xfs: validate btree records on retrieval

So we don't check the validity of records as we walk the btree. When
there are corrupt records in the free space btree (e.g. zero
startblock/length or beyond EOAG) we just blindly use it and things
go bad from there. That leads to assert failures on debug kernels
like this:

XFS: Assertion failed: fs_is_ok, file: fs/xfs/libxfs/xfs_alloc.c, line: 450
....
Call Trace:
 xfs_alloc_fixup_trees+0x368/0x5c0
 xfs_alloc_ag_vextent_near+0x79a/0xe20
 xfs_alloc_ag_vextent+0x1d3/0x330
 xfs_alloc_vextent+0x5e9/0x870

Or crashes like this:

XFS (loop0): xfs_buf_find: daddr 0x7fb28 out of range, EOFS 0x8000
.....
BUG: unable to handle kernel NULL pointer dereference at 00000000000000c8
....
Call Trace:
 xfs_bmap_add_extent_hole_real+0x67d/0x930
 xfs_bmapi_write+0x934/0xc90
 xfs_da_grow_inode_int+0x27e/0x2f0
 xfs_dir2_grow_inode+0x55/0x130
 xfs_dir2_sf_to_block+0x94/0x5d0
 xfs_dir2_sf_addname+0xd0/0x590
 xfs_dir_createname+0x168/0x1a0
 xfs_rename+0x658/0x9b0

By checking that free space records pulled from the trees are
within the valid range, we catch many of these corruptions before
they can do damage.

This is a generic btree record checking deficiency. We need to
validate the records we fetch from all the different btrees before
we use them to catch corruptions like this.

This patch results in a corrupt record emitting an error message and
returning -EFSCORRUPTED, and the higher layers catch that and abort:

 XFS (loop0): Size Freespace BTree record corruption in AG 0 detected!
 XFS (loop0): start block 0x0 block count 0x0
 XFS (loop0): Internal error xfs_trans_cancel at line 1012 of file fs/xfs/xfs_trans.c.  Caller xfs_create+0x42a/0x670
 .....
 Call Trace:
  dump_stack+0x85/0xcb
  xfs_trans_cancel+0x19f/0x1c0
  xfs_create+0x42a/0x670
  xfs_generic_create+0x1f6/0x2c0
  vfs_create+0xf9/0x180
  do_mknodat+0x1f9/0x210
  do_syscall_64+0x5a/0x180
  entry_SYSCALL_64_after_hwframe+0x49/0xbe
.....
 XFS (loop0): xfs_do_force_shutdown(0x8) called from line 1013 of file fs/xfs/xfs_trans.c.  Return address = ffffffff81500868
 XFS (loop0): Corruption of in-memory data detected.  Shutting down filesystem

Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: push corruption -> ESTALE conversion to xfs_nfs_get_inode()

In xfs_imap_to_bp(), we convert a -EFSCORRUPTED error to -EINVAL if
we are doing an untrusted lookup. This is done because we need
failed filehandle lookups to report -ESTALE to the caller, and it
does this by converting -EINVAL and -ENOENT errors to -ESTALE.

The squashing of EFSCORRUPTED in imap_to_bp makes it impossible for
for xfs_iget(UNTRUSTED) callers to determine the difference between
"inode does not exist" and "corruption detected during lookup". We
realy need that distinction in places calling xfS_iget(UNTRUSTED),
so move the filehandle error case handling all the way out to
xfs_nfs_get_inode() where it is needed.

Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Carlos Maiolino <cmaiolino@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: verify root inode more thoroughly

When looking up the root inode at mount time, we don't actually do
any verification to check that the inode is allocated and accounted
for correctly in the INOBT. Make the checks on the root inode more
robust by making it an untrusted lookup. This forces the inode
lookup to use the inode btree to verify the inode is allocated
and mapped correctly to disk. This will also have the effect of
catching a significant number of AGI/INOBT related corruptions in
AG 0 at mount time.

Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Carlos Maiolino <cmaiolino@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: verify COW extent size hint is valid in inode verifier

There are rules for vald extent size hints. We enforce them when
applications set them, but fuzzers violate those rules and that
screws us over. Validate COW extent size hint rules in the inode
verifier to catch this.

Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Carlos Maiolino <cmaiolino@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: verify extent size hint is valid in inode verifier

There are rules for vald extent size hints. We enforce them when
applications set them, but fuzzers violate those rules and that
screws us over.

This results in alignment assertion failures when setting up
allocations such as this in direct IO:

XFS: Assertion failed: ap->length, file: fs/xfs/libxfs/xfs_bmap.c, line: 3432
....
Call Trace:
 xfs_bmap_btalloc+0x415/0x910
 xfs_bmapi_write+0x71c/0x12e0
 xfs_iomap_write_direct+0x2a9/0x420
 xfs_file_iomap_begin+0x4dc/0xa70
 iomap_apply+0x43/0x100
 iomap_file_buffered_write+0x62/0x90
 xfs_file_buffered_aio_write+0xba/0x300
 __vfs_write+0xd5/0x150
 vfs_write+0xb6/0x180
 ksys_write+0x45/0xa0
 do_syscall_64+0x5a/0x180
 entry_SYSCALL_64_after_hwframe+0x49/0xbe

And from xfs_db:

core.extsize = 10380288

Which is not an integer multiple of the block size, and so violates
Rule #7 for setting extent size hints. Validate extent size hint
rules in the inode verifier to catch this.

Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Carlos Maiolino <cmaiolino@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: catch bad stripe alignment configurations

When stripe alignments are invalid, data alignment algorithms in the
allocator may not work correctly. Ensure we catch superblocks with
invalid stripe alignment setups at mount time. These data alignment
mismatches are now detected at mount time like this:

XFS (loop0): SB stripe unit sanity check failed
XFS (loop0): Metadata corruption detected at xfs_sb_read_verify+0xab/0x110, xfs_sb block 0xffffffffffffffff
XFS (loop0): Unmount and run xfs_repair
XFS (loop0): First 128 bytes of corrupted metadata buffer:
0000000091c2de02: 58 46 53 42 00 00 10 00 00 00 00 00 00 00 10 00  XFSB............
0000000023bff869: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00  ................
00000000cdd8c893: 17 32 37 15 ff ca 46 3d 9a 17 d3 33 04 b5 f1 a2  .27...F=...3....
000000009fd2844f: 00 00 00 00 00 00 00 04 00 00 00 00 00 00 06 d0  ................
0000000088e9b0bb: 00 00 00 00 00 00 06 d1 00 00 00 00 00 00 06 d2  ................
00000000ff233a20: 00 00 00 01 00 00 10 00 00 00 00 01 00 00 00 00  ................
000000009db0ac8b: 00 00 03 60 e1 34 02 00 08 00 00 02 00 00 00 00  ...`.4..........
00000000f7022460: 00 00 00 00 00 00 00 00 0c 09 0b 01 0c 00 00 19  ................
XFS (loop0): SB validate failed with error -117.

And the mount fails.

Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Carlos Maiolino <cmaiolino@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

iomap: fsync swap files before iterating mappings

Swap files require that all the file mapping metadata be stable on disk.
It is insufficient to flush dirty pages in the page cache because that
won't necessarily result in filesystems pushing all their metadata out
to disk.  Therefore, call fsync from iomap_swapfile_activate.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Jan Kara <jack@suse.cz>

xfs: use xfs_trans_getsb in xfs_sync_sb_buf

Use xfs_trans_getsb rather than reaching right in for
mp->m_sb_bp; I think this is more correct, and it facilitates
building this libxfs code in userspace as well.

Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: don't assert on corrupted unlinked inode list

Use the per-ag inode number verifiers to detect corrupt lists and error
out, instead of using ASSERTs.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>

xfs: explicitly pass buffer size to xfs_corruption_error

Explicitly pass the buffer length to xfs_corruption_error() instead of
assuming XFS_CORRUPTION_DUMP_LEN so that we avoid dumping off the end
of the buffer.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>

xfs: don't assert when on-disk btree pointers are garbage

Don't ASSERT when we encounter bad on-disk btree pointers in the debug
check functions.  Log the error to leave breadcrumbs and let the upper
layers deal with it.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>

xfs: strengthen btree pointer checks before use

Instead of ASSERTing on null btree pointers in xfs_btree_ptr_to_daddr,
use the new block number verifiers to ensure that the btree pointer
doesn't point to any sensitive areas (AG headers, past-EOFS) and return
-EFSCORRUPTED if this is the case.  Remove the ASSERT because on-disk
corruptions shouldn't trigger ASSERTs.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>

xfs: introduce xfs_btree_debug_check_ptr

Make xfs_btree_check_ptr a non-debug function and introduce a new _debug
version that only runs when #ifdef DEBUG.   This will enable us to reuse
the checking logic with other parts of the btree code.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>

xfs: check directory bestfree information in the verifier

Create a variant of xfs_dir2_data_freefind that is suitable for use in a
verifier.  Because _freefind is called by the verifier, we simply
duplicate the _freefind function, convert the ASSERTs to return
__this_address, and modify the verifier to call our new function.  Once
we've made it impossible for directory blocks with bad bestfree data to
make it into the filesystem we can remove the DEBUG code from the
regular _freefind function.

Underlying argument: corruption of on-disk metadata should return
-EFSCORRUPTED instead of blowing ASSERTs.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>

xfs: don't return garbage buffers in xfs_da3_node_read

If we're reading a node in a dir/attr btree and the buffer comes off the
disk with a magic number we don't recognize, don't ASSERT and don't set
a garbage buffer type (0 also triggers ASSERTs).  Instead, report the
corruption, release the buffer, and return -EFSCORRUPTED because that's
what the dabtree is -- corrupt.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>

xfs: don't ASSERT on short form btree root pointer of zero

Don't ASSERT if the short form btree root pointer is zero.  Now that we
use xfs_verify_agbno to check all short form btree pointers, we'll let
that log the error and pass it to the upper layers.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>

