Thanks to Christian Brauner for the reminder and Jeff Layton for the fix suggestion. I will base my patch on vfs.fixes and resend the patch v4 after fixing it.
> -----原始郵件-----
> 發件人: "Jeff Layton" <jlayton@xxxxxxxxxx>
> 發送時間:2024-11-23 22:01:08 (星期六)
> 收件人: "Hao-ran Zheng" <zhenghaoran@xxxxxxxxxxx>, viro@xxxxxxxxxxxxxxxxxx, brauner@xxxxxxxxxx, jack@xxxxxxx, linux-fsdevel@xxxxxxxxxxxxxxx, linux-kernel@xxxxxxxxxxxxxxx
> 抄送: baijiaju1990@xxxxxxxxx, 21371365@xxxxxxxxxxx
> 主題: Re: [PATCH v3] fs: Fix data race in inode_set_ctime_to_ts
>
> On Fri, 2024-11-22 at 21:06 +0800, Hao-ran Zheng wrote:
> > A data race may occur when the function `inode_set_ctime_to_ts()` and
> > the function `inode_get_ctime_sec()` are executed concurrently. When
> > two threads call `aio_read` and `aio_write` respectively, they will
> > be distributed to the read and write functions of the corresponding
> > file system respectively. Taking the btrfs file system as an example,
> > the `btrfs_file_read_iter` and `btrfs_file_write_iter` functions are
> > finally called. These two functions created a data race when they
> > finally called `inode_get_ctime_sec()` and `inode_set_ctime_to_ns()`.
> > The specific call stack that appears during testing is as follows:
> >
> > ============DATA_RACE============
> > btrfs_delayed_update_inode+0x1f61/0x7ce0 [btrfs]
> > btrfs_update_inode+0x45e/0xbb0 [btrfs]
> > btrfs_dirty_inode+0x2b8/0x530 [btrfs]
> > btrfs_update_time+0x1ad/0x230 [btrfs]
> > touch_atime+0x211/0x440
> > filemap_read+0x90f/0xa20
> > btrfs_file_read_iter+0xeb/0x580 [btrfs]
> > aio_read+0x275/0x3a0
> > io_submit_one+0xd22/0x1ce0
> > __se_sys_io_submit+0xb3/0x250
> > do_syscall_64+0xc1/0x190
> > entry_SYSCALL_64_after_hwframe+0x77/0x7f
> > ============OTHER_INFO============
> > btrfs_write_check+0xa15/0x1390 [btrfs]
> > btrfs_buffered_write+0x52f/0x29d0 [btrfs]
> > btrfs_do_write_iter+0x53d/0x1590 [btrfs]
> > btrfs_file_write_iter+0x41/0x60 [btrfs]
> > aio_write+0x41e/0x5f0
> > io_submit_one+0xd42/0x1ce0
> > __se_sys_io_submit+0xb3/0x250
> > do_syscall_64+0xc1/0x190
> > entry_SYSCALL_64_after_hwframe+0x77/0x7f
> >
> > To address this issue, it is recommended to add WRITE_ONCE
> > when writing the `inode->i_ctime_sec` variable.and add
> > READ_ONCE when reading in function `inode_get_ctime_sec()`
> > and `inode_get_ctime_nsec()`.
> >
> > Signed-off-by: Hao-ran Zheng <zhenghaoran@xxxxxxxxxxx>
> > ---
> > V2 -> V3: Added READ_ONCE in inode_get_ctime_nsec() and addressed review comments
> > V1 -> V2: Added READ_ONCE in inode_get_ctime_sec()
> > ---
> > include/linux/fs.h | 8 ++++----
> > 1 file changed, 4 insertions(+), 4 deletions(-)
> >
> > diff --git a/include/linux/fs.h b/include/linux/fs.h
> > index 3559446279c1..c18f9a9ee5e7 100644
> > --- a/include/linux/fs.h
> > +++ b/include/linux/fs.h
> > @@ -1655,12 +1655,12 @@ static inline struct timespec64 inode_set_mtime(struct inode *inode,
> >
> > static inline time64_t inode_get_ctime_sec(const struct inode *inode)
> > {
> > - return inode->i_ctime_sec;
> > + return READ_ONCE(inode->i_ctime_sec);
> > }
> >
> > static inline long inode_get_ctime_nsec(const struct inode *inode)
> > {
> > - return inode->i_ctime_nsec;
> > + return READ_ONCE(inode->i_ctime_nsec);
> > }
> >
> > static inline struct timespec64 inode_get_ctime(const struct inode *inode)
> > @@ -1674,8 +1674,8 @@ static inline struct timespec64 inode_get_ctime(const struct inode *inode)
> > static inline struct timespec64 inode_set_ctime_to_ts(struct inode *inode,
> > struct timespec64 ts)
> > {
> > - inode->i_ctime_sec = ts.tv_sec;
> > - inode->i_ctime_nsec = ts.tv_nsec;
> > + WRITE_ONCE(inode->i_ctime_sec, ts.tv_sec);
> > + WRITE_ONCE(inode->i_ctime_nsec, ts.tv_nsec);
> > return ts;
> > }
> >
>
> Looks reasonable. There are also bare fetches and stores of the
> i_ctime_sec field in inode_set_ctime_current(). Do we need something
> like this in addition to the above?
>
>
> diff --git a/fs/inode.c b/fs/inode.c
> index b13b778257ae..d869ee6f1c6b 100644
> --- a/fs/inode.c
> +++ b/fs/inode.c
> @@ -2788,7 +2788,7 @@ struct timespec64 inode_set_ctime_current(struct inode *inode)
> */
> cns = smp_load_acquire(&inode->i_ctime_nsec);
> if (cns & I_CTIME_QUERIED) {
> - struct timespec64 ctime = { .tv_sec = inode->i_ctime_sec,
> + struct timespec64 ctime = { .tv_sec = READ_ONCE(inode->i_ctime_sec),
> .tv_nsec = cns & ~I_CTIME_QUERIED };
>
> if (timespec64_compare(&now, &ctime) <= 0) {
> @@ -2809,7 +2809,7 @@ struct timespec64 inode_set_ctime_current(struct inode *inode)
> /* Try to swap the nsec value into place. */
> if (try_cmpxchg(&inode->i_ctime_nsec, &cur, now.tv_nsec)) {
> /* If swap occurred, then we're (mostly) done */
> - inode->i_ctime_sec = now.tv_sec;
> + WRITE_ONCE(inode->i_ctime_sec, now.tv_sec);
> trace_ctime_ns_xchg(inode, cns, now.tv_nsec, cur);
> mgtime_counter_inc(mg_ctime_swaps);
> } else {
> @@ -2824,7 +2824,7 @@ struct timespec64 inode_set_ctime_current(struct inode *inode)
> goto retry;
> }
> /* Otherwise, keep the existing ctime */
> - now.tv_sec = inode->i_ctime_sec;
> + now.tv_sec = READ_ONCE(inode->i_ctime_sec);
> now.tv_nsec = cur & ~I_CTIME_QUERIED;
> }
> out:
