platform/kernel/linux-rpi.git
7 years agoMerge branch 'xfs-4.10-misc-fixes-3' into for-next
Dave Chinner [Wed, 7 Dec 2016 06:42:30 +0000 (17:42 +1100)]
Merge branch 'xfs-4.10-misc-fixes-3' into for-next

7 years agoxfs: use rhashtable to track buffer cache
Lucas Stach [Wed, 7 Dec 2016 06:36:36 +0000 (17:36 +1100)]
xfs: use rhashtable to track buffer cache

On filesystems with a lot of metadata and in metadata intensive workloads
xfs_buf_find() is showing up at the top of the CPU cycles trace. Most of
the CPU time is spent on CPU cache misses while traversing the rbtree.

As the buffer cache does not need any kind of ordering, but fast lookups
a hashtable is the natural data structure to use. The rhashtable
infrastructure provides a self-scaling hashtable implementation and
allows lookups to proceed while the table is going through a resize
operation.

This reduces the CPU-time spent for the lookups to 1/3 even for small
filesystems with a relatively small number of cached buffers, with
possibly much larger gains on higher loaded filesystems.

[dchinner: reduce minimum hash size to an acceptable size for large
   filesystems with many AGs with no active use.]
[dchinner: remove stale rbtree asserts.]
[dchinner: use xfs_buf_map for compare function argument.]
[dchinner: make functions static.]
[dchinner: remove redundant comments.]

Signed-off-by: Lucas Stach <dev@lynxeye.de>
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: optimise CRC updates
Dave Chinner [Mon, 5 Dec 2016 03:40:32 +0000 (14:40 +1100)]
xfs: optimise CRC updates

Nick Piggin reported that the CRC overhead in an fsync heavy
workload was higher than expected on a Power8 machine. Part of this
was to do with the fact that the power8 CRC implementation is not
efficient for CRC lengths of less than 512 bytes, and so the way we
split the CRCs over the CRC field means a lot of the CRCs are
reduced to being less than than optimal size.

To optimise this, change the CRC update mechanism to zero the CRC
field first, and then compute the CRC in one pass over the buffer
and write the result back into the buffer. We can do this safely
because anything writing a CRC has exclusive access to the buffer
the CRC is being calculated over.

We leave the CRC verify code the same - it still splits the CRC
calculation - because we do not want read-only operations modifying
the underlying buffer. This is because read-only operations may not
have an exclusive access to the buffer guaranteed, and so temporary
modifications could leak out to to other processes accessing the
buffer concurrently.

Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: make xfs btree stats less huge
Dave Chinner [Mon, 5 Dec 2016 03:38:58 +0000 (14:38 +1100)]
xfs: make xfs btree stats less huge

Embedding a switch statement in every btree stats inc/add adds a lot
of code overhead to the core btree infrastructure paths. Stats are
supposed to be small and lightweight, but the btree stats have
become big and bloated as we've added more btrees. It needs fixing
because the reflink code will just add more overhead again.

Convert the v2 btree stats to arrays instead of independent
variables, and instead use the type to index the specific btree
array via an enum. This allows us to use array based indexing
to update the stats, rather than having to derefence variables
specific to the btree type.

If we then wrap the xfsstats structure in a union and place uint32_t
array beside it, and calculate the correct btree stats array base
array index when creating a btree cursor,  we can easily access
entries in the stats structure without having to switch names based
on the btree type.

We then replace with the switch statement with a simple set of stats
wrapper macros, resulting in a significant simplification of the
btree stats code, and:

   text    data     bss     dec     hex filename
  48905     144       8   49057    bfa1 fs/xfs/libxfs/xfs_btree.o.old
  36793     144       8   36945    9051 fs/xfs/libxfs/xfs_btree.o

it reduces the core btree infrastructure code size by close to 25%!

Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: don't cap maximum dedupe request length
Darrick J. Wong [Mon, 5 Dec 2016 01:38:57 +0000 (12:38 +1100)]
xfs: don't cap maximum dedupe request length

After various discussions on linux-fsdevel, it has been decided that it
is not necessary to cap the length of a dedupe request, and that
correctly-written userspace client programs will be able to absorb the
change.  Therefore, remove the length clamping behavior.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: don't allow di_size with high bit set
Darrick J. Wong [Mon, 5 Dec 2016 01:38:38 +0000 (12:38 +1100)]
xfs: don't allow di_size with high bit set

The on-disk field di_size is used to set i_size, which is a signed
integer of loff_t.  If the high bit of di_size is set, we'll end up with
a negative i_size, which will cause all sorts of problems.  Since the
VFS won't let us create a file with such length, we should catch them
here in the verifier too.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: error out if trying to add attrs and anextents > 0
Darrick J. Wong [Mon, 5 Dec 2016 01:38:11 +0000 (12:38 +1100)]
xfs: error out if trying to add attrs and anextents > 0

We shouldn't assert if somehow we end up trying to add an attr fork to
an inode that apparently already has attr extents because this is an
indication of on-disk corruption.  Instead, return an error code to
userspace.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: don't crash if reading a directory results in an unexpected hole
Darrick J. Wong [Mon, 5 Dec 2016 01:37:47 +0000 (12:37 +1100)]
xfs: don't crash if reading a directory results in an unexpected hole

