2012/6/5 Dave Chinner <david@xxxxxxxxxxxxx>:
> On Thu, May 24, 2012 at 02:35:38PM +0200, Jan Kara wrote:
>> On Sat 19-05-12 11:40:24, Dave Chinner wrote:
>> > So let's step back a moment and have a look at how we've got here.
>> > The problem is that we've optimised ourselves into a corner with the
>> > way we handle page cache truncation - we don't need mmap
>> > serialisation because of the combination of i_size and page locks
>> > mean we can detect truncated pages safely at page fault time. With
>> > hole punching, we don't have that i_size safety blanket, and so we
>> > need some other serialisation mechanism to safely detect whether a
>> > page is valid or not at any given point in time.
>> > Because it needs to serialise against IO operations, we need a
>> > sleeping lock of some kind, and it can't be the existing IO lock.
>> > And now we are looking at needing a new lock for hole punching, I'm
>> > really wondering if the i_size/page lock truncation optimisation
>> > should even continue to exist. i.e. replace it with a single
>> > mechanism that works for both hole punching, truncation and other
>> > functions that require exclusive access or exclusion against
>> > modifications to the mapping tree.
>> > But this is only one of the problems in this area.The way I see it
>> > is that we have many kludges in the area of page invalidation w.r.t.
>> > different types of IO, the page cache and mmap, especially when we
>> > take into account direct IO. What we are seeing here is we need
>> > some level of _mapping tree exclusion_ between:
>> > 1. mmap vs hole punch (broken)
>> > 2. mmap vs truncate (i_size/page lock)
>> > 3. mmap vs direct IO (non-existent)
>> > 4. mmap vs buffered IO (page lock)
>> > 5. writeback vs truncate (i_size/page lock)
>> > 6. writeback vs hole punch (page lock, possibly broken)
>> > 7. direct IO vs buffered IO (racy - flush cache before/after DIO)
>> Yes, this is a nice summary of the most interesting cases. For
>> here are the remaining cases:
>> 8. mmap vs writeback (page lock)
>> 9. writeback vs direct IO (as direct IO vs buffered IO)
>> 10. writeback vs buffered IO (page lock)
>> 11. direct IO vs truncate (dio_wait)
>> 12. direct IO vs hole punch (dio_wait)
>> 13. buffered IO vs truncate (i_mutex for writes, i_size/page lock for reads)
>> 14. buffered IO vs hole punch (fs dependent, broken for ext4)
>> 15. truncate vs hole punch (fs dependent)
>> 16. mmap vs mmap (page lock)
>> 17. writeback vs writeback (page lock)
>> 18. direct IO vs direct IO (i_mutex or fs dependent)
>> 19. buffered IO vs buffered IO (i_mutex for writes, page lock for reads)
>> 20. truncate vs truncate (i_mutex)
>> 21. punch hole vs punch hole (fs dependent)
I think we have even the xip cases here.