When a node is removed that held a PW/EX lock, the
existing master node should invalidate the lvb on the
resource due to the purged lock.

Previously, the existing master node was invalidating
the lvb if it found only NL/CR locks on the resource
during recovery for the removed node.  This could lead
to cases where it invalidated the lvb and shouldn't
have, or cases where it should have invalidated and
didn't.

When recovery selects a *new* master node for a
resource, and that new master finds only NL/CR locks
on the resource after lock recovery, it should
invalidate the lvb.  This case was handled correctly
(but was incorrectly applied to the existing master
case also.)

When a process exits while holding a PW/EX lock,
the lvb on the resource should be invalidated.
This was not happening.

The lvb contents and VALNOTVALID flag should be
recovered before granting locks in recovery so that
the recovered lvb state is provided in the callback.
The lvb was being recovered after the lock was granted.
Signed-off-by: NDavid Teigland <teigland@redhat.com>

I don't know exactly how, but in some cases, a dir
record is not removed, or a new one is created when
it shouldn't be.  The result is that the dir node
lookup returns a master node where the rsb does not
exist.  In this case, The master node will repeatedly
return -EBADR for requests, and the lock requests will
be stuck.

Until all possible ways for this to happen can be
eliminated, a simple and effective way to recover from
this situation is for the supposed master node to send
a standard remove message to the dir node when it
receives a request for a resource it has no rsb for.
Signed-off-by: NDavid Teigland <teigland@redhat.com>

The process of rebuilding locks on a new master during
recovery could re-order the locks on the convert queue,
creating an "in place" conversion deadlock that would
not be resolved.  Fix this by not considering queue
order when granting conversions after recovery.
Signed-off-by: NDavid Teigland <teigland@redhat.com>

It was possible for a remove message on an old
rsb to be sent after a lookup message on a new
rsb, where the rsbs were for the same resource
name.  This could lead to a missing directory
entry for the new rsb.

It is fixed by keeping a copy of the resource
name being removed until after the remove has
been sent.  A lookup checks if this in-progress
remove matches the name it is looking up.
Signed-off-by: NDavid Teigland <teigland@redhat.com>

Remove the dir hash table (dirtbl), and use
the rsb hash table (rsbtbl) as the resource
directory.  It has always been an unnecessary
duplication of information.

This improves efficiency by using a single rsbtbl
lookup in many cases where both rsbtbl and dirtbl
lookups were needed previously.

This eliminates the need to handle cases of rsbtbl
and dirtbl being out of sync.

In many cases there will be memory savings because
the dir hash table no longer exists.
Signed-off-by: NDavid Teigland <teigland@redhat.com>

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: NDavid Teigland <teigland@redhat.com>

Change some existing error/debug messages to
collect more useful information, and add
some new error/debug messages to address
recently found problems.
Signed-off-by: NDavid Teigland <teigland@redhat.com>

If the rsb is found in the "keep" tree, but is
not the right type (i.e. not MASTER), we can
return immediately with the result.  There's
no point in going on to search the "toss" list
as if we hadn't found it.
Signed-off-by: NDavid Teigland <teigland@redhat.com>

An outstanding remote operation (an lkb on the "waiter"
list) could sometimes miss being resent during recovery.
The decision was based on the lkb_nodeid field, which
could have changed during an earlier aborted recovery,
so it no longer represents the actual remote destination.
The lkb_wait_nodeid is always the actual remote node,
so it is the best value to use.
Signed-off-by: NDavid Teigland <teigland@redhat.com>

The QUECVT flag should not prevent conversions from
being granted immediately when the convert queue is
empty.
Signed-off-by: NDavid Teigland <teigland@redhat.com>

The function used to find an rsb during directory
recovery was searching the single linear list of
rsb's.  This wasted a lot of time compared to
using the standard hash table to find the rsb.
Signed-off-by: NDavid Teigland <teigland@redhat.com>

Change the linked lists to rb_tree's in the rsb
hash table to speed up searches.  Slow rsb searches
were having a large impact on gfs2 performance due
to the large number of dlm locks gfs2 uses.
Signed-off-by: NBob Peterson <rpeterso@redhat.com>
Signed-off-by: NDavid Teigland <teigland@redhat.com>

Instead of creating our own kthread (dlm_astd) to deliver
callbacks for all lockspaces, use a per-lockspace workqueue
to deliver the callbacks.  This eliminates complications and
slowdowns from many lockspaces sharing the same thread.
Signed-off-by: NDavid Teigland <teigland@redhat.com>

gfs2 recently began using this feature heavily,
creating more debug output than we want to see.
Signed-off-by: NDavid Teigland <teigland@redhat.com>

By pre-allocating rsb structs before searching the hash
table, they can be inserted immediately.  This avoids
always having to repeat the search when adding the struct
to hash list.

This also adds space to the rsb struct for a max resource
name, so an rsb allocation can be used by any request.
The constant size also allows us to finally use a slab
for the rsb structs.
Signed-off-by: NDavid Teigland <teigland@redhat.com>

This is simpler and quicker than the hash table, and
avoids needing to search the hash list for every new
lkid to check if it's used.
Signed-off-by: NDavid Teigland <teigland@redhat.com>

The gfp and size args were switched.
Signed-off-by: NDavid Teigland <teigland@redhat.com>

In fs/dlm/lock.c in the dlm_scan_waiters() function there are 3 small
issues:

1) There's no need to test the return value of the allocation and do a
memset if is succeedes. Just use kzalloc() to obtain zeroed memory.

2) Since kfree() handles NULL pointers gracefully, the test of
'warned' against NULL before the kfree() after the loop is completely
pointless. Remove it.

3) The arguments to kmalloc() (now kzalloc()) were swapped. Thanks to
Dr. David Alan Gilbert for pointing this out.
Signed-off-by: NJesper Juhl <jj@chaosbits.net>
Signed-off-by: NDavid Teigland <teigland@redhat.com>

kmalloc a stub message struct during recovery instead of sharing the
struct in the lockspace.  This leaves the lockspace stub_ms only for
faking downconvert replies, where it is never modified and sharing
is not a problem.

Also improve the debug messages in the same recovery function.
Signed-off-by: NDavid Teigland <teigland@redhat.com>

Add an option (disabled by default) to print a warning message
when a lock has been waiting a configurable amount of time for
a reply message from another node.  This is mainly for debugging.
Signed-off-by: NDavid Teigland <teigland@redhat.com>

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

Change how callbacks are recorded for locks.  Previously, information
about multiple callbacks was combined into a couple of variables that
indicated what the end result should be.  In some situations, we
could not tell from this combined state what the exact sequence of
callbacks were, and would end up either delivering the callbacks in
the wrong order, or suppress redundant callbacks incorrectly.  This
new approach records all the data for each callback, leaving no
uncertainty about what needs to be delivered.
Signed-off-by: NDavid Teigland <teigland@redhat.com>

When converting a lock, an lkb is in the granted state and also being used
to request a new state. In the case that the conversion was a "try 1cb"
type which has failed, and if the new state was incompatible with the old
state, a callback was being generated to the requesting node. This is
incorrect as callbacks should only be sent to all the other nodes holding
blocking locks. The requesting node should receive the normal (failed)
response to its "try 1cb" conversion request only.

This was discovered while debugging a performance problem on GFS2, however
this fix also speeds up GFS as well. In the GFS2 case the performance gain
is over 10x for cases of write activity to an inode whose glock is cached
on another, idle (wrt that glock) node.

(comment added, dct)
Signed-off-by: NSteven Whitehouse <swhiteho@redhat.com>
Tested-by: NAbhijith Das <adas@redhat.com>
Signed-off-by: NDavid Teigland <teigland@redhat.com>

Smatch complains because "lkb" is never NULL.  Looking at it, the original
code actually adds the new element to the end of the list fine, so we can
just get rid of the if condition.  This code is four years old and no one
has complained so it must work.
Signed-off-by: NDan Carpenter <error27@gmail.com>
Signed-off-by: NDavid Teigland <teigland@redhat.com>

include cleanup: Update gfp.h and slab.h includes to prepare for breaking implicit slab.h inclusion from percpu.h

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: NTejun Heo <tj@kernel.org>
Guess-its-ok-by: NChristoph Lameter <cl@linux-foundation.org>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>

The bast mode that appears in the debugfs output should be
useful on both master and process nodes.  lkb_highbast is
currently printed, and is only useful on the master node.
lkb_bastmode is only useful on the process node.  This
patch sets lkb_bastmode on the master node as well, and
uses that value in the debugfs print.
Signed-off-by: NDavid Teigland <teigland@redhat.com>

When the lock master processes a successful operation (request,
convert, cancel, or unlock), it will process the effects of the
change before sending the reply for the operation.  The "effects"
of the operation are:

- blocking callbacks (basts) for any newly granted locks
- waiting or converting locks that can now be granted

The cast is queued on the local node when the reply from the lock
master is received.  This means that a lock holder can receive a
bast for a lock mode that is doesn't yet know has been granted.
Signed-off-by: NDavid Teigland <teigland@redhat.com>

When both blocking and completion callbacks are queued for lock,
the dlm would always deliver the completion callback (cast) first.
In some cases the blocking callback (bast) is queued before the
cast, though, and should be delivered first.  This patch keeps
track of the order in which they were queued and delivers them
in that order.

This patch also keeps track of the granted mode in the last cast
and eliminates the following bast if the bast mode is compatible
with the preceding cast mode.  This happens when a remotely mastered
lock is demoted, e.g. EX->NL, in which case the local node queues
a cast immediately after sending the demote message.  In this way
a cast can be queued for a mode, e.g. NL, that makes an in-transit
bast extraneous.
Signed-off-by: NDavid Teigland <teigland@redhat.com>

Replace all GFP_KERNEL and ls_allocation with GFP_NOFS.
ls_allocation would be GFP_KERNEL for userland lockspaces
and GFP_NOFS for file system lockspaces.

It was discovered that any lockspaces on the system can
affect all others by triggering memory reclaim in the
file system which could in turn call back into the dlm
to acquire locks, deadlocking dlm threads that were
shared by all lockspaces, like dlm_recv.
Signed-off-by: NDavid Teigland <teigland@redhat.com>

  CC [M]  fs/dlm/lock.o
fs/dlm/lock.c: In function ‘find_rsb’:
fs/dlm/lock.c:438: warning: ‘r’ may be used uninitialized in this function

Since r is used on the error path to set r_ret, set it to NULL.
Signed-off-by: NSteven Whitehouse <swhiteho@redhat.com>
Signed-off-by: NDavid Teigland <teigland@redhat.com>

Return immediately from dlm_unlock(CANCEL) if the lock is
granted and not being converted; there's nothing to cancel.
Signed-off-by: NDavid Teigland <teigland@redhat.com>

When a conversion completes successfully and finds that a cancel
of the convert is still in progress (which is now a moot point),
preemptively clear the state associated with outstanding cancel.
That state could cause a subsequent conversion to be ignored.

Also, improve the consistency and content of error and debug
messages in this area.
Signed-off-by: NDavid Teigland <teigland@redhat.com>

The rwlock is almost always used in write mode, so there's no reason
to not use a spinlock instead.
Signed-off-by: NDavid Teigland <teigland@redhat.com>

Record the time the latest blocking callback was queued for
a lock.  This will be used for debugging in combination with
lock queue timestamp changes in the previous patch.
Signed-off-by: NDavid Teigland <teigland@redhat.com>

Use ktime instead of jiffies for timestamping lkb's.  Also stamp the
time on every lkb whenever it's added to a resource queue, instead of
just stamping locks subject to timeouts.  This will allow us to use
timestamps more widely for debugging all locks.
Signed-off-by: NDavid Teigland <teigland@redhat.com>

The lkb bastmode value is set in the context of processing the
lock, and read by the dlm_astd thread.  Because it's accessed
in these two separate contexts, the writing/reading ought to
be done under a lock.  This is simple to do by setting it and
reading it when the lkb is added to and removed from dlm_astd's
callback list which is properly locked.
Signed-off-by: NDavid Teigland <teigland@redhat.com>

gcc 4.3.0 correctly emits the following warning.
search_rsb_list does not *r_ret if no dlm_rsb is found
and _search_rsb may pass the uninitialized value upstream
on the error path when both calls to search_rsb_list
return non-zero error.

The fix sets *r_ret to NULL on search_rsb_list's not-found path.
Signed-off-by: NBenny Halevy <bhalevy@panasas.com>
Signed-off-by: NDavid Teigland <teigland@redhat.com>

The fix in commit 36509258 was addressing
the case of a granted PR lock with waiting PR and CW locks.  It's a
special case that requires forcing a CW bast.  However, that forced CW
bast was incorrectly applying to a second condition where the granted
lock was CW.  So, the holder of a CW lock could receive an extraneous CW
bast instead of a PR bast.  This fix narrows the original special case to
what was intended.
Signed-off-by: NDavid Teigland <teigland@redhat.com>

When a NOQUEUE request fails, the rsb res_master field is unnecessarily
reset to -1, instead of leaving the valid master setting in place.  We
want to save the looked-up master values while the rsb is on the "toss
list" so that another lookup can be avoided if the rsb is soon reused.
The fix is to simply leave res_master value alone.
Signed-off-by: NDavid Teigland <teigland@redhat.com>

dlm_print_rsb() can now become static.
Signed-off-by: NAdrian Bunk <bunk@kernel.org>
Signed-off-by: NDavid Teigland <teigland@redhat.com>