On Sat 18-03-23 16:51:00, Zhihao Cheng wrote:
> > Hi, Jan
> > > On Fri 17-03-23 09:35:53, Zhihao Cheng wrote:
> > > > Following process makes i_disksize exceed i_size:
> > > >
> > > > generic_perform_write
> > > > copied = iov_iter_copy_from_user_atomic(len) // copied < len
> > > > ext4_da_write_end
> > > > | ext4_update_i_disksize
> > > > | new_i_size = pos + copied;
> > > > | WRITE_ONCE(EXT4_I(inode)->i_disksize, newsize) // update i_disksize
> > > > | generic_write_end
> > > > | copied = block_write_end(copied, len) // copied = 0
> > > > | if (unlikely(copied < len))
> > > > | if (!PageUptodate(page))
> > > > | copied = 0;
> > > > | if (pos + copied > inode->i_size) // return false
> > > > if (unlikely(copied == 0))
> > > > goto again;
> > > > if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
> > > > status = -EFAULT;
> > > > break;
> > > > }
> > > >
> > > > We get i_disksize greater than i_size here, which could trigger WARNING
> > > > check 'i_size_read(inode) < EXT4_I(inode)->i_disksize' while doing dio:
> > > >
> > > > ext4_dio_write_iter
> > > > iomap_dio_rw
> > > > __iomap_dio_rw // return err, length is not aligned to 512
> > > > ext4_handle_inode_extension
> > > > WARN_ON_ONCE(i_size_read(inode) < EXT4_I(inode)->i_disksize) // Oops
> > > >
> > > > WARNING: CPU: 2 PID: 2609 at fs/ext4/file.c:319
> > > > CPU: 2 PID: 2609 Comm: aa Not tainted 6.3.0-rc2
> > > > RIP: 0010:ext4_file_write_iter+0xbc7
> > > > Call Trace:
> > > > vfs_write+0x3b1
> > > > ksys_write+0x77
> > > > do_syscall_64+0x39
> > > >
> > > > Fix it by putting block_write_end() before i_disksize updating just
> > > > like ext4_write_end() does.
> > > >
> > > > Fetch a reproducer in [Link].
> > > >
> > > > Link: https://bugzilla.kernel.org/show_bug.cgi?id=217209
> > > > Fixes: 64769240bd07f ("ext4: Add delayed allocation support in
> > > > data=writeback mode")
> > > > Signed-off-by: Zhihao Cheng <chengzhihao1@xxxxxxxxxx>
> > >
> > > Good catch (although practically this will hardly have any negative
> > > effect). But rather than opencoding generic_write_end() I'd do:
> > >
> > > if (unlikely(copied < len) && !PageUptodate(page))
> > > copied = 0;
> > >
> > > at the beginning of ext4_da_write_end() and that should solve these
> > > problems as well?
> > >
> >
> > Yes, your suggestion looks good, and I think we can put the checking
> > just after ext4_write_inline_data_end(Line 3150)? On the one hand, we
> > can pass original 'copied' value in trace_ext4_da_write_end(), one the
> > other hand, ext4_write_inline_data_end() already has this checking.
> > .
>
> BTW, I want send another patch as follows:
> diff --git a/fs/ext4/inode.c b/fs/ext4/inode.c
> index bf0b7dea4900..570a687ae847 100644
> --- a/fs/ext4/inode.c
> +++ b/fs/ext4/inode.c
> @@ -3149,7 +3149,7 @@ static int ext4_da_write_end(struct file *file,
> return ext4_write_inline_data_end(inode, pos, len, copied,
> page);
>
> start = pos & (PAGE_SIZE - 1);
> - end = start + copied - 1;
> + end = start + (copied ? copied - 1 : copied);
>
> /*
> * Since we are holding inode lock, we are sure i_disksize <=
> @@ -3167,7 +3167,7 @@ static int ext4_da_write_end(struct file *file,
> * ext4_da_write_inline_data_end().
> */
> new_i_size = pos + copied;
> - if (copied && new_i_size > inode->i_size &&
> + if (new_i_size > inode->i_size &&
> ext4_da_should_update_i_disksize(page, end))
> ext4_update_i_disksize(inode, new_i_size);
>
> This modification handle unconsistent i_size and i_disksize imported by
> ea51d132dbf9 ("ext4: avoid hangs in ext4_da_should_update_i_disksize()").
>
> Paritially written may display a fake inode size for user, for example:
>
>
>
> i_disksize=1
>
> generic_perform_write
>
> copied = iov_iter_copy_from_user_atomic(len) // copied = 0
>
> ext4_da_write_end // skip updating i_disksize
>
> generic_write_end
>
> if (pos + copied > inode->i_size) { // 10 + 0 > 1, true
>
> i_size_write(inode, pos + copied); // i_size = 10
>
> }
>
>
>
> 0 1 10 4096
>
> |_|_______|_________..._____|
>
> | |
>
> i_size pos
>
>
>
> Now, user see the i_size is 10 (i_disksize is still 1). After inode
>
> destroyed, user will get the i_size is 1 read from disk.
OK, but shouldn't we rather change generic_write_end() to not increase
i_size if no write happened? Because that is what seems somewhat
problematic...
Honza
--
Jan Kara <jack@xxxxxxxx>
SUSE Labs, CR
