cascardo/linux.git
7 years agoxfs: increase XFS_BTREE_MAXLEVELS to fit the rmapbt
Darrick J. Wong [Wed, 3 Aug 2016 01:29:42 +0000 (11:29 +1000)]
xfs: increase XFS_BTREE_MAXLEVELS to fit the rmapbt

By my calculations, a 1,073,741,824 block AG with a 1k block size
can attain a maximum height of 9.  Assuming a record size of 24
bytes, a key/ptr size of 44 bytes, and half-full btree nodes, we'd
need 53,687,092 blocks for the records and ~6 million blocks for the
keys.  That requires a btree of height 9 based on the following
derivation:

Block size = 1024b
sblock CRC header = 56b
== 1024-56 = 968 bytes for tree data

rmapbt record = 24b
== 40 records per leaf block

rmapbt ptr/key = 44b
== 22 ptr/keys per block

Worst case, each block is half full, so 20 records and 11 ptrs per block.

1073741824 rmap records / 20 records per block
== 53687092 leaf blocks

53687092 leaves / 11 ptrs per block
== 4880645 level 1 blocks
== 443695 level 2 blocks
== 40336 level 3 blocks
== 3667 level 4 blocks
== 334 level 5 blocks
== 31 level 6 blocks
== 3 level 7 blocks
== 1 level 8 block

Signed-off-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 tracepoints and error injection for deferred extent freeing
Darrick J. Wong [Wed, 3 Aug 2016 01:26:33 +0000 (11:26 +1000)]
xfs: add tracepoints and error injection for deferred extent freeing

Add a couple of tracepoints for the deferred extent free operation and
a site for injecting errors while finishing the operation.  This makes
it easier to debug deferred ops and test log redo.

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: refactor redo intent item processing
Darrick J. Wong [Wed, 3 Aug 2016 01:23:49 +0000 (11:23 +1000)]
xfs: refactor redo intent item processing

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: 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: rename flist/free_list to dfops
Darrick J. Wong [Wed, 3 Aug 2016 01:19:29 +0000 (11:19 +1000)]
xfs: rename flist/free_list to dfops

Mechanical change of flist/free_list to dfops, since they're now
deferred ops, not just a freeing list.

Signed-off-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: change xfs_bmap_{finish,cancel,init,free} -> xfs_defer_*
Darrick J. Wong [Wed, 3 Aug 2016 01:18:10 +0000 (11:18 +1000)]
xfs: change xfs_bmap_{finish,cancel,init,free} -> xfs_defer_*

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: 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: rework xfs_bmap_free callers to use xfs_defer_ops
Darrick J. Wong [Wed, 3 Aug 2016 01:15:38 +0000 (11:15 +1000)]
xfs: rework xfs_bmap_free callers to use xfs_defer_ops

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: 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: enable the xfs_defer mechanism to process extents to free
Darrick J. Wong [Wed, 3 Aug 2016 01:14:35 +0000 (11:14 +1000)]
xfs: enable the xfs_defer mechanism to process extents to free

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: 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: clean up typedef usage in the EFI/EFD handling code
Darrick J. Wong [Wed, 3 Aug 2016 01:13:47 +0000 (11:13 +1000)]
xfs: clean up typedef usage in the EFI/EFD handling code

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: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: add tracepoints for the deferred ops mechanism
Darrick J. Wong [Wed, 3 Aug 2016 01:13:02 +0000 (11:13 +1000)]
xfs: add tracepoints for the deferred ops mechanism

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: 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: move deferred operations into a separate file
Darrick J. Wong [Wed, 3 Aug 2016 01:12:25 +0000 (11:12 +1000)]
xfs: move deferred operations into a separate file

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: 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 btree owner change into a separate visit-blocks function
Darrick J. Wong [Wed, 3 Aug 2016 01:10:55 +0000 (11:10 +1000)]
xfs: refactor btree owner change into a separate visit-blocks function

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: Darrick J. Wong <darrick.wong@oracle.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: introduce interval queries on btrees
Darrick J. Wong [Wed, 3 Aug 2016 01:10:21 +0000 (11:10 +1000)]
xfs: introduce interval queries on btrees

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: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: support btrees with overlapping intervals for keys
Darrick J. Wong [Wed, 3 Aug 2016 01:08:36 +0000 (11:08 +1000)]
xfs: support btrees with overlapping intervals for keys

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: 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 function pointers for get/update keys to the btree
Darrick J. Wong [Wed, 3 Aug 2016 01:03:38 +0000 (11:03 +1000)]
xfs: add function pointers for get/update keys to the btree

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: 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: during btree split, save new block key & ptr for future insertion
Darrick J. Wong [Wed, 3 Aug 2016 01:02:39 +0000 (11:02 +1000)]
xfs: during btree split, save new block key & ptr for future insertion

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: Darrick J. Wong <darrick.wong@oracle.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 *stat=1 after iroot realloc
Darrick J. Wong [Wed, 3 Aug 2016 01:01:25 +0000 (11:01 +1000)]
xfs: set *stat=1 after iroot realloc

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: Darrick J. Wong <darrick.wong@oracle.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: fix locking of the rt bitmap/summary inodes
Darrick J. Wong [Wed, 3 Aug 2016 01:00:42 +0000 (11:00 +1000)]
xfs: fix locking of the rt bitmap/summary inodes

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: Darrick J. Wong <darrick.wong@oracle.com>
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 attr shortform structure alignment on cris
Darrick J. Wong [Wed, 3 Aug 2016 00:59:42 +0000 (10:59 +1000)]
xfs: fix attr shortform structure alignment on cris

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: Darrick J. Wong <darrick.wong@oracle.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: in _attrlist_by_handle, copy the cursor back to userspace
Darrick J. Wong [Wed, 3 Aug 2016 00:58:53 +0000 (10:58 +1000)]
xfs: in _attrlist_by_handle, copy the cursor back to userspace

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: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoMerge branch 'xfs-4.8-misc-fixes-4' into for-next
Dave Chinner [Fri, 22 Jul 2016 04:10:56 +0000 (14:10 +1000)]
Merge branch 'xfs-4.8-misc-fixes-4' into for-next

7 years agoxfs: remove EXPERIMENTAL tag from sparse inode feature
Dave Chinner [Fri, 22 Jul 2016 04:10:18 +0000 (14:10 +1000)]
xfs: remove EXPERIMENTAL tag from sparse inode feature

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: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Eric Sandeen <sandeen@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: bufferhead chains are invalid after end_page_writeback
Dave Chinner [Thu, 21 Jul 2016 23:56:38 +0000 (09:56 +1000)]
xfs: bufferhead chains are invalid after end_page_writeback

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: Dave Chinner <dchinner@redhat.com>
Reported-and-Tested-by: Calvin Owens <calvinowens@fb.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: allocate log vector buffers outside CIL context lock
Dave Chinner [Thu, 21 Jul 2016 23:52:35 +0000 (09:52 +1000)]
xfs: allocate log vector buffers outside CIL context lock

