Re: [RFC 4/9] gfs2: Fix mmap + page fault deadlocks (part 1)

[Al Viro]

On Wed, Jun 02, 2021 at 01:16:32PM +0200, Andreas Gruenbacher wrote:

> Well, iomap_file_buffered_write() does that by using
> iov_iter_fault_in_readable() and iov_iter_copy_from_user_atomic() as
> in iomap_write_actor(), but the read and direct I/O side doesn't seem
> to have equivalents. I suspect we can't just wrap
> generic_file_read_iter() and iomap_dio_rw() calls in
> pagefault_disable().

And it will have zero effect on O_DIRECT case, so you get the same
deadlocks right back.  Because there you hit
	iomap_dio_bio_actor()
		bio_iov_iter_get_pages()
			....
				get_user_pages_fast()
					....
						faultin_page()
							handle_mm_fault()
and at no point had CPU hit an exception, so disable_pagefault() will
have no effect whatsoever.  You can bloody well hit gfs2 readpage/mkwrite
if the destination is in mmapped area of some GFS2 file.  Do that
while holding GFS2 locks and you are fucked.

No amount of prefaulting will protect you, BTW - it might make the
deadlock harder to reproduce, but that's it.