In xfs_dir3_data_read, we can encounter the situation where err == 0 and
*bpp == NULL if the given bno offset happens to be a hole; this leads to
a crash if we try to set the buffer type after the _da_read_buf call.
Holes can happen due to corrupt or malicious entries in the bmbt data,
so be a little more careful when we're handling buffers.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: complain if we don't get nextents bmap records
Darrick J. Wong [Mon, 5 Dec 2016 01:36:56 +0000 (12:36 +1100)]
xfs: complain if we don't get nextents bmap records

When reading into memory all extents of a btree-format inode fork,
complain if the number of extents we find is not the same as the number
of extents reported in the inode core.  This is needed to stop an IO
action from accessing the garbage areas of the in-core fork.

[dchinner: removed redundant assert]

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: check for bogus values in btree block headers
Darrick J. Wong [Mon, 5 Dec 2016 01:33:54 +0000 (12:33 +1100)]
xfs: check for bogus values in btree block headers

When we're reading a btree block, make sure that what we retrieved
matches the owner and level; and has a plausible number of records.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: forbid AG btrees with level == 0
Darrick J. Wong [Mon, 5 Dec 2016 01:32:50 +0000 (12:32 +1100)]
xfs: forbid AG btrees with level == 0

There is no such thing as a zero-level AG btree since even a single-node
zero-records btree has one level.  Btree cursor constructors read
cur_nlevels straight from disk and then access things like
cur_bufs[cur_nlevels - 1] which is /really/ bad if cur_nlevels is zero!
Therefore, strengthen the verifiers to prevent this possibility.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: several xattr functions can be void
Eric Sandeen [Mon, 5 Dec 2016 01:32:14 +0000 (12:32 +1100)]
xfs: several xattr functions can be void

There are a handful of xattr functions which now return
nothing but zero.  They can be made void, chased through calling
functions, and error handling etc can be removed.

Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: handle cow fork in xfs_bmap_trace_exlist
Eric Sandeen [Mon, 5 Dec 2016 01:32:00 +0000 (12:32 +1100)]
xfs: handle cow fork in xfs_bmap_trace_exlist

By inspection, xfs_bmap_trace_exlist isn't handling cow forks,
and will trace the data fork instead.

Fix this by setting state appropriately if whichfork
== XFS_COW_FORK.

()___()
< @ @ >
 |   |
 {o_o}
  (|)

Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: pass state not whichfork to trace_xfs_extlist
Eric Sandeen [Mon, 5 Dec 2016 01:31:50 +0000 (12:31 +1100)]
xfs: pass state not whichfork to trace_xfs_extlist

When xfs_bmap_trace_exlist called trace_xfs_extlist,
it sent in the "whichfork" var instead of the bmap "state"
as expected (even though state was already set up for this
purpose).

As a result, the xfs_bmap_class in tracing code used
"whichfork" not state in xfs_iext_state_to_fork(), and got
the wrong ifork pointer.  It all goes downhill from
there, including an ASSERT when ifp_bytes is empty
by the time it reaches xfs_iext_get_ext():

XFS: Assertion failed: idx < ifp->if_bytes / sizeof(xfs_bmbt_rec_t)

Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: Move AGI buffer type setting to xfs_read_agi
Eric Sandeen [Mon, 5 Dec 2016 01:31:31 +0000 (12:31 +1100)]
xfs: Move AGI buffer type setting to xfs_read_agi

We've missed properly setting the buffer type for
an AGI transaction in 3 spots now, so just move it
into xfs_read_agi() and set it if we are in a transaction
to avoid the problem in the future.

This is similar to how it is done in i.e. the dir3
and attr3 read functions.

Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: set AGI buffer type in xlog_recover_clear_agi_bucket
Eric Sandeen [Mon, 5 Dec 2016 01:31:06 +0000 (12:31 +1100)]
xfs: set AGI buffer type in xlog_recover_clear_agi_bucket

xlog_recover_clear_agi_bucket didn't set the
type to XFS_BLFT_AGI_BUF, so we got a warning during log
replay (or an ASSERT on a debug build).

    XFS (md0): Unknown buffer type 0!
    XFS (md0): _xfs_buf_ioapply: no ops on block 0xaea8802/0x1

Fix this, as was done in f19b872b for 2 other locations
with the same problem.

cc: <stable@vger.kernel.org> # 3.10 to current
Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoMerge branch 'iomap-4.10-directio' into for-next
Dave Chinner [Wed, 30 Nov 2016 03:39:29 +0000 (14:39 +1100)]
Merge branch 'iomap-4.10-directio' into for-next

7 years agoxfs: use iomap_dio_rw
Christoph Hellwig [Wed, 30 Nov 2016 03:37:15 +0000 (14:37 +1100)]
xfs: use iomap_dio_rw

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: Christoph Hellwig <hch@lst.de>
Tested-by: Jens Axboe <axboe@fb.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoiomap: implement direct I/O
Christoph Hellwig [Wed, 30 Nov 2016 03:36:01 +0000 (14:36 +1100)]
iomap: implement direct I/O