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: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Brian Foster <bfoster@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agolibxfs: directory node splitting does not have an extra block
Dave Chinner [Thu, 21 Jul 2016 23:51:05 +0000 (09:51 +1000)]
libxfs: directory node splitting does not have an extra block

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: Dave Chinner <dchinner@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: remove dax code from object file when disabled
Arnd Bergmann [Thu, 21 Jul 2016 23:50:55 +0000 (09:50 +1000)]
xfs: remove dax code from object file when disabled

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: Arnd Bergmann <arnd@arndb.de>
Fixes: 16d4d43595b4 ("xfs: split direct I/O and DAX path")
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: skip dirty pages in ->releasepage()
Brian Foster [Thu, 21 Jul 2016 23:50:38 +0000 (09:50 +1000)]
xfs: skip dirty pages in ->releasepage()

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: Dave Chinner <david@fromorbit.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 agoMerge branch 'xfs-4.8-dir2-sf-fixes' into for-next
Dave Chinner [Wed, 20 Jul 2016 01:54:59 +0000 (11:54 +1000)]
Merge branch 'xfs-4.8-dir2-sf-fixes' into for-next

7 years agoMerge branch 'xfs-4.8-split-dax-dio' into for-next
Dave Chinner [Wed, 20 Jul 2016 01:54:37 +0000 (11:54 +1000)]
Merge branch 'xfs-4.8-split-dax-dio' into for-next

7 years agoMerge branch 'xfs-4.8-buf-fixes' into for-next
Dave Chinner [Wed, 20 Jul 2016 01:53:35 +0000 (11:53 +1000)]
Merge branch 'xfs-4.8-buf-fixes' into for-next

7 years agoMerge branch 'xfs-4.8-misc-fixes-3' into for-next
Dave Chinner [Wed, 20 Jul 2016 01:51:08 +0000 (11:51 +1000)]
Merge branch 'xfs-4.8-misc-fixes-3' into for-next

7 years agoxfs: remove __arch_pack
Christoph Hellwig [Wed, 20 Jul 2016 01:48:46 +0000 (11:48 +1000)]
xfs: remove __arch_pack

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: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: kill xfs_dir2_inou_t
Christoph Hellwig [Wed, 20 Jul 2016 01:48:31 +0000 (11:48 +1000)]
xfs: kill xfs_dir2_inou_t

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: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: kill xfs_dir2_sf_off_t
Christoph Hellwig [Wed, 20 Jul 2016 01:47:21 +0000 (11:47 +1000)]
xfs: kill xfs_dir2_sf_off_t

Just use an array of two unsigned chars directly to avoid problems
with architectures that pad the size of structures.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: split direct I/O and DAX path
Christoph Hellwig [Wed, 20 Jul 2016 01:38:55 +0000 (11:38 +1000)]
xfs: split direct I/O and DAX path

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: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: direct calls in the direct I/O path
Christoph Hellwig [Wed, 20 Jul 2016 01:38:01 +0000 (11:38 +1000)]
xfs: direct calls in the direct I/O path

We control both the callers and callees of ->direct_IO, so remove the
indirect calls.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: stop using generic_file_read_iter for direct I/O
Christoph Hellwig [Wed, 20 Jul 2016 01:36:57 +0000 (11:36 +1000)]
xfs: stop using generic_file_read_iter for direct I/O

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: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: split xfs_file_read_iter into buffered and direct I/O helpers
Christoph Hellwig [Wed, 20 Jul 2016 01:35:42 +0000 (11:35 +1000)]
xfs: split xfs_file_read_iter into buffered and direct I/O helpers

Similar to what we did on the write side a while ago.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: remove s_maxbytes enforcement in xfs_file_read_iter
Christoph Hellwig [Wed, 20 Jul 2016 01:31:53 +0000 (11:31 +1000)]
xfs: remove s_maxbytes enforcement in xfs_file_read_iter

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: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: kill ioflags
Christoph Hellwig [Wed, 20 Jul 2016 01:31:42 +0000 (11:31 +1000)]
xfs: kill ioflags

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: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: don't pass ioflags around in the ioctl path
Christoph Hellwig [Wed, 20 Jul 2016 01:29:35 +0000 (11:29 +1000)]
xfs: don't pass ioflags around in the ioctl path

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: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: track and serialize in-flight async buffers against unmount
Brian Foster [Wed, 20 Jul 2016 01:15:28 +0000 (11:15 +1000)]
xfs: track and serialize in-flight async buffers against unmount

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: Brian Foster <bfoster@redhat.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: exclude never-released buffers from buftarg I/O accounting
Brian Foster [Wed, 20 Jul 2016 01:13:43 +0000 (11:13 +1000)]
xfs: exclude never-released buffers from buftarg I/O accounting

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: 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 reset b_retries to 0 on every failure
Eric Sandeen [Wed, 20 Jul 2016 00:54:09 +0000 (10:54 +1000)]
xfs: don't reset b_retries to 0 on every failure

With the code as it stands today, b_retries never increments because
it gets reset to 0 in the error callback.

Remove that, and fix a similar problem where the first retry time
was constantly being overwritten, which defeated the timeout tunable
as well.  We now only set first retry time if a non-zero timeout is
set, to match the behavior of only incrementing retries if a retry
value is set.

This way max retries & timeouts consistently take effect after a
tunable is set, rather than acting retroactively on a buffer which
has failed at some point in the past and has accumulated state from
those prior failures.

Thanks to dchinner for talking through this with me.

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 agoxfs: remove extraneous buffer flag changes
Eric Sandeen [Wed, 20 Jul 2016 00:53:22 +0000 (10:53 +1000)]
xfs: remove extraneous buffer flag changes

Fix up a couple places where extra flag manipulation occurs.

In the first case we clear XBF_ASYNC and then immediately reset it -
so don't bother clearing in the first place.

In the 2nd case we are at a point in the function where the buffer
must already be async, so there is no need to reset it.

Add consistent spacing around the " | " while we're at it.

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 agoxfs: fix xfs_error_get_cfg for negative errnos
Eric Sandeen [Wed, 20 Jul 2016 00:48:51 +0000 (10:48 +1000)]
xfs: fix xfs_error_get_cfg for negative errnos

xfs_error_get_cfg() is called with bp->b_error as an arg, which is
negative, so the switch statement won't ever find any matches.

This results in only the default error handler having any effect, as
EIO/ENOSPC/ENODEV get ignored due to the wrong sign.

It seems simplest to always flip the error sign to positive, so that
we can handle either negative errors in bp->b_error, or possibly a
positive errno via something like xfs_error_get_cfg(EIO) - this
future-proofs the function.

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 agoxfs: remove the magic numbers in xfs_btree_block-related len macros
Hou Tao [Wed, 20 Jul 2016 00:43:11 +0000 (10:43 +1000)]
xfs: remove the magic numbers in xfs_btree_block-related len macros

replace the magic numbers by offsetof(...) and sizeof(...), and add two
extra checks on xfs_check_ondisk_structs()

[dchinner: renamed header structures to be more descriptive]

Signed-off-by: Hou Tao <houtao1@huawei.com>
Reviewed-by: Carlos Maiolino <cmaiolino@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: indentation fix in xfs_btree_get_iroot()
Kaho Ng [Wed, 20 Jul 2016 00:37:50 +0000 (10:37 +1000)]
xfs: indentation fix in xfs_btree_get_iroot()

The indentation in this function is different from the other functions.
Those spacebars are converted to tabs to improve readability.

