Instead of creating a worker for each cwq and putting it into the
shared pool, manage per-cpu workers dynamically.

Works aren't supposed to be cpu cycle hogs and maintaining just enough
concurrency to prevent work processing from stalling due to lack of
processing context is optimal.  gcwq keeps the number of concurrent
active workers to minimum but no less.  As long as there's one or more
running workers on the cpu, no new worker is scheduled so that works
can be processed in batch as much as possible but when the last
running worker blocks, gcwq immediately schedules new worker so that
the cpu doesn't sit idle while there are works to be processed.

gcwq always keeps at least single idle worker around.  When a new
worker is necessary and the worker is the last idle one, the worker
assumes the role of "manager" and manages the worker pool -
ie. creates another worker.  Forward-progress is guaranteed by having
dedicated rescue workers for workqueues which may be necessary while
creating a new worker.  When the manager is having problem creating a
new worker, mayday timer activates and rescue workers are summoned to
the cpu and execute works which might be necessary to create new
workers.

Trustee is expanded to serve the role of manager while a CPU is being
taken down and stays down.  As no new works are supposed to be queued
on a dead cpu, it just needs to drain all the existing ones.  Trustee
continues to try to create new workers and summon rescuers as long as
there are pending works.  If the CPU is brought back up while the
trustee is still trying to drain the gcwq from the previous offlining,
the trustee will kill all idles ones and tell workers which are still
busy to rebind to the cpu, and pass control over to gcwq which assumes
the manager role as necessary.

Concurrency managed worker pool reduces the number of workers
drastically.  Only workers which are necessary to keep the processing
going are created and kept.  Also, it reduces cache footprint by
avoiding unnecessarily switching contexts between different workers.

Please note that this patch does not increase max_active of any
workqueue.  All workqueues can still only process one work per cpu.
Signed-off-by: NTejun Heo <tj@kernel.org>

Implement worker_{set|clr}_flags() to manipulate worker flags.  These
are currently simple wrappers but logics to track the current worker
state and the current level of concurrency will be added.
Signed-off-by: NTejun Heo <tj@kernel.org>

Use gcwq->worklist instead of cwq->worklist and break the strict
association between a cwq and its worker.  All works queued on a cpu
are queued on gcwq->worklist and processed by any available worker on
the gcwq.

As there no longer is strict association between a cwq and its worker,
whether a work is executing can now only be determined by calling
[__]find_worker_executing_work().

After this change, the only association between a cwq and its worker
is that a cwq puts a worker into shared worker pool on creation and
kills it on destruction.  As all workqueues are still limited to
max_active of one, this means that there are always at least as many
workers as active works and thus there's no danger for deadlock.

The break of strong association between cwqs and workers requires
somewhat clumsy changes to current_is_keventd() and
destroy_workqueue().  Dynamic worker pool management will remove both
clumsy changes.  current_is_keventd() won't be necessary at all as the
only reason it exists is to avoid queueing a work from a work which
will be allowed just fine.  The clumsy part of destroy_workqueue() is
added because a worker can only be destroyed while idle and there's no
guarantee a worker is idle when its wq is going down.  With dynamic
pool management, workers are not associated with workqueues at all and
only idle ones will be submitted to destroy_workqueue() so the code
won't be necessary anymore.
Signed-off-by: NTejun Heo <tj@kernel.org>

With gcwq managing all the workers and work->data pointing to the last
gcwq it was on, non-reentrance can be easily implemented by checking
whether the work is still running on the previous gcwq on queueing.
Implement it.
Signed-off-by: NTejun Heo <tj@kernel.org>

To implement non-reentrant workqueue, the last gcwq a work was
executed on must be reliably obtainable as long as the work structure
is valid even if the previous workqueue has been destroyed.

To achieve this, work->data will be overloaded to carry the last cpu
number once execution starts so that the previous gcwq can be located
reliably.  This means that cwq can't be obtained from work after
execution starts but only gcwq.

Implement set_work_{cwq|cpu}(), get_work_[g]cwq() and
clear_work_data() to set work data to the cpu number when starting
execution, access the overloaded work data and clear it after
cancellation.

queue_delayed_work_on() is updated to preserve the last cpu while
in-flight in timer and other callers which depended on getting cwq
from work after execution starts are converted to depend on gcwq
instead.

* Anton Blanchard fixed compile error on powerpc due to missing
  linux/threads.h include.
Signed-off-by: NTejun Heo <tj@kernel.org>
Cc: Anton Blanchard <anton@samba.org>

Now that all the workers are tracked by gcwq, we can find which worker
is executing a work from gcwq.  Implement find_worker_executing_work()
and make worker track its current_cwq so that we can find things the
other way around.  This will be used to implement non-reentrant wqs.
Signed-off-by: NTejun Heo <tj@kernel.org>

Reimplement st (single thread) workqueue so that it's friendly to
shared worker pool.  It was originally implemented by confining st
workqueues to use cwq of a fixed cpu and always having a worker for
the cpu.  This implementation isn't very friendly to shared worker
pool and suboptimal in that it ends up crossing cpu boundaries often.

Reimplement st workqueue using dynamic single cpu binding and
cwq->limit.  WQ_SINGLE_THREAD is replaced with WQ_SINGLE_CPU.  In a
single cpu workqueue, at most single cwq is bound to the wq at any
given time.  Arbitration is done using atomic accesses to
wq->single_cpu when queueing a work.  Once bound, the binding stays
till the workqueue is drained.

Note that the binding is never broken while a workqueue is frozen.
This is because idle cwqs may have works waiting in delayed_works
queue while frozen.  On thaw, the cwq is restarted if there are any
delayed works or unbound otherwise.

When combined with max_active limit of 1, single cpu workqueue has
exactly the same execution properties as the original single thread
workqueue while allowing sharing of per-cpu workers.
Signed-off-by: NTejun Heo <tj@kernel.org>

Reimplement CPU hotplugging support using trustee thread.  On CPU
down, a trustee thread is created and each step of CPU down is
executed by the trustee and workqueue_cpu_callback() simply drives and
waits for trustee state transitions.

CPU down operation no longer waits for works to be drained but trustee
sticks around till all pending works have been completed.  If CPU is
brought back up while works are still draining,
workqueue_cpu_callback() tells trustee to step down and tell workers
to rebind to the cpu.

As it's difficult to tell whether cwqs are empty if it's freezing or
frozen, trustee doesn't consider draining to be complete while a gcwq
is freezing or frozen (tracked by new GCWQ_FREEZING flag).  Also,
workers which get unbound from their cpu are marked with WORKER_ROGUE.

Trustee based implementation doesn't bring any new feature at this
point but it will be used to manage worker pool when dynamic shared
worker pool is implemented.
Signed-off-by: NTejun Heo <tj@kernel.org>

Implement worker states.  After created, a worker is STARTED.  While a
worker isn't processing a work, it's IDLE and chained on
gcwq->idle_list.  While processing a work, a worker is BUSY and
chained on gcwq->busy_hash.  Also, gcwq now counts the number of all
workers and idle ones.

worker_thread() is restructured to reflect state transitions.
cwq->more_work is removed and waking up a worker makes it check for
events.  A worker is killed by setting DIE flag while it's IDLE and
waking it up.

This gives gcwq better visibility of what's going on and allows it to
find out whether a work is executing quickly which is necessary to
have multiple workers processing the same cwq.
Signed-off-by: NTejun Heo <tj@kernel.org>

There is one gcwq (global cwq) per each cpu and all cwqs on an cpu
point to it.  A gcwq contains a lock to be used by all cwqs on the cpu
and an ida to give IDs to workers belonging to the cpu.

This patch introduces gcwq, moves worker_ida into gcwq and make all
cwqs on the same cpu use the cpu's gcwq->lock instead of separate
locks.  gcwq->ida is now protected by gcwq->lock too.
Signed-off-by: NTejun Heo <tj@kernel.org>

Currently, workqueue freezing is implemented by marking the worker
freezeable and calling try_to_freeze() from dispatch loop.
Reimplement it using cwq->limit so that the workqueue is frozen
instead of the worker.

* workqueue_struct->saved_max_active is added which stores the
  specified max_active on initialization.

* On freeze, all cwq->max_active's are quenched to zero.  Freezing is
  complete when nr_active on all cwqs reach zero.

* On thaw, all cwq->max_active's are restored to wq->saved_max_active
  and the worklist is repopulated.

This new implementation allows having single shared pool of workers
per cpu.
Signed-off-by: NTejun Heo <tj@kernel.org>

Add cwq->nr_active, cwq->max_active and cwq->delayed_work.  nr_active
counts the number of active works per cwq.  A work is active if it's
flushable (colored) and is on cwq's worklist.  If nr_active reaches
max_active, new works are queued on cwq->delayed_work and activated
later as works on the cwq complete and decrement nr_active.

cwq->max_active can be specified via the new @max_active parameter to
__create_workqueue() and is set to 1 for all workqueues for now.  As
each cwq has only single worker now, this double queueing doesn't
cause any behavior difference visible to its users.

This will be used to reimplement freeze/thaw and implement shared
worker pool.
Signed-off-by: NTejun Heo <tj@kernel.org>

A work is linked to the next one by having WORK_STRUCT_LINKED bit set
and these links can be chained.  When a linked work is dispatched to a
worker, all linked works are dispatched to the worker's newly added
->scheduled queue and processed back-to-back.

Currently, as there's only single worker per cwq, having linked works
doesn't make any visible behavior difference.  This change is to
prepare for multiple shared workers per cpu.
Signed-off-by: NTejun Heo <tj@kernel.org>

Separate out worker thread related information to struct worker from
struct cpu_workqueue_struct and implement helper functions to deal
with the new struct worker.  The only change which is visible outside
is that now workqueue worker are all named "kworker/CPUID:WORKERID"
where WORKERID is allocated from per-cpu ida.

This is in preparation of concurrency managed workqueue where shared
multiple workers would be available per cpu.
Signed-off-by: NTejun Heo <tj@kernel.org>

Reimplement workqueue flushing using color coded works.  wq has the
current work color which is painted on the works being issued via
cwqs.  Flushing a workqueue is achieved by advancing the current work
colors of cwqs and waiting for all the works which have any of the
previous colors to drain.

Currently there are 16 possible colors, one is reserved for no color
and 15 colors are useable allowing 14 concurrent flushes.  When color
space gets full, flush attempts are batched up and processed together
when color frees up, so even with many concurrent flushers, the new
implementation won't build up huge queue of flushers which has to be
processed one after another.

Only works which are queued via __queue_work() are colored.  Works
which are directly put on queue using insert_work() use NO_COLOR and
don't participate in workqueue flushing.  Currently only works used
for work-specific flush fall in this category.

This new implementation leaves only cleanup_workqueue_thread() as the
user of flush_cpu_workqueue().  Just make its users use
flush_workqueue() and kthread_stop() directly and kill
cleanup_workqueue_thread().  As workqueue flushing doesn't use barrier
request anymore, the comment describing the complex synchronization
around it in cleanup_workqueue_thread() is removed together with the
function.

This new implementation is to allow having and sharing multiple
workers per cpu.

Please note that one more bit is reserved for a future work flag by
this patch.  This is to avoid shifting bits and updating comments
later.
Signed-off-by: NTejun Heo <tj@kernel.org>

work->data field is used for two purposes.  It points to cwq it's
queued on and the lower bits are used for flags.  Currently, two bits
are reserved which is always safe as 4 byte alignment is guaranteed on
every architecture.  However, future changes will need more flag bits.

On SMP, the percpu allocator is capable of honoring larger alignment
(there are other users which depend on it) and larger alignment works
just fine.  On UP, percpu allocator is a thin wrapper around
kzalloc/kfree() and don't honor alignment request.

This patch introduces WORK_STRUCT_FLAG_BITS and implements
alloc/free_cwqs() which guarantees max(1 << WORK_STRUCT_FLAG_BITS,
__alignof__(unsigned long long) alignment both on SMP and UP.  On SMP,
simply wrapping percpu allocator is enough.  On UP, extra space is
allocated so that cwq can be aligned and the original pointer can be
stored after it which is used in the free path.

* Alignment problem on UP is reported by Michal Simek.
Signed-off-by: NTejun Heo <tj@kernel.org>
Cc: Christoph Lameter <cl@linux-foundation.org>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Reported-by: NMichal Simek <michal.simek@petalogix.com>

Worker management is about to be overhauled.  Simplify things by
removing cpu_populated_map, creating workers for all possible cpus and
making single threaded workqueues behave more like multi threaded
ones.

After this patch, all cwqs are always initialized, all workqueues are
linked on the workqueues list and workers for all possibles cpus
always exist.  This also makes CPU hotplug support simpler - checking
->cpus_allowed before processing works in worker_thread() and flushing
cwqs on CPU_POST_DEAD are enough.

While at it, make get_cwq() always return the cwq for the specified
cpu, add target_cwq() for cases where single thread distinction is
necessary and drop all direct usage of per_cpu_ptr() on wq->cpu_wq.
Signed-off-by: NTejun Heo <tj@kernel.org>

Strip tracing code from workqueue and remove workqueue tracing.  This
is temporary measure till concurrency managed workqueue is complete.
Signed-off-by: NTejun Heo <tj@kernel.org>
Cc: Frederic Weisbecker <fweisbec@gmail.com>

Separate out process_one_work() out of run_workqueue().  This patch
doesn't cause any behavior change.
Signed-off-by: NTejun Heo <tj@kernel.org>

Work flags are about to see more traditional mask handling.  Define
WORK_STRUCT_*_BIT as the bit position constant and redefine
WORK_STRUCT_* as bit masks.  Also, make WORK_STRUCT_STATIC_* flags
conditional

While at it, re-define these constants as enums and use
WORK_STRUCT_STATIC instead of hard-coding 2 in
WORK_DATA_STATIC_INIT().
Signed-off-by: NTejun Heo <tj@kernel.org>

Currently, __create_workqueue_key() takes @singlethread and
@freezeable paramters and store them separately in workqueue_struct.
Merge them into a single flags parameter and field and use
WQ_FREEZEABLE and WQ_SINGLE_THREAD.
Signed-off-by: NTejun Heo <tj@kernel.org>

Make the following updates in preparation of concurrency managed
workqueue.  None of these changes causes any visible behavior
difference.

* Add comments and adjust indentations to data structures and several
  functions.

* Rename wq_per_cpu() to get_cwq() and swap the position of two
  parameters for consistency.  Convert a direct per_cpu_ptr() access
  to wq->cpu_wq to get_cwq().

* Add work_static() and Update set_wq_data() such that it sets the
  flags part to WORK_STRUCT_PENDING | WORK_STRUCT_STATIC if static |
  @extra_flags.

* Move santiy check on work->entry emptiness from queue_work_on() to
  __queue_work() which all queueing paths share.

* Make __queue_work() take @cpu and @wq instead of @cwq.

* Restructure flush_work() and __create_workqueue_key() to make them
  easier to modify.
Signed-off-by: NTejun Heo <tj@kernel.org>

With stop_machine() converted to use cpu_stop, RT workqueue doesn't
have any user left.  Kill RT workqueue support.
Signed-off-by: NTejun Heo <tj@kernel.org>

By the previous modification, the cpu notifier can return encapsulate
errno value.  This converts the cpu notifiers for kernel/*.c
Signed-off-by: NAkinobu Mita <akinobu.mita@gmail.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: NAndrew Morton <akpm@linux-foundation.org>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

In short: change cancel_work_sync(work) to mark this work as "never
queued" upon return.

When cancel_work_sync(work) succeeds, we know that this work can't be
queued or running, and since we own WORK_STRUCT_PENDING nobody can change
the bits in work->data under us. This means we can also clear the "cwq"
part along with _PENDING bit lockless before return, unless the work is
queued nobody can assume get_wq_data() is stable even under cwq->lock.

This change can speedup the subsequent cancel/flush requests, and as
Dmitry pointed out this simplifies the usage of work_struct's which
can be queued on different workqueues. Consider this pseudo code from
the input subsystem:

	struct workqueue_struct *WQ;
	struct work_struct *WORK;

	for (;;) {
		WQ = create_workqueue();
		...
		if (condition())
			queue_work(WQ, WORK);
		...
		cancel_work_sync(WORK);
		destroy_workqueue(WQ);
	}

If condition() returns T and then F, cancel_work_sync() will crash the
kernel because WORK->data still points to the already destroyed workqueue.
With this patch the code like above becomes correct.
Suggested-by: NDmitry Torokhov <dmitry.torokhov@gmail.com>
Signed-off-by: NOleg Nesterov <oleg@redhat.com>
Signed-off-by: NTejun Heo <tj@kernel.org>

This patch (as1319) adds kerneldoc and a pointed warning to
flush_scheduled_work().
Signed-off-by: NAlan Stern <stern@rowland.harvard.edu>
Signed-off-by: NTejun Heo <tj@kernel.org>

flush_delayed_work() always uses keventd_wq for re-queueing,
but it should use the workqueue this dwork was queued on.
Signed-off-by: NOleg Nesterov <oleg@redhat.com>
Signed-off-by: NTejun Heo <tj@kernel.org>

Commit 65a64464 ("HWPOISON: Allow
schedule_on_each_cpu() from keventd") which allows schedule_on_each_cpu()
to be called from keventd added a race condition.  schedule_on_each_cpu()
may race with cpu hotplug and end up executing the function twice on a
cpu.

Fix it by moving direct execution into the section protected with
get/put_online_cpus().  While at it, update code such that direct
execution is done after works have been scheduled for all other cpus and
drop unnecessary cpu != orig test from flush loop.
Signed-off-by: NTejun Heo <tj@kernel.org>
Cc: Andi Kleen <ak@linux.intel.com>
Acked-by: NOleg Nesterov <oleg@redhat.com>
Cc: Ingo Molnar <mingo@redhat.com>
Signed-off-by: NAndrew Morton <akpm@linux-foundation.org>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

Add debugobject support to track the life time of work_structs.

While at it, remove duplicate definition of
INIT_DELAYED_WORK_ON_STACK().
Signed-off-by: NThomas Gleixner <tglx@linutronix.de>
Signed-off-by: NTejun Heo <tj@kernel.org>

Right now when calling schedule_on_each_cpu() from keventd there
is a deadlock because it tries to schedule a work item on the current CPU
too. This happens via lru_add_drain_all() in hwpoison.

Just call the function for the current CPU in this case. This is actually
faster too.

Debugging with Fengguang Wu & Max Asbock
Signed-off-by: NAndi Kleen <ak@linux.intel.com>

It basically turns a delayed work into an immediate work, and then waits
for it to finish, thus allowing you to force (and wait for) an immediate
flush of a delayed work.

We'll want to use this in the tty layer to clean up tty_flush_to_ldisc().
Acked-by: NOleg Nesterov <oleg@redhat.com>
[ Fixed to use 'del_timer_sync()' as noted by Oleg ]
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

It basically turns a delayed work into an immediate work, and then waits
for it to finish.

Removes kthread/workqueue priority boost, they increase worst-case
desktop latencies.
Signed-off-by: NMike Galbraith <efault@gmx.de>
Acked-by: NPeter Zijlstra <a.p.zijlstra@chello.nl>
LKML-Reference: <1252486344.28645.18.camel@marge.simson.net>
Signed-off-by: NIngo Molnar <mingo@elte.hu>

Two important aspects of the schedule_work() function are not
yet documented:

 - that it is allowed to pass a struct work_struct * to this
   function that is already on the kernel-global workqueue;

 - the meaning of its return value.

The patch below documents both aspects.
Signed-off-by: NBart Van Assche <bart.vanassche@gmail.com>
Cc: "Greg Kroah-Hartman" <gregkh@suse.de>
Cc: Andrew Morton <akpm@linux-foundation.org>
LKML-Reference: <200907301900.54202.bart.vanassche@gmail.com>
Signed-off-by: NIngo Molnar <mingo@elte.hu>

v3: zhaolei@cn.fujitsu.com: Change TRACE_EVENT definition to new format
    introduced by Steven Rostedt: consolidate trace and trace_event headers
v2: kosaki@jp.fujitsu.com: print the function names instead of addr, and zap
    the work addr
v1: zhaolei@cn.fujitsu.com: Make workqueue tracepoints use TRACE_EVENT macro

TRACE_EVENT is a more generic way to define tracepoints.
Doing so adds these new capabilities to the tracepoints:

  - zero-copy and per-cpu splice() tracing
  - binary tracing without printf overhead
  - structured logging records exposed under /debug/tracing/events
  - trace events embedded in function tracer output and other plugins
  - user-defined, per tracepoint filter expressions

Then, this patch converts DEFINE_TRACE to TRACE_EVENT in workqueue related
tracepoints.

[ Impact: expand workqueue tracer to events tracing ]
Signed-off-by: NZhao Lei <zhaolei@cn.fujitsu.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Tom Zanussi <tzanussi@gmail.com>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: NKOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Signed-off-by: NFrederic Weisbecker <fweisbec@gmail.com>

Impact: circular locking bugfix

The various implemetnations and proposed implemetnations of work_on_cpu()
are vulnerable to various deadlocks because they all used queues of some
form.

Unrelated pieces of kernel code thus gained dependencies wherein if one
work_on_cpu() caller holds a lock which some other work_on_cpu() callback
also takes, the kernel could rarely deadlock.

Fix this by creating a short-lived kernel thread for each work_on_cpu()
invokation.

This is not terribly fast, but the only current caller of work_on_cpu() is
pci_call_probe().

It would be nice to find some other way of doing the node-local
allocations in the PCI probe code so that we can zap work_on_cpu()
altogether.  The code there is rather nasty.  I can't think of anything
simple at this time...

Cc: Ingo Molnar <mingo@elte.hu>
Signed-off-by: NAndrew Morton <akpm@linux-foundation.org>
Signed-off-by: NRusty Russell <rusty@rustcorp.com.au>

