On Fri, 7 Feb 2020 at 16:38, Qian Cai <cai@xxxxxx> wrote: > > 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? The effect is the same, so yes. If you have the writer side split the write, but have a concurrent load, the observed value will appear "teared". Similar if the reader side splits the reads (the more obvious case). > > > > 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 > > >