Signed-off-by: Kaho Ng <ngkaho1234@gmail.com>
Reviewed-by: Carlos Maiolino <cmaiolino@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: don't allow negative error tags
Dan Carpenter [Wed, 20 Jul 2016 00:37:13 +0000 (10:37 +1000)]
xfs: don't allow negative error tags

Errors go from zero which means no error to XFS_ERRTAG_MAX (22).  My
static checker complains that xfs_errortag_add() puts an upper bound on
this but not a lower bound.  Let's fix it by making it unsigned.

Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: fix type confusion in xfs_ioc_swapext
Jann Horn [Wed, 20 Jul 2016 00:30:30 +0000 (10:30 +1000)]
xfs: fix type confusion in xfs_ioc_swapext

When calling fdget() in xfs_ioc_swapext(), we need to verify that
the file descriptors passed into the ioctl point to XFS inodes
before we start operations on them. If we don't do this, we could be
referencing arbitrary kernel memory as an XFS inode. THis could lead
to memory corruption and/or performing locking operations on
attacker-chosen structures in kernel memory.

[dchinner: rewrite commit message ]
[dchinner: add comment explaining new check ]

Signed-off-by: Jann Horn <jann@thejh.net>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoMerge branch 'xfs-4.8-misc-fixes-2' into for-next
Dave Chinner [Tue, 21 Jun 2016 01:55:13 +0000 (11:55 +1000)]
Merge branch 'xfs-4.8-misc-fixes-2' into for-next

7 years agoxfs: refactor btree maxlevels computation
Darrick J. Wong [Tue, 21 Jun 2016 01:53:28 +0000 (11:53 +1000)]
xfs: refactor btree maxlevels computation

Create a common function to calculate the maximum height of a per-AG
btree.  This will eventually be used by the rmapbt and refcountbt
code to calculate appropriate maxlevels values for each.  This is
important because the verifiers and the transaction block
reservations depend on accurate estimates of how many blocks are
needed to satisfy a btree split.

We were mistakenly using the max bnobt height for all the btrees,
which creates a dangerous situation since the larger records and
keys in an rmapbt make it very possible that the rmapbt will be
taller than the bnobt and so we can run out of transaction block
reservation.

Signed-off-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: convert list of extents to free into a regular list
Darrick J. Wong [Tue, 21 Jun 2016 01:53:28 +0000 (11:53 +1000)]
xfs: convert list of extents to free into a regular list

In struct xfs_bmap_free, convert the open-coded free extent list to
a regular list, then use list_sort to sort it prior to processing.

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: separate freelist fixing into a separate helper
Dave Chinner [Tue, 21 Jun 2016 01:53:28 +0000 (11:53 +1000)]
xfs: separate freelist fixing into a separate helper

Break up xfs_free_extent() into a helper that fixes the freelist.
This helper will be used subsequently to ensure the freelist during
deferred rmap processing.

[darrick: refactor to put this at the head of the patchset]

Signed-off-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: rearrange xfs_bmap_add_free parameters
Darrick J. Wong [Tue, 21 Jun 2016 01:53:28 +0000 (11:53 +1000)]
xfs: rearrange xfs_bmap_add_free parameters

This is already in xfsprogs' libxfs, so port it to the kernel.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
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: check for a valid error_tag in errortag_add
Darrick J. Wong [Tue, 21 Jun 2016 01:53:28 +0000 (11:53 +1000)]
xfs: check for a valid error_tag in errortag_add

Currently we don't check the error_tag when someone's trying to set up
error injection testing.  If userspace passes in a value we don't know
about, send back an error.  This will help xfstests to _notrun a test
that uses error injection to test things like log replay.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: enable buffer deadlock postmortem diagnosis via ftrace
Darrick J. Wong [Tue, 21 Jun 2016 01:53:28 +0000 (11:53 +1000)]
xfs: enable buffer deadlock postmortem diagnosis via ftrace

Create a second buf_trylock tracepoint so that we can distinguish
between a successful and a failed trylock.  With this piece, we can
use a script to look at the ftrace output to detect buffer deadlocks.

[dchinner: update to if/else as per hch's suggestion]

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: check offsets of variable length structures
Darrick J. Wong [Tue, 21 Jun 2016 01:53:28 +0000 (11:53 +1000)]
xfs: check offsets of variable length structures

Some of the directory/attr structures contain variable-length objects,
so the enclosing structure doesn't have a meaningful fixed size at
compile time.  We can check the offsets of the members before the
variable-length member, so do those.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: refactor xfs_reserve_blocks() to handle ENOSPC correctly
Brian Foster [Tue, 21 Jun 2016 01:53:28 +0000 (11:53 +1000)]
xfs: refactor xfs_reserve_blocks() to handle ENOSPC correctly

xfs_reserve_blocks() is responsible to update the XFS reserved block
pool count at mount time or based on user request. When the caller
requests to increase the reserve pool, blocks must be allocated from
the global counters such that they are no longer available for
general purpose use. If the requested reserve pool size is too
large, XFS reserves what blocks are available. The implementation
requires looking at the percpu counters and making an educated guess
as to how many blocks to try and allocate from xfs_mod_fdblocks(),
which can return -ENOSPC if the guess was not accurate due to
counters being modified in parallel.

xfs_reserve_blocks() retries the guess in this scenario until the
allocation succeeds or it is determined that there is no space
available in the fs. While not easily reproducible in the current
form, the retry code doesn't actually work correctly if
xfs_mod_fdblocks() actually fails. The problem is that the percpu
calculations use the m_resblks counter to determine how many blocks
to allocate, but unconditionally update m_resblks before the block
allocation has actually succeeded.  Therefore, if xfs_mod_fdblocks()
fails, the code jumps to the retry label and uses the already
updated m_resblks value to determine how many blocks to try and
allocate. If the percpu counters previously suggested that the
entire request was available, fdblocks_delta could end up set to 0.
In that case, m_resblks is updated to the requested value, yet no
blocks have been reserved at all.

Refactor xfs_reserve_blocks() to use an explicit loop and make the
code easier to follow. Since we have to drop the spinlock across the
xfs_mod_fdblocks() call, use a delta value for m_resblks as well and
only apply the delta once allocation succeeds.

[dchinner: convert to do {} while() loop]

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: cancel eofblocks background trimming on remount read-only
Brian Foster [Tue, 21 Jun 2016 01:53:28 +0000 (11:53 +1000)]
xfs: cancel eofblocks background trimming on remount read-only

The filesystem quiesce sequence performs the operations necessary to
drain all background work, push pending transactions through the log
infrastructure and wait on I/O resulting from the final AIL push. We
have had reports of remount,ro hangs in xfs_log_quiesce() ->
xfs_wait_buftarg(), however, and some instrumentation code to detect
transaction commits at this point in the quiesce sequence has inculpated
the eofblocks background scanner as a cause.

