Re: [RFC v2 0/9] Replacing the readpages a_op

[Dave Chinner <david@fromorbit.com>]

On Thu, Jan 23, 2020 at 11:21:01AM +0100, Jan Kara wrote:
> On Thu 23-01-20 10:47:40, Dave Chinner wrote:
> > On Tue, Jan 21, 2020 at 01:48:45PM -0800, Matthew Wilcox wrote:
> > > On Tue, Jan 21, 2020 at 12:36:27PM +0100, Jan Kara wrote:
> > > > > v2: Chris asked me to show what this would look like if we just have
> > > > > the implementation look up the pages in the page cache, and I managed
> > > > > to figure out some things I'd done wrong last time.  It's even simpler
> > > > > than v1 (net 104 lines deleted).
> > > > 
> > > > I have an unfinished patch series laying around that pulls the ->readpage
> > > > / ->readpages API in somewhat different direction so I'd like to discuss
> > > > whether it's possible to solve my problem using your API. The problem I
> > > > have is that currently some operations such as hole punching can race with
> > > > ->readpage / ->readpages like:
> > > > 
> > > > CPU0						CPU1
> > > > fallocate(fd, FALLOC_FL_PUNCH_HOLE, off, len)
> > > >   filemap_write_and_wait_range()
> > > >   down_write(inode->i_rwsem);
> > > >   truncate_pagecache_range();
> > 
> > shouldn't fallocate be holding EXT4_I(inode)->i_mmap_sem before it
> > truncates the page cache? Otherwise it's not serialised against
> > page faults. Looks at code ... oh, it does hold the i_mmap_sem in
> > write mode, so....
> 
> Yes.
> 
> > > > 						readahead(fd, off, len)
> > > > 						  creates pages in page cache
> > > > 						  looks up block mapping
> > > >   removes blocks from inode and frees them
> > > > 						  issues bio
> > > > 						    - reads stale data -
> > > > 						      potential security
> > > > 						      issue
> > 
> > .... I'm not sure that this race condition should exist anymore
> > as readahead should not run until the filesystem drops it's inode
> > and mmap locks after the entire extent freeing operation is
> > complete...
> 
> Not for XFS but for all the other filesystems see below..
> 
> > > > Now how I wanted to address this is that I'd change the API convention for
> > > > ->readpage() so that we call it with the page unlocked and the function
> > > > would lock the page, check it's still OK, and do what it needs. And this
> > > > will allow ->readpage() and also ->readpages() to grab lock
> > > > (EXT4_I(inode)->i_mmap_sem in case of ext4) to synchronize with hole punching
> > > > while we are adding pages to page cache and mapping underlying blocks.
> > > > 
> > > > Now your API makes even ->readpages() (actually ->readahead) called with
> > > > pages locked so that makes this approach problematic because of lock
> > > > inversions. So I'd prefer if we could keep the situation that ->readpages /
> > > > ->readahead gets called without any pages in page cache locked...
> > > 
> > > I'm not a huge fan of that approach because it increases the number of
> > > atomic ops (right now, we __SetPageLocked on the page before adding it
> > > to i_pages).  Holepunch is a rather rare operation while readpage and
> > > readpages/readahead are extremely common, so can we make holepunch take
> > > a lock that will prevent new readpage(s) succeeding?
> > > 
> > > I have an idea to move the lock entries from DAX to being a generic page
> > > cache concept.  That way, holepunch could insert lock entries into the
> > > pagecache to cover the range being punched, and readpage(s) would either
> > > skip lock entries or block on them.
> > > 
> > > Maybe there's a better approach though.
> > 
> > Can we step back for a moment and look at how we already serialise
> > readahead against truncate/hole punch? While the readahead code
> > itself doesn't serialise against truncate, in all cases we should be
> > running through the filesystem at a higher layer and provides the
> > truncate/holepunch serialisation before we get to the readahead
> > code.
> > 
> > The read() syscall IO path:
> > 
> >   read()
> >     ->read_iter()
> >       filesystem takes truncate serialisation lock
> >       generic_file_read_iter()
> >         generic_file_buffered_read()
> > 	  page_cache_sync_readahead()
> > 	    ....
> > 	  page_cache_async_readahead()
> > 	    ....
> >       .....
> >       filesystem drops truncate serialisation lock
> 
> Yes, this is the scheme XFS uses. But ext4 and other filesystems use a
> scheme where read is serialized against truncate only by page locks and
> i_size checks.

Right, which I've characterised in the past as a "performance over
correctness" hack that has resulted in infrastructure that can only
protect against truncate and not general page invalidation
operations. It's also been the source of many, many truncate bugs
over the past 20 years and has never worked for hole punch, yet....

> Which works for truncate but is not enough for hole
> punching. And locking read(2) and readahead(2) in all these filesystem with
> i_rwsem is going to cause heavy regressions with mixed read-write workloads
> and unnecessarily so because we don't need to lock reads against writes,
> just against truncate or hole punching.

.... performance over correctness is still used as the justification
for keeping this ancient, troublesome architecture in place even
though it cannot support the requirements of modern filesystems.

Indeed, this problem had to be fixed with DAX, and the ext4 DAX read
path uses shared inode locks. And now the ext4 DIO path uses shared
inode locking, too. Only the ext4 buffered read path does not use
shared read locks to avoid this race condition.

I don't have a "maintain the existing mixed rw performance" solution
for you here, except to say that I'm working (slowly) towards fixing
this problem on XFS with range locking for IO instead of using the
i_rwsem. That fixes so many other concurrency issues, too, such as
allowing fallocate() and ftruncate() to run concurrently with
non-overlapping IO to the same file. IOWs, moving IO locking away
from the page cache will allow much greater concurrency and
performance for a much wider range of filesystem operations than
trying to screw around with individual page locks.

> So I wanted to use i_mmap_sem for the serialization of the read path against
> truncate. But due to lock ordering with mmap_sem and because reads do take
> page faults to copy data it is not straightforward - hence my messing with
> ->readpage().

Another manifestation of the age old "can't take the same lock in
the syscall IO path and the page fault IO path" problem. We've
previously looked hacking about readpages in XFS, too, but it was
too complex, didn't improve performance, and in general such changes
were heading in the opposite direction we have been moving the IO
path infrastructure towards with iomap.

i.e. improving buffered IO efficiency requires use to reduce the
amount of work we do per page, not increase it....

> Now that I'm thinking about it, there's also a possibility of
> introducing yet another rwsem into the inode that would rank above
> mmap_sem and be used to serialize ->read_iter and ->fadvise against
> truncate. But having three rwsems in the inode for serialization seems a
> bit too convoluted for my taste.

Yup, that's called "falling off the locking cliff". :/

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com