Re: [Cluster-devel] [PATCH 03/12] filemap: update ki_pos in generic_perform_write

[Al Viro <viro@zeniv.linux.org.uk>]

On Sun, Aug 27, 2023 at 08:41:22PM +0100, Al Viro wrote:
> On Thu, Jun 01, 2023 at 04:58:55PM +0200, Christoph Hellwig wrote:
> > All callers of generic_perform_write need to updated ki_pos, move it into
> > common code.
> 
> > @@ -4034,7 +4037,6 @@ ssize_t __generic_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
> >  		endbyte = pos + status - 1;
> >  		err = filemap_write_and_wait_range(mapping, pos, endbyte);
> >  		if (err == 0) {
> > -			iocb->ki_pos = endbyte + 1;
> >  			written += status;
> >  			invalidate_mapping_pages(mapping,
> >  						 pos >> PAGE_SHIFT,
> > @@ -4047,8 +4049,6 @@ ssize_t __generic_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
> >  		}
> >  	} else {
> >  		written = generic_perform_write(iocb, from);
> > -		if (likely(written > 0))
> > -			iocb->ki_pos += written;
> >  	}
> >  out:
> >  	return written ? written : err;
> 
> [another late reply, sorry]
> 
> That part is somewhat fishy - there's a case where you return a positive value
> and advance ->ki_pos by more than that amount.  I really wonder if all callers
> of ->write_iter() are OK with that.  Consider e.g. this:
> 
> ssize_t ksys_write(unsigned int fd, const char __user *buf, size_t count)
> {
>         struct fd f = fdget_pos(fd);
>         ssize_t ret = -EBADF;
> 
>         if (f.file) {
>                 loff_t pos, *ppos = file_ppos(f.file);
>                 if (ppos) {
>                         pos = *ppos;   
>                         ppos = &pos;
>                 }
>                 ret = vfs_write(f.file, buf, count, ppos);
>                 if (ret >= 0 && ppos)
>                         f.file->f_pos = pos;
>                 fdput_pos(f);
>         }
> 
>         return ret;
> }
> 
> ssize_t vfs_write(struct file *file, const char __user *buf, size_t count, loff_t *pos)
> {
>         ssize_t ret;
> 
>         if (!(file->f_mode & FMODE_WRITE))
>                 return -EBADF;
>         if (!(file->f_mode & FMODE_CAN_WRITE))
>                 return -EINVAL;
>         if (unlikely(!access_ok(buf, count)))
>                 return -EFAULT;
> 
>         ret = rw_verify_area(WRITE, file, pos, count);
>         if (ret)
>                 return ret;
>         if (count > MAX_RW_COUNT)
>                 count =  MAX_RW_COUNT;
>         file_start_write(file);
>         if (file->f_op->write)
>                 ret = file->f_op->write(file, buf, count, pos);
>         else if (file->f_op->write_iter)
>                 ret = new_sync_write(file, buf, count, pos);
>         else   
>                 ret = -EINVAL;
>         if (ret > 0) {
>                 fsnotify_modify(file);
>                 add_wchar(current, ret);
>         }
>         inc_syscw(current);
>         file_end_write(file);
>         return ret;
> }
> 
> static ssize_t new_sync_write(struct file *filp, const char __user *buf, size_t len, loff_t *ppos)
> {
>         struct kiocb kiocb;
>         struct iov_iter iter;
>         ssize_t ret; 
> 
>         init_sync_kiocb(&kiocb, filp);
>         kiocb.ki_pos = (ppos ? *ppos : 0);
>         iov_iter_ubuf(&iter, ITER_SOURCE, (void __user *)buf, len);
> 
>         ret = call_write_iter(filp, &kiocb, &iter);
>         BUG_ON(ret == -EIOCBQUEUED);
>         if (ret > 0 && ppos)
>                 *ppos = kiocb.ki_pos;
>         return ret;
> } 
> 
> Suppose ->write_iter() ends up doing returning a positive value smaller than
> the increment of kiocb.ki_pos.  What do we get?  ret is positive, so
> kiocb.ki_pos gets copied into *ppos, which is ksys_write's pos and there
> we copy it into file->f_pos.
> 
> Is it really OK to have write() return 4096 and advance the file position
> by 16K?  AFAICS, userland wouldn't get any indication of something
> odd going on - just a short write to a regular file, with followup write
> of remaining 12K getting quietly written in the range 16K..28K.
> 
> I don't remember what POSIX says about that, but it would qualify as
> nasty surprise for any userland program - sure, one can check fsync()
> results before closing the sucker and see if everything looks fine,
> but the way it's usually discussed could easily lead to assumption that
> (synchronous) O_DIRECT writes would not be affected by anything of that
> sort.

IOW, I suspect that the right thing to do would be something along the lines
of

direct_write_fallback(): on error revert the ->ki_pos update from buffered write

If we fail filemap_write_and_wait_range() on the range the buffered write went
into, we only report the "number of bytes which we direct-written", to quote
the comment in there.  Which is fine, but buffered write has already advanced
iocb->ki_pos, so we need to roll that back.  Otherwise we end up with e.g.
write(2) advancing position by more than the amount it reports having written.

Fixes: 182c25e9c157 "filemap: update ki_pos in generic_perform_write"
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
---
diff --git a/fs/libfs.c b/fs/libfs.c
index 5b851315eeed..712c57828c0e 100644
--- a/fs/libfs.c
+++ b/fs/libfs.c
@@ -1646,6 +1646,7 @@ ssize_t direct_write_fallback(struct kiocb *iocb, struct iov_iter *iter,
 		 * We don't know how much we wrote, so just return the number of
 		 * bytes which were direct-written
 		 */
+		iocb->ki_pos -= buffered_written;
 		if (direct_written)
 			return direct_written;
 		return err;