Refactor the EFI intent item recovery (and cancellation) functions
into a general function that scans the AIL and an intent item type
specific handler.  Move the function that recovers a single EFI item
into the extent free item code.  We'll want the generalized function
when we start wiring up more redo item types.

Furthermore, ensure that log recovery only replays the redo items
that were in the AIL prior to recovery by checking the item LSN
against the largest LSN seen during log scanning.  As written this
should never happen, but we can be defensive anyway.
Signed-off-by: NDarrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Mechanical change of flist/free_list to dfops, since they're now
deferred ops, not just a freeing list.
Signed-off-by: NDarrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Drop the compatibility shims that we were using to integrate the new
deferred operation mechanism into the existing code.  No new code.
Signed-off-by: NDarrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Restructure everything that used xfs_bmap_free to use xfs_defer_ops
instead.  For now we'll just remove the old symbols and play some
cpp magic to make it work; in the next patch we'll actually rename
everything.
Signed-off-by: NDarrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Connect the xfs_defer mechanism with the pieces that we'll need to
handle deferred extent freeing.  We'll wire up the existing code to
our new deferred mechanism later.
Signed-off-by: NDarrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Replace structure typedefs with struct xfs_foo_* in the EFI/EFD
handling code in preparation to move it over to deferred ops.
Signed-off-by: NDarrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Add tracepoints for the internals of the deferred ops mechanism
and tracepoint classes for clients of the dops, to make debugging
easier.
Signed-off-by: NDarrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

All the code around struct xfs_bmap_free basically implements a
deferred operation framework through which we can roll transactions
(to unlock buffers and avoid violating lock order rules) while
managing all the necessary log redo items.  Previously we only used
this code to free extents after some sort of mapping operation, but
with the advent of rmap and reflink, we suddenly need to do more than
that.

With that in mind, xfs_bmap_free really becomes a deferred ops control
structure.  Rename the structure and move the deferred ops into their
own file to avoid further bloating of the bmap code.
Signed-off-by: NDarrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Refactor the btree_change_owner function into a more generic apparatus
which visits all blocks in a btree.  We'll use this in a subsequent
patch for counting btree blocks for AG reservations.
Signed-off-by: NDarrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Create a function to enable querying of btree records mapping to a
range of keys.  This will be used in subsequent patches to allow
querying the reverse mapping btree to find the extents mapped to a
range of physical blocks, though the generic code can be used for
any range query.

The overlapped query range function needs to use the btree get_block
helper because the root block could be an inode, in which case
bc_bufs[nlevels-1] will be NULL.
Signed-off-by: NDarrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NDave Chinner <david@fromorbit.com>

On a filesystem with both reflink and reverse mapping enabled, it's
possible to have multiple rmap records referring to the same blocks on
disk.  When overlapping intervals are possible, querying a classic
btree to find all records intersecting a given interval is inefficient
because we cannot use the left side of the search interval to filter
out non-matching records the same way that we can use the existing
btree key to filter out records coming after the right side of the
search interval.  This will become important once we want to use the
rmap btree to rebuild BMBTs, or implement the (future) fsmap ioctl.

(For the non-overlapping case, we can perform such queries trivially
by starting at the left side of the interval and walking the tree
until we pass the right side.)

Therefore, extend the btree code to come closer to supporting
intervals as a first-class record attribute.  This involves widening
the btree node's key space to store both the lowest key reachable via
the node pointer (as the btree does now) and the highest key reachable
via the same pointer and teaching the btree modifying functions to
keep the highest-key records up to date.

This behavior can be turned on via a new btree ops flag so that btrees
that cannot store overlapping intervals don't pay the overhead costs
in terms of extra code and disk format changes.

When we're deleting a record in a btree that supports overlapped
interval records and the deletion results in two btree blocks being
joined, we defer updating the high/low keys until after all possible
joining (at higher levels in the tree) have finished.  At this point,
the btree pointers at all levels have been updated to remove the empty
blocks and we can update the low and high keys.

When we're doing this, we must be careful to update the keys of all
node pointers up to the root instead of stopping at the first set of
keys that don't need updating.  This is because it's possible for a
single deletion to cause joining of multiple levels of tree, and so
we need to update everything going back to the root.

The diff_two_keys functions return < 0, 0, or > 0 if key1 is less than,
equal to, or greater than key2, respectively.  This is consistent
with the rest of the kernel and the C library.

In btree_updkeys(), we need to evaluate the force_all parameter before
running the key diff to avoid reading uninitialized memory when we're
forcing a key update.  This happens when we've allocated an empty slot
at level N + 1 to point to a new block at level N and we're in the
process of filling out the new keys.
Signed-off-by: NDarrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: NDave Chinner <dchinner@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Add some function pointers to bc_ops to get the btree keys for
leaf and node blocks, and to update parent keys of a block.
Convert the _btree_updkey calls to use our new pointer, and
modify the tree shape changing code to call the appropriate
get_*_keys pointer instead of _btree_copy_keys because the
overlapping btree has to calculate high key values.
Signed-off-by: NDarrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

When a btree block has to be split, we pass the new block's ptr from
xfs_btree_split() back to xfs_btree_insert() via a pointer parameter;
however, we pass the block's key through the cursor's record.  It is a
little weird to "initialize" a record from a key since the non-key
attributes will have garbage values.

When we go to add support for interval queries, we have to be able to
pass the lowest and highest keys accessible via a pointer.  There's no
clean way to pass this back through the cursor's record field.
Therefore, pass the key directly back to xfs_btree_insert() the same
way that we pass the btree_ptr.

As a bonus, we no longer need init_rec_from_key and can drop it from the
codebase.
Signed-off-by: NDarrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NDave Chinner <david@fromorbit.com>

If we make the inode root block of a btree unfull by expanding the
root, we must set *stat to 1 to signal success, rather than leaving
it uninitialized.
Signed-off-by: NDarrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NDave Chinner <david@fromorbit.com>

When we're deleting realtime extents, we need to lock the summary
inode in case we need to update the summary info to prevent an assert
on the rsumip inode lock on a debug kernel.  While we're at it, fix
the locking annotations so that we avoid triggering lockdep warnings.
Signed-off-by: NDarrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Apparently cris doesn't require structure stride to align with the
largest type in the struct, so list[0] isn't at offset 4 like it is
everywhere else.  Fix this... insofar as existing XFSes on cris are
screwed.
Signed-off-by: NDarrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NDave Chinner <david@fromorbit.com>

When we're iterating inode xattrs by handle, we have to copy the
cursor back to userspace so that a subsequent invocation actually
retrieves subsequent contents.
Signed-off-by: NDarrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Been around for long enough now, hasn't caused any regression test
failures in the past 3 months, so it's time to make it a fully
supported feature.
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NEric Sandeen <sandeen@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

In xfs_finish_page_writeback(), we have a loop that looks like this:

        do {
                if (off < bvec->bv_offset)
                        goto next_bh;
                if (off > end)
                        break;
                bh->b_end_io(bh, !error);
next_bh:
                off += bh->b_size;
        } while ((bh = bh->b_this_page) != head);

The b_end_io function is end_buffer_async_write(), which will call
end_page_writeback() once all the buffers have marked as no longer
under IO.  This issue here is that the only thing currently
protecting both the bufferhead chain and the page from being
reclaimed is the PageWriteback state held on the page.

While we attempt to limit the loop to just the buffers covered by
the IO, we still read from the buffer size and follow the next
pointer in the bufferhead chain. There is no guarantee that either
of these are valid after the PageWriteback flag has been cleared.
Hence, loops like this are completely unsafe, and result in
use-after-free issues. One such problem was caught by Calvin Owens
with KASAN:

.....
 INFO: Freed in 0x103fc80ec age=18446651500051355200 cpu=2165122683 pid=-1
  free_buffer_head+0x41/0x90
  __slab_free+0x1ed/0x340
  kmem_cache_free+0x270/0x300
  free_buffer_head+0x41/0x90
  try_to_free_buffers+0x171/0x240
  xfs_vm_releasepage+0xcb/0x3b0
  try_to_release_page+0x106/0x190
  shrink_page_list+0x118e/0x1a10
  shrink_inactive_list+0x42c/0xdf0
  shrink_zone_memcg+0xa09/0xfa0
  shrink_zone+0x2c3/0xbc0
.....
 Call Trace:
  <IRQ>  [<ffffffff81e8b8e4>] dump_stack+0x68/0x94
  [<ffffffff8153a995>] print_trailer+0x115/0x1a0
  [<ffffffff81541174>] object_err+0x34/0x40
  [<ffffffff815436e7>] kasan_report_error+0x217/0x530
  [<ffffffff81543b33>] __asan_report_load8_noabort+0x43/0x50
  [<ffffffff819d651f>] xfs_destroy_ioend+0x3bf/0x4c0
  [<ffffffff819d69d4>] xfs_end_bio+0x154/0x220
  [<ffffffff81de0c58>] bio_endio+0x158/0x1b0
  [<ffffffff81dff61b>] blk_update_request+0x18b/0xb80
  [<ffffffff821baf57>] scsi_end_request+0x97/0x5a0
  [<ffffffff821c5558>] scsi_io_completion+0x438/0x1690
  [<ffffffff821a8d95>] scsi_finish_command+0x375/0x4e0
  [<ffffffff821c3940>] scsi_softirq_done+0x280/0x340


Where the access is occuring during IO completion after the buffer
had been freed from direct memory reclaim.

Prevent use-after-free accidents in this end_io processing loop by
pre-calculating the loop conditionals before calling bh->b_end_io().
The loop is already limited to just the bufferheads covered by the
IO in progress, so the offset checks are sufficient to prevent
accessing buffers in the chain after end_page_writeback() has been
called by the the bh->b_end_io() callout.

Yet another example of why Bufferheads Must Die.

cc: <stable@vger.kernel.org> # 4.7
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reported-and-Tested-by: NCalvin Owens <calvinowens@fb.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

One of the problems we currently have with delayed logging is that
under serious memory pressure we can deadlock memory reclaim. THis
occurs when memory reclaim (such as run by kswapd) is reclaiming XFS
inodes and issues a log force to unpin inodes that are dirty in the
CIL.

The CIL is pushed, but this will only occur once it gets the CIL
context lock to ensure that all committing transactions are complete
and no new transactions start being committed to the CIL while the
push switches to a new context.

The deadlock occurs when the CIL context lock is held by a
committing process that is doing memory allocation for log vector
buffers, and that allocation is then blocked on memory reclaim
making progress. Memory reclaim, however, is blocked waiting for
a log force to make progress, and so we effectively deadlock at this
point.

To solve this problem, we have to move the CIL log vector buffer
allocation outside of the context lock so that memory reclaim can
always make progress when it needs to force the log. The problem
with doing this is that a CIL push can take place while we are
determining if we need to allocate a new log vector buffer for
an item and hence the current log vector may go away without
warning. That means we canot rely on the existing log vector being
present when we finally grab the context lock and so we must have a
replacement buffer ready to go at all times.

To ensure this, introduce a "shadow log vector" buffer that is
always guaranteed to be present when we gain the CIL context lock
and format the item. This shadow buffer may or may not be used
during the formatting, but if the log item does not have an existing
log vector buffer or that buffer is too small for the new
modifications, we swap it for the new shadow buffer and format
the modifications into that new log vector buffer.

The result of this is that for any object we modify more than once
in a given CIL checkpoint, we double the memory required
to track dirty regions in the log. For single modifications then
we consume the shadow log vectorwe allocate on commit, and that gets
consumed by the checkpoint. However, if we make multiple
modifications, then the second transaction commit will allocate a
shadow log vector and hence we will end up with double the memory
usage as only one of the log vectors is consumed by the CIL
checkpoint. The remaining shadow vector will be freed when th elog
item is freed.

This can probably be optimised in future - access to the shadow log
vector is serialised by the object lock (as opposited to the active
log vector, which is controlled by the CIL context lock) and so we
can probably free shadow log vector from some objects when the log
item is marked clean on removal from the AIL.
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

xfsprogs source commit 4280e59dcbc4cd8e01585efe788a68eb378048e8

xfs_da3_split() has to handle all three versions of the
directory/attribute btree structure. The attr tree is v1, the dir
tre is v2 or v3. The main difference between the v1 and v2/3 trees
is the way tree nodes are split - in the v1 tree we can require a
double split to occur because the object to be inserted may be
larger than the space made by splitting a leaf. In this case we need
to do a double split - one to split the full leaf, then another to
allocate an empty leaf block in the correct location for the new
entry.  This does not happen with dir (v2/v3) formats as the objects
being inserted are always guaranteed to fit into the new space in
the split blocks.

Indeed, for directories they *may* be an extra block on this buffer
pointer. However, it's guaranteed not to be a leaf block (i.e. a
directory data block) - the directory code only ever places hash
index or free space blocks in this pointer (as a cursor of
sorts), and so to use it as a directory data block will immediately
corrupt the directory.

The problem is that the code assumes that there may be extra blocks
that we need to link into the tree once we've split the root, but
this is not true for either dir or attr trees, because the extra
attr block is always consumed by the last node split before we split
the root. Hence the linking in an extra block is always wrong at the
root split level, and this manifests itself in repair as a directory
corruption in a repaired directory, leaving the directory rebuild
incomplete.

This is a dir v2 zero-day bug - it was in the initial dir v2 commit
that was made back in February 1998.

Fix this by ensuring the linking of the blocks after the root split
never tries to make use of the extra blocks that may be held in the
cursor. They are held there for other purposes and should never be
touched by the root splitting code.
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NDave Chinner <david@fromorbit.com>

We check IS_DAX(inode) before calling either xfs_file_dax_read or
xfs_file_dax_write, and this will lead the call being optimized out at
compile time when CONFIG_FS_DAX is disabled.

However, the two functions are marked STATIC, so they become global
symbols when CONFIG_XFS_DEBUG is set, leaving us with two unused global
functions that call into an undefined function and a broken "allmodconfig"
build:

fs/built-in.o: In function `xfs_file_dax_read':
fs/xfs/xfs_file.c:348: undefined reference to `dax_do_io'
fs/built-in.o: In function `xfs_file_dax_write':
fs/xfs/xfs_file.c:758: undefined reference to `dax_do_io'

Marking the two functions 'static noinline' instead of 'STATIC' will let
the compiler drop the symbols when there are no callers but avoid the
implicit inlining.
Signed-off-by: NArnd Bergmann <arnd@arndb.de>
Fixes: 16d4d435 ("xfs: split direct I/O and DAX path")
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NDave Chinner <david@fromorbit.com>

XFS has had scattered reports of delalloc blocks present at
->releasepage() time. This results in a warning with a stack trace
similar to the following:

 ...
 Call Trace:
  [<ffffffffa23c5b8f>] dump_stack+0x63/0x84
  [<ffffffffa20837a7>] warn_slowpath_common+0x97/0xe0
  [<ffffffffa208380a>] warn_slowpath_null+0x1a/0x20
  [<ffffffffa2326caf>] xfs_vm_releasepage+0x10f/0x140
  [<ffffffffa218c680>] ? page_mkclean_one+0xd0/0xd0
  [<ffffffffa218d3a0>] ? anon_vma_prepare+0x150/0x150
  [<ffffffffa21521c2>] try_to_release_page+0x32/0x50
  [<ffffffffa2166b2e>] shrink_active_list+0x3ce/0x3e0
  [<ffffffffa21671c7>] shrink_lruvec+0x687/0x7d0
  [<ffffffffa21673ec>] shrink_zone+0xdc/0x2c0
  [<ffffffffa2168539>] kswapd+0x4f9/0x970
  [<ffffffffa2168040>] ? mem_cgroup_shrink_node_zone+0x1a0/0x1a0
  [<ffffffffa20a0d99>] kthread+0xc9/0xe0
  [<ffffffffa20a0cd0>] ? kthread_stop+0x100/0x100
  [<ffffffffa26b404f>] ret_from_fork+0x3f/0x70
  [<ffffffffa20a0cd0>] ? kthread_stop+0x100/0x100

This occurs because it is possible for shrink_active_list() to send
pages marked dirty to ->releasepage() when certain buffer_head threshold
conditions are met. shrink_active_list() doesn't check the page dirty
state apparently to handle an old ext3 corner case where in some cases
clean pages would not have the dirty bit cleared, thus it is up to the
filesystem to determine how to handle the page.

XFS currently handles the delalloc case properly, but this behavior
makes the warning spurious. Update the XFS ->releasepage() handler to
explicitly skip dirty pages. Retain the existing delalloc/unwritten
checks so we continue to warn if such buffers exist on clean pages when
they shouldn't.
Diagnosed-by: NDave Chinner <david@fromorbit.com>
Signed-off-by: NBrian Foster <bfoster@redhat.com>
Reviewed-by: NDave Chinner <dchinner@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Instead we always declare struct xfs_dir2_sf_hdr as packed.  That's
the expected layout, and while most major architectures do the packing
by default the new structure size and offset checker showed that not
only the ARM old ABI got this wrong, but various minor embedded
architectures did as well.

[Verified that no code change on x86-64 results from this change]
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NDave Chinner <dchinner@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

And use an array of unsigned char values directly to avoid problems
with architectures that pad the size of structures.  This also gets
rid of the xfs_dir2_ino4_t and xfs_dir2_ino8_t types, and introduces
new constants for the size of 4 and 8 bytes as well as the size
difference between the two.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NDave Chinner <dchinner@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Just use an array of two unsigned chars directly to avoid problems
with architectures that pad the size of structures.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NDave Chinner <dchinner@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

So far the DAX code overloaded the direct I/O code path.  There is very little
in common between the two, and untangling them allows to clean up both variants.

As a side effect we also get separate trace points for both I/O types.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NDave Chinner <dchinner@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

We control both the callers and callees of ->direct_IO, so remove the
indirect calls.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NDave Chinner <dchinner@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

XFS already implement it's own flushing of the pagecache because it
implements proper synchronization for direct I/O reads.  This means
calling generic_file_read_iter for direct I/O is rather useless,
as it doesn't do much but updating the atime and iocb position for
us.  This also gets rid of the buffered I/O fallback that isn't used
for XFS.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NDave Chinner <dchinner@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Similar to what we did on the write side a while ago.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NDave Chinner <dchinner@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

All the three low-level read implementations that we might call already
take care of not overflowing the maximum supported bytes, no need to
duplicate it here.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NDave Chinner <dchinner@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Now that we have the direct I/O kiocb flag there is no real need to sample
the value inside of XFS, and the invis flag was always just partially used
and isn't worth keeping this infrastructure around for.   This also splits
the read tracepoint into buffered vs direct as we've done for writes a long
time ago.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NDave Chinner <dchinner@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Instead check the file pointer for the invisble I/O flag directly, and
use the chance to drop redundant arguments from the xfs_ioc_space
prototype.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NDave Chinner <dchinner@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Newly allocated XFS metadata buffers are added to the LRU once the hold
count is released, which typically occurs after I/O completion. There is
no other mechanism at current that tracks the existence or I/O state of
a new buffer. Further, readahead I/O tends to be submitted
asynchronously by nature, which means the I/O can remain in flight and
actually complete long after the calling context is gone. This means
that file descriptors or any other holds on the filesystem can be
released, allowing the filesystem to be unmounted while I/O is still in
flight. When I/O completion occurs, core data structures may have been
freed, causing completion to run into invalid memory accesses and likely
to panic.

This problem is reproduced on XFS via directory readahead. A filesystem
is mounted, a directory is opened/closed and the filesystem immediately
unmounted. The open/close cycle triggers a directory readahead that if
delayed long enough, runs buffer I/O completion after the unmount has
completed.

To address this problem, add a mechanism to track all in-flight,
asynchronous buffers using per-cpu counters in the buftarg. The buffer
is accounted on the first I/O submission after the current reference is
acquired and unaccounted once the buffer is returned to the LRU or
freed. Update xfs_wait_buftarg() to wait on all in-flight I/O before
walking the LRU list. Once in-flight I/O has completed and the workqueue
has drained, all new buffers should have been released onto the LRU.
Signed-off-by: NBrian Foster <bfoster@redhat.com>
Reviewed-by: NDave Chinner <dchinner@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

The upcoming buftarg I/O accounting mechanism maintains a count of
all buffers that have undergone I/O in the current hold-release
cycle.  Certain buffers associated with core infrastructure (e.g.,
the xfs_mount superblock buffer, log buffers) are never released,
however. This means that accounting I/O submission on such buffers
elevates the buftarg count indefinitely and could lead to lockup on
unmount.

Define a new buffer flag to explicitly exclude buffers from buftarg
I/O accounting. Set the flag on the superblock and associated log
buffers.
Signed-off-by: NBrian Foster <bfoster@redhat.com>
Reviewed-by: NDave Chinner <dchinner@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>