbd2]
> __ext4_journal_get_write_access+0x50/0x90 [ext4]
> ext4_mb_mark_diskspace_used+0x158/0x620 [ext4]
> ext4_mb_new_blocks+0x54f/0xca0 [ext4]
> ext4_ind_map_blocks+0xc79/0x1b40 [ext4]
> ext4_map_blocks+0x3b4/0x950 [ext4]
> _ext4_get_block+0xfc/0x270 [ext4]
> ext4_get_block+0x3b/0x50 [ext4]
> __block_write_begin_int+0x22e/0xae0
> __block_write_begin+0x39/0x50
> ext4_write_begin+0x388/0xb50 [ext4]
> generic_perform_write+0x15d/0x290
> ext4_buffered_write_iter+0x11f/0x210 [ext4]
> ext4_file_write_iter+0xce/0x9e0 [ext4]
> new_sync_write+0x29c/0x3b0
> __vfs_write+0x92/0xa0
> vfs_write+0x103/0x260
> ksys_write+0x9d/0x130
> __x64_sys_write+0x4c/0x60
> do_syscall_64+0x91/0xb05
> entry_SYSCALL_64_after_hwframe+0x49/0xbe
>
> 5 locks held by fsync04/25724:
> #0: ffff99f9911093f8 (sb_writers#13){.+.+}, at: vfs_write+0x21c/0x260
> #1: ffff99f9db4c0348 (&sb->s_type->i_mutex_key#15){+.+.}, at: ext4_buffered_write_iter+0x65/0x210 [ext4]
> #2: ffff99f5e7dfcf58 (jbd2_handle){++++}, at: start_this_handle+0x1c1/0x9d0 [jbd2]
> #3: ffff99f9db4c0168 (&ei->i_data_sem){++++}, at: ext4_map_blocks+0x176/0x950 [ext4]
> #4: ffffffff99086b40 (rcu_read_lock){....}, at: jbd2_write_access_granted+0x4e/0x250 [jbd2]
> irq event stamp: 1407125
> hardirqs last enabled at (1407125): [<ffffffff980da9b7>] __find_get_block+0x107/0x790
> hardirqs last disabled at (1407124): [<ffffffff980da8f9>] __find_get_block+0x49/0x790
> softirqs last enabled at (1405528): [<ffffffff98a0034c>] __do_softirq+0x34c/0x57c
> softirqs last disabled at (1405521): [<ffffffff97cc67a2>] irq_exit+0xa2/0xc0
>
> Reported by Kernel Concurrency Sanitizer on:
> CPU: 68 PID: 25724 Comm: fsync04 Tainted: G L 5.6.0-rc2-next-20200221+ #7
> Hardware name: HPE ProLiant DL385 Gen10/ProLiant DL385 Gen10, BIOS A40 07/10/2019
>
> The plain reads are outside of jh->b_state_lock critical section which result
> in data races. Fix them by adding pairs of READ|WRITE_ONCE().
>
> Signed-off-by: Qian Cai <cai@lca.pw>

Yeah, makes sense. You can add:

Reviewed-by: Jan Kara <jack@suse.cz>

Honza

> ---
> fs/jbd2/transaction.c | 8 ++++----
> 1 file changed, 4 insertions(+), 4 deletions(-)
>
> diff --git a/fs/jbd2/transaction.c b/fs/jbd2/transaction.c
> index 2dd848a743ed..c5f7f6d0f33b 100644
> --- a/fs/jbd2/transaction.c
> +++ b/fs/jbd2/transaction.c
> @@ -1152,8 +1152,8 @@ static bool jbd2_write_access_granted(handle_t *handle, struct buffer_head *bh,
> /* For undo access buffer must have data copied */
> if (undo && !jh->b_committed_data)
> goto out;
> - if (jh->b_transaction != handle->h_transaction &&
> - jh->b_next_transaction != handle->h_transaction)
> + if (READ_ONCE(jh->b_transaction) != handle->h_transaction &&
> + READ_ONCE(jh->b_next_transaction) != handle->h_transaction)
> goto out;
> /*
> * There are two reasons for the barrier here:
> @@ -2565,8 +2565,8 @@ bool __jbd2_journal_refile_buffer(struct journal_head *jh)
> * our jh reference and thus __jbd2_journal_file_buffer() must not
> * take a new one.
> */
> - jh->b_transaction = jh->b_next_transaction;
> - jh->b_next_transaction = NULL;
> + WRITE_ONCE(jh->b_transaction, jh->b_next_transaction);
> + WRITE_ONCE(jh->b_next_transaction, NULL);
> if (buffer_freed(bh))
> jlist = BJ_Forget;
> else if (jh->b_modified)
> --
> 2.21.0 (Apple Git-122.2)
>
--
Jan Kara <jack@suse.com>
SUSE Labs, CR

\
 
 \ /
  Last update: 2020-02-24 10:12    [W:0.045 / U:0.084 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site