xfs: btree lookup shouldn't ASSERT on empty btree nodes

If a btree lookup encounters an empty btree node or an empty btree leaf
on a multi-level btree, that's evidence of a corrupt on-disk btree.
Therefore, we should return -EFSCORRUPTED to the upper levels, not an
ASSERT failure.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>

xfs: xfs_alloc_get_rec should return EFSCORRUPTED for obvious bnobt corruption

Return -EFSCORRUPTED when the bnobt/cntbt return obviously corrupt
values, rather than letting them bounce around in the internal code.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>

xfs: remove redundant ASSERT on insufficient bestfree length in _leaf_addname

In xfs_dir2_leaf_addname we ASSERT if the length of the unused space
described by bestfree[0] is less the amount of space we wish to consume.
Immediately after it is a call to xfs_dir2_data_use_free where the
offset parameter is offset of the unused space and the length parameter
is the amount of space we wish to consume.  Both values (and the unused
space pointer) are passed into xfs_dir2_data_check_free, which also
validates that the region of unused space is big enough to cover the
space we wish to consume.  This is effectively the same check that the
ASSERT covers, and since a check failure results in a corruption message
being logged we can remove the ASSERT.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>

xfs: don't assert when reporting on-disk corruption while loading btree

Don't bother ASSERTing when we're already going to log and return the
corruption status.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>

xfs: don't forbid setting dax flag on directories if device doesn't dax

On a directory, the DAX flag is merely a hint that files created in the
directory should have the DAX flag set at creation time.  We don't care
if the underlying device supports DAX or not because directory metadata
are always cached in DRAM.  We don't care if new files get the flag even
if the device doesn't support DAX because we always check for DAX
support before setting the VFS flag (S_DAX).

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>

xfs: verify AGI unlinked list contains valid blocks

The heads of tha AGI unlinked list are only scanned on debug
kernels when the verifier runs. Change that to always scan the heads
and validate that the inode numbers are valid.

Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

fs: use ->is_partially_uptodate in page_cache_seek_hole_data

This way the implementation doesn't depend on buffer_head internals.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Andreas Gruenbacher <agruenba@redhat.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

fs: remove the buffer_unwritten check in page_seek_hole_data

We only call into this function through the iomap iterators, so we already
know the buffer is unwritten.  In addition to that we always require the
uptodate flag that is ORed with the result anyway.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Andreas Gruenbacher <agruenba@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

fs: move page_cache_seek_hole_data to iomap.c

This function is only used by the iomap code, depends on being called
from it, and will soon stop poking into buffer head internals.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Andreas Gruenbacher <agruenba@redhat.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: use iomap_bmap

Switch to the iomap based bmap implementation to get rid of one of the
last users of xfs_get_blocks.

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

iomap: add an iomap-based bmap implementation

This adds a simple iomap-based implementation of the legacy ->bmap
interface.  Note that we can't easily add checks for rt or reflink
files, so these will have to remain in the callers.  This interface
just needs to die..

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

iomap: add a iomap_sector helper

Factor the repeated calculation of the on-disk sector for a given logical
block into a littler helper.

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

iomap: use __bio_add_page in iomap_dio_zero

We don't need any merging logic, and this also replaces a BUG_ON with a
WARN_ON_ONCE inside __bio_add_page for the impossible overflow condition.

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

iomap: move IOMAP_F_BOUNDARY to gfs2

Just define a range of fs specific flags and use that in gfs2 instead of
exposing this internal flag globally.

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

iomap: fix the comment describing IOMAP_NOWAIT

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

iomap: inline data should be an iomap type, not a flag

Inline data is fundamentally different from our normal mapped case in that
it doesn't even have a block address.  So instead of having a flag for it
it should be an entirely separate iomap range type.

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

mm: split ->readpages calls to avoid non-contiguous pages lists

That way file systems don't have to go spotting for non-contiguous pages
and work around them.  It also kicks off I/O earlier, allowing it to
finish earlier and reduce latency.

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

mm: return an unsigned int from __do_page_cache_readahead

We never return an error, so switch to returning an unsigned int.  Most
callers already did implicit casts to an unsigned type, and the one that
didn't can be simplified now.

Suggested-by: Matthew Wilcox <willy@infradead.org>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

mm: give the 'ret' variable a better name __do_page_cache_readahead

It counts the number of pages acted on, so name it nr_pages to make that
obvious.

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

block: add a lower-level bio_add_page interface

For the upcoming removal of buffer heads in XFS we need to keep track of
the number of outstanding writeback requests per page.  For this we need
to know if bio_add_page merged a region with the previous bvec or not.
Instead of adding additional arguments this refactors bio_add_page to
be implemented using three lower level helpers which users like XFS can
use directly if they care about the merge decisions.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Jens Axboe <axboe@kernel.dk>
Reviewed-by: Ming Lei <ming.lei@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: fix error handling in xfs_refcount_insert()

generic/475 fired an assert failure just after the filesystem was
shut down:

XFS: Assertion failed: fs_is_ok, file: fs/xfs/libxfs/xfs_refcount.c, line: 182
.....
Call Trace:
 xfs_refcount_insert+0x151/0x190
 xfs_refcount_adjust_extents.constprop.11+0x9c/0x470
 xfs_refcount_adjust.constprop.10+0xb0/0x270
 xfs_refcount_finish_one+0x25a/0x420
 xfs_trans_log_finish_refcount_update+0x2a/0x40
 xfs_refcount_update_finish_item+0x35/0xa0
 xfs_defer_finish+0x15e/0x4d0
 xfs_reflink_remap_extent+0x1bc/0x610
 xfs_reflink_remap_blocks+0x6e/0x280
 xfs_reflink_remap_range+0x311/0x530
 vfs_clone_file_range+0x119/0x200
 ....

If xfs_btree_insert() returns an error, the corruption check fires
instead of passing the error back the caller. The corruption check
should be after we've checked for an error, not before, thereby
avoiding assert failures if the filesystem shuts down during a
refcount btree record insert.

Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: fix xfs_rtalloc_rec units

All the realtime allocation functions deal with space on the rtdev in
units of realtime extents.  However, struct xfs_rtalloc_rec confusingly
uses the word 'block' in the name, even though they're really extents.

Fix the naming problem and fix all the unit handling problems in the two
existing users.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Allison Henderson <allison.henderson@oracle.com>
Reviewed-by: Bill O'Donnell <billodo@redhat.com>

xfs: strengthen rtalloc query range checks

Strengthen the rtalloc range query checks to make sure that the keys do
not run off the end of the realtime device inappropriately.  Note that
the query range functions require units of rt extents, not blocks,
despite the type name.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Allison Henderson <allison.henderson@oracle.com>
Reviewed-by: Bill O'Donnell <billodo@redhat.com>

xfs: xfs_rtbuf_get should check the bmapi_read results

The xfs_rtbuf_get function should check the block mapping it gets back
from bmapi_read.  If there are no mappings or the mapping isn't a real
extent, we should return -EFSCORRUPTED rather than trying to read a
garbage value.  We also require realtime bitmap blocks to be real,
written allocations.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Allison Henderson <allison.henderson@oracle.com>
Reviewed-by: Bill O'Donnell <billodo@redhat.com>

xfs: xfs_rtword_t should be unsigned, not signed

xfs_rtword_t is used for bit manipulations in the realtime bitmap file.
Since we're performing bit shifts with this type, we don't want sign
extension and we don't want to be left shifting negative quantities
because that's undefined behavior.

This also shuts up these UBSAN warnings:
UBSAN: Undefined behaviour in fs/xfs/libxfs/xfs_rtbitmap.c:833:48
signed integer overflow:
-2147483648 - 1 cannot be represented in type 'int'

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Allison Henderson <allison.henderson@oracle.com>
Reviewed-by: Bill O'Donnell <billodo@redhat.com>

dax: change bdev_dax_supported() to support boolean returns

The function return values are confusing with the way the function is
named. We expect a true or false return value but it actually returns
0/-errno.  This makes the code very confusing. Changing the return values
to return a bool where if DAX is supported then return true and no DAX
support returns false.

Signed-off-by: Dave Jiang <dave.jiang@intel.com>
Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

fs: allow per-device dax status checking for filesystems

Change bdev_dax_supported so it takes a bdev parameter.  This enables
multi-device filesystems like xfs to check that a dax device can work for
the particular filesystem.  Once that's in place, actually fix all the
parts of XFS where we need to be able to distinguish between datadev and
rtdev.

This patch fixes the problem where we screw up the dax support checking
in xfs if the datadev and rtdev have different dax capabilities.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
[rez: Re-added __bdev_dax_supported() for !CONFIG_FS_DAX cases]
Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com>
Reviewed-by: Eric Sandeen <sandeen@redhat.com>

xfs: repair superblocks

If one of the backup superblocks is found to differ seriously from
superblock 0, write out a fresh copy from the in-core sb.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Allison Henderson <allison.henderson@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>

xfs: add helpers to attach quotas to inodes

Add a helper routine to attach quota information to inodes that are
about to undergo repair.  If that fails, we need to schedule a
quotacheck for the next mount but allow the corrupted metadata repair to
continue.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Allison Henderson <allison.henderson@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>

xfs: recover AG btree roots from rmap data

Add a helper function to help us recover btree roots from the rmap data.
Callers pass in a list of rmap owner codes, buffer ops, and magic
numbers.  We iterate the rmap records looking for owner matches, and
then read the matching blocks to see if the magic number & uuid match.
If so, we then read-verify the block, and if that passes then we retain
a pointer to the block with the highest level, assuming that by the end
of the call we will have found the root.  This will be used to reset the
AGF/AGI btree root fields during their rebuild procedures.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>

xfs: add helpers to dispose of old btree blocks after a repair

Now that we've plumbed in the ability to construct a list of dead btree
blocks following a repair, add more helpers to dispose of them.  This is
done by examining the rmapbt -- if the btree was the only owner we can
free the block, otherwise it's crosslinked and we can only remove the
rmapbt record.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>

xfs: add helpers to collect and sift btree block pointers during repair

Add some helpers to assemble a list of fs block extents.  Generally,
repair functions will iterate the rmapbt to make a list (1) of all
extents owned by the nominal owner of the metadata structure; then they
will iterate all other structures with the same rmap owner to make a
list (2) of active blocks; and finally we have a subtraction function to
subtract all the blocks in (2) from (1), with the result that (1) is now
a list of blocks that were owned by the old btree and must be disposed.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>

xfs: add helpers to allocate and initialize fresh btree roots

Add a pair of helper functions to allocate and initialize fresh btree
roots.  The repair functions will use these as part of recreating
corrupted metadata.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Allison Henderson <allison.henderson@oracle.com>

xfs: add helpers to deal with transaction allocation and rolling

For repairs, we need to reserve at least as many blocks as we think
we're going to need to rebuild the data structure, and we're going to
need some helpers to roll transactions while maintaining locks on the AG
headers so that other threads cannot wander into the middle of a repair.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Allison Henderson <allison.henderson@oracle.com>

xfs: grab the per-ag structure whenever relevant

Grab and hold the per-AG data across a scrub run whenever relevant.
This helps us avoid repeated trips through rcu and the radix tree
in the repair code.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>

fs: xfs: Change return type to vm_fault_t

Use new return type vm_fault_t for fault handlers.

Signed-off-by: Souptick Joarder <jrdr.linux@gmail.com>
Reviewed-by: Matthew Wilcox <mawilcox@microsoft.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: fix inobt magic number check

In commit a6a781a58befcbd467c ("xfs: have buffer verifier functions
report failing address") the bad magic number return was ported
incorrectly.

Fixes: a6a781a58befcbd467ce843af4eaca3906aa1f08
Reported-by: syzbot+08ab33be0178b76851c8@syzkaller.appspotmail.com
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Eric Sandeen <sandeen@redhat.com>

fs: clear writeback errors in inode_init_always

In inode_init_always(), we clear the inode mapping flags, which clears
any retained error (AS_EIO, AS_ENOSPC) bits.  Unfortunately, we do not
also clear wb_err, which means that old mapping errors can leak through
to new inodes.

This is crucial for the XFS inode allocation path because we recycle old
in-core inodes and we do not want error state from an old file to leak
into the new file.  This bug was discovered by running generic/036 and
generic/047 in a loop and noticing that the EIOs generated by the
collision of direct and buffered writes in generic/036 would survive the
remount between 036 and 047, and get reported to the fsyncs (on
different files!) in generic/047.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Jeff Layton <jlayton@kernel.org>
Reviewed-by: Brian Foster <bfoster@redhat.com>

iomap: don't allow holes in swapfiles

generic_swapfile_activate() doesn't allow holes, so we should be
consistent here. This is also a bit safer: if the user creates a
swapfile with, say, truncate -s $SIZE followed by mkswap, they should
really get an error and not much less swap space than they expected.
swapon(8) will error out before calling swapon(2) if the file has holes,
anyways.

Fixes: 9d93388b0afe ("iomap: add a swapfile activation function")
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Omar Sandoval <osandov@fb.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

iomap: provide more useful errors for invalid swap files

Currently, for an invalid swap file, we print the same error message
regardless of the reason. This isn't very useful for an admin, who will
likely want to know why exactly they can't use their swap file. So,
let's add specific error messages for each reason, and also move the
bdev check after the flags checks, since the latter are more
fundamental.

Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Omar Sandoval <osandov@fb.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: implement online get/set fs label

The GET ioctl is trivial, just return the current label.

The SET ioctl is more involved:
It transactionally modifies the superblock to write a new filesystem
label to the primary super.

A new variant of xfs_sync_sb then writes the superblock buffer
immediately to disk so that the change is visible from userspace.

It then invalidates any page cache that userspace might have previously
read on the block device so that i.e. blkid can see the change
immediately, and updates all secondary superblocks as userspace relable
does.

Signed-off-by: Eric Sandeen <sandeen@redhat.com>
[darrick: use dchinner's new xfs_update_secondary_sbs function]
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

fs: copy BTRFS_IOC_[SG]ET_FSLABEL to vfs

This retains 256 chars as the maximum size through the interface, which
is the btrfs limit and AFAIK exceeds any other filesystem's maximum
label size.

This just copies the ioctl for now and leaves it in place for btrfs
for the time being.  A later patch will allow btrfs to use the new
common ioctl definition, but it may be sent after this is merged.

(Note, Reviewed-by's were originally given for the combined vfs+btrfs
patch, some license taken here.)

Signed-off-by: Eric Sandeen <sandeen@redhat.com>
Reviewed-by: Andreas Dilger <adilger@dilger.ca>
Reviewed-by: David Sterba <dsterba@suse.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: factor the ag length extension code into libxfs

Growfs currently manually codes the extension of the last AG in a
filesytem during the growfs process. Factor that out of the growfs
code and move it into libxfs along with teh rest of the AG header
modification code.

Signed-Off-By: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: move growfs core to libxfs

So it can be shared with userspace (e.g. mkfs) easily.

Signed-Off-By: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: rework secondary superblock updates in growfs

Right now we wait until we've committed changes to the primary
superblock before we initialise any of the new secondary
superblocks. This means that if we have any write errors for new
secondary superblocks we end up with garbage in place rather than
zeros or even an "in progress" superblock to indicate a grow
operation is being done.

To ensure we can write the secondary superblocks, initialise them
earlier in the same loop that initialises the AG headers. We stamp
the new secondary superblocks here with the old geometry, but set
the "sb_inprogress" field to indicate that updates are being done to
the superblock so they cannot be used.  This will result in the
secondary superblock fields being updated or triggering errors that
will abort the grow before we commit any permanent changes.

This also means we can change the update mechanism of the secondary
superblocks.  We know that we are going to wholly overwrite the
information in the struct xfs_sb in the buffer, so there's no point
reading it from disk. Just allocate an uncached buffer, zero it in
memory, stamp the new superblock structure in it and write it out.
If we fail to write it out, then we'll leave the existing sb (old or
new w/ inprogress) on disk for repair to deal with later.

Signed-Off-By: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: separate secondary sb update in growfs

This happens after all the transactions to update the superblock
occur, and errors need to be handled slightly differently. Seperate
out the code into it's own function, and clean up the error goto
stack in the core growfs code as it is now much simpler.

Signed-Off-By: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: make imaxpct changes in growfs separate

When growfs changes the imaxpct value of the filesystem, it runs
through all the "change size" growfs code, whether it needs to or
not. Separate out changing imaxpct into it's own function and
transaction to simplify the rest of the growfs code.

Signed-Off-By: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: turn ag header initialisation into a table driven operation

There's still more cookie cutter code in setting up each AG header.
Separate all the variables into a simple structure and iterate a
table of header definitions to initialise everything.

Signed-Off-By: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: factor ag btree root block initialisation

Cookie cutter code, easily factored.

Signed-Off-By: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: convert growfs AG header init to use buffer lists

We currently write all new AG headers synchronously, which can be
slow for large grow operations. All we really need to do is ensure
all the headers are on disk before we run the growfs transaction, so
convert this to a buffer list and a delayed write operation. We
block waiting for the delayed write buffer submission to complete,
so this will fulfill the requirement to have all the buffers written
correctly before proceeding.

Signed-Off-By: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: factor out AG header initialisation from growfs core

The intialisation of new AG headers is mostly common with the
userspace mkfs code and growfs in the kernel, so start factoring it
out so we can move it to libxfs and use it in both places.

Signed-Off-By: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: one-shot cached buffers

For the new growfs work, we want to ensure that we serialise
secondary superblock updates with other operations (e.g. scrub)
correctly, but we don't want to cache the buffers for long term
reuse. We need cached buffers for serialisation, however.

To solve this, introduce a "oneshot" buffer which will be marshalled
through the cache but then released once the last current reference
goes away. If the buffer is already cached, then we ignore the
"one-shot" behaviour and leave the buffer in the state it was prior
to the one-shot command being run. This means we don't perturb
either the working set or existing cached buffer state by a one-shot
operation.

Signed-Off-By: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: implement the metadata repair ioctl flag

Plumb in the pieces necessary to make the "scrub" subfunction of
the scrub ioctl actually work.  This means that we make the IFLAG_REPAIR
flag to the scrub ioctl actually do something, and we add an errortag
knob so that xfstests can force the kernel to rebuild a metadata
structure even if there's nothing wrong with it.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>

xfs: create tracepoints for online repair

These tracepoints will be used to debug the online repair routines.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>

xfs: teach xfs_bmapi_remap to accept some bmapi flags

Teach xfs_bmapi_remap how to map in unwritten extent and to skip rmap
updates.  This enables us to rebuild real and unwritten extents from the
rmapbt.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>

xfs: make xfs_bmapi_remapi work with attribute forks

Add a new flags argument to xfs_bmapi_remapi so that we can pass BMAPI
flags into the function.  This enables us to pass in BMAPI_ATTRFORK so
that we can remap things into the attribute fork.  Eventually the
online repair code will use this to rebuild attribute forks, so make it
non-static.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>

xfs: hoist xfs_scrub_agfl_walk to libxfs as xfs_agfl_walk

This function is basically a generic AGFL block iterator, so promote it
to libxfs ahead of online repair wanting to use it.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>

xfs: avoid ABBA deadlock when scrubbing parent pointers

In normal operation, the XFS convention is to take an inode's iolock
and then allocate a transaction.  However, when scrubbing parent inodes
this is inverted -- we allocated the transaction to do the scrub, and
now we're trying to grab the parent's iolock.  This can lead to ABBA
deadlocks: some thread grabbed the parent's iolock and is waiting for
space for a transaction while our parent scrubber is sitting on a
transaction trying to get the parent's iolock.

Therefore, convert all iolock attempts to use trylock; if that fails,
they can use the existing mechanisms to back off and try again.

The ABBA deadlock didn't happen with a non-repair scrub because the
transactions don't reserve any space, but repair scrubs require
reservation in order to update metadata.  However, any other concurrent
metadata update (e.g. directory create in the parent) could also induce
this deadlock with the parent scrubber.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>

xfs: scrub the data fork of the realtime inodes

The realtime bitmap and summary inodes live on the metadata device, so
we can scrub their data forks with the regular scrubbers.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>

xfs: quota scrub should use bmapbtd scrubber

Replace the quota scrubber's open-coded data fork scrubber with a
redirected call to the bmapbtd scrubber.  This strengthens the quota
scrub to include all the cross-referencing that it does.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>

xfs: don't continue scrub if already corrupt

If we've already decided that something is corrupt, we might as well
abort all the loops and exit as quickly as possible.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>

xfs: refactor quota limits initialization

Replace all the if (!error) weirdness with helper functions that follow
our regular coding practices, and factor out the ternary expression soup.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>

xfs: superblock scrub should use short-lived buffers

Secondary superblocks are rarely used, so create a helper to read a
given non-primary AG's superblock and ensure that it won't stick around
hogging memory.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>

xfs: skip scrub xref if corruption already noted

Don't bother looking for cross-referencing problems if the metadata is
already corrupt or we've already found a cross-referencing problem.
Since we added a helper function for flags testing, convert existing
users to use it.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>

xfs: clear sb->s_fs_info on mount failure

We recently had an oops reported on a 4.14 kernel in
xfs_reclaim_inodes_count() where sb->s_fs_info pointed to garbage
and so the m_perag_tree lookup walked into lala land.

Essentially, the machine was under memory pressure when the mount
was being run, xfs_fs_fill_super() failed after allocating the
xfs_mount and attaching it to sb->s_fs_info. It then cleaned up and
freed the xfs_mount, but the sb->s_fs_info field still pointed to
the freed memory. Hence when the superblock shrinker then ran
it fell off the bad pointer.

With the superblock shrinker problem fixed at teh VFS level, this
stale s_fs_info pointer is still a problem - we use it
unconditionally in ->put_super when the superblock is being torn
down, and hence we can still trip over it after a ->fill_super
call failure. Hence we need to clear s_fs_info if
xfs-fs_fill_super() fails, and we need to check if it's valid in
the places it can potentially be dereferenced after a ->fill_super
failure.

Signed-Off-By: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: add mount delay debug option

Similar to log_recovery_delay, this delay occurs between the VFS
superblock being initialised and the xfs_mount being fully
initialised. It also poisons the per-ag radix tree node so that it
can be used for triggering shrinker races during mount
such as the following:

<run memory pressure workload in background>

$ cat dirty-mount.sh
#! /bin/bash

umount -f /dev/pmem0
mkfs.xfs -f /dev/pmem0
mount /dev/pmem0 /mnt/test
rm -f /mnt/test/foo
xfs_io -fxc "pwrite 0 4k" -c fsync -c "shutdown" /mnt/test/foo
umount /dev/pmem0

# let's crash it now!
echo 30 > /sys/fs/xfs/debug/mount_delay
mount /dev/pmem0 /mnt/test
echo 0 > /sys/fs/xfs/debug/mount_delay
umount /dev/pmem0
$ sudo ./dirty-mount.sh
.....
[   60.378118] CPU: 3 PID: 3577 Comm: fs_mark Tainted: G      D W        4.16.0-rc5-dgc #440
[   60.378120] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.10.2-1 04/01/2014
[   60.378124] RIP: 0010:radix_tree_next_chunk+0x76/0x320
[   60.378127] RSP: 0018:ffffc9000276f4f8 EFLAGS: 00010282
[   60.383670] RAX: a5a5a5a5a5a5a5a4 RBX: 0000000000000010 RCX: 000000000000001a
[   60.385277] RDX: 0000000000000000 RSI: ffffc9000276f540 RDI: 0000000000000000
[   60.386554] RBP: 0000000000000000 R08: 0000000000000000 R09: a5a5a5a5a5a5a5a5
[   60.388194] R10: 0000000000000006 R11: 0000000000000001 R12: ffffc9000276f598
[   60.389288] R13: 0000000000000040 R14: 0000000000000228 R15: ffff880816cd6458
[   60.390827] FS:  00007f5c124b9740(0000) GS:ffff88083fc00000(0000) knlGS:0000000000000000
[   60.392253] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[   60.393423] CR2: 00007f5c11bba0b8 CR3: 000000035580e001 CR4: 00000000000606e0
[   60.394519] Call Trace:
[   60.395252]  radix_tree_gang_lookup_tag+0xc4/0x130
[   60.395948]  xfs_perag_get_tag+0x37/0xf0
[   60.396522]  xfs_reclaim_inodes_count+0x32/0x40
[   60.397178]  xfs_fs_nr_cached_objects+0x11/0x20
[   60.397837]  super_cache_count+0x35/0xc0
[   60.399159]  shrink_slab.part.66+0xb1/0x370
[   60.400194]  shrink_node+0x7e/0x1a0
[   60.401058]  try_to_free_pages+0x199/0x470
[   60.402081]  __alloc_pages_slowpath+0x3a1/0xd20
[   60.403729]  __alloc_pages_nodemask+0x1c3/0x200
[   60.404941]  cache_grow_begin+0x20b/0x2e0
[   60.406164]  fallback_alloc+0x160/0x200
[   60.407088]  kmem_cache_alloc+0x111/0x4e0
[   60.408038]  ? xfs_buf_rele+0x61/0x430
[   60.408925]  kmem_zone_alloc+0x61/0xe0
[   60.409965]  xfs_inode_alloc+0x24/0x1d0
.....

Signed-Off-By: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

xfs: factor out nodiscard helpers

The changes to skip discards of speculative preallocation and
unwritten extents introduced several new wrapper functions through
the bunmapi -> extent free codepath to reduce churn in all of the
associated callers. In several cases, these wrappers simply toggle a
single flag to skip or not skip discards for the resulting blocks.

The explicit _nodiscard() wrappers for such an isolated set of
callers is a bit overkill. Kill off these wrappers and replace with
the calls to the underlying functions in the contexts that need to
control discard behavior. Retain the wrappers that preserve the
original calling conventions to serve the original purpose of
reducing code churn.

This is a refactoring patch and does not change behavior.

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: Darrick J. Wong <darrick.wong@oracle.com>

iomap: add a swapfile activation function

Add a new iomap_swapfile_activate function so that filesystems can
activate swap files without having to use the obsolete and slow bmap
function.  This enables XFS to support fallocate'd swap files and
swap files on realtime devices.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Jan Kara <jack@suse.cz>

xfs: halt auto-reclamation activities while rebuilding rmap

Rebuilding the reverse-mapping tree requires us to quiesce all inodes in
the filesystem, so we must stop background reclamation of post-EOF and
CoW prealloc blocks.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>

xfs: add BMAPI_NORMAP flag to perform block remapping without updating rmapbt

Add a new flag, XFS_BMAPI_NORMAP, which will perform file block
remapping without updating the rmapbt.  This will be used by the repair
code to reconstruct bmbts from the rmapbt, in which case we don't want
the rmapbt update.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>

xfs: add repair helpers for the reference count btree

Add a couple of functions to the refcount btree and generic btree code
that will be used to repair the refcountbt.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>

xfs: add repair helpers for the reverse mapping btree

Add a couple of functions to the reverse mapping btree that will be used
to repair the rmapbt.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>

xfs: expose various functions to repair code

Expose various helpers that the repair code will want to use.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>

xfs: add helpers to calculate btree size

Add a bunch of helper functions that calculate the sizes of various
btrees.  These will be used to repair btrees and btree headers.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>

xfs: refactor scrub transaction allocation function

Since the transaction allocation helper is about to become more complex,
move it to common.c and remove the redundant parameters.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>

xfs: btree scrub should check minrecs

Strengthen the btree block header checks to detect the number of records
being less than the btree type's minimum record count.  Certain blocks
are allowed to violate this constraint -- specifically any btree block
at the top of the tree can have fewer than minrecs records.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>

xfs: clean up scrub usage of KM_NOFS

All scrub code runs in transaction context, which means that memory
allocations are automatically run in PF_MEMALLOC_NOFS context.  It's
therefore unnecessary to pass in KM_NOFS to allocation routines, so
clean them all out.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>

xfs: avoid ilock games in the quota scrubber

Refactor the quota scrubber to take the quotaofflock and grab the quota
inode in the setup function so that we can treat quota in the same
"scrub in the context of this inode" (i.e. sc->ip) manner as we treat
any other inode.  We do have to drop the quota inode's ILOCK_EXCL to use
dqiterate, but since dquots have their own individual locks the ILOCK
wasn't helping us anyway.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>

xfs: refactor dquot iteration

Create a helper function to iterate all the dquots of a given type in
the system, and refactor the dquot scrub to use it.  This will get more
use in the quota repair code.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>

xfs: rename on-disk dquot counter zap functions

The function 'xfs_qm_dqiterate' doesn't iterate dquots at all, it
iterates all dquot blocks of a quota inode and clears the counters.
Therefore, change the name to something more descriptive so that we can
introduce a real dquot iterator later.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>

xfs: replace XFS_QMOPT_DQALLOC with a simple boolean

DQALLOC is only ever used with xfs_qm_dqget*, and the only flag that the
_dqget family of functions cares about is DQALLOC.  Therefore, change
it to a boolean 'can alloc?' flag for the dqget interfaces where that
makes sense.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>

xfs: remove direct calls to _qm_dqread

The quota initialization code needs an "uncached" variant of _dqget to
read in default quota limits and timers before the dquot cache is fully
set up.  We've already split up _dqget into its component pieces so
create a fourth variant to address this need, and make dqread internal
to xfs_dquot.c again.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>

xfs: refactor xfs_qm_dqtobp and xfs_qm_dqalloc

Separate the disk dquot read and allocation functionality into
two helper functions, then refactor dqread to call them directly.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>

xfs: refactor incore dquot initialization functions

Create two incore dquot initialization functions that will help us to
disentangle dqget and dqread.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>