Originally-From: Dave Chinner <dchinner@redhat.com>

Now we have all the surrounding call infrastructure in place, we can
start filling out the rmap btree implementation. Start with the
on-disk btree format; add everything needed to read, write and
manipulate rmap btree blocks. This prepares the way for adding the
btree operations implementation.

[darrick: record owner and offset info in rmap btree]
[darrick: fork, bmbt and unwritten state in rmap btree]
[darrick: flags are a separate field in xfs_rmap_irec]
[darrick: calculate maxlevels separately]
[darrick: move the 'unwritten' bit into unused parts of rm_offset]
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Signed-off-by: NDarrick J. Wong <darrick.wong@oracle.com>
Reviewed-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Originally-From: Dave Chinner <dchinner@redhat.com>

XFS reserves a small amount of space in each AG for the minimum
number of free blocks needed for operation. Adding the rmap btree
increases the number of reserved blocks, but it also increases the
complexity of the calculation as the free inode btree is optional
(like the rmbt).

Rather than calculate the prealloc blocks every time we need to
check it, add a function to calculate it at mount time and store it
in the struct xfs_mount, and convert the XFS_PREALLOC_BLOCKS macro
just to use the xfs-mount variable directly.
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

If we take "retry forever" literally on metadata IO errors, we can
hang at unmount, once it retries those writes forever. This is the
default behavior, unfortunately.

Add an error configuration option for this behavior and default it
to "fail" so that an unmount will trigger actuall errors, a shutdown
and allow the unmount to succeed. It will be noisy, though, as it
will log the errors and shutdown that occurs.

To fix this, we need to mark the filesystem as being in the process
of unmounting. Do this with a mount flag that is added at the
appropriate time (i.e. before the blocking AIL sync). We also need
to add this flag if mount fails after the initial phase of log
recovery has been run.
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Signed-off-by: NCarlos Maiolino <cmaiolino@redhat.com>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

now most of the infrastructure is in place, we can start adding
support for configuring specific errors such as ENODEV, ENOSPC, EIO,
etc. Add these error configurations and configure them all to have
appropriate behaviours. That is, all will be configured to retry
forever by default, except for ENODEV, which is an unrecoverable
error, so it will be configured to not retry on error
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Signed-off-by: NCarlos Maiolino <cmaiolino@redhat.com>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

On reception of an error, we can fail immediately, perform some
bound amount of retries or retry indefinitely. The current behaviour
we have is to retry forever.

However, we'd like the ability to choose how long the filesystem
should try after an error, it can either fail immediately, retry a
few times, or retry forever. This is implemented by using
max_retries sysfs attribute, to hold the amount of times we allow
the filesystem to retry after an error. Being -1 a special case
where the filesystem will retry indefinitely.

Add both a maximum retry count and a retry timeout so that we can
bound by time and/or physical IO attempts.

Finally, plumb these into xfs_buf_iodone error processing so that
the error behaviour follows the selected configuration.
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Signed-off-by: NCarlos Maiolino <cmaiolino@redhat.com>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

With the error configuration handle for async metadata write errors
in place, we can now add initial support to the IO error processing
in xfs_buf_iodone_error().

Add an infrastructure function to look up the configuration handle,
and rearrange the error handling to prepare the way for different
error handling conigurations to be used.
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Signed-off-by: NCarlos Maiolino <cmaiolino@redhat.com>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Now we have the basic infrastructure, add the first error class so
we can build up the infrastructure in a meaningful way. Add the
metadata async write IO error class and sysfs entry, and introduce a
default configuration that matches the existing "retry forever"
behavior for async write metadata buffers.
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Signed-off-by: NCarlos Maiolino <cmaiolino@redhat.com>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

We need to be able to change the way XFS behaviours in error
conditions depending on the type of underlying storage. This is
necessary for handling non-traditional block devices with extended
error cases, such as thin provisioned devices that can return ENOSPC
as an IO error.

Introduce the basic sysfs infrastructure needed to define and
configure error behaviours. This is done to be generic enough to
extend to configuring behaviour in other error conditions, such as
ENOMEM, which also has different desired behaviours according to
machine configuration.
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Signed-off-by: NCarlos Maiolino <cmaiolino@redhat.com>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} macros were introduced *long* time
ago with promise that one day it will be possible to implement page
cache with bigger chunks than PAGE_SIZE.

