On Wed, Sep 16, 2020 at 05:09:49PM -0400, Qian Cai wrote: > On Sat, 2020-09-12 at 00:55 +0100, Al Viro wrote: > > On Fri, Sep 11, 2020 at 05:59:04PM -0400, Qian Cai wrote: > > > Super easy to reproduce on today's mainline by just fuzzing for a few > > > minutes > > > on virtiofs (if it ever matters). Any thoughts? > > > > Usually happens when ->direct_IO() fucks up and reports the wrong amount > > of data written/read. We had several bugs like that in the past - see > > e.g. 85128b2be673 (fix nfs O_DIRECT advancing iov_iter too much). > > > > Had there been any recent O_DIRECT-related patches on the filesystems > > involved? > > This is only reproducible using FUSE/virtiofs so far, so I will stare at > fuse_direct_IO() until someone can beat me to it. What happens there is that it tries to play with iov_iter_truncate() in ->direct_IO() without a corresponding iov_iter_reexpand(). Could you check if the following helps? Signed-off-by: Al Viro <viro@xxxxxxxxxxxxxxxxxx> --- diff --git a/fs/fuse/file.c b/fs/fuse/file.c index 6611ef3269a8..d3eec2e11975 100644 --- a/fs/fuse/file.c +++ b/fs/fuse/file.c @@ -3095,7 +3095,7 @@ fuse_direct_IO(struct kiocb *iocb, struct iov_iter *iter) loff_t pos = 0; struct inode *inode; loff_t i_size; - size_t count = iov_iter_count(iter); + size_t count = iov_iter_count(iter), shortened; loff_t offset = iocb->ki_pos; struct fuse_io_priv *io; @@ -3111,7 +3111,8 @@ fuse_direct_IO(struct kiocb *iocb, struct iov_iter *iter) if (offset >= i_size) return 0; iov_iter_truncate(iter, fuse_round_up(ff->fc, i_size - offset)); - count = iov_iter_count(iter); + shortened = count - iov_iter_count(iter); + count -= shortened; } io = kmalloc(sizeof(struct fuse_io_priv), GFP_KERNEL); @@ -3177,6 +3178,8 @@ fuse_direct_IO(struct kiocb *iocb, struct iov_iter *iter) else if (ret < 0 && offset + count > i_size) fuse_do_truncate(file); } + if (shortened) + iov_iter_reexpand(iter, shortened); return ret; }