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: Christoph Hellwig <hch@lst.de>
Reviewed-by: Kent Overstreet <kent.overstreet@gmail.com>
Tested-by: Jens Axboe <axboe@fb.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agofs: make sb_init_dio_done_wq available outside of direct-io.c
Christoph Hellwig [Wed, 30 Nov 2016 03:33:53 +0000 (14:33 +1100)]
fs: make sb_init_dio_done_wq available outside of direct-io.c

We want to use the per-sb completion workqueue from the new iomap
direct I/O code.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Tested-by: Jens Axboe <axboe@fb.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: remove i_iolock and use i_rwsem in the VFS inode instead
Christoph Hellwig [Wed, 30 Nov 2016 03:33:25 +0000 (14:33 +1100)]
xfs: remove i_iolock and use i_rwsem in the VFS inode instead

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: Christoph Hellwig <hch@lst.de>
Tested-by: Jens Axboe <axboe@fb.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agolocking/lockdep: Provide a type check for lock_is_held
Peter Zijlstra [Wed, 30 Nov 2016 03:32:25 +0000 (14:32 +1100)]
locking/lockdep: Provide a type check for lock_is_held

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: Christoph Hellwig <hch@lst.de>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Tested-by: Jens Axboe <axboe@fb.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoblock: add bio_iov_iter_get_pages()
Kent Overstreet [Mon, 31 Oct 2016 17:59:24 +0000 (11:59 -0600)]
block: add bio_iov_iter_get_pages()

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: Kent 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: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
(cherry picked from commit 9cd56d916aa481ce8f56d9c5302a6ed90c2e0b5f)

7 years agoMerge branch 'xfs-4.10-misc-fixes-2' into iomap-4.10-directio
Dave Chinner [Wed, 30 Nov 2016 01:49:38 +0000 (12:49 +1100)]
Merge branch 'xfs-4.10-misc-fixes-2' into iomap-4.10-directio

7 years agoMerge branch 'xfs-4.10-misc-fixes-2' into for-next
Dave Chinner [Mon, 28 Nov 2016 04:06:03 +0000 (15:06 +1100)]
Merge branch 'xfs-4.10-misc-fixes-2' into for-next

7 years agoxfs: pass post-eof speculative prealloc blocks to bmapi
Brian Foster [Mon, 28 Nov 2016 03:57:42 +0000 (14:57 +1100)]
xfs: pass post-eof speculative prealloc blocks to bmapi

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: Brian Foster <bfoster@redhat.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: clean up cow fork reservation and tag inodes correctly
Brian Foster [Mon, 28 Nov 2016 03:57:42 +0000 (14:57 +1100)]
xfs: clean up cow fork reservation and tag inodes correctly

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: Brian Foster <bfoster@redhat.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: track preallocation separately in xfs_bmapi_reserve_delalloc()
Brian Foster [Mon, 28 Nov 2016 03:57:42 +0000 (14:57 +1100)]
xfs: track preallocation separately in xfs_bmapi_reserve_delalloc()

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: Brian Foster <bfoster@redhat.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: always succeed when deduping zero bytes
Darrick J. Wong [Mon, 28 Nov 2016 03:57:42 +0000 (14:57 +1100)]
xfs: always succeed when deduping zero bytes

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: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agofs: xfs: libxfs: constify xfs_nameops structures
Bhumika Goyal [Mon, 28 Nov 2016 03:57:42 +0000 (14:57 +1100)]
fs: xfs: libxfs: constify xfs_nameops structures

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: Bhumika Goyal <bhumirks@gmail.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agofs: xfs: xfs_icreate_item: constify xfs_item_ops structure
Bhumika Goyal [Mon, 28 Nov 2016 03:57:42 +0000 (14:57 +1100)]
fs: xfs: xfs_icreate_item: constify xfs_item_ops structure

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: Bhumika Goyal <bhumirks@gmail.com>
Reviewed-by: Eric Sandeen <sandeen@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: factor rmap btree size into the indlen calculations
Darrick J. Wong [Mon, 28 Nov 2016 03:57:42 +0000 (14:57 +1100)]
xfs: factor rmap btree size into the indlen calculations

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: Eryu Guan <eguan@redhat.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: add XBF_XBF_NO_IOACCT to buf trace output
Eric Sandeen [Mon, 28 Nov 2016 03:57:42 +0000 (14:57 +1100)]
xfs: add XBF_XBF_NO_IOACCT to buf trace output

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: Eric Sandeen <sandeen@redhat.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoMerge branch 'xfs-4.10-extent-lookup' into for-next
Dave Chinner [Thu, 24 Nov 2016 00:41:59 +0000 (11:41 +1100)]
Merge branch 'xfs-4.10-extent-lookup' into for-next

7 years agoxfs: remove NULLEXTNUM
Christoph Hellwig [Thu, 24 Nov 2016 00:40:32 +0000 (11:40 +1100)]
xfs: remove NULLEXTNUM

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: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: remove xfs_bmap_search_extents
Christoph Hellwig [Thu, 24 Nov 2016 00:40:14 +0000 (11:40 +1100)]
xfs: remove xfs_bmap_search_extents

Now that all users are gone.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: use new extent lookup helpers in xfs_reflink_end_cow
Christoph Hellwig [Thu, 24 Nov 2016 00:39:50 +0000 (11:39 +1100)]
xfs: use new extent lookup helpers in xfs_reflink_end_cow