This promise never materialized.  And unlikely will.

We have many places where PAGE_CACHE_SIZE assumed to be equal to
PAGE_SIZE.  And it's constant source of confusion on whether
PAGE_CACHE_* or PAGE_* constant should be used in a particular case,
especially on the border between fs and mm.

Global switching to PAGE_CACHE_SIZE != PAGE_SIZE would cause to much
breakage to be doable.

Let's stop pretending that pages in page cache are special.  They are
not.

The changes are pretty straight-forward:

 - <foo> << (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>;

 - <foo> >> (PAGE_CACHE_SHIFT - PAGE_SHIFT) -> <foo>;

 - PAGE_CACHE_{SIZE,SHIFT,MASK,ALIGN} -> PAGE_{SIZE,SHIFT,MASK,ALIGN};

 - page_cache_get() -> get_page();

 - page_cache_release() -> put_page();

This patch contains automated changes generated with coccinelle using
script below.  For some reason, coccinelle doesn't patch header files.
I've called spatch for them manually.

The only adjustment after coccinelle is revert of changes to
PAGE_CAHCE_ALIGN definition: we are going to drop it later.

There are few places in the code where coccinelle didn't reach.  I'll
fix them manually in a separate patch.  Comments and documentation also
will be addressed with the separate patch.

virtual patch

@@
expression E;
@@
- E << (PAGE_CACHE_SHIFT - PAGE_SHIFT)
+ E

@@
expression E;
@@
- E >> (PAGE_CACHE_SHIFT - PAGE_SHIFT)
+ E

@@
@@
- PAGE_CACHE_SHIFT
+ PAGE_SHIFT

@@
@@
- PAGE_CACHE_SIZE
+ PAGE_SIZE

@@
@@
- PAGE_CACHE_MASK
+ PAGE_MASK

@@
expression E;
@@
- PAGE_CACHE_ALIGN(E)
+ PAGE_ALIGN(E)

@@
expression E;
@@
- page_cache_get(E)
+ get_page(E)

@@
expression E;
@@
- page_cache_release(E)
+ put_page(E)
Signed-off-by: NKirill A. Shutemov <kirill.shutemov@linux.intel.com>
Acked-by: NMichal Hocko <mhocko@suse.com>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

Add a DEBUG mode-only sysfs knob to enable forced buffered write
failure. An additional side effect of this mode is brute force killing
of delayed allocation blocks in the range of the write. The latter is
the prime motiviation behind this patch, as userspace test
infrastructure requires a reliable mechanism to create and split
delalloc extents without causing extent conversion.

Certain fallocate operations (i.e., zero range) were used for this in
the past, but the implementations have changed such that delalloc
extents are flushed and converted to real blocks, rendering the test
useless.
Signed-off-by: NBrian Foster <bfoster@redhat.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NDave Chinner <david@fromorbit.com>

inode32/inode64 allocator behavior with respect to mount, remount
and growfs is a little tricky.

The inode32 mount option should only enable the inode32 allocator
heuristics if the filesystem is large enough for 64-bit inodes to
exist.  Today, it has this behavior on the initial mount, but a
remount with inode32 unconditionally changes the allocation
heuristics, even for a small fs.

Also, an inode32 mounted small filesystem should transition to the
inode32 allocator if the filesystem is subsequently grown to a
sufficient size.  Today that does not happen.

This patch consolidates xfs_set_inode32 and xfs_set_inode64 into a
single new function, and moves the "is the maximum inode number big
enough to matter" test into that function, so it doesn't rely on the
caller to get it right - which remount did not do, previously.
Signed-off-by: NEric Sandeen <sandeen@redhat.com>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Old leftovers.
Signed-off-by: NEric Sandeen <sandeen@redhat.com>
Reviewed-by: NDave Chinner <dchinner@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Don't leak the UUID table when the module is unloaded.
(Found with kmemleak.)
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>

To enable DAX to do atomic allocation of zeroed extents, we need to
drive the block zeroing deep into the allocator. Because
xfs_bmapi_write() can return merged extents on allocation that were
only partially allocated (i.e. requested range spans allocated and
hole regions, allocation into the hole was contiguous), we cannot
zero the extent returned from xfs_bmapi_write() as that can
overwrite existing data with zeros.

