| Message ID | 20210423171010.12-1-jack@suse.cz |
|---|---|
| Headers | show |
| Series | fs: Hole punch vs page cache filling races | expand |
On Fri, Apr 23, 2021 at 07:29:31PM +0200, Jan Kara wrote: > Currently, serializing operations such as page fault, read, or readahead > against hole punching is rather difficult. The basic race scheme is > like: > > fallocate(FALLOC_FL_PUNCH_HOLE) read / fault / .. > truncate_inode_pages_range() > <create pages in page > cache here> > <update fs block mapping and free blocks> > > Now the problem is in this way read / page fault / readahead can > instantiate pages in page cache with potentially stale data (if blocks > get quickly reused). Avoiding this race is not simple - page locks do > not work because we want to make sure there are *no* pages in given > range. One of the things I've had in mind for a while is moving the DAX locked entry concept into the page cache proper. It would avoid creating the new semaphore, at the cost of taking the i_pages lock twice (once to insert the entries that cover the range, and once to delete the entries). It'd have pretty much the same effect, though -- read/fault/... would block until the entry was deleted from the page cache.
On Fri, Apr 23, 2021 at 07:29:31PM +0200, Jan Kara wrote: > Currently, serializing operations such as page fault, read, or readahead > against hole punching is rather difficult. The basic race scheme is > like: > > fallocate(FALLOC_FL_PUNCH_HOLE) read / fault / .. > truncate_inode_pages_range() > <create pages in page > cache here> > <update fs block mapping and free blocks> > > Now the problem is in this way read / page fault / readahead can > instantiate pages in page cache with potentially stale data (if blocks > get quickly reused). Avoiding this race is not simple - page locks do > not work because we want to make sure there are *no* pages in given > range. inode->i_rwsem does not work because page fault happens under > mmap_sem which ranks below inode->i_rwsem. Also using it for reads makes > the performance for mixed read-write workloads suffer. > > So create a new rw_semaphore in the address_space - invalidate_lock - > that protects adding of pages to page cache for page faults / reads / > readahead. ..... > diff --git a/fs/inode.c b/fs/inode.c > index a047ab306f9a..43596dd8b61e 100644 > --- a/fs/inode.c > +++ b/fs/inode.c > @@ -191,6 +191,9 @@ int inode_init_always(struct super_block *sb, struct inode *inode) > mapping_set_gfp_mask(mapping, GFP_HIGHUSER_MOVABLE); > mapping->private_data = NULL; > mapping->writeback_index = 0; > + init_rwsem(&mapping->invalidate_lock); > + lockdep_set_class(&mapping->invalidate_lock, > + &sb->s_type->invalidate_lock_key); > inode->i_private = NULL; > inode->i_mapping = mapping; > INIT_HLIST_HEAD(&inode->i_dentry); /* buggered by rcu freeing */ Oh, lockdep. That might be a problem here. The XFS_MMAPLOCK has non-trivial lockdep annotations so that it is tracked as nesting properly against the IOLOCK and the ILOCK. When you end up using xfs_ilock(XFS_MMAPLOCK..) to lock this, XFS will add subclass annotations to the lock and they are going to be different to the locking that the VFS does. We'll see this from xfs_lock_two_inodes() (e.g. in xfs_swap_extents()) and xfs_ilock2_io_mmap() during reflink oper..... Oooooh. The page cache copy done when breaking a shared extent needs to lock out page faults on both the source and destination, but it still needs to be able to populate the page cache of both the source and destination file..... .... and vfs_dedupe_file_range_compare() has to be able to read pages from both the source and destination file to determine that the contents are identical and that's done while we hold the XFS_MMAPLOCK exclusively so the compare is atomic w.r.t. all other user data modification operations being run.... I now have many doubts that this "serialise page faults by locking out page cache instantiation" method actually works as a generic mechanism. It's not just page cache invalidation that relies on being able to lock out page faults: copy-on-write and deduplication both require the ability to populate the page cache with source data while page faults are locked out so the data can be compared/copied atomically with the extent level manipulations and so user data modifications cannot occur until the physical extent manipulation operation has completed. Having only just realised this is a problem, no solution has immediately popped into my mind. I'll chew on it over the weekend, but I'm not hopeful at this point... Cheers, Dave.
On Sat, Apr 24, 2021 at 08:07:51AM +1000, Dave Chinner wrote: > I've got to spend time now reconstructing the patchset into a single > series because the delivery has been spread across three different > mailing lists and so hit 3 different procmail filters. I'll comment > on the patches once I've reconstructed the series and read through > it as a whole... $ b4 mbox 20210423171010.12-1-jack@suse.cz Looking up https://lore.kernel.org/r/20210423171010.12-1-jack%40suse.cz Grabbing thread from lore.kernel.org/ceph-devel 6 messages in the thread Saved ./20210423171010.12-1-jack@suse.cz.mbx
On Sat, Apr 24, 2021 at 12:51:49AM +0100, Matthew Wilcox wrote: > On Sat, Apr 24, 2021 at 08:07:51AM +1000, Dave Chinner wrote: > > I've got to spend time now reconstructing the patchset into a single > > series because the delivery has been spread across three different > > mailing lists and so hit 3 different procmail filters. I'll comment > > on the patches once I've reconstructed the series and read through > > it as a whole... > > $ b4 mbox 20210423171010.12-1-jack@suse.cz > Looking up https://lore.kernel.org/r/20210423171010.12-1-jack%40suse.cz > Grabbing thread from lore.kernel.org/ceph-devel > 6 messages in the thread > Saved ./20210423171010.12-1-jack@suse.cz.mbx Yikes. Just send them damn mails. Or switch the lists to NNTP, but don't let the people who are reviewing your patches do stupid work with weird tools.