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: use new extent lookup helpers in xfs_reflink_cancel_cow_blocks
Christoph Hellwig [Thu, 24 Nov 2016 00:39:50 +0000 (11:39 +1100)]
xfs: use new extent lookup helpers in xfs_reflink_cancel_cow_blocks

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: use new extent lookup helpers in xfs_reflink_trim_irec_to_next_cow
Christoph Hellwig [Thu, 24 Nov 2016 00:39:50 +0000 (11:39 +1100)]
xfs: use new extent lookup helpers in xfs_reflink_trim_irec_to_next_cow

And remove the unused return value.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: cleanup xfs_reflink_find_cow_mapping
Christoph Hellwig [Thu, 24 Nov 2016 00:39:49 +0000 (11:39 +1100)]
xfs: cleanup xfs_reflink_find_cow_mapping

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: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: use new extent lookup helpers in __xfs_reflink_reserve_cow
Christoph Hellwig [Thu, 24 Nov 2016 00:39:49 +0000 (11:39 +1100)]
xfs: use new extent lookup helpers in __xfs_reflink_reserve_cow

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: use new extent lookup helpers xfs_file_iomap_begin_delay
Christoph Hellwig [Thu, 24 Nov 2016 00:39:44 +0000 (11:39 +1100)]
xfs: use new extent lookup helpers xfs_file_iomap_begin_delay

And only lookup the previous extent inside xfs_iomap_prealloc_size
if we actually need it.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: remove prev argument to xfs_bmapi_reserve_delalloc
Christoph Hellwig [Thu, 24 Nov 2016 00:39:44 +0000 (11:39 +1100)]
xfs: remove prev argument to xfs_bmapi_reserve_delalloc

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: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: use new extent lookup helpers in __xfs_bunmapi
Christoph Hellwig [Thu, 24 Nov 2016 00:39:44 +0000 (11:39 +1100)]
xfs: use new extent lookup helpers in __xfs_bunmapi

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: use new extent lookup helpers in xfs_bmapi_write
Christoph Hellwig [Thu, 24 Nov 2016 00:39:43 +0000 (11:39 +1100)]
xfs: use new extent lookup helpers in xfs_bmapi_write

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: use new extent lookup helpers in xfs_bmapi_read
Christoph Hellwig [Thu, 24 Nov 2016 00:39:43 +0000 (11:39 +1100)]
xfs: use new extent lookup helpers in xfs_bmapi_read

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: cleanup xfs_bmap_last_before
Christoph Hellwig [Thu, 24 Nov 2016 00:39:38 +0000 (11:39 +1100)]
xfs: cleanup xfs_bmap_last_before

Rewrite the function using xfs_iext_lookup_extent and xfs_iext_get_extent,
and massage the flow into something easily understandable.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: new inode extent list lookup helpers
Christoph Hellwig [Thu, 24 Nov 2016 00:39:32 +0000 (11:39 +1100)]
xfs: new inode extent list lookup helpers

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: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoMerge branch 'xfs-4.10-misc-fixes-1' into for-next
Dave Chinner [Wed, 9 Nov 2016 23:29:43 +0000 (10:29 +1100)]
Merge branch 'xfs-4.10-misc-fixes-1' into for-next

7 years agoMerge branch 'xfs-4.10-libxfs-cleanups' into for-next
Dave Chinner [Wed, 9 Nov 2016 23:29:29 +0000 (10:29 +1100)]
Merge branch 'xfs-4.10-libxfs-cleanups' into for-next

7 years agoMerge branch 'dax-4.10-iomap-pmd' into for-next
Dave Chinner [Wed, 9 Nov 2016 23:29:06 +0000 (10:29 +1100)]
Merge branch 'dax-4.10-iomap-pmd' into for-next

7 years agodax: Introduce IOMAP_FAULT flag
Jan Kara [Wed, 9 Nov 2016 23:26:50 +0000 (10:26 +1100)]
dax: Introduce IOMAP_FAULT flag

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: Jan Kara <jack@suse.cz>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: fix unbalanced inode reclaim flush locking
Brian Foster [Wed, 9 Nov 2016 21:23:22 +0000 (08:23 +1100)]
xfs: fix unbalanced inode reclaim flush locking

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: Zorro Lang <zlang@redhat.com>
Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: check minimum block size for CRC filesystems
Darrick J. Wong [Wed, 9 Nov 2016 01:11:12 +0000 (12:11 +1100)]
xfs: check minimum block size for CRC filesystems

Check the minimum block size on v5 filesystems.

