On Mon, Aug 11, 2014 at 04:35:00PM +0200, Jan Kara wrote:

On Mon 11-08-14 10:13:08, Matthew Wilcox wrote:

quoted

On Mon, Aug 11, 2014 at 10:51:47AM +0200, Jan Kara wrote:

quoted

So I'm afraid we'll have to find some other way to synchronize
page faults and truncate / punch hole in DAX.

What if we don't?  If we hit the race (which is vanishingly unlikely with
real applications), the consequence is simply that after a truncate, a
file may be left with one or two blocks allocated somewhere after i_size.
As I understand it, that's not a real problem; they're temporarily
unavailable for allocation but will be freed on file removal or the next
truncation of that file.

  You mean if you won't have any locking between page fault and truncate?
You can have:
a) extending truncate making forgotten blocks with non-zeros visible
b) filesystem corruption due to doubly used blocks (block will be freed
from the truncated file and thus can be reallocated but it will still be
accessible via mmap from the truncated file).

  So not a good idea.

Not *no* locking ... just no locking around get_block, like in v7.
So check i_size, call get_block, lock i_mmap_mutex, re-check i_size,
insert mapping if i_size is OK, drop i_mmap_mutex.  As long as get_block()
has enough locking of its own against set_size and concurrent calls
to get_block(), I don't think we can get visible non-zeroes or double
allocation.

quoted

I'm also still considering the possibility of having truncate-down block
until all mmaps that extend after the new i_size have been removed ...

  Hum, I'm not sure how you would do that with current locking scheme and
wait for all page faults on that range to finish but maybe you have some
good idea :)

While it can be blocked with i_dio_count currently, this would be a more
complicated thing to do ...

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>