While higher level remount code generally prevents user modifications by
the time the filesystem has made it to xfs_log_quiesce(), the background
scanner may still be alive and can perform pending work at any time. If
this occurs between the xfs_log_force() and xfs_wait_buftarg() calls
within xfs_log_quiesce(), this can lead to an indefinite lockup in
xfs_wait_buftarg().

To prevent this problem, cancel the background eofblocks scan worker
during the remount read-only quiesce sequence. This suspends background
trimming when a filesystem is remounted read-only. This is only done in
the remount path because the freeze codepath has already locked out new
transactions by the time the filesystem attempts to quiesce (and thus
waiting on an active work item could deadlock). Kick the eofblocks
worker to pick up where it left off once an fs is remounted back to
read-write.

Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Eric Sandeen <sandeen@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoMerge branch 'xfs-4.8-iomap-write' into for-next
Dave Chinner [Tue, 21 Jun 2016 00:10:38 +0000 (10:10 +1000)]
Merge branch 'xfs-4.8-iomap-write' into for-next

7 years agoMerge branch 'fs-4.8-iomap-infrastructure' into for-next
Dave Chinner [Tue, 21 Jun 2016 00:10:19 +0000 (10:10 +1000)]
Merge branch 'fs-4.8-iomap-infrastructure' into for-next

7 years agoxfs: kill xfs_zero_remaining_bytes
Christoph Hellwig [Tue, 21 Jun 2016 00:02:23 +0000 (10:02 +1000)]
xfs: kill xfs_zero_remaining_bytes

Instead punch the whole first, and the use the our zeroing helper
to punch out the edge blocks.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: split xfs_free_file_space in manageable pieces
Christoph Hellwig [Tue, 21 Jun 2016 00:00:55 +0000 (10:00 +1000)]
xfs: split xfs_free_file_space in manageable pieces

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: use xfs_zero_range in xfs_zero_eof
Christoph Hellwig [Mon, 20 Jun 2016 23:57:26 +0000 (09:57 +1000)]
xfs: use xfs_zero_range in xfs_zero_eof

We now skip holes in it, so no need to have the caller do it as well.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: handle 64-bit length in xfs_iozero
Christoph Hellwig [Mon, 20 Jun 2016 23:56:26 +0000 (09:56 +1000)]
xfs: handle 64-bit length in xfs_iozero

We'll want to use this code for large offsets now that we're
skipping holes and unwritten extents efficiently.  Also rename it to
xfs_zero_range to be a bit more descriptive, and tell the caller if
we actually did any zeroing.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: use iomap infrastructure for DAX zeroing
Christoph Hellwig [Mon, 20 Jun 2016 23:55:18 +0000 (09:55 +1000)]
xfs: use iomap infrastructure for DAX zeroing

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: use iomap fiemap implementation
Christoph Hellwig [Mon, 20 Jun 2016 23:54:53 +0000 (09:54 +1000)]
xfs: use iomap fiemap implementation

Note that this removes support for the untested FIEMAP_FLAG_XATTR.  It
could be added relatively easily with iomap ops for the attr fork, but
without test coverage I don't feel safe doing this.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: remove buffered write support from __xfs_get_blocks
Christoph Hellwig [Mon, 20 Jun 2016 23:53:45 +0000 (09:53 +1000)]
xfs: remove buffered write support from __xfs_get_blocks

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: implement iomap based buffered write path
Christoph Hellwig [Mon, 20 Jun 2016 23:53:44 +0000 (09:53 +1000)]
xfs: implement iomap based buffered write path

Convert XFS to use the new iomap based multipage write path. This involves
implementing the ->iomap_begin and ->iomap_end methods, and switching the
buffered file write, page_mkwrite and xfs_iozero paths to the new iomap
helpers.

With this change __xfs_get_blocks will never be used for buffered writes,
and the code handling them can be removed.

Based on earlier code from Dave Chinner.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: reorder zeroing and flushing sequence in truncate
Christoph Hellwig [Mon, 20 Jun 2016 23:52:47 +0000 (09:52 +1000)]
xfs: reorder zeroing and flushing sequence in truncate

Currently zeroing out blocks and waiting for writeout is a bit of a mess in
truncate.  This patch gives it a clear order in preparation for the iomap
path:

 (1) we first wait for any direct I/O to complete to prevent any races
     for it
 (2) we then perform the actual zeroing, and only use the truncate_page
     helpers for truncating down.  The truncate up case already is
     handled by the separate call to xfs_zero_eof.
 (3) only then we write back dirty data, as zeroing block may cause
     dirty pages when using either xfs_zero_eof or the new iomap
     infrastructure.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agoxfs: make xfs_bmbt_to_iomap available outside of xfs_pnfs.c
Christoph Hellwig [Mon, 20 Jun 2016 23:52:47 +0000 (09:52 +1000)]
xfs: make xfs_bmbt_to_iomap available outside of xfs_pnfs.c

And ensure it works for RT subvolume files an set the block device,
both of which will be needed to be able to use the function in the
buffered write path.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agofs: iomap based fiemap implementation
Christoph Hellwig [Mon, 20 Jun 2016 23:38:45 +0000 (09:38 +1000)]
fs: iomap based fiemap implementation

Add a simple fiemap implementation based on iomap_ops, partially based
on a previous implementation from Bob Peterson <rpeterso@redhat.com>.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agofs: support DAX based iomap zeroing
Christoph Hellwig [Mon, 20 Jun 2016 23:31:39 +0000 (09:31 +1000)]
fs: support DAX based iomap zeroing

This avoid needing a separate inefficient get_block based DAX zero_range
implementation in file systems.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agofs: introduce iomap infrastructure
Christoph Hellwig [Mon, 20 Jun 2016 23:23:11 +0000 (09:23 +1000)]
fs: introduce iomap infrastructure

Add infrastructure for multipage buffered writes.  This is implemented
using an main iterator that applies an actor function to a range that
can be written.

This infrastucture is used to implement a buffered write helper, one
to zero file ranges and one to implement the ->page_mkwrite VM
operations.  All of them borrow a fair amount of code from fs/buffers.
for now by using an internal version of __block_write_begin that
gets passed an iomap and builds the corresponding buffer head.

The file system is gets a set of paired ->iomap_begin and ->iomap_end
calls which allow it to map/reserve a range and get a notification
once the write code is finished with it.

Based on earlier code from Dave Chinner.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
7 years agofs: move struct iomap from exportfs.h to a separate header
Christoph Hellwig [Mon, 20 Jun 2016 23:22:39 +0000 (09:22 +1000)]
fs: move struct iomap from exportfs.h to a separate header

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
8 years agoxfs: reduce lock hold times in buffer writeback
Dave Chinner [Wed, 1 Jun 2016 07:38:15 +0000 (17:38 +1000)]
xfs: reduce lock hold times in buffer writeback

When we have a lot of metadata to flush from the AIL, the buffer
list can get very long. The current submission code tries to batch
submission to optimise IO order of the metadata (i.e. ascending
block order) to maximise block layer merging or IO to adjacent
metadata blocks.

Unfortunately, the method used can result in long lock times
occurring as buffers locked early on in the buffer list might not be
dispatched until the end of the IO licst processing. This is because
sorting does not occur util after the buffer list has been processed
and the buffers that are going to be submitted are locked. Hence
when the buffer list is several thousand buffers long, the lock hold
times before IO dispatch can be significant.

To fix this, sort the buffer list before we start trying to lock and
submit buffers. This means we can now submit buffers immediately
after they are locked, allowing merging to occur immediately on the
plug and dispatch to occur as quickly as possible. This means there
is minimal delay between locking the buffer and IO submission
occuring, hence reducing the worst case lock hold times seen during
delayed write buffer IO submission signficantly.

Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Carlos Maiolino <cmaiolino@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
8 years agoxfs: define XFS_IOC_FREEZE even if FIFREEZE is defined
Christoph Hellwig [Wed, 1 Jun 2016 07:38:15 +0000 (17:38 +1000)]
xfs: define XFS_IOC_FREEZE even if FIFREEZE is defined

And the same for XFS_IOC_THAW.  Just because we now have a common
version of the ioctl we still need to provide the old name for it
for anyone using those.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Eric Sandeen <sandeen@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
8 years agoxfs: make several functions static
Eric Sandeen [Wed, 1 Jun 2016 07:38:15 +0000 (17:38 +1000)]
xfs: make several functions static

Al Viro noticed that xfs_lock_inodes should be static, and
that led to ... a few more.

These are just the easy ones, others require moving functions
higher in source files, so that's not done here to keep
this review simple.

Signed-off-by: Eric Sandeen <sandeen@sandeen.net>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Dave Chinner <david@fromorbit.com>
8 years agoxfs: remove spurious shutdown type check from xfs_bmap_finish()
Brian Foster [Wed, 1 Jun 2016 07:38:15 +0000 (17:38 +1000)]
xfs: remove spurious shutdown type check from xfs_bmap_finish()

The static checker reports that after commit 8d99fe92fed0 ("xfs: fix
efi/efd error handling to avoid fs shutdown hangs"), the code has been
reworked such that error == -EFSCORRUPTED is not possible in this
codepath.

Remove the spurious error check and just use SHUTDOWN_META_IO_ERROR
unconditionally.

Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Carlos Maiolino <cmaiolino@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Dave Chinner <david@fromorbit.com>
8 years agoxfs: fix broken multi-fsb buffer logging
Brian Foster [Wed, 1 Jun 2016 07:38:12 +0000 (17:38 +1000)]
xfs: fix broken multi-fsb buffer logging

Multi-block buffers are logged based on buffer offset in
xfs_trans_log_buf(). xfs_buf_item_log() ultimately walks each mapping in
the buffer and marks the associated range to be logged in the
xfs_buf_log_format bitmap for that mapping. This code is broken,
however, in that it marks the actual buffer offsets of the associated
range in each bitmap rather than shifting to the byte range for that
particular mapping.

For example, on a 4k fsb fs, buffer offset 4096 refers to the first byte
of the second mapping in the buffer. This means byte 0 of the second log
format bitmap should be tagged as dirty. Instead, the current code marks
byte offset 4096 of the second log format bitmap, which is invalid and
potentially out of range of the mapping.

As a result of this, the log item format code invoked at transaction
commit time is not be able to correctly identify what parts of the
buffer to copy into log vectors. This can lead to NULL log vector
pointer dereferences in CIL push context if the item format code was not
able to locate any dirty ranges at all. This crash has been reproduced
on a 4k FSB filesystem using 16k directory blocks where an unlink
operation happened not to log anything in the first block of the
mapping. The logged offsets were all over 4k, marked as such in the
subsequent log format mappings, and thus left the transaction with an
xfs_log_item that is marked DIRTY but without any logged regions.

Further, even when the logged regions are marked correctly in the buffer
log format bitmaps, the format code doesn't copy the correct ranges of
the buffer into the log. This means that any logged region beyond the
first block of a multi-block buffer is subject to corruption after a
crash and log recovery sequence. This is due to a failure to convert the
mapping bm_len field from basic blocks to bytes in the buffer offset
tracking code in xfs_buf_item_format().

Update xfs_buf_item_log() to convert buffer offsets to segment relative
offsets when logging multi-block buffers. This ensures that the modified
regions of a buffer are logged correctly and avoids the aforementioned
crash. Also update xfs_buf_item_format() to correctly track the source
offset into the buffer for the log vector formatting code. This ensures
that the correct data is copied into the log.

Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Eric Sandeen <sandeen@redhat.com>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Dave Chinner <david@fromorbit.com>
8 years agoLinux 4.7-rc1 v4.7-rc1
Linus Torvalds [Sun, 29 May 2016 16:29:24 +0000 (09:29 -0700)]
Linux 4.7-rc1

8 years agohash_string: Fix zero-length case for !DCACHE_WORD_ACCESS
George Spelvin [Sun, 29 May 2016 12:05:56 +0000 (08:05 -0400)]
hash_string: Fix zero-length case for !DCACHE_WORD_ACCESS

The self-test was updated to cover zero-length strings; the function
needs to be updated, too.

Reported-by: Geert Uytterhoeven <geert@linux-m68k.org>
Signed-off-by: George Spelvin <linux@sciencehorizons.net>
Fixes: fcfd2fbf22d2 ("fs/namei.c: Add hashlen_string() function")
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agoRename other copy of hash_string to hashlen_string
George Spelvin [Sun, 29 May 2016 05:26:41 +0000 (01:26 -0400)]
Rename other copy of hash_string to hashlen_string

The original name was simply hash_string(), but that conflicted with a
function with that name in drivers/base/power/trace.c, and I decided
that calling it "hashlen_" was better anyway.

But you have to do it in two places.

[ This caused build errors for architectures that don't define
  CONFIG_DCACHE_WORD_ACCESS   - Linus ]

Signed-off-by: George Spelvin <linux@sciencehorizons.net>
Reported-by: Guenter Roeck <linux@roeck-us.net>
Fixes: fcfd2fbf22d2 ("fs/namei.c: Add hashlen_string() function")
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agohpfs: implement the show_options method
Mikulas Patocka [Tue, 24 May 2016 20:49:18 +0000 (22:49 +0200)]
hpfs: implement the show_options method

The HPFS filesystem used generic_show_options to produce string that is
displayed in /proc/mounts.  However, there is a problem that the options
may disappear after remount.  If we mount the filesystem with option1
and then remount it with option2, /proc/mounts should show both option1
and option2, however it only shows option2 because the whole option
string is replaced with replace_mount_options in hpfs_remount_fs.

To fix this bug, implement the hpfs_show_options function that prints
options that are currently selected.

Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
Cc: stable@vger.kernel.org
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agoaffs: fix remount failure when there are no options changed
Mikulas Patocka [Tue, 24 May 2016 20:48:33 +0000 (22:48 +0200)]
affs: fix remount failure when there are no options changed

Commit c8f33d0bec99 ("affs: kstrdup() memory handling") checks if the
kstrdup function returns NULL due to out-of-memory condition.

However, if we are remounting a filesystem with no change to
filesystem-specific options, the parameter data is NULL.  In this case,
kstrdup returns NULL (because it was passed NULL parameter), although no
out of memory condition exists.  The mount syscall then fails with
ENOMEM.

This patch fixes the bug.  We fail with ENOMEM only if data is non-NULL.

The patch also changes the call to replace_mount_options - if we didn't
pass any filesystem-specific options, we don't call
replace_mount_options (thus we don't erase existing reported options).

Fixes: c8f33d0bec99 ("affs: kstrdup() memory handling")
Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
Cc: stable@vger.kernel.org # v4.1+
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agohpfs: fix remount failure when there are no options changed
Mikulas Patocka [Tue, 24 May 2016 20:47:00 +0000 (22:47 +0200)]
hpfs: fix remount failure when there are no options changed

Commit ce657611baf9 ("hpfs: kstrdup() out of memory handling") checks if
the kstrdup function returns NULL due to out-of-memory condition.

However, if we are remounting a filesystem with no change to
filesystem-specific options, the parameter data is NULL.  In this case,
kstrdup returns NULL (because it was passed NULL parameter), although no
out of memory condition exists.  The mount syscall then fails with
ENOMEM.

This patch fixes the bug.  We fail with ENOMEM only if data is non-NULL.

The patch also changes the call to replace_mount_options - if we didn't
pass any filesystem-specific options, we don't call
replace_mount_options (thus we don't erase existing reported options).

Fixes: ce657611baf9 ("hpfs: kstrdup() out of memory handling")
Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
Cc: stable@vger.kernel.org
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agoMerge branch 'upstream' of git://git.linux-mips.org/pub/scm/ralf/upstream-linus
Linus Torvalds [Sat, 28 May 2016 23:41:39 +0000 (16:41 -0700)]
Merge branch 'upstream' of git://git.linux-mips.org/ralf/upstream-linus

Pull more MIPS updates from Ralf Baechle:
 "This is the secondnd batch of MIPS patches for 4.7. Summary:

  CPS:
   - Copy EVA configuration when starting secondary VPs.

  EIC:
   - Clear Status IPL.

  Lasat:
   - Fix a few off by one bugs.

  lib:
   - Mark intrinsics notrace.  Not only are the intrinsics
     uninteresting, it would cause infinite recursion.

  MAINTAINERS:
   - Add file patterns for MIPS BRCM device tree bindings.
   - Add file patterns for mips device tree bindings.

  MT7628:
   - Fix MT7628 pinmux typos.
   - wled_an pinmux gpio.
   - EPHY LEDs pinmux support.

  Pistachio:
   - Enable KASLR

  VDSO:
   - Build microMIPS VDSO for microMIPS kernels.
   - Fix aliasing warning by building with `-fno-strict-aliasing' for
     debugging but also tracing them might result in recursion.

  Misc:
   - Add missing FROZEN hotplug notifier transitions.
   - Fix clk binding example for varioius PIC32 devices.
   - Fix cpu interrupt controller node-names in the DT files.
   - Fix XPA CPU feature separation.
   - Fix write_gc0_* macros when writing zero.
   - Add inline asm encoding helpers.
   - Add missing VZ accessor microMIPS encodings.
   - Fix little endian microMIPS MSA encodings.
   - Add 64-bit HTW fields and fix its configuration.
   - Fix sigreturn via VDSO on microMIPS kernel.
   - Lots of typo fixes.
   - Add definitions of SegCtl registers and use them"

* 'upstream' of git://git.linux-mips.org/pub/scm/ralf/upstream-linus: (49 commits)
  MIPS: Add missing FROZEN hotplug notifier transitions
  MIPS: Build microMIPS VDSO for microMIPS kernels
  MIPS: Fix sigreturn via VDSO on microMIPS kernel
  MIPS: devicetree: fix cpu interrupt controller node-names
  MIPS: VDSO: Build with `-fno-strict-aliasing'
  MIPS: Pistachio: Enable KASLR
  MIPS: lib: Mark intrinsics notrace
  MIPS: Fix 64-bit HTW configuration
  MIPS: Add 64-bit HTW fields
  MAINTAINERS: Add file patterns for mips device tree bindings
  MAINTAINERS: Add file patterns for mips brcm device tree bindings
  MIPS: Simplify DSP instruction encoding macros
  MIPS: Add missing tlbinvf/XPA microMIPS encodings
  MIPS: Fix little endian microMIPS MSA encodings
  MIPS: Add missing VZ accessor microMIPS encodings
  MIPS: Add inline asm encoding helpers
  MIPS: Spelling fix lets -> let's
  MIPS: VR41xx: Fix typo
  MIPS: oprofile: Fix typo
  MIPS: math-emu: Fix typo
  ...

8 years agofs: fix binfmt_aout.c build error
Guenter Roeck [Sat, 28 May 2016 22:26:02 +0000 (15:26 -0700)]
fs: fix binfmt_aout.c build error

Various builds (such as i386:allmodconfig) fail with

  fs/binfmt_aout.c:133:2: error: expected identifier or '(' before 'return'
  fs/binfmt_aout.c:134:1: error: expected identifier or '(' before '}' token

[ Oops. My bad, I had stupidly thought that "allmodconfig" covered this
  on x86-64 too, but it obviously doesn't.  Egg on my face.  - Linus ]

Fixes: 5d22fc25d4fc ("mm: remove more IS_ERR_VALUE abuses")
Signed-off-by: Guenter Roeck <linux@roeck-us.net>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
8 years agoMerge branch 'hash' of git://ftp.sciencehorizons.net/linux
Linus Torvalds [Sat, 28 May 2016 23:15:25 +0000 (16:15 -0700)]
Merge branch 'hash' of git://ftp.sciencehorizons.net/linux

Pull string hash improvements from George Spelvin:
 "This series does several related things:

   - Makes the dcache hash (fs/namei.c) useful for general kernel use.

     (Thanks to Bruce for noticing the zero-length corner case)

   - Converts the string hashes in <linux/sunrpc/svcauth.h> to use the
     above.

   - Avoids 64-bit multiplies in hash_64() on 32-bit platforms.  Two
     32-bit multiplies will do well enough.

   - Rids the world of the bad hash multipliers in hash_32.

     This finishes the job started in commit 689de1d6ca95 ("Minimal
     fix-up of bad hashing behavior of hash_64()")

     The vast majority of Linux architectures have hardware support for
     32x32-bit multiply and so derive no benefit from "simplified"
     multipliers.

     The few processors that do not (68000, h8/300 and some models of
     Microblaze) have arch-specific implementations added.  Those
     patches are last in the series.

   - Overhauls the dcache hash mixing.

     The patch in commit 0fed3ac866ea ("namei: Improve hash mixing if
     CONFIG_DCACHE_WORD_ACCESS") was an off-the-cuff suggestion.
     Replaced with a much more careful design that's simultaneously
     faster and better.  (My own invention, as there was noting suitable
     in the literature I could find.  Comments welcome!)