[dchinner: cleaned up XFS_MIN_CRC_BLOCKSIZE check]

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: provide helper for counting extents from if_bytes
Eric Sandeen [Tue, 8 Nov 2016 01:59:42 +0000 (12:59 +1100)]
xfs: provide helper for counting extents from if_bytes

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: Eric Sandeen <sandeen@redhat.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: fix up xfs_swap_extent_forks inline extent handling
Eric Sandeen [Tue, 8 Nov 2016 01:55:18 +0000 (12:55 +1100)]
xfs: fix up xfs_swap_extent_forks inline extent handling

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: Eric Sandeen <sandeen@redhat.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: don't BUG() on mixed direct and mapped I/O
Brian Foster [Tue, 8 Nov 2016 01:54:14 +0000 (12:54 +1100)]
xfs: don't BUG() on mixed direct and mapped I/O

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: Brian Foster <bfoster@redhat.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: don't skip cow forks w/ delalloc blocks in cowblocks scan
Brian Foster [Tue, 8 Nov 2016 01:53:33 +0000 (12:53 +1100)]
xfs: don't skip cow forks w/ delalloc blocks in cowblocks scan

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: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: check return value of _trans_reserve_quota_nblks
Darrick J. Wong [Tue, 8 Nov 2016 00:59:26 +0000 (11:59 +1100)]
xfs: check return value of _trans_reserve_quota_nblks

Check the return value of xfs_trans_reserve_quota_nblks for errors.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: move dir_ino_validate declaration per xfsprogs
Darrick J. Wong [Tue, 8 Nov 2016 00:59:12 +0000 (11:59 +1100)]
xfs: move dir_ino_validate declaration per xfsprogs

Move the declaration of _dir_ino_validate out of the private
dir2 header file into the public one, since xfsprogs did that
for the benefit of xfs_repair.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: don't call xfs_sb_quota_from_disk twice
Eric Sandeen [Tue, 8 Nov 2016 00:58:55 +0000 (11:58 +1100)]
xfs: don't call xfs_sb_quota_from_disk twice

Source xfsprogs commit: ee3754254e8c186c99b6cdd4d59f741759d04acb

Kernel commit 5ef828c4 ("xfs: avoid false quotacheck after unclean
shutdown") made xfs_sb_from_disk() also call xfs_sb_quota_from_disk
by default.

However, when this was merged to libxfs, existing separate
calls to libxfs_sb_quota_from_disk remained, and calling it
twice in a row on a V4 superblock leads to issues, because:

        if (sbp->sb_qflags & XFS_PQUOTA_ACCT)  {
...
                sbp->sb_pquotino = sbp->sb_gquotino;
                sbp->sb_gquotino = NULLFSINO;

and after the second call, we have set both pquotino and gquotino
to NULLFSINO.

Fix this by making it safe to call twice, and also remove the extra
calls to libxfs_sb_quota_from_disk.

This is only spotted when running xfstests with "-m crc=0" because
the sb_from_disk change came about after V5 became default, and
the above behavior only exists on a V4 superblock.

Reported-by: Eryu Guan <eguan@redhat.com>
Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Reviewed-by: Carlos Maiolino <cmaiolino@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agolibxfs: clean up _dir2_data_freescan
Darrick J. Wong [Tue, 8 Nov 2016 00:56:51 +0000 (11:56 +1100)]
libxfs: clean up _dir2_data_freescan

Refactor the implementations of xfs_dir2_data_freescan into a
routine that takes the raw directory block parameters and
a second function that figures out the raw parameters from the
directory inode.  This enables us to use the exact same code
for both userspace and the kernel, since repair knows exactly
which directory block geometry parameters it needs.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agolibxfs: fix xfs_attr_shortform_bytesfit declaration
Darrick J. Wong [Tue, 8 Nov 2016 00:56:20 +0000 (11:56 +1100)]
libxfs: fix xfs_attr_shortform_bytesfit declaration

Change the xfs_attr_shortform_bytesfit declaration to have
struct xfs_inode to avoid tripping up the libxfs-diff scanner.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Eric Sandeen <sandeen@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agolibxfs: fix whitespace problems
Darrick J. Wong [Tue, 8 Nov 2016 00:56:13 +0000 (11:56 +1100)]
libxfs: fix whitespace problems

Fix some whitespace problems that trip up my libxfs-diff script.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Eric Sandeen <sandeen@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agolibxfs: synchronize dinode_verify with userspace
Darrick J. Wong [Tue, 8 Nov 2016 00:56:06 +0000 (11:56 +1100)]
libxfs: synchronize dinode_verify with userspace

The userspace version of _dinode_verify takes a raw inode number
instead of an inode itself.  Since neither version actually needs
the inode, port the changes to the kernel.  This will also reduce
the libxfs diff noise.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agolibxfs: convert ushort to unsigned short
Darrick J. Wong [Tue, 8 Nov 2016 00:55:48 +0000 (11:55 +1100)]
libxfs: convert ushort to unsigned short

Since xfsprogs dropped ushort in favor of unsigned short, do that
here too.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agodax: remove "depends on BROKEN" from FS_DAX_PMD
Ross Zwisler [Tue, 8 Nov 2016 00:35:16 +0000 (11:35 +1100)]
dax: remove "depends on BROKEN" from FS_DAX_PMD

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: Ross Zwisler <ross.zwisler@linux.intel.com>
Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: use struct iomap based DAX PMD fault path
Ross Zwisler [Tue, 8 Nov 2016 00:35:02 +0000 (11:35 +1100)]
xfs: use struct iomap based DAX PMD fault path

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: Ross Zwisler <ross.zwisler@linux.intel.com>
Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agodax: add struct iomap based DAX PMD support
Ross Zwisler [Tue, 8 Nov 2016 00:34:45 +0000 (11:34 +1100)]
dax: add struct iomap based DAX PMD support

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: Ross Zwisler <ross.zwisler@linux.intel.com>
Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agodax: move put_(un)locked_mapping_entry() in dax.c
Ross Zwisler [Tue, 8 Nov 2016 00:33:44 +0000 (11:33 +1100)]
dax: move put_(un)locked_mapping_entry() in dax.c

No functional change.