Hence we have to drive the extent zeroing into the allocation code,
prior to where we merge the extents into the BMBT and return the
resultant map. This means we need to propagate this need down to
the xfs_alloc_vextent() and issue the block zeroing at this point.

While this functionality is being introduced for DAX, there is no
reason why it is specific to DAX - we can per-zero blocks during the
allocation transaction on any type of device. It's just slow (and
usually slower than unwritten allocation and conversion) on
traditional block devices so doesn't tend to get used. We can,
however, hook hardware zeroing optimisations via sb_issue_zeroout()
to this operation, so it may be useful in future and hence the
"allocate zeroed blocks" API needs to be implementation neutral.
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

This patch implements per-filesystem stats objects in sysfs. It
depends on the application of the previous patch series that
develops the infrastructure to support both xfs global stats and
xfs per-fs stats in sysfs.

Stats objects are instantiated when an xfs filesystem is mounted
and deleted on unmount. With this patch, the stats directory is
created and populated with the familiar stats and stats_clear files.
Example:
        /sys/fs/xfs/sda9/stats/stats
        /sys/fs/xfs/sda9/stats/stats_clear

With this patch, the individual counts within the new per-fs
stats file(s) remain at zero. Functions that use the the macros
to increment, decrement, and add-to the per-fs stats counts will
be covered in a separate new patch to follow this one. Note that
the counts within the global stats file (/sys/fs/xfs/stats/stats)
advance normally and can be cleared as it was prior to this patch.

[dchinner: move setup/teardown to xfs_fs_{fill|put}_super() so
it is down before/after any path that uses the per-mount stats. ]
Signed-off-by: NBill O'Donnell <billodo@redhat.com>
Reviewed-by: NEric Sandeen <sandeen@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Add initial DAX support to XFS. To do this we need a new mount
option to turn DAX on filesystem, and we need to propagate this into
the inode flags whenever an inode is instantiated so that the
per-inode checks throughout the code Do The Right Thing.
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

xfs_ialloc_ag_select() iterates through the allocation groups looking
for free inodes or free space to determine whether to allow an inode
allocation to proceed. If no free inodes are available, it assumes that
an AG must have an extent longer than mp->m_ialloc_blks.

Sparse inode chunk support currently allows for allocations smaller than
the traditional inode chunk size specified in m_ialloc_blks. The current
minimum sparse allocation is set in the superblock sb_spino_align field
at mkfs time. Create a new m_ialloc_min_blks field in xfs_mount and use
this to represent the minimum supported allocation size for inode
chunks. Initialize m_ialloc_min_blks at mount time based on whether
sparse inodes are supported.
Signed-off-by: NBrian Foster <bfoster@redhat.com>
Reviewed-by: NDave Chinner <dchinner@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Now that there are no users of the bitfield based incore superblock
modification API, just remove the whole damn lot of it, including
all the bitfield definitions. This finally removes a lot of cruft
that has been around for a long time.

Credit goes to Christoph Hellwig for providing a great patch
connecting all the dots to enale us to do this. This patch is
derived from that work.
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Introduce helper functions for modifying fields in the superblock
into xfs_trans.c, the only caller of xfs_mod_incore_sb_batch().  We
can then use these directly in xfs_trans_unreserve_and_mod_sb() and
so remove another user of the xfs_mode_incore_sb() API without
losing any functionality or scalability of the transaction commit
code..

Based on a patch from Christoph Hellwig.
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Add a new helper to modify the incore counter of free realtime
extents. This matches the helpers used for inode and data block
counters, and removes a significant users of the xfs_mod_incore_sb()
interface.

Based on a patch originally from Christoph Hellwig.
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Now that the in-core superblock infrastructure has been replaced with
generic per-cpu counters, we don't need it anymore. Nuke it from
orbit so we are sure that it won't haunt us again...
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

XFS has hand-rolled per-cpu counters for the superblock since before
there was any generic implementation. The free block counter is
special in that it is used for ENOSPC detection outside transaction
contexts for for delayed allocation. This means that the counter
needs to be accurate at zero. The current per-cpu counter code jumps
through lots of hoops to ensure we never run past zero, but we don't
need to make all those jumps with the generic counter
implementation.

The generic counter implementation allows us to pass a "batch"
threshold at which the addition/subtraction to the counter value
will be folded back into global value under lock. We can use this
feature to reduce the batch size as we approach 0 in a very similar
manner to the existing counters and their rebalance algorithm. If we
use a batch size of 1 as we approach 0, then every addition and
subtraction will be done against the global value and hence allow
accurate detection of zero threshold crossing.

Hence we can replace the handrolled, accurate-at-zero counters with
generic percpu counters.

Note: this removes just enough of the icsb infrastructure to compile
without warnings. The rest will go in subsequent commits.
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

XFS has hand-rolled per-cpu counters for the superblock since before
there was any generic implementation. The free inode counter is not
used for any limit enforcement - the per-AG free inode counters are
used during allocation to determine if there are inode available for
allocation.

Hence we don't need any of the complexity of the hand-rolled
counters and we can simply replace them with generic per-cpu
counters similar to the inode counter.

This version introduces a xfs_mod_ifree() helper function from
Christoph Hellwig.
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

XFS has hand-rolled per-cpu counters for the superblock since before
there was any generic implementation. There are some warts around
the  use of them for the inode counter as the hand rolled counter is
designed to be accurate at zero, but has no specific accurracy at
any other value. This design causes problems for the maximum inode
count threshold enforcement, as there is no trigger that balances
the counters as they get close tothe maximum threshold.

Instead of designing new triggers for balancing, just replace the
handrolled per-cpu counter with a generic counter.  This enables us
to update the counter through the normal superblock modification
funtions, but rather than do that we add a xfs_mod_icount() helper
function (from Christoph Hellwig) and keep the percpu counter
outside the superblock in the struct xfs_mount.

This means we still need to initialise the per-cpu counter
specifically when we read the superblock, and vice versa when we
log/write it, but it does mean that we don't need to change any
other code.
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Add operations to export pNFS block layouts from an XFS filesystem.  See
the previous commit adding the operations for an explanation of them.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NDave Chinner <dchinner@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

We now have several superblock loggin functions that are identical
except for the transaction reservation and whether it shoul dbe a
synchronous transaction or not. Consolidate these all into a single
function, a single reserveration and a sync flag and call it
xfs_sync_sb().

Also, xfs_mod_sb() is not really a modification function - it's the
operation of logging the superblock buffer. hence change the name of
it to reflect this.

Note that we have to change the mp->m_update_flags that are passed
around at mount time to a boolean simply to indicate a superblock
update is needed.
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

When we log changes to the superblock, we first have to write them
to the on-disk buffer, and then log that. Right now we have a
complex bitfield based arrangement to only write the modified field
to the buffer before we log it.

This used to be necessary as a performance optimisation because we
logged the superblock buffer in every extent or inode allocation or
freeing, and so performance was extremely important. We haven't done
this for years, however, ever since the lazy superblock counters
pulled the superblock logging out of the transaction commit
fast path.

Hence we have a bunch of complexity that is not necessary that makes
writing the in-core superblock to disk much more complex than it
needs to be. We only need to log the superblock now during
management operations (e.g. during mount, unmount or quota control
operations) so it is not a performance critical path anymore.

As such, remove the complex field based logging mechanism and
replace it with a simple conversion function similar to what we use
for all other on-disk structures.

This means we always log the entirity of the superblock, but again
because we rarely modify the superblock this is not an issue for log
bandwidth or CPU time. Indeed, if we do log the superblock
frequently, delayed logging will minimise the impact of this
overhead.

[Fixed gquota/pquota inode sharing regression noticed by bfoster.]
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

More on-disk format consolidation.  A few declarations that weren't on-disk
format related move into better suitable spots.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NDave Chinner <dchinner@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

The expectation since the introduction the lazy superblock counters is
that the counters are synced and superblock logged appropriately as part
of the filesystem freeze sequence. This does not occur, however, due to
the logic in xfs_fs_writable() that prevents progress when the fs is in
any state other than SB_UNFROZEN.

While this is a bug, it has not been exposed to date because the last
thing XFS does during freeze is dirty the log. The log recovery process
recalculates the counters from AGI/AGF metadata to ensure everything is
correct. Therefore should a crash occur while an fs is frozen, the
subsequent log recovery puts everything back in order. See the following
commit for reference:

	92821e2b [XFS] Lazy Superblock Counters

We might not always want to rely on dirtying the log on a frozen fs.
Modify xfs_log_sbcount() to proceed when the filesystem is freezing but
not once the freeze process has completed. Modify xfs_fs_writable() to
accept the minimum freeze level for which modifications should be
blocked to support various codepaths.
Signed-off-by: NBrian Foster <bfoster@redhat.com>
Reviewed-by: NDave Chinner <dchinner@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

The xfslogd workqueue is a global, single-job workqueue for buffer ioend
processing. This means we allow for a single work item at a time for all
possible XFS mounts on a system. fsstress testing in loopback XFS over
XFS configurations has reproduced xfslogd deadlocks due to the single
threaded nature of the queue and dependencies introduced between the
separate XFS instances by online discard (-o discard).

Discard over a loopback device converts the discard request to a hole
punch (fallocate) on the underlying file. Online discard requests are
issued synchronously and from xfslogd context in XFS, hence the xfslogd
workqueue is blocked in the upper fs waiting on a hole punch request to
be servied in the lower fs. If the lower fs issues I/O that depends on
xfslogd to complete, both filesystems end up hung indefinitely. This is
reproduced reliabily by generic/013 on XFS->loop->XFS test devices with
the '-o discard' mount option.

Further, docker implementations appear to use this kind of configuration
for container instance filesystems by default (container fs->dm->
loop->base fs) and therefore are subject to this deadlock when running
on XFS.

Replace the global xfslogd workqueue with a per-mount variant. This
guarantees each mount access to a single worker and prevents deadlocks
due to inter-fs dependencies introduced by discard. Since the queue is
only responsible for buffer iodone processing at this point in time,
rename xfslogd to xfs-buf.
Signed-off-by: NBrian Foster <bfoster@redhat.com>
Reviewed-by: NDave Chinner <dchinner@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Embed a base kobject into xfs_mount. This creates a kobject associated
with each XFS mount and a subdirectory in sysfs with the name of the
filesystem. The subdirectory lifecycle matches that of the mount. Also
add the new xfs_sysfs.[c,h] source files with some XFS sysfs
infrastructure to facilitate attribute creation.

Note that there are currently no attributes exported as part of the
xfs_mount kobject. It exists solely to serve as a per-mount container
for child objects.
Signed-off-by: NBrian Foster <bfoster@redhat.com>
Reviewed-by: NDave Chinner <dchinner@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

The directory code has a dependency on the struct xfs_mount to
supply the directory block geometry. Block size, block log size,
and other parameters are pre-caclulated in the struct xfs_mount or
access directly from the superblock embedded in the struct
xfs_mount.

Extract all of this geometry information out of the struct xfs_mount
and superblock and place it into a new struct xfs_da_geometry
defined by the directory code. Allocate and initialise it at mount
time, and attach it to the struct xfs_mount so it canbe passed back
into the directory code appropriately rather than using the struct
xfs_mount.
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NBrian Foster <bfoster@redhat.com>
Signed-off-by: NDave Chinner <david@fromorbit.com>

v5 filesystems use 512 byte inodes as a minimum, so read inodes in
clusters that are effectively half the size of a v4 filesystem with
256 byte inodes. For v5 fielsystems, scale the inode cluster size
with the size of the inode so that we keep a constant 32 inodes per
cluster ratio for all inode IO.

This only works if mkfs.xfs sets the inode alignment appropriately
for larger inode clusters, so this functionality is made conditional
on mkfs doing the right thing. xfs_repair needs to know about
the inode alignment changes, too.

Wall time:
	create	bulkstat	find+stat	ls -R	unlink
v4	237s	161s		173s		201s	299s
v5	235s	163s		205s		 31s	356s
patched	234s	160s		182s		 29s	317s

System time:
	create	bulkstat	find+stat	ls -R	unlink
v4	2601s	2490s		1653s		1656s	2960s
v5	2637s	2497s		1681s		  20s	3216s
patched	2613s	2451s		1658s		  20s	3007s

So, wall time same or down across the board, system time same or
down across the board, and cache hit rates all improve except for
the ls -R case which is a pure cold cache directory read workload
on v5 filesystems...

So, this patch removes most of the performance and CPU usage
differential between v4 and v5 filesystems on traversal related
workloads.

Note: while this patch is currently for v5 filesystems only, there
is no reason it can't be ported back to v4 filesystems.  This hasn't
been done here because bringing the code back to v4 requires
forwards and backwards kernel compatibility testing.  i.e. to
deterine if older kernels(*) do the right thing with larger inode
alignments but still only using 8k inode cluster sizes. None of this
testing and validation on v4 filesystems has been done, so for the
moment larger inode clusters is limited to v5 superblocks.

(*) a current default config v4 filesystem should mount just fine on
2.6.23 (when lazy-count support was introduced), and so if we change
the alignment emitted by mkfs without a feature bit then we have to
make sure it works properly on all kernels since 2.6.23. And if we
allow it to be changed when the lazy-count bit is not set, then it's
all kernels since v2 logs were introduced that need to be tested for
compatibility...
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NEric Sandeen <sandeen@redhat.com>
Signed-off-by: NBen Myers <bpm@sgi.com>

The remaining non-vectorised code for the directory structure is the
node format blocks. This is shared with the attribute tree, and so
is slightly more complex to vectorise.

Introduce a "non-directory" directory ops structure that is attached
to all non-directory inodes so that attribute operations can be
vectorised for all inodes.

Once we do this, we can vectorise all the da btree operations.
Because this patch adds more infrastructure than it removes the
binary size does not decrease:

   text    data     bss     dec     hex filename
 794490   96802    1096  892388   d9de4 fs/xfs/xfs.o.orig
 792986   96802    1096  890884   d9804 fs/xfs/xfs.o.p1
 792350   96802    1096  890248   d9588 fs/xfs/xfs.o.p2
 789293   96802    1096  887191   d8997 fs/xfs/xfs.o.p3
 789005   96802    1096  886903   d8997 fs/xfs/xfs.o.p4
 789061   96802    1096  886959   d88af fs/xfs/xfs.o.p5
 789733   96802    1096  887631   d8b4f fs/xfs/xfs.o.p6
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NBen Myers <bpm@sgi.com>
Signed-off-by: NBen Myers <bpm@sgi.com>

Lots of the dir code now goes through switches to determine what is
the correct on-disk format to parse. It generally involves a
"xfs_sbversion_hasfoo" check, deferencing the superblock version and
feature fields and hence touching several cache lines per operation
in the process. Some operations do multiple checks because they nest
conditional operations and they don't pass the information in a
direct fashion between each other.

Hence, add an ops vector to the xfs_inode structure that is
configured when the inode is initialised to point to all the correct
decode and encoding operations.  This will significantly reduce the
branchiness and cacheline footprint of the directory object decoding
and encoding.

This is the first patch in a series of conversion patches. It will
introduce the ops structure, the setup of it and add the first
operation to the vector. Subsequent patches will convert directory
ops one at a time to keep the changes simple and obvious.

Just this patch shows the benefit of such an approach on code size.
Just converting the two shortform dir operations as this patch does
decreases the built binary size by ~1500 bytes:

$ size fs/xfs/xfs.o.orig fs/xfs/xfs.o.p1
   text    data     bss     dec     hex filename
 794490   96802    1096  892388   d9de4 fs/xfs/xfs.o.orig
 792986   96802    1096  890884   d9804 fs/xfs/xfs.o.p1
$

That's a significant decrease in the instruction cache footprint of
the directory code for such a simple change, and indicates that this
approach is definitely worth pursuing further.
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NBen Myers <bpm@sgi.com>