Re: [PATCH v2] ext4: fix a data race in EXT4_I(inode)->i_disksize

From: Qian Cai
Date: Fri Feb 07 2020 - 10:38:54 EST


On Fri, 2020-02-07 at 16:12 +0100, Marco Elver wrote:
> On Fri, 7 Feb 2020 at 15:29, Qian Cai <cai@xxxxxx> wrote:
> >
> > EXT4_I(inode)->i_disksize could be accessed concurrently as noticed by
> > KCSAN,
> >
> > BUG: KCSAN: data-race in ext4_write_end [ext4] / ext4_writepages [ext4]
> >
> > write to 0xffff91c6713b00f8 of 8 bytes by task 49268 on cpu 127:
> > ext4_write_end+0x4e3/0x750 [ext4]
> > ext4_update_i_disksize at fs/ext4/ext4.h:3032
> > (inlined by) ext4_update_inode_size at fs/ext4/ext4.h:3046
> > (inlined by) ext4_write_end at fs/ext4/inode.c:1287
> > generic_perform_write+0x208/0x2a0
> > 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+0x91/0xb47
> > entry_SYSCALL_64_after_hwframe+0x49/0xbe
> >
> > read to 0xffff91c6713b00f8 of 8 bytes by task 24872 on cpu 37:
> > ext4_writepages+0x10ac/0x1d00 [ext4]
> > mpage_map_and_submit_extent at fs/ext4/inode.c:2468
> > (inlined by) ext4_writepages at fs/ext4/inode.c:2772
> > do_writepages+0x5e/0x130
> > __writeback_single_inode+0xeb/0xb20
> > writeback_sb_inodes+0x429/0x900
> > __writeback_inodes_wb+0xc4/0x150
> > wb_writeback+0x4bd/0x870
> > wb_workfn+0x6b4/0x960
> > process_one_work+0x54c/0xbe0
> > worker_thread+0x80/0x650
> > kthread+0x1e0/0x200
> > ret_from_fork+0x27/0x50
> >
> > Reported by Kernel Concurrency Sanitizer on:
> > CPU: 37 PID: 24872 Comm: kworker/u261:2 Tainted: G W O L 5.5.0-next-20200204+ #5
> > Hardware name: HPE ProLiant DL385 Gen10/ProLiant DL385 Gen10, BIOS A40 07/10/2019
> > Workqueue: writeback wb_workfn (flush-7:0)
> >
> > Since only the read is operating as lockless (outside of the
> > "i_data_sem"), load tearing could introduce a logic bug. Fix it by
> > adding READ_ONCE() for the read and WRITE_ONCE() for the write.
> >
> > Signed-off-by: Qian Cai <cai@xxxxxx>
> > ---
> >
> > v2: also add WRITE_ONCE() which is recommended even for fixing load tearing.
>
> Just a note: I keep seeing 'load tearing' mentioned as the only reason:
>
> - The WRITE_ONCE avoids store-tearing (and other optimizations).
>
> - We're not only interested in avoiding load/store tearing. There
> are plenty other compiler optimizations that can break concurrent
> code: https://lwn.net/Articles/793253/

I also realized that from that article, store tearing is strictly from multiple
concurrent writers. However, in the sense of without the WRITE_ONCE() here,
compilers could still have 2 store instructions, so

CPU0: CPU1:
store #1
read
store #2

which was not mentioned in that article. I called it also load tearing, but
maybe you will call that store tearing. Do I understand correctly?

>
> Thanks,
> -- Marco
>
>
> > fs/ext4/ext4.h | 2 +-
> > fs/ext4/inode.c | 2 +-
> > 2 files changed, 2 insertions(+), 2 deletions(-)
> >
> > diff --git a/fs/ext4/ext4.h b/fs/ext4/ext4.h
> > index 9a2ee2428ecc..8329ccc82fa9 100644
> > --- a/fs/ext4/ext4.h
> > +++ b/fs/ext4/ext4.h
> > @@ -3029,7 +3029,7 @@ static inline void ext4_update_i_disksize(struct inode *inode, loff_t newsize)
> > !inode_is_locked(inode));
> > down_write(&EXT4_I(inode)->i_data_sem);
> > if (newsize > EXT4_I(inode)->i_disksize)
> > - EXT4_I(inode)->i_disksize = newsize;
> > + WRITE_ONCE(EXT4_I(inode)->i_disksize, newsize);
> > up_write(&EXT4_I(inode)->i_data_sem);
> > }
> >
> > diff --git a/fs/ext4/inode.c b/fs/ext4/inode.c
> > index 3313168b680f..6f9862bf63f1 100644
> > --- a/fs/ext4/inode.c
> > +++ b/fs/ext4/inode.c
> > @@ -2465,7 +2465,7 @@ static int mpage_map_and_submit_extent(handle_t *handle,
> > * truncate are avoided by checking i_size under i_data_sem.
> > */
> > disksize = ((loff_t)mpd->first_page) << PAGE_SHIFT;
> > - if (disksize > EXT4_I(inode)->i_disksize) {
> > + if (disksize > READ_ONCE(EXT4_I(inode)->i_disksize)) {
> > int err2;
> > loff_t i_size;
> >
> > --
> > 1.8.3.1
> >