On Sat 24-04-21 09:04:49, Dave Chinner wrote: > On Fri, Apr 23, 2021 at 07:29:31PM +0200, Jan Kara wrote: > > Currently, serializing operations such as page fault, read, or readahead > > against hole punching is rather difficult. The basic race scheme is > > like: > > > > fallocate(FALLOC_FL_PUNCH_HOLE) read / fault / .. > > truncate_inode_pages_range() > > <create pages in page > > cache here> > > <update fs block mapping and free blocks> > > > > Now the problem is in this way read / page fault / readahead can > > instantiate pages in page cache with potentially stale data (if blocks > > get quickly reused). Avoiding this race is not simple - page locks do > > not work because we want to make sure there are *no* pages in given > > range. inode->i_rwsem does not work because page fault happens under > > mmap_sem which ranks below inode->i_rwsem. Also using it for reads makes > > the performance for mixed read-write workloads suffer. > > > > So create a new rw_semaphore in the address_space - invalidate_lock - > > that protects adding of pages to page cache for page faults / reads / > > readahead. > ..... > > diff --git a/fs/inode.c b/fs/inode.c > > index a047ab306f9a..43596dd8b61e 100644 > > --- a/fs/inode.c > > +++ b/fs/inode.c > > @@ -191,6 +191,9 @@ int inode_init_always(struct super_block *sb, struct inode *inode) > > mapping_set_gfp_mask(mapping, GFP_HIGHUSER_MOVABLE); > > mapping->private_data = NULL; > > mapping->writeback_index = 0; > > + init_rwsem(&mapping->invalidate_lock); > > + lockdep_set_class(&mapping->invalidate_lock, > > + &sb->s_type->invalidate_lock_key); > > inode->i_private = NULL; > > inode->i_mapping = mapping; > > INIT_HLIST_HEAD(&inode->i_dentry); /* buggered by rcu freeing */ > > Oh, lockdep. That might be a problem here. > > The XFS_MMAPLOCK has non-trivial lockdep annotations so that it is > tracked as nesting properly against the IOLOCK and the ILOCK. When > you end up using xfs_ilock(XFS_MMAPLOCK..) to lock this, XFS will > add subclass annotations to the lock and they are going to be > different to the locking that the VFS does. > > We'll see this from xfs_lock_two_inodes() (e.g. in > xfs_swap_extents()) and xfs_ilock2_io_mmap() during reflink > oper..... Thanks for the pointer. I was kind of wondering what lockdep nesting games XFS plays but then forgot to look into details. Anyway, I've preserved the nesting annotations in XFS and fstests run on XFS passed without lockdep complaining so there isn't at least an obvious breakage. Also as far as I'm checking the code XFS usage in and lock nesting of MMAPLOCK should be compatible with the nesting VFS enforces (also see below)... > Oooooh. The page cache copy done when breaking a shared extent needs > to lock out page faults on both the source and destination, but it > still needs to be able to populate the page cache of both the source > and destination file..... > > .... and vfs_dedupe_file_range_compare() has to be able to read > pages from both the source and destination file to determine that > the contents are identical and that's done while we hold the > XFS_MMAPLOCK exclusively so the compare is atomic w.r.t. all other > user data modification operations being run.... So I started wondering why fstests passed when reading this :) The reason is that vfs_dedupe_get_page() does not use standard page cache filling path (neither readahead API nor filemap_read()), instead it uses read_mapping_page() and so gets into page cache filling path below the level at which we get invalidate_lock and thus everything works as it should. So read_mapping_page() is similar to places like e.g. block_truncate_page() or block_write_begin() which may end up filling in page cache contents but they rely on upper layers to already hold appropriate locks. I'll add a comment to read_mapping_page() about this. Once all filesystems are converted to use invalidate_lock, I also want to add WARN_ON_ONCE() to various places verifying that invalidate_lock is held as it should... > I now have many doubts that this "serialise page faults by locking > out page cache instantiation" method actually works as a generic > mechanism. It's not just page cache invalidation that relies on > being able to lock out page faults: copy-on-write and deduplication > both require the ability to populate the page cache with source data > while page faults are locked out so the data can be compared/copied > atomically with the extent level manipulations and so user data > modifications cannot occur until the physical extent manipulation > operation has completed. Hum, that is a good point. So there are actually two different things you want to block at different places: 1) You really want to block page cache instantiation for operations such as hole punch as that operation mutates data and thus contents would become stale. 2) You want to block page cache *modification* for operations such as dedupe while keeping page cache in place. This is somewhat different requirement but invalidate_lock can, in principle, cover it as well. Basically we just need to keep invalidate_lock usage in .page_mkwrite helpers. The question remains whether invalidate_lock is still a good name with this usage in mind and I probably need to update a documentation to reflect this usage. Honza -- Jan Kara <jack@suse.com> SUSE Labs, CR