Save the effort of allocating, reading and writing
the lvb for most glocks that do not use it.
Signed-off-by: NDavid Teigland <teigland@redhat.com>
Signed-off-by: NSteven Whitehouse <swhiteho@redhat.com>

When unmounting, gfs2 does a full dlm_unlock operation on every
cached lock.  This can create a very large amount of work and can
take a long time to complete.  However, the vast majority of these
dlm unlock operations are unnecessary because after all the unlocks
are done, gfs2 leaves the dlm lockspace, which automatically clears
the locks of the leaving node, without unlocking each one individually.
So, gfs2 can skip explicit dlm unlocks, and use dlm_release_lockspace to
remove the locks implicitly.  The one exception is when the lock's lvb is
being used.  In this case, dlm_unlock is called because it may update the
lvb of the resource.
Signed-off-by: NDavid Teigland <teigland@redhat.com>
Signed-off-by: NSteven Whitehouse <swhiteho@redhat.com>

[Editorial: This is a nit, but has been a minor irritation for a long time:]

This patch renames glops structure item for go_xmote_th to go_sync.
The functionality is unchanged; it's just for readability.
Signed-off-by: NBob Peterson <rpeterso@redhat.com>
Signed-off-by: NSteven Whitehouse <swhiteho@redhat.com>

Two of the bug traps here could really be warnings. The others are
converted from BUG() to GLOCK_BUG_ON() since we'll most likely
need to know the glock state in order to debug any issues which
arise. As a result of this, __dump_glock has to be renamed and
is no longer static.
Signed-off-by: NSteven Whitehouse <swhiteho@redhat.com>

Function add_to_queue was checking may_grant for the passed-in
holder for every iteration of its gh2 loop. Now it only checks it
once at the beginning to see if a try lock is futile.
Signed-off-by: NBob Peterson <rpeterso@redhat.com>
Signed-off-by: NSteven Whitehouse <swhiteho@redhat.com>

Function gfs2_glock_dq_wait called two-line function wait_on_demote,
so they were combined.
Signed-off-by: NBob Peterson <rpeterso@redhat.com>
Signed-off-by: NSteven Whitehouse <swhiteho@redhat.com>

Function gfs2_glock_wait only called function wait_on_holder and
returned its return code, so they were combined for readability.
Signed-off-by: NBob Peterson <rpeterso@redhat.com>
Signed-off-by: NSteven Whitehouse <swhiteho@redhat.com>

Since function gfs2_glock_schedule_for_reclaim is only two
significant lines, we can eliminate it, simplifying the code
and making it more readable.
Signed-off-by: NBob Peterson <rpeterso@redhat.com>
Signed-off-by: NSteven Whitehouse <swhiteho@redhat.com>

This places a limit on the buffer size for archs with larger
PAGE_SIZE.
Signed-off-by: NSteven Whitehouse <swhiteho@redhat.com>
Reported-by: NEric Dumazet <eric.dumazet@gmail.com>

Make use of the newly added seq_vprintf() function.
Signed-off-by: NSteven Whitehouse <swhiteho@redhat.com>
Reported-by: NEric Dumazet <eric.dumazet@gmail.com>
Acked-by: NAl Viro <viro@ZenIV.linux.org.uk>

Instead of reading in the resource groups when gfs2 is checking
for free space to allocate from, gfs2 can store the necessary infromation
in the resource group's lvb.  Also, instead of searching for unlinked
inodes in every resource group that's checked for free space, gfs2 can
store the number of unlinked but inodes in the lvb, and only check for
unlinked inodes if it will find some.

The first time a resource group is locked, the lvb must initialized.
Since this involves counting the unlinked inodes in the resource group,
this takes a little extra time.  But after that, if the resource group
is locked with GL_SKIP, the buffer head won't be read in unless it's
actually needed.

Enabling the resource groups lvbs is done via the rgrplvb mount option.  If
this option isn't set, the lvbs will still be set and updated, but they won't
be verfied or used by the filesystem.  To safely turn on this option, all of
the nodes mounting the filesystem must be running code with this patch, and
the filesystem must have been completely unmounted since they were updated.
Signed-off-by: NBenjamin Marzinski <bmarzins@redhat.com>
Signed-off-by: NSteven Whitehouse <swhiteho@redhat.com>

For the glocks and glstats seq_files, which are exposed via debugfs
we should cache the most recent hash bucket, along with the offset
into that bucket. This allows us to restart from that point, rather
than having to begin at the beginning each time.

This is an idea from Eric Dumazet, however I've slightly extended it
so that if the position from which we are due to start is at any
point beyond the last cached point, we start from the last cached
point, plus whatever is the appropriate offset. I don't really expect
people to be lseeking around these files, but if they did so with only
positive offsets, then we'd still get some of the benefit of using a
cached offset.

With my simple test of around 200k entries in the file, I'm seeing
an approx 10x speed up.

Cc: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: NSteven Whitehouse <swhiteho@redhat.com>

As per Al Viro's suggestion, this increases the buffer size used
for these two files. This provides a speed up of slightly less than
8x (i.e. proportional to the buffer size) for cases when we have
large numbers of glocks.

Cc: Al Viro <viro@ZenIV.linux.org.uk>
Signed-off-by: NSteven Whitehouse <swhiteho@redhat.com>

The stats are divided into two sets: those relating to the
super block and those relating to an individual glock. The
super block stats are done on a per cpu basis in order to
try and reduce the overhead of gathering them. They are also
further divided by glock type.

In the case of both the super block and glock statistics,
the same information is gathered in each case. The super
block statistics are used to provide default values for
most of the glock statistics, so that newly created glocks
should have, as far as possible, a sensible starting point.

The statistics are divided into three pairs of mean and
variance, plus two counters. The mean/variance pairs are
smoothed exponential estimates and the algorithm used is
one which will be very familiar to those used to calculation
of round trip times in network code.

The three pairs of mean/variance measure the following
things:

 1. DLM lock time (non-blocking requests)
 2. DLM lock time (blocking requests)
 3. Inter-request time (again to the DLM)

A non-blocking request is one which will complete right
away, whatever the state of the DLM lock in question. That
currently means any requests when (a) the current state of
the lock is exclusive (b) the requested state is either null
or unlocked or (c) the "try lock" flag is set. A blocking
request covers all the other lock requests.

There are two counters. The first is there primarily to show
how many lock requests have been made, and thus how much data
has gone into the mean/variance calculations. The other counter
is counting queueing of holders at the top layer of the glock
code. Hopefully that number will be a lot larger than the number
of dlm lock requests issued.

So why gather these statistics? There are several reasons
we'd like to get a better idea of these timings:

1. To be able to better set the glock "min hold time"
2. To spot performance issues more easily
3. To improve the algorithm for selecting resource groups for
allocation (to base it on lock wait time, rather than blindly
using a "try lock")
Due to the smoothing action of the updates, a step change in
some input quantity being sampled will only fully be taken
into account after 8 samples (or 4 for the variance) and this
needs to be carefully considered when interpreting the
results.

Knowing both the time it takes a lock request to complete and
the average time between lock requests for a glock means we
can compute the total percentage of the time for which the
node is able to use a glock vs. time that the rest of the
cluster has its share. That will be very useful when setting
the lock min hold time.

The other point to remember is that all times are in
nanoseconds. Great care has been taken to ensure that we
measure exactly the quantities that we want, as accurately
as possible. There are always inaccuracies in any
measuring system, but I hope this is as accurate as we
can reasonably make it.
Signed-off-by: NSteven Whitehouse <swhiteho@redhat.com>

This patch fixes a narrow race window between the glock ref count
hitting zero and glocks being removed from the lru_list.
Signed-off-by: NSteven Whitehouse <swhiteho@redhat.com>

This new method of managing recovery is an alternative to
the previous approach of using the userland gfs_controld.

- use dlm slot numbers to assign journal id's
- use dlm recovery callbacks to initiate journal recovery
- use a dlm lock to determine the first node to mount fs
- use a dlm lock to track journals that need recovery
Signed-off-by: NDavid Teigland <teigland@redhat.com>
Signed-off-by: NSteven Whitehouse <swhiteho@redhat.com>

This patch is a performance improvement for GFS2 in a clustered
environment. It makes the glock hold time self-adjusting.
Signed-off-by: NBob Peterson <rpeterso@redhat.com>
Signed-off-by: NSteven Whitehouse <swhiteho@redhat.com>

Change each shrinker's API by consolidating the existing parameters into
shrink_control struct.  This will simplify any further features added w/o
touching each file of shrinker.

[akpm@linux-foundation.org: fix build]
[akpm@linux-foundation.org: fix warning]
[kosaki.motohiro@jp.fujitsu.com: fix up new shrinker API]
[akpm@linux-foundation.org: fix xfs warning]
[akpm@linux-foundation.org: update gfs2]
Signed-off-by: NYing Han <yinghan@google.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Minchan Kim <minchan.kim@gmail.com>
Acked-by: NPavel Emelyanov <xemul@openvz.org>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Mel Gorman <mel@csn.ul.ie>
Acked-by: NRik van Riel <riel@redhat.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Dave Hansen <dave@linux.vnet.ibm.com>
Cc: Steven Whitehouse <swhiteho@redhat.com>
Signed-off-by: NAndrew Morton <akpm@linux-foundation.org>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

This patch fixes a race in the GFS2 glock state machine that may
result in lockups.  The symptom is that all nodes but one will
hang, waiting for a particular glock.  All the holder records
will have the "W" (Waiting) bit set.  The other node will
typically have the glock stuck in Exclusive mode (EX) with no
holder records, but the dinode will be cached.  In other words,
an entry with "I:" will appear in the glock dump for that glock,
but nothing else.

The race has to do with the glock "Pending Demote" bit, which
can be set, then immediately reset, thus losing the fact that
another node needs the glock.  The sequence of events is:

1. Something schedules the glock workqueue (e.g. glock request from fs)
2. The glock workqueue gets to the point between the test of the reply pending
bit and the spin lock:

        if (test_and_clear_bit(GLF_REPLY_PENDING, &gl->gl_flags)) {
                finish_xmote(gl, gl->gl_reply);
                drop_ref = 1;
        }
        down_read(&gfs2_umount_flush_sem);         <---- i.e. here
        spin_lock(&gl->gl_spin);

3. In comes (a) the reply to our EX lock request setting GLF_REPLY_PENDING and
            (b) the demote request which sets GLF_PENDING_DEMOTE

4. The following test is executed:

        if (test_and_clear_bit(GLF_PENDING_DEMOTE, &gl->gl_flags) &&
            gl->gl_state != LM_ST_UNLOCKED &&
            gl->gl_demote_state != LM_ST_EXCLUSIVE) {

This resets the pending demote flag, and gl->gl_demote_state is not equal to
exclusive, however because the reply from the dlm arrived after we checked for
the GLF_REPLY_PENDING flag, gl->gl_state is still equal to unlocked, so
although we reset the GLF_PENDING_DEMOTE flag, we didn't then set the
GLF_DEMOTE flag or reinstate the GLF_PENDING_DEMOTE_FLAG.

The patch closes the timing window by only transitioning the
"Pending demote" bit to the "demote" flag once we know the
other conditions (not unlocked and not exclusive) are met.
Signed-off-by: NBob Peterson <rpeterso@redhat.com>
Signed-off-by: NSteven Whitehouse <swhiteho@redhat.com>

Previously we marked all locks being promoted to a higher mode
with the try flag to avoid any potential deadlocks issues. The
DLM is able to detect these and report them in way that GFS2 can
deal with them correctly. So we can just request the required mode
and wait for a response without needing to perform this check.
Signed-off-by: NSteven Whitehouse <swhiteho@redhat.com>

Now that the whole dcache_hash_bucket crap is gone, go all the way and
also remove the weird locking layering violations for locking the hash
buckets.  Add hlist_bl_lock/unlock helpers to move the locking into the
list abstraction instead of requiring each caller to open code it.
After all allowing for the bit locks is the whole point of these helpers
over the plain hlist variant.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

This patch adds writeback_control to writing back the AIL
list. This means that we can then take advantage of the
information we get in ->write_inode() in order to set off
some pre-emptive writeback.

In addition, the AIL code is cleaned up a bit to make it
a bit simpler to understand.

There is still more which can usefully be done in this area,
but this is a good start at least.
Signed-off-by: NSteven Whitehouse <swhiteho@redhat.com>

The GLF_LRU flag introduced in the previous patch can be
used to check if a glock is on the lru list when a new
holder is queued and if so remove it, without having first
to get the lru_lock.

The main purpose of this patch however is to optimise the
glocks left over when an inode at end of life is being
evicted. Previously such glocks were left with the GLF_LFLUSH
flag set, so that when reclaimed, each one required a log flush.
This patch resets the GLF_LFLUSH flag when there is nothing
left to flush thus preventing later log flushes as glocks are
reused or demoted.

In order to do this, we need to keep track of the number of
revokes which are outstanding, and also to clear the GLF_LFLUSH
bit after a log commit when only revokes have been processed.
Signed-off-by: NSteven Whitehouse <swhiteho@redhat.com>

This adds support for two new flags. One keeps track of whether
the glock is on the LRU list or not. The other isn't really a
flag as such, but an indication of whether the glock has an
attached object or not. This indication is reported without
any locking, which is ok since we do not dereference the object
pointer but merely report whether it is NULL or not.

Also, this fixes one place where a tracepoint was missing, which
was at the point we remove deallocated blocks from the journal.
Signed-off-by: NSteven Whitehouse <swhiteho@redhat.com>

Rather than allowing the glocks to be scheduled for possible
reclaim as soon as they have exited the journal, this patch
delays their entry to the list until the glocks in question
are no longer in use.

This means that we will rely on the vm for writeback of all
dirty data and metadata from now on. When glocks are added
to the lru list they should be freeable much faster since all
the I/O required to free them should have already been completed.

This should lead to much better I/O patterns under low memory
conditions.
Signed-off-by: NSteven Whitehouse <swhiteho@redhat.com>

Fixes generated by 'codespell' and manually reviewed.
Signed-off-by: NLucas De Marchi <lucas.demarchi@profusion.mobi>

As per RCU glock patch review comments, don't use the _raw
version of this function here.
Signed-off-by: NSteven Whitehouse <swhiteho@redhat.com>
Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Reviewed-by: NPaul E. McKenney <paulmck@linux.vnet.ibm.com>

This is a small patch that optimizes multiple glock dequeue
operations.  It changes the unlock order to be more efficient
and makes it easier for lock debugging tools to unravel.  It
also eliminates the need for the temp variable x, although
that would likely be optimized out.
Signed-off-by: NBob Peterson <rpeterso@redhat.com>
Signed-off-by: NSteven Whitehouse <swhiteho@redhat.com>

This patch fixes a race in deallocating glocks which was introduced
in the RCU glock patch. We need to ensure that the glock count is
kept correct even in the case that there is a race to add a new
glock into the hash table. Also, to avoid having to wait for an
RCU grace period, the glock counter can be decremented before
call_rcu() is called.
Signed-off-by: NSteven Whitehouse <swhiteho@redhat.com>

There are two spellings in use for 'freeze' + 'able' - 'freezable' and
'freezeable'.  The former is the more prominent one.  The latter is
mostly used by workqueue and in a few other odd places.  Unify the
spelling to 'freezable'.
Signed-off-by: NTejun Heo <tj@kernel.org>
Reported-by: NAlan Stern <stern@rowland.harvard.edu>
Acked-by: N"Rafael J. Wysocki" <rjw@sisk.pl>
Acked-by: NGreg Kroah-Hartman <gregkh@suse.de>
Acked-by: NDmitry Torokhov <dtor@mail.ru>
Cc: David Woodhouse <dwmw2@infradead.org>
Cc: Alex Dubov <oakad@yahoo.com>
Cc: "David S. Miller" <davem@davemloft.net>
Cc: Steven Whitehouse <swhiteho@redhat.com>

Somehow this tracepoint landed up in the wrong place. This moves it
to where it should be.
Signed-off-by: NSteven Whitehouse <swhiteho@redhat.com>

This has a number of advantages:

 - Reduces contention on the hash table lock
 - Makes the code smaller and simpler
 - Should speed up glock dumps when under load
 - Removes ref count changing in examine_bucket
 - No longer need hash chain lock in glock_put() in common case

There are some further changes which this enables and which
we may do in the future. One is to look at using SLAB_RCU,
and another is to look at using a per-cpu counter for the
per-sb glock counter, since that is touched twice in the
lifetime of each glock (but only used at umount time).
Signed-off-by: NSteven Whitehouse <swhiteho@redhat.com>
Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com>

We can only merge the fields into a bitfield if the locking
rules for them are the same. In this case gl_spin covers all
of the fields (write side) but a couple of them are used
with GLF_LOCK as the read side lock, which should be ok
since we know that the field in question won't be changing
at the time.

The gl_req setting has to be done earlier (in glock.c) in order
to place it under gl_spin. The gl_reply setting also has to be
brought under gl_spin in order to comply with the new rules.

This saves 4*sizeof(unsigned int) per glock.
Signed-off-by: NSteven Whitehouse <swhiteho@redhat.com>
Cc: Bob Peterson <rpeterso@redhat.com>

The DLM never returns -EAGAIN in response to dlm_lock(), and even
if it did, the test in gdlm_lock() was wrong anyway. Once that
test is removed, it is possible to greatly simplify this code
by simply using a "normal" error return code (0 for success).

We then no longer need the LM_OUT_ASYNC return code which can
be removed.
Signed-off-by: NSteven Whitehouse <swhiteho@redhat.com>

Using %pV reduces the number of printk calls and
eliminates any possible message interleaving from
other printk calls.
Signed-off-by: NJoe Perches <joe@perches.com>
Signed-off-by: NSteven Whitehouse <swhiteho@redhat.com>

Signed-off-by: NJoe Perches <joe@perches.com>
Signed-off-by: NSteven Whitehouse <swhiteho@redhat.com>

The WQ_RESCUER flag should only be used internally to the
workqueue implementation.
Signed-off-by: NSteven Whitehouse <swhiteho@redhat.com>
Acked-by: NTejun Heo <tj@kernel.org>

This area of the code has always been a bit delicate due to the
subtleties of lock ordering. The problem is that for "normal"
alloc/dealloc, we always grab the inode locks first and the rgrp lock
later.

In order to ensure no races in looking up the unlinked, but still
allocated inodes, we need to hold the rgrp lock when we do the lookup,
which means that we can't take the inode glock.

The solution is to borrow the technique already used by NFS to solve
what is essentially the same problem (given an inode number, look up
the inode carefully, checking that it really is in the expected
state).

We cannot do that directly from the allocation code (lock ordering
again) so we give the job to the pre-existing delete workqueue and
carry on with the allocation as normal.

If we find there is no space, we do a journal flush (required anyway
if space from a deallocation is to be released) which should block
against the pending deallocations, so we should always get the space
back.
Signed-off-by: NSteven Whitehouse <swhiteho@redhat.com>

The tests further down the recovery function relating to
unlocking the journal need to be updated to match the
intial test. Also, a test in the umount code which was
surplus to requirements has been removed. Umounting
spectator mounts now works correctly, as expected.
Signed-off-by: NSteven Whitehouse <swhiteho@redhat.com>

The recovery workqueue can be freezable since
we want it to finish what it is doing if the system is to
be frozen (although why you'd want to freeze a cluster node
is beyond me since it will result in it being ejected from
the cluster). It does still make sense for single node
GFS2 filesystems though.

The glock workqueue will benefit from being able to run more
work items concurrently. A test running postmark shows
improved performance and multi-threaded workloads are likely
to benefit even more. It needs to be high priority because
the latency directly affects the latency of filesystem glock
operations.

The delete workqueue is similar to the recovery workqueue in
that it must not get blocked by memory allocations, and may
run for a long time.

Potentially other GFS2 threads might also be converted to
workqueues, but I'll leave that for a later patch.
Signed-off-by: NSteven Whitehouse <swhiteho@redhat.com>
Acked-by: NTejun Heo <tj@kernel.org>