There is no reason anymore for not issuing device integrity
operations when teh filesystem requires ordering or data integrity
guarantees. We should always issue cache flushes and FUA writes
where necessary and let the underlying storage optimise them as
necessary for correct integrity operation.
Signed-Off-By: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NDave Chinner <david@fromorbit.com>

When we create a new attribute, we first create a shortform
attribute, and try to fit the new attribute into it.
If that fails, we copy the (empty) attribute into a leaf attribute,
and do the copy again.  Thus there can be a transient state where
we have an empty leaf attribute.

If we encounter this during log replay, the verifier will fail.
So add a test to ignore this part of the leaf attr verification
during log replay.

Thanks as usual to dchinner for spotting the problem.
Signed-off-by: NEric Sandeen <sandeen@redhat.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Straight switch over to using iomap for direct I/O - we already have the
non-COW dio path in write_begin for DAX and files with extent size hints,
so nothing to add there.  The COW path is ported over from the old
get_blocks version and a bit of a mess, but I have some work in progress
to make it look more like the buffered I/O COW path.

This gets rid of xfs_get_blocks_direct and the last caller of
xfs_get_blocks with the create flag set, so all that code can be removed.

Last but not least I've removed a comment in xfs_filemap_fault that
refers to xfs_get_blocks entirely instead of updating it - while the
reference is correct, the whole DAX fault path looks different than
the non-DAX one, so it seems rather pointless.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Tested-by: NJens Axboe <axboe@fb.com>
Reviewed-by: NDarrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

This adds a full fledget direct I/O implementation using the iomap
interface. Full fledged in this case means all features are supported:
AIO, vectored I/O, any iov_iter type including kernel pointers, bvecs
and pipes, support for hole filling and async apending writes.  It does
not mean supporting all the warts of the old generic code.  We expect
i_rwsem to be held over the duration of the call, and we expect to
maintain i_dio_count ourselves, and we pass on any kinds of mapping
to the file system for now.

The algorithm used is very simple: We use iomap_apply to iterate over
the range of the I/O, and then we use the new bio_iov_iter_get_pages
helper to lock down the user range for the size of the extent.
bio_iov_iter_get_pages can currently lock down twice as many pages as
the old direct I/O code did, which means that we will have a better
batch factor for everything but overwrites of badly fragmented files.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NKent Overstreet <kent.overstreet@gmail.com>
Tested-by: NJens Axboe <axboe@fb.com>
Reviewed-by: NDarrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

We want to use the per-sb completion workqueue from the new iomap
direct I/O code.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Tested-by: NJens Axboe <axboe@fb.com>
Reviewed-by: NDarrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

This patch drops the XFS-own i_iolock and uses the VFS i_rwsem which
recently replaced i_mutex instead.  This means we only have to take
one lock instead of two in many fast path operations, and we can
also shrink the xfs_inode structure.  Thanks to the xfs_ilock family
there is very little churn, the only thing of note is that we need
to switch to use the lock_two_directory helper for taking the i_rwsem
on two inodes in a few places to make sure our lock order matches
the one used in the VFS.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Tested-by: NJens Axboe <axboe@fb.com>
Reviewed-by: NDarrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Christoph requested lockdep_assert_held() variants that distinguish
between held-for-read or held-for-write.

Provide:

  int lock_is_held_type(struct lockdep_map *lock, int read)

which takes the same argument as lock_acquire(.read) and matches it to
the held_lock instance.

Use of this function should be gated by the debug_locks variable. When
that is 0 the return value of the lock_is_held_type() function is
undefined. This is done to allow both negative and positive tests for
holding locks.

By default we provide (positive) lockdep_assert_held{,_exclusive,_read}()
macros.
Requested-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NPeter Zijlstra (Intel) <peterz@infradead.org>
Tested-by: NJens Axboe <axboe@fb.com>
Reviewed-by: NDarrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

This is a helper that pins down a range from an iov_iter and adds it to
a bio without requiring a separate memory allocation for the page array.
It will be used for upcoming direct I/O implementations for block devices
and iomap based file systems.
Signed-off-by: NKent Overstreet <kent.overstreet@gmail.com>
[hch: ported to the iov_iter interface, renamed and added comments.
      All blame should be directed to me and all fame should go to Kent
      after this!]
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NJens Axboe <axboe@fb.com>

(cherry picked from commit 9cd56d916aa481ce8f56d9c5302a6ed90c2e0b5f)

xfs_file_iomap_begin_delay() implements post-eof speculative
preallocation by extending the block count of the requested delayed
allocation. Now that xfs_bmapi_reserve_delalloc() has been updated to
handle prealloc blocks separately and tag the inode, update
xfs_file_iomap_begin_delay() to use the new parameter and rely on the
former to tag the inode.

Note that this patch does not change behavior.
Signed-off-by: NBrian Foster <bfoster@redhat.com>
Reviewed-by: NDave Chinner <dchinner@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

COW fork reservation is implemented via delayed allocation. The code is
modeled after the traditional delalloc allocation code, but is slightly
different in terms of how preallocation occurs. Rather than post-eof
speculative preallocation, COW fork preallocation is implemented via a
COW extent size hint that is designed to minimize fragmentation as a
reflinked file is split over time.

xfs_reflink_reserve_cow() still uses logic that is oriented towards
dealing with post-eof speculative preallocation, however, and is stale
or not necessarily correct. First, the EOF alignment to the COW extent
size hint is implemented in xfs_bmapi_reserve_delalloc() (which does so
correctly by aligning the start and end offsets) and so is not necessary
in xfs_reflink_reserve_cow(). The backoff and retry logic on ENOSPC is
also ineffective for the same reason, as xfs_bmapi_reserve_delalloc()
will simply perform the same allocation request on the retry. Finally,
since the COW extent size hint aligns the start and end offset of the
range to allocate, the end_fsb != orig_end_fsb logic is not sufficient.
Indeed, if a write request happens to end on an aligned offset, it is
possible that we do not tag the inode for COW preallocation even though
xfs_bmapi_reserve_delalloc() may have preallocated at the start offset.

Kill the unnecessary, duplicate code in xfs_reflink_reserve_cow().
Remove the inode tag logic as well since xfs_bmapi_reserve_delalloc()
has been updated to tag the inode correctly.
Signed-off-by: NBrian Foster <bfoster@redhat.com>
Reviewed-by: NDave Chinner <dchinner@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Speculative preallocation is currently processed entirely by the callers
of xfs_bmapi_reserve_delalloc(). The caller determines how much
preallocation to include, adjusts the extent length and passes down the
resulting request.

While this works fine for post-eof speculative preallocation, it is not
as reliable for COW fork preallocation. COW fork preallocation is
implemented via the cowextszhint, which aligns the start offset as well
as the length of the extent. Further, it is difficult for the caller to
accurately identify when preallocation occurs because the returned
extent could have been merged with neighboring extents in the fork.

To simplify this situation and facilitate further COW fork preallocation
enhancements, update xfs_bmapi_reserve_delalloc() to take a separate
preallocation parameter to incorporate into the allocation request. The
preallocation blocks value is tacked onto the end of the request and
adjusted to accommodate neighboring extents and extent size limits.
Since xfs_bmapi_reserve_delalloc() now knows precisely how much
preallocation was included in the allocation, it can also tag the inodes
appropriately to support preallocation reclaim.

Note that xfs_bmapi_reserve_delalloc() callers are not yet updated to
use the preallocation mechanism. This patch should not change behavior
outside of correctly tagging reflink inodes when start offset
preallocation occurs (which the caller does not handle correctly).
Signed-off-by: NBrian Foster <bfoster@redhat.com>
Reviewed-by: NDave Chinner <dchinner@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

It turns out that btrfs and xfs had differing interpretations of what
to do when the dedupe length is zero.  Change xfs to follow btrfs'
semantics so that the userland interface is consistent.
Signed-off-by: NDarrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: NDave Chinner <dchinner@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Declare the structure xfs_nameops as const as it is only stored in the
m_dirnameops field of a xfs_mount structure. This field is of type
const struct xfs_nameops *, so xfs_nameops structures having this
property can be declared as const.
Done using Coccinelle:
@r1 disable optional_qualifier @
identifier i;
position p;
@@
static struct xfs_nameops i@p = {...};

@ok1@
identifier r1.i;
position p;
struct xfs_mount mp;
@@
mp.m_dirnameops=&i@p

@bad@
position p!={r1.p,ok1.p};
identifier r1.i;
@@
i@p

@depends on !bad disable optional_qualifier@
identifier r1.i;
@@
static
+const
struct xfs_nameops i={...};

@depends on !bad disable optional_qualifier@
identifier r1.i;
@@
+const
struct xfs_nameops i;

File size before:
   text	   data	    bss	    dec	    hex	filename
   5302	     85	      0	   5387	   150b	fs/xfs/libxfs/xfs_dir2.o

File size after:
   text	   data	    bss	    dec	    hex	filename
   5318	     69	      0	   5387	   150b	fs/xfs/libxfs/xfs_dir2.o
Signed-off-by: NBhumika Goyal <bhumirks@gmail.com>
Reviewed-by: NDave Chinner <dchinner@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Declare the structure xfs_item_ops as const as it is only passed as an
argument to the function xfs_log_item_init. As this argument is of type
const struct xfs_item_ops *, so xfs_item_ops structures having this
property can be declared as const.
Done using Coccinelle:

@r1 disable optional_qualifier @
identifier i;
position p;
@@
static struct xfs_item_ops i@p = {...};

@ok1@
identifier r1.i;
position p;
expression e1,e2,e3;
@@
xfs_log_item_init(e1,e2,e3,&i@p)

@bad@
position p!={r1.p,ok1.p};
identifier r1.i;
@@
i@p

@depends on !bad disable optional_qualifier@
identifier r1.i;
@@
static
+const
struct xfs_item_ops i={...};

@depends on !bad disable optional_qualifier@
identifier r1.i;
@@
+const
struct xfs_item_ops i;

File size before:
   text	   data	    bss	    dec	    hex	filename
    737	     64	      8	    809	    329	fs/xfs/xfs_icreate_item.o

File size after:
   text	   data	    bss	    dec	    hex	filename
    801	      0	      8	    809	    329	fs/xfs/xfs_icreate_item.o
Signed-off-by: NBhumika Goyal <bhumirks@gmail.com>
Reviewed-by: NEric Sandeen <sandeen@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

When we're estimating the amount of space it's going to take to satisfy
a delalloc reservation, we need to include the space that we might need
to grow the rmapbt.  This helps us to avoid running out of space later
when _iomap_write_allocate needs more space than we reserved.  Eryu Guan
observed this happening on generic/224 when sunit/swidth were set.
Reported-by: NEryu Guan <eguan@redhat.com>
Signed-off-by: NDarrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: NDave Chinner <dchinner@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

When XBF_NO_IOACCT got added, it missed the translation
in XFS_BUF_FLAGS, so we see "0x8" in trace output rather
than the flag name.  Fix it.
Signed-off-by: NEric Sandeen <sandeen@redhat.com>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

We only ever set a field to this constant for an impossible to reach
error case in xfs_bmap_search_extents.  That functions has been removed,
so we can remove the constant as well.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Now that all users are gone.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

And remove the unused return value.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Use xfs_iext_lookup_extent to look up the extent, drop a useless check,
drop a unneeded return value and clean up the general style a little bit.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

And only lookup the previous extent inside xfs_iomap_prealloc_size
if we actually need it.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

We can easily lookup the previous extent for the cases where we need it,
which saves the callers from looking it up for us later in the series.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Rewrite the function using xfs_iext_lookup_extent and xfs_iext_get_extent,
and massage the flow into something easily understandable.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

xfs_iext_lookup_extent looks up a single extent at the passed in offset,
and returns the extent covering the area, or the one behind it in case
of a hole, as well as the index of the returned extent in arguments,
as well as a simple bool as return value that is set to false if no
extent could be found because the offset is behind EOF.  It is a simpler
replacement for xfs_bmap_search_extent that leaves looking up the rarely
needed previous extent to the caller and has a nicer calling convention.

xfs_iext_get_extent is a helper for iterating over the extent list,
it takes an extent index as input, and returns the extent at that index
in it's expanded form in an argument if it exists.  The actual return
value is a bool whether the index is valid or not.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Introduce a flag telling iomap operations whether they are handling a
fault or other IO. That may influence behavior wrt inode size and
similar things.
Signed-off-by: NJan Kara <jack@suse.cz>
Reviewed-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Filesystem shutdown testing on an older distro kernel has uncovered an
imbalanced locking pattern for the inode flush lock in
xfs_reclaim_inode(). Specifically, there is a double unlock sequence
between the call to xfs_iflush_abort() and xfs_reclaim_inode() at the
"reclaim:" label.

This actually does not cause obvious problems on current kernels due to
the current flush lock implementation. Older kernels use a counting
based flush lock mechanism, however, which effectively breaks the lock
indefinitely when an already unlocked flush lock is repeatedly unlocked.
Though this only currently occurs on filesystem shutdown, it has
reproduced the effect of elevating an fs shutdown to a system-wide crash
or hang.

As it turns out, the flush lock is not actually required for the reclaim
logic in xfs_reclaim_inode() because by that time we have already cycled
the flush lock once while holding ILOCK_EXCL. Therefore, remove the
additional flush lock/unlock cycle around the 'reclaim:' label and
update branches into this label to release the flush lock where
appropriate. Add an assert to xfs_ifunlock() to help prevent future
occurences of the same problem.
Reported-by: NZorro Lang <zlang@redhat.com>
Signed-off-by: NBrian Foster <bfoster@redhat.com>
Reviewed-by: NDave Chinner <dchinner@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

The open-coded pattern:

ifp->if_bytes / (uint)sizeof(xfs_bmbt_rec_t)

is all over the xfs code; provide a new helper
xfs_iext_count(ifp) to count the number of inline extents
in an inode fork.

[dchinner: pick up several missed conversions]
Signed-off-by: NEric Sandeen <sandeen@redhat.com>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

There have been several reports over the years of NULL pointer
dereferences in xfs_trans_log_inode during xfs_fsr processes,
when the process is doing an fput and tearing down extents
on the temporary inode, something like:

BUG: unable to handle kernel NULL pointer dereference at 0000000000000018
PID: 29439  TASK: ffff880550584fa0  CPU: 6   COMMAND: "xfs_fsr"
    [exception RIP: xfs_trans_log_inode+0x10]
 #9 [ffff8800a57bbbe0] xfs_bunmapi at ffffffffa037398e [xfs]
#10 [ffff8800a57bbce8] xfs_itruncate_extents at ffffffffa0391b29 [xfs]
#11 [ffff8800a57bbd88] xfs_inactive_truncate at ffffffffa0391d0c [xfs]
#12 [ffff8800a57bbdb8] xfs_inactive at ffffffffa0392508 [xfs]
#13 [ffff8800a57bbdd8] xfs_fs_evict_inode at ffffffffa035907e [xfs]
#14 [ffff8800a57bbe00] evict at ffffffff811e1b67
#15 [ffff8800a57bbe28] iput at ffffffff811e23a5
#16 [ffff8800a57bbe58] dentry_kill at ffffffff811dcfc8
#17 [ffff8800a57bbe88] dput at ffffffff811dd06c
#18 [ffff8800a57bbea8] __fput at ffffffff811c823b
#19 [ffff8800a57bbef0] ____fput at ffffffff811c846e
#20 [ffff8800a57bbf00] task_work_run at ffffffff81093b27
#21 [ffff8800a57bbf30] do_notify_resume at ffffffff81013b0c
#22 [ffff8800a57bbf50] int_signal at ffffffff8161405d

As it turns out, this is because the i_itemp pointer, along
with the d_ops pointer, has been overwritten with zeros
when we tear down the extents during truncate.  When the in-core
inode fork on the temporary inode used by xfs_fsr was originally
set up during the extent swap, we mistakenly looked at di_nextents
to determine whether all extents fit inline, but this misses extents
generated by speculative preallocation; we should be using if_bytes
instead.

This mistake corrupts the in-memory inode, and code in
xfs_iext_remove_inline eventually gets bad inputs, causing
it to memmove and memset incorrect ranges; this became apparent
because the two values in ifp->if_u2.if_inline_ext[1] contained
what should have been in d_ops and i_itemp; they were memmoved due
to incorrect array indexing and then the original locations
were zeroed with memset, again due to an array overrun.

Fix this by properly using i_df.if_bytes to determine the number
of extents, not di_nextents.

Thanks to dchinner for looking at this with me and spotting the
root cause.

Cc: stable@vger.kernel.org
Signed-off-by: NEric Sandeen <sandeen@redhat.com>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

We've had reports of generic/095 causing XFS to BUG() in
__xfs_get_blocks() due to the existence of delalloc blocks on a
direct I/O read. generic/095 issues a mix of various types of I/O,
including direct and memory mapped I/O to a single file. This is
clearly not supported behavior and is known to lead to such
problems. E.g., the lack of exclusion between the direct I/O and
write fault paths means that a write fault can allocate delalloc
blocks in a region of a file that was previously a hole after the
direct read has attempted to flush/inval the file range, but before
it actually reads the block mapping. In turn, the direct read
discovers a delalloc extent and cannot proceed.

While the appropriate solution here is to not mix direct and memory
mapped I/O to the same regions of the same file, the current
BUG_ON() behavior is probably overkill as it can crash the entire
system.  Instead, localize the failure to the I/O in question by
returning an error for a direct I/O that cannot be handled safely
due to delalloc blocks. Be careful to allow the case of a direct
write to post-eof delalloc blocks. This can occur due to speculative
preallocation and is safe as post-eof blocks are not accompanied by
dirty pages in pagecache (conversely, preallocation within eof must
have been zeroed, and thus dirtied, before the inode size could have
been increased beyond said blocks).

Finally, provide an additional warning if a direct I/O write occurs
while the file is memory mapped. This may not catch all problematic
scenarios, but provides a hint that some known-to-be-problematic I/O
methods are in use.
Signed-off-by: NBrian Foster <bfoster@redhat.com>
Reviewed-by: NDave Chinner <dchinner@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

The cowblocks background scanner currently clears the cowblocks tag
for inodes without any real allocations in the cow fork. This
excludes inodes with only delalloc blocks in the cow fork. While we
might never expect to clear delalloc blocks from the cow fork in the
background scanner, it is not necessarily correct to clear the
cowblocks tag from such inodes.

For example, if the background scanner happens to process an inode
between a buffered write and writeback, the scanner catches the
inode in a state after delalloc blocks have been allocated to the
cow fork but before the delalloc blocks have been converted to real
blocks by writeback. The background scanner then incorrectly clears
the cowblocks tag, even if part of the aforementioned delalloc
reservation will not be remapped to the data fork (i.e., extra
blocks due to the cowextsize hint). This means that any such
additional blocks in the cow fork might never be reclaimed by the
background scanner and could persist until the inode itself is
reclaimed.

To address this problem, only skip and clear inodes without any cow
fork allocations whatsoever from the background scanner. While we
generally do not want to cancel delalloc reservations from the
background scanner, the pagecache dirty check following the
cowblocks check should prevent that situation. If we do end up with
delalloc cow fork blocks without a dirty address space mapping, this
is probably an indication that something has gone wrong and the
blocks should be reclaimed, as they may never be converted to a real
allocation.
Signed-off-by: NBrian Foster <bfoster@redhat.com>
Reviewed-by: NDarrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Now that DAX PMD faults are once again working and are now participating in
DAX's radix tree locking scheme, allow their config option to be enabled.
Signed-off-by: NRoss Zwisler <ross.zwisler@linux.intel.com>
Reviewed-by: NJan Kara <jack@suse.cz>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Switch xfs_filemap_pmd_fault() from using dax_pmd_fault() to the new and
improved dax_iomap_pmd_fault().  Also, now that it has no more users,
remove xfs_get_blocks_dax_fault().
Signed-off-by: NRoss Zwisler <ross.zwisler@linux.intel.com>
Reviewed-by: NJan Kara <jack@suse.cz>
Signed-off-by: NDave Chinner <david@fromorbit.com>

DAX PMDs have been disabled since Jan Kara introduced DAX radix tree based
locking.  This patch allows DAX PMDs to participate in the DAX radix tree
based locking scheme so that they can be re-enabled using the new struct
iomap based fault handlers.

There are currently three types of DAX 4k entries: 4k zero pages, 4k DAX
mappings that have an associated block allocation, and 4k DAX empty
entries.  The empty entries exist to provide locking for the duration of a
given page fault.

This patch adds three equivalent 2MiB DAX entries: Huge Zero Page (HZP)
entries, PMD DAX entries that have associated block allocations, and 2 MiB
DAX empty entries.

Unlike the 4k case where we insert a struct page* into the radix tree for
4k zero pages, for HZP we insert a DAX exceptional entry with the new
RADIX_DAX_HZP flag set.  This is because we use a single 2 MiB zero page in
every 2MiB hole mapping, and it doesn't make sense to have that same struct
page* with multiple entries in multiple trees.  This would cause contention
on the single page lock for the one Huge Zero Page, and it would break the
page->index and page->mapping associations that are assumed to be valid in
many other places in the kernel.

One difficult use case is when one thread is trying to use 4k entries in
radix tree for a given offset, and another thread is using 2 MiB entries
for that same offset.  The current code handles this by making the 2 MiB
user fall back to 4k entries for most cases.  This was done because it is
the simplest solution, and because the use of 2MiB pages is already
opportunistic.

If we were to try to upgrade from 4k pages to 2MiB pages for a given range,
we run into the problem of how we lock out 4k page faults for the entire
2MiB range while we clean out the radix tree so we can insert the 2MiB
entry.  We can solve this problem if we need to, but I think that the cases
where both 2MiB entries and 4K entries are being used for the same range
will be rare enough and the gain small enough that it probably won't be
worth the complexity.
Signed-off-by: NRoss Zwisler <ross.zwisler@linux.intel.com>
Reviewed-by: NJan Kara <jack@suse.cz>
Signed-off-by: NDave Chinner <david@fromorbit.com>