   - Modify the hash_name() loop to skip the initial HASH_MIX().  This
     would let us salt the hash if we ever wanted to.

   - Sort out partial_name_hash().

     The hash function is declared as using a long state, even though
     it's truncated to 32 bits at the end and the extra internal state
     contributes nothing to the result.  And some callers do odd things:

      - fs/hfs/string.c only allocates 32 bits of state
      - fs/hfsplus/unicode.c uses it to hash 16-bit unicode symbols not bytes

   - Modify bytemask_from_count to handle inputs of 1..sizeof(long)
     rather than 0..sizeof(long)-1.  This would simplify users other
     than full_name_hash"

  Special thanks to Bruce Fields for testing and finding bugs in v1.  (I
  learned some humbling lessons about "obviously correct" code.)

  On the arch-specific front, the m68k assembly has been tested in a
  standalone test harness, I've been in contact with the Microblaze
  maintainers who mostly don't care, as the hardware multiplier is never
  omitted in real-world applications, and I haven't heard anything from
  the H8/300 world"

* 'hash' of git://ftp.sciencehorizons.net/linux:
  h8300: Add <asm/hash.h>
  microblaze: Add <asm/hash.h>
  m68k: Add <asm/hash.h>
  <linux/hash.h>: Add support for architecture-specific functions
  fs/namei.c: Improve dcache hash function
  Eliminate bad hash multipliers from hash_32() and  hash_64()
  Change hash_64() return value to 32 bits
  <linux/sunrpc/svcauth.h>: Define hash_str() in terms of hashlen_string()
  fs/namei.c: Add hashlen_string() function
  Pull out string hash to <linux/stringhash.h>

8 years agoh8300: Add <asm/hash.h>
George Spelvin [Wed, 25 May 2016 18:19:49 +0000 (14:19 -0400)]
h8300: Add <asm/hash.h>

This will improve the performance of hash_32() and hash_64(), but due
to complete lack of multi-bit shift instructions on H8, performance will
still be bad in surrounding code.

Designing H8-specific hash algorithms to work around that is a separate
project.  (But if the maintainers would like to get in touch...)

Signed-off-by: George Spelvin <linux@sciencehorizons.net>
Cc: Yoshinori Sato <ysato@users.sourceforge.jp>
Cc: uclinux-h8-devel@lists.sourceforge.jp
8 years agomicroblaze: Add <asm/hash.h>
George Spelvin [Wed, 25 May 2016 15:06:09 +0000 (11:06 -0400)]
microblaze: Add <asm/hash.h>

Microblaze is an FPGA soft core that can be configured various ways.

If it is configured without a multiplier, the standard __hash_32()
will require a call to __mulsi3, which is a slow software loop.

Instead, use a shift-and-add sequence for the constant multiply.
GCC knows how to do this, but it's not as clever as some.

Signed-off-by: George Spelvin <linux@sciencehorizons.net>
Cc: Alistair Francis <alistair.francis@xilinx.com>
Cc: Michal Simek <michal.simek@xilinx.com>
8 years agom68k: Add <asm/hash.h>
George Spelvin [Thu, 26 May 2016 15:36:19 +0000 (11:36 -0400)]
m68k: Add <asm/hash.h>

This provides a multiply by constant GOLDEN_RATIO_32 = 0x61C88647
for the original mc68000, which lacks a 32x32-bit multiply instruction.

Yes, the amount of optimization effort put in is excessive. :-)

Shift-add chain found by Yevgen Voronenko's Hcub algorithm at
http://spiral.ece.cmu.edu/mcm/gen.html

Signed-off-by: George Spelvin <linux@sciencehorizons.net>
Cc: Geert Uytterhoeven <geert@linux-m68k.org>
Cc: Greg Ungerer <gerg@linux-m68k.org>
Cc: Andreas Schwab <schwab@linux-m68k.org>
Cc: Philippe De Muyter <phdm@macq.eu>
Cc: linux-m68k@lists.linux-m68k.org
8 years ago<linux/hash.h>: Add support for architecture-specific functions
George Spelvin [Fri, 27 May 2016 02:11:51 +0000 (22:11 -0400)]
<linux/hash.h>: Add support for architecture-specific functions

This is just the infrastructure; there are no users yet.

This is modelled on CONFIG_ARCH_RANDOM; a CONFIG_ symbol declares
the existence of <asm/hash.h>.

That file may define its own versions of various functions, and define
HAVE_* symbols (no CONFIG_ prefix!) to suppress the generic ones.

Included is a self-test (in lib/test_hash.c) that verifies the basics.
It is NOT in general required that the arch-specific functions compute
the same thing as the generic, but if a HAVE_* symbol is defined with
the value 1, then equality is tested.

Signed-off-by: George Spelvin <linux@sciencehorizons.net>
Cc: Geert Uytterhoeven <geert@linux-m68k.org>
Cc: Greg Ungerer <gerg@linux-m68k.org>
Cc: Andreas Schwab <schwab@linux-m68k.org>
Cc: Philippe De Muyter <phdm@macq.eu>
Cc: linux-m68k@lists.linux-m68k.org
Cc: Alistair Francis <alistai@xilinx.com>
Cc: Michal Simek <michal.simek@xilinx.com>
Cc: Yoshinori Sato <ysato@users.sourceforge.jp>
Cc: uclinux-h8-devel@lists.sourceforge.jp
8 years agofs/namei.c: Improve dcache hash function
George Spelvin [Mon, 23 May 2016 11:43:58 +0000 (07:43 -0400)]
fs/namei.c: Improve dcache hash function

Patch 0fed3ac866 improved the hash mixing, but the function is slower
than necessary; there's a 7-instruction dependency chain (10 on x86)
each loop iteration.

Word-at-a-time access is a very tight loop (which is good, because
link_path_walk() is one of the hottest code paths in the entire kernel),
and the hash mixing function must not have a longer latency to avoid
slowing it down.

There do not appear to be any published fast hash functions that:
1) Operate on the input a word at a time, and
2) Don't need to know the length of the input beforehand, and
3) Have a single iterated mixing function, not needing conditional
   branches or unrolling to distinguish different loop iterations.

One of the algorithms which comes closest is Yann Collet's xxHash, but
that's two dependent multiplies per word, which is too much.

The key insights in this design are:

1) Barring expensive ops like multiplies, to diffuse one input bit
   across 64 bits of hash state takes at least log2(64) = 6 sequentially
   dependent instructions.  That is more cycles than we'd like.
2) An operation like "hash ^= hash << 13" requires a second temporary
   register anyway, and on a 2-operand machine like x86, it's three
   instructions.
3) A better use of a second register is to hold a two-word hash state.
   With careful design, no temporaries are needed at all, so it doesn't
   increase register pressure.  And this gets rid of register copying
   on 2-operand machines, so the code is smaller and faster.
4) Using two words of state weakens the requirement for one-round mixing;
   we now have two rounds of mixing before cancellation is possible.
5) A two-word hash state also allows operations on both halves to be
   done in parallel, so on a superscalar processor we get more mixing
   in fewer cycles.

I ended up using a mixing function inspired by the ChaCha and Speck
round functions.  It is 6 simple instructions and 3 cycles per iteration
(assuming multiply by 9 can be done by an "lea" instruction):

x ^= *input++;
y ^= x; x = ROL(x, K1);
x += y; y = ROL(y, K2);
y *= 9;

Not only is this reversible, two consecutive rounds are reversible:
if you are given the initial and final states, but not the intermediate
state, it is possible to compute both input words.  This means that at
least 3 words of input are required to create a collision.

(It also has the property, used by hash_name() to avoid a branch, that
it hashes all-zero to all-zero.)

The rotate constants K1 and K2 were found by experiment.  The search took
a sample of random initial states (I used 1023) and considered the effect
of flipping each of the 64 input bits on each of the 128 output bits two
rounds later.  Each of the 8192 pairs can be considered a biased coin, and
adding up the Shannon entropy of all of them produces a score.

The best-scoring shifts also did well in other tests (flipping bits in y,
trying 3 or 4 rounds of mixing, flipping all 64*63/2 pairs of input bits),
so the choice was made with the additional constraint that the sum of the
shifts is odd and not too close to the word size.

The final state is then folded into a 32-bit hash value by a less carefully
optimized multiply-based scheme.  This also has to be fast, as pathname
components tend to be short (the most common case is one iteration!), but
there's some room for latency, as there is a fair bit of intervening logic
before the hash value is used for anything.

(Performance verified with "bonnie++ -s 0 -n 1536:-2" on tmpfs.  I need
a better benchmark; the numbers seem to show a slight dip in performance
between 4.6.0 and this patch, but they're too noisy to quote.)

Special thanks to Bruce fields for diligent testing which uncovered a
nasty fencepost error in an earlier version of this patch.

[checkpatch.pl formatting complaints noted and respectfully disagreed with.]

Signed-off-by: George Spelvin <linux@sciencehorizons.net>
Tested-by: J. Bruce Fields <bfields@redhat.com>
8 years agoEliminate bad hash multipliers from hash_32() and hash_64()
George Spelvin [Fri, 27 May 2016 03:00:23 +0000 (23:00 -0400)]
Eliminate bad hash multipliers from hash_32() and  hash_64()

The "simplified" prime multipliers made very bad hash functions, so get rid
of them.  This completes the work of 689de1d6ca.

To avoid the inefficiency which was the motivation for the "simplified"
multipliers, hash_64() on 32-bit systems is changed to use a different
algorithm.  It makes two calls to hash_32() instead.

drivers/media/usb/dvb-usb-v2/af9015.c uses the old GOLDEN_RATIO_PRIME_32
for some horrible reason, so it inherits a copy of the old definition.

Signed-off-by: George Spelvin <linux@sciencehorizons.net>
Cc: Antti Palosaari <crope@iki.fi>
Cc: Mauro Carvalho Chehab <m.chehab@samsung.com>
8 years agoChange hash_64() return value to 32 bits
George Spelvin [Fri, 27 May 2016 02:22:01 +0000 (22:22 -0400)]
Change hash_64() return value to 32 bits

That's all that's ever asked for, and it makes the return
type of hash_long() consistent.

It also allows (upcoming patch) an optimized implementation
of hash_64 on 32-bit machines.

I tried adding a BUILD_BUG_ON to ensure the number of bits requested
was never more than 32 (most callers use a compile-time constant), but
adding <linux/bug.h> to <linux/hash.h> breaks the tools/perf compiler
unless tools/perf/MANIFEST is updated, and understanding that code base
well enough to update it is too much trouble.  I did the rest of an
allyesconfig build with such a check, and nothing tripped.

Signed-off-by: George Spelvin <linux@sciencehorizons.net>
8 years ago<linux/sunrpc/svcauth.h>: Define hash_str() in terms of hashlen_string()
George Spelvin [Fri, 20 May 2016 17:31:33 +0000 (13:31 -0400)]
<linux/sunrpc/svcauth.h>: Define hash_str() in terms of hashlen_string()

Finally, the first use of previous two patches: eliminate the
separate ad-hoc string hash functions in the sunrpc code.

Now hash_str() is a wrapper around hash_string(), and hash_mem() is
likewise a wrapper around full_name_hash().

Note that sunrpc code *does* call hash_mem() with a zero length, which
is why the previous patch needed to handle that in full_name_hash().
(Thanks, Bruce, for finding that!)

This also eliminates the only caller of hash_long which asks for
more than 32 bits of output.

The comment about the quality of hashlen_string() and full_name_hash()
is jumping the gun by a few patches; they aren't very impressive now,
but will be improved greatly later in the series.

Signed-off-by: George Spelvin <linux@sciencehorizons.net>
Tested-by: J. Bruce Fields <bfields@redhat.com>
Acked-by: J. Bruce Fields <bfields@redhat.com>
Cc: Jeff Layton <jlayton@poochiereds.net>
Cc: linux-nfs@vger.kernel.org
8 years agofs/namei.c: Add hashlen_string() function
George Spelvin [Fri, 20 May 2016 12:41:37 +0000 (08:41 -0400)]
fs/namei.c: Add hashlen_string() function

We'd like to make more use of the highly-optimized dcache hash functions
throughout the kernel, rather than have every subsystem create its own,
and a function that hashes basic null-terminated strings is required
for that.

(The name is to emphasize that it returns both hash and length.)

It's actually useful in the dcache itself, specifically d_alloc_name().
Other uses in the next patch.

full_name_hash() is also tweaked to make it more generally useful:
1) Take a "char *" rather than "unsigned char *" argument, to
   be consistent with hash_name().
2) Handle zero-length inputs.  If we want more callers, we don't want
   to make them worry about corner cases.

Signed-off-by: George Spelvin <linux@sciencehorizons.net>
8 years agoPull out string hash to <linux/stringhash.h>
George Spelvin [Fri, 20 May 2016 11:26:00 +0000 (07:26 -0400)]
Pull out string hash to <linux/stringhash.h>

... so they can be used without the rest of <linux/dcache.h>

The hashlen_* macros will make sense next patch.

Signed-off-by: George Spelvin <linux@sciencehorizons.net>
8 years agoMerge branch 'i2c/for-next' of git://git.kernel.org/pub/scm/linux/kernel/git/wsa...
Linus Torvalds [Sat, 28 May 2016 19:38:50 +0000 (12:38 -0700)]
Merge branch 'i2c/for-next' of git://git./linux/kernel/git/wsa/linux

Pull i2c fix from Wolfram Sang:
 "A fix for a regression introduced yesterday.

  The regression didn't show up here locally because I did not have
  PAGE_POISONING enabled.  And buildbots discovered this only after it
  hit your tree.  Thanks to Dan for the quick response"

* 'i2c/for-next' of git://git.kernel.org/pub/scm/linux/kernel/git/wsa/linux:
  i2c: dev: use after free in detach