The "nodir" mode (statically assign master nodes instead
of using the resource directory) has always been highly
experimental, and never seriously used. This commit
fixes a number of problems, making nodir much more usable.
- Major change to recovery: recover all locks and restart
all in-progress operations after recovery. In some
cases it's not possible to know which in-progess locks
to recover, so recover all. (Most require recovery
in nodir mode anyway since rehashing changes most
master nodes.)
- Change the way nodir mode is enabled, from a command
line mount arg passed through gfs2, into a sysfs
file managed by dlm_controld, consistent with the
other config settings.
- Allow recovering MSTCPY locks on an rsb that has not
yet been turned into a master copy.
- Ignore RCOM_LOCK and RCOM_LOCK_REPLY recovery messages
from a previous, aborted recovery cycle. Base this
on the local recovery status not being in the state
where any nodes should be sending LOCK messages for the
current recovery cycle.
- Hold rsb lock around dlm_purge_mstcpy_locks() because it
may run concurrently with dlm_recover_master_copy().
- Maintain highbast on process-copy lkb's (in addition to
the master as is usual), because the lkb can switch
back and forth between being a master and being a
process copy as the master node changes in recovery.
- When recovering MSTCPY locks, flag rsb's that have
non-empty convert or waiting queues for granting
at the end of recovery. (Rename flag from LOCKS_PURGED
to RECOVER_GRANT and similar for the recovery function,
because it's not only resources with purged locks
that need grant a grant attempt.)
- Replace a couple of unnecessary assertion panics with
error messages.
Signed-off-by: David Teigland <teigland@redhat.com>
This patch instructs DLM to prevent an "in place" conversion, where the
lock just stays on the granted queue, and instead forces the conversion to
the back of the convert queue. This is done on upward conversions only.
This is useful in cases where, for example, a lock is frequently needed in
PR on one node, but another node needs it temporarily in EX to update it.
This may happen, for example, when the rindex is being updated by gfs2_grow.
The gfs2_grow needs to have the lock in EX, but the other nodes need to
re-read it to retrieve the updates. The glock is already granted in PR on
the non-growing nodes, so this prevents them from continually re-granting
the lock in PR, and forces the EX from gfs2_grow to go through.
Signed-off-by: Bob Peterson <rpeterso@redhat.com>
Signed-off-by: Steven 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: Steven 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: David Teigland <teigland@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
These new callbacks notify the dlm user about lock recovery.
GFS2, and possibly others, need to be aware of when the dlm
will be doing lock recovery for a failed lockspace member.
In the past, this coordination has been done between dlm and
file system daemons in userspace, which then direct their
kernel counterparts. These callbacks allow the same
coordination directly, and more simply.
Signed-off-by: David Teigland <teigland@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: Steven 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: Steven 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: Steven 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: Steven Whitehouse <swhiteho@redhat.com>
GFS2's idea of which return codes it needs to handle was based
upon those listed in dlm.h. Those didn't cover all the possible
codes and listed some which never happen. This updates GFS2 to
handle all the codes which can actually be returned from the
DLM under various circumstances.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
percpu.h is included by sched.h and module.h and thus ends up being
included when building most .c files. percpu.h includes slab.h which
in turn includes gfp.h making everything defined by the two files
universally available and complicating inclusion dependencies.
percpu.h -> slab.h dependency is about to be removed. Prepare for
this change by updating users of gfp and slab facilities include those
headers directly instead of assuming availability. As this conversion
needs to touch large number of source files, the following script is
used as the basis of conversion.
http://userweb.kernel.org/~tj/misc/slabh-sweep.py
The script does the followings.
* Scan files for gfp and slab usages and update includes such that
only the necessary includes are there. ie. if only gfp is used,
gfp.h, if slab is used, slab.h.
* When the script inserts a new include, it looks at the include
blocks and try to put the new include such that its order conforms
to its surrounding. It's put in the include block which contains
core kernel includes, in the same order that the rest are ordered -
alphabetical, Christmas tree, rev-Xmas-tree or at the end if there
doesn't seem to be any matching order.
* If the script can't find a place to put a new include (mostly
because the file doesn't have fitting include block), it prints out
an error message indicating which .h file needs to be added to the
file.
The conversion was done in the following steps.
1. The initial automatic conversion of all .c files updated slightly
over 4000 files, deleting around 700 includes and adding ~480 gfp.h
and ~3000 slab.h inclusions. The script emitted errors for ~400
files.
2. Each error was manually checked. Some didn't need the inclusion,
some needed manual addition while adding it to implementation .h or
embedding .c file was more appropriate for others. This step added
inclusions to around 150 files.
3. The script was run again and the output was compared to the edits
from #2 to make sure no file was left behind.