The static functions put_locked_mapping_entry() and
put_unlocked_mapping_entry() will soon be used in error cases in
grab_mapping_entry(), so move their definitions above this function.

Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com>
Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agodax: move RADIX_DAX_* defines to dax.h
Ross Zwisler [Tue, 8 Nov 2016 00:33:35 +0000 (11:33 +1100)]
dax: move RADIX_DAX_* defines to dax.h

The RADIX_DAX_* defines currently mostly live in fs/dax.c, with just
RADIX_DAX_ENTRY_LOCK being in include/linux/dax.h so it can be used in
mm/filemap.c.  When we add PMD support, though, mm/filemap.c will also need
access to the RADIX_DAX_PTE type so it can properly construct a 4k sized
empty entry.

Instead of shifting the defines between dax.c and dax.h as they are
individually used in other code, just move them wholesale to dax.h so
they'll be available when we need them.

Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agodax: dax_iomap_fault() needs to call iomap_end()
Ross Zwisler [Tue, 8 Nov 2016 00:33:26 +0000 (11:33 +1100)]
dax: dax_iomap_fault() needs to call iomap_end()

Currently iomap_end() doesn't do anything for DAX page faults for both ext2
and XFS.  ext2_iomap_end() just checks for a write underrun, and
xfs_file_iomap_end() checks to see if it needs to finish a delayed
allocation.  However, in the future iomap_end() calls might be needed to
make sure we have balanced allocations, locks, etc.  So, add calls to
iomap_end() with appropriate error handling to dax_iomap_fault().

Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com>
Suggested-by: Jan Kara <jack@suse.cz>
Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agodax: add dax_iomap_sector() helper function
Ross Zwisler [Tue, 8 Nov 2016 00:33:09 +0000 (11:33 +1100)]
dax: add dax_iomap_sector() helper function

To be able to correctly calculate the sector from a file position and a
struct iomap there is a complex little bit of logic that currently happens
in both dax_iomap_actor() and dax_iomap_fault().  This will need to be
repeated yet again in the DAX PMD fault handler when it is added, so break
it out into a helper function.

Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agodax: correct dax iomap code namespace
Ross Zwisler [Tue, 8 Nov 2016 00:32:46 +0000 (11:32 +1100)]
dax: correct dax iomap code namespace

The recently added DAX functions that use the new struct iomap data
structure were named iomap_dax_rw(), iomap_dax_fault() and
iomap_dax_actor().  These are actually defined in fs/dax.c, though, so
should be part of the "dax" namespace and not the "iomap" namespace.
Rename them to dax_iomap_rw(), dax_iomap_fault() and dax_iomap_actor()
respectively.

Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com>
Suggested-by: Dave Chinner <david@fromorbit.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agodax: remove dax_pmd_fault()
Ross Zwisler [Tue, 8 Nov 2016 00:32:35 +0000 (11:32 +1100)]
dax: remove dax_pmd_fault()

dax_pmd_fault() is the old struct buffer_head + get_block_t based 2 MiB DAX
fault handler.  This fault handler has been disabled for several kernel
releases, and support for PMDs will be reintroduced using the struct iomap
interface instead.

Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agodax: coordinate locking for offsets in PMD range
Ross Zwisler [Tue, 8 Nov 2016 00:32:20 +0000 (11:32 +1100)]
dax: coordinate locking for offsets in PMD range

DAX radix tree locking currently locks entries based on the unique
combination of the 'mapping' pointer and the pgoff_t 'index' for the entry.
This works for PTEs, but as we move to PMDs we will need to have all the
offsets within the range covered by the PMD to map to the same bit lock.
To accomplish this, for ranges covered by a PMD entry we will instead lock
based on the page offset of the beginning of the PMD entry.  The 'mapping'
pointer is still used in the same way.

Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agodax: consistent variable naming for DAX entries
Ross Zwisler [Tue, 8 Nov 2016 00:32:12 +0000 (11:32 +1100)]
dax: consistent variable naming for DAX entries

No functional change.

Consistently use the variable name 'entry' instead of 'ret' for DAX radix
tree entries.  This was already happening in most of the code, so update
get_unlocked_mapping_entry(), grab_mapping_entry() and
dax_unlock_mapping_entry().

Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agodax: remove the last BUG_ON() from fs/dax.c
Ross Zwisler [Tue, 8 Nov 2016 00:32:00 +0000 (11:32 +1100)]
dax: remove the last BUG_ON() from fs/dax.c

Don't take down the kernel if we get an invalid 'from' and 'length'
argument pair.  Just warn once and return an error.

Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agodax: make 'wait_table' global variable static
Ross Zwisler [Tue, 8 Nov 2016 00:31:44 +0000 (11:31 +1100)]
dax: make 'wait_table' global variable static

The global 'wait_table' variable is only used within fs/dax.c, and
generates the following sparse warning:

fs/dax.c:39:19: warning: symbol 'wait_table' was not declared. Should it be static?

Make it static so it has scope local to fs/dax.c, and to make sparse happy.

Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoext2: remove support for DAX PMD faults
Ross Zwisler [Tue, 8 Nov 2016 00:31:33 +0000 (11:31 +1100)]
ext2: remove support for DAX PMD faults

DAX PMD support was added via the following commit:

commit e7b1ea2ad658 ("ext2: huge page fault support")

I believe this path to be untested as ext2 doesn't reliably provide block
allocations that are aligned to 2MiB.  In my testing I've been unable to
get ext2 to actually fault in a PMD.  It always fails with a "pfn
unaligned" message because the sector returned by ext2_get_block() isn't
aligned.

I've tried various settings for the "stride" and "stripe_width" extended
options to mkfs.ext2, without any luck.

Since we can't reliably get PMDs, remove support so that we don't have an
untested code path that we may someday traverse when we happen to get an
aligned block allocation.  This should also make 4k DAX faults in ext2 a
bit faster since they will no longer have to call the PMD fault handler
only to get a response of VM_FAULT_FALLBACK.

Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agodax: remove buffer_size_valid()
Ross Zwisler [Tue, 8 Nov 2016 00:31:14 +0000 (11:31 +1100)]
dax: remove buffer_size_valid()

Now that ext4 properly sets bh.b_size when we call get_block() for a hole,
rely on that value and remove the buffer_size_valid() sanity check.

Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com>
Reviewed-by: Jan Kara <jack@suse.cz>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoext4: tell DAX the size of allocation holes
Ross Zwisler [Tue, 8 Nov 2016 00:30:58 +0000 (11:30 +1100)]
ext4: tell DAX the size of allocation holes

When DAX calls _ext4_get_block() and the file offset points to a hole we
currently don't set bh->b_size.  This is current worked around via
buffer_size_valid() in fs/dax.c.

_ext4_get_block() has the hole size information from ext4_map_blocks(), so
populate bh->b_size so we can remove buffer_size_valid() in a later patch.

Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com>
Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: defer should abort intent items if the trans roll fails
Darrick J. Wong [Mon, 24 Oct 2016 03:21:18 +0000 (14:21 +1100)]
xfs: defer should abort intent items if the trans roll fails

If the deferred ops transaction roll fails, we need to abort the intent
items if we haven't already logged a done item for it, regardless of
whether or not the deferred ops has had a transaction committed.  Dave
found this while running generic/388.

Move the tracepoint to make it easier to track object lifetimes.

Reported-by: Dave Chinner <david@fromorbit.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: clear cowblocks tag when cow fork is emptied
Brian Foster [Mon, 24 Oct 2016 03:21:08 +0000 (14:21 +1100)]
xfs: clear cowblocks tag when cow fork is emptied

The background cowblocks scan job takes care of scanning for inodes with
potentially lingering blocks in the cow fork and clearing them out. If
the background scanner reclaims the cow fork blocks, however, it doesn't
immediately clear the cowblocks tag from the inode. Instead, the inode
remains tagged until the background scanner comes around again,
discovers the inode cow fork has no blocks, clears the tag and fires the
trace_xfs_inode_free_cowblocks_invalid() tracepoint to indicate that the
inode may have been incorrectly tagged.

This is not a major functional problem as the tag is ultimately cleared.
Nonetheless, clear the tag when an inode cow fork is explicitly emptied
to avoid the extra round trip through the background scanner and
spurious "invalid" tracepoint.

Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: fix up inode cowblocks tracking tracepoints
Brian Foster [Mon, 24 Oct 2016 03:21:00 +0000 (14:21 +1100)]
xfs: fix up inode cowblocks tracking tracepoints

These calls are still using the eofblocks tracepoints. The cowblocks
equivalents are already defined, we just aren't actually calling them.

Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agofs: Do to trim high file position bits in iomap_page_mkwrite_actor
Jan Kara [Mon, 24 Oct 2016 03:20:25 +0000 (14:20 +1100)]
fs: Do to trim high file position bits in iomap_page_mkwrite_actor

iomap_page_mkwrite_actor() calls __block_write_begin_int() with position
masked as pos & ~PAGE_MASK which is equivalent to pos & (PAGE_SIZE-1).
Thus it masks off high bits of file position. However
__block_write_begin_int() expects full file position on input. This does
not cause any visible issues because all __block_write_begin_int()
really cares about are low file position bits but still it is a bug
waiting to happen.

Signed-off-by: Jan Kara <jack@suse.cz>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: remove xfs_bunmapi_cow
Christoph Hellwig [Thu, 20 Oct 2016 04:54:59 +0000 (15:54 +1100)]
xfs: remove xfs_bunmapi_cow

Since no one uses it anymore.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: optimize xfs_reflink_end_cow
Christoph Hellwig [Thu, 20 Oct 2016 04:54:45 +0000 (15:54 +1100)]
xfs: optimize xfs_reflink_end_cow

Instead of doing a full extent list search for each extent that is
to be deleted using xfs_bmapi_read and then doing another one inside
of xfs_bunmapi_cow use the same scheme that xfs_bumapi uses:  look
up the last extent to be deleted and then use the extent index to
walk downward until we are outside the range to be deleted.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: optimize xfs_reflink_cancel_cow_blocks
Christoph Hellwig [Thu, 20 Oct 2016 04:54:31 +0000 (15:54 +1100)]
xfs: optimize xfs_reflink_cancel_cow_blocks

Rewrite xfs_reflink_cancel_cow_blocks so that we only do a search for
the first extent in the extent list and then iterate over the remaining
extents using the extent index, passing the extent we operate on
directly to xfs_bmap_del_extent_delay or xfs_bmap_del_extent_cow instead
of going through xfs_bunmapi and doing yet another extent list lookup.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: refactor xfs_bunmapi_cow
Christoph Hellwig [Thu, 20 Oct 2016 04:54:14 +0000 (15:54 +1100)]
xfs: refactor xfs_bunmapi_cow

Split out two helpers for deleting delayed or real extents from the COW fork.
This allows to call them directly from xfs_reflink_cow_end_io once that
function is refactored to iterate the extent tree.  It will also allow
to reuse the delalloc deletion from xfs_bunmapi in the future.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: optimize writes to reflink files
Christoph Hellwig [Thu, 20 Oct 2016 04:53:50 +0000 (15:53 +1100)]
xfs: optimize writes to reflink files

Instead of reserving space as the first thing in write_begin move it past
reading the extent in the data fork.  That way we only have to read from
the data fork once and can reuse that information for trimming the extent
to the shared/unshared boundary.  Additionally this allows to easily
limit the actual write size to said boundary, and avoid a roundtrip on the
ilock.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: don't bother looking at the refcount tree for reads
Christoph Hellwig [Thu, 20 Oct 2016 04:53:32 +0000 (15:53 +1100)]
xfs: don't bother looking at the refcount tree for reads

There is no need to trim an extent into a shared or non-shared one, or
report any flags for plain old reads.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: handle "raw" delayed extents xfs_reflink_trim_around_shared
Christoph Hellwig [Thu, 20 Oct 2016 04:52:00 +0000 (15:52 +1100)]
xfs: handle "raw" delayed extents xfs_reflink_trim_around_shared

Delalloc extents in the extent list contain the number of reserved
indirect blocks in their startblock value and don't use the magic
DELAYSTARTBLOCK constant.  Ensure that xfs_reflink_trim_around_shared
handles them properly by checking for isnullstartblock().

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: add xfs_trim_extent
Darrick J. Wong [Thu, 20 Oct 2016 04:51:50 +0000 (15:51 +1100)]
xfs: add xfs_trim_extent

This helpers allows to trim an extent to a subset of it's original range
while making sure the block numbers in it remain valid,

In the future xfs_trim_extent and xfs_bmapi_trim_map should probably be
merged in some form.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
[hch: split from a previous patch from Darrick, moved around and added
 support for "raw" delayed extents"]
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoiomap: add IOMAP_REPORT
Christoph Hellwig [Thu, 20 Oct 2016 04:51:28 +0000 (15:51 +1100)]
iomap: add IOMAP_REPORT

This allows the file system to tell a FIEMAP from a read operation, and thus
avoids the need to report flags that aren't actually used in the read path.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: merge xfs_reflink_remap_range and xfs_file_share_range
Christoph Hellwig [Thu, 20 Oct 2016 04:50:07 +0000 (15:50 +1100)]
xfs: merge xfs_reflink_remap_range and xfs_file_share_range

There is no clear division of responsibility between those functions, so
just merge them into one to keep the code simple.  Also move
xfs_file_wait_for_io to xfs_reflink.c together with its only caller.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: remove xfs_file_wait_for_io
Christoph Hellwig [Thu, 20 Oct 2016 04:49:55 +0000 (15:49 +1100)]
xfs: remove xfs_file_wait_for_io

filemap_write_and_wait_range operates on full pages, so there is no
need for the rounding operations.  Additionally this allows us to
micro-optimize by skipping the second inode_dio_wait for a
intra-file clone.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: move inode locking from xfs_reflink_remap_range to xfs_file_share_range
Christoph Hellwig [Thu, 20 Oct 2016 04:49:19 +0000 (15:49 +1100)]
xfs: move inode locking from xfs_reflink_remap_range to xfs_file_share_range

We need the iolock protection to stabilizie the IS_SWAPFILE and
IS_IMMUTABLE values, as well as preventing new buffered writers
re-dirtying the file data that we just wrote out.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: fix the same_inode check in xfs_file_share_range
Christoph Hellwig [Thu, 20 Oct 2016 04:49:03 +0000 (15:49 +1100)]
xfs: fix the same_inode check in xfs_file_share_range

The VFS i_ino is an unsigned long, while XFS inode numbers are 64-bit
wide, so checking i_ino for equality could lead to rate false positives
on 32-bit architectures.  Just compare the inode pointers themselves
to be safe.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: remove the same fs check from xfs_file_share_range
Christoph Hellwig [Thu, 20 Oct 2016 04:48:54 +0000 (15:48 +1100)]
xfs: remove the same fs check from xfs_file_share_range

The VFS already does the check, and the placement of this duplicate
is in the way of the following locking rework.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>