Hugh pointed me to ext4 hole punching code which is clearly missing some
locking. But looking at the code more deeply I realized I don't see
anything preventing the following race in XFS or ext4:
punch_hole(file, 0, 4096)
addr = mmap(file);
addr = 1
^^ writeably fault a page
remove file blocks
write out file
^^ interesting things can
happen because we expect blocks under the first page to be allocated /
reserved but they are not...
I'm pretty sure ext4 has this problem, I'm not completely sure whether
XFS has something to protect against such race but I don't see anything.
It's not easy to protect against these races. For truncate, i_size protects
us against similar races but for hole punching we don't have any such
mechanism. One way to avoid the race would be to hold mmap_sem while we are
invalidating the page cache and punching hole but that sounds a bit ugly.
Alternatively we could just have some special lock (rwsem?) held during
page_mkwrite() (for reading) and during whole hole punching (for writing)
to serialize these two operations.
Another alternative, which doesn't really look more appealing, is to go
page-by-page and always free corresponding blocks under page lock.
Any other ideas or thoughts?
Jan Kara <jack@xxxxxxx>
SUSE Labs, CR