4. Several build tests were done and a couple of problems were fixed.
e.g. lib/decompress_*.c used malloc/free() wrappers around slab
APIs requiring slab.h to be added manually.
5. The script was run on all .h files but without automatically
editing them as sprinkling gfp.h and slab.h inclusions around .h
files could easily lead to inclusion dependency hell. Most gfp.h
inclusion directives were ignored as stuff from gfp.h was usually
wildly available and often used in preprocessor macros. Each
slab.h inclusion directive was examined and added manually as
necessary.
6. percpu.h was updated not to include slab.h.
7. Build test were done on the following configurations and failures
were fixed. CONFIG_GCOV_KERNEL was turned off for all tests (as my
distributed build env didn't work with gcov compiles) and a few
more options had to be turned off depending on archs to make things
build (like ipr on powerpc/64 which failed due to missing writeq).
* x86 and x86_64 UP and SMP allmodconfig and a custom test config.
* powerpc and powerpc64 SMP allmodconfig
* sparc and sparc64 SMP allmodconfig
* ia64 SMP allmodconfig
* s390 SMP allmodconfig
* alpha SMP allmodconfig
* um on x86_64 SMP allmodconfig
8. percpu.h modifications were reverted so that it could be applied as
a separate patch and serve as bisection point.
Given the fact that I had only a couple of failures from tests on step
6, I'm fairly confident about the coverage of this conversion patch.
If there is a breakage, it's likely to be something in one of the arch
headers which should be easily discoverable easily on most builds of
the specific arch.
Signed-off-by: Tejun Heo <tj@kernel.org>
Guess-its-ok-by: Christoph Lameter <cl@linux-foundation.org>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Since the start of GFS2, an "extra" inode has been used to store
the metadata belonging to each inode. The only reason for using
this inode was to have an extra address space, the other fields
were unused. This means that the memory usage was rather inefficient.
The reason for keeping each inode's metadata in a separate address
space is that when glocks are requested on remote nodes, we need to
be able to efficiently locate the data and metadata which relating
to that glock (inode) in order to sync or sync and invalidate it
(depending on the remotely requested lock mode).
This patch adds a new type of glock, which has in addition to
its normal fields, has an address space. This applies to all
inode and rgrp glocks (but to no other glock types which remain
as before). As a result, we no longer need to have the second
inode.
This results in three major improvements:
1. A saving of approx 25% of memory used in caching inodes
2. A removal of the circular dependency between inodes and glocks
3. No confusion between "normal" and "metadata" inodes in super.c
Although the first of these is the more immediately apparent, the
second is just as important as it now enables a number of clean
ups at umount time. Those will be the subject of future patches.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Although all glocks are, by the time of the umount glock wait,
scheduled for demotion, some of them haven't made it far
enough through the process for the original set of waiting
code to wait for them.
This extends the ref count to the whole glock lifetime in order
to ensure that the waiting does catch all glocks. It does make
it a bit more invasive, but it seems the only sensible solution
at the moment.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This patch adds a wait on umount between the point at which we
dispose of all glocks and the point at which we unmount the
lock protocol. This ensures that we've received all the replies
to our unlock requests before we stop the locking.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
Reported-by: Fabio M. Di Nitto <fdinitto@redhat.com>
After calling out to the dlm, GFS2 sets the new state of a glock to
gl_target in gdlm_ast(). However, gl_target is not always the lock
state that was requested. If a conversion from shared to exclusive
fails, finish_xmote() will call do_xmote() with LM_ST_UNLOCKED, instead
of gl->gl_target, so that it can reacquire the lock in exlusive the next
time around. In this case, setting the lock to gl_target in gdlm_ast()
will make GFS2 think that it has the glock in exclusive mode, when
really, it doesn't have the glock locked at all. This patch adds a new
field to the gfs2_glock structure, gl_req, to track the mode that was
requested.
Signed-off-by: Benjamin Marzinski <bmarzins@redhat.com>
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
This is the big patch that I've been working on for some time
now. There are many reasons for wanting to make this change
such as:
o Reducing overhead by eliminating duplicated fields between structures
o Simplifcation of the code (reduces the code size by a fair bit)
o The locking interface is now the DLM interface itself as proposed
some time ago.
o Fewer lookups of glocks when processing replies from the DLM
o Fewer memory allocations/deallocations for each glock
o Scope to do further optimisations in the future (but this patch is
more than big enough for now!)
Please note that (a) this patch relates to the lock_dlm module and
not the DLM itself, that is still a separate module; and (b) that
we retain the ability to build GFS2 as a standalone single node
filesystem with out requiring the DLM.
This patch needs a lot of testing, hence my keeping it I restarted
my -git tree after the last merge window. That way, this has the maximum
exposure before its merged. This is (modulo a few minor bug fixes) the
same patch that I've been posting on and off the the last three months
and its passed a number of different tests so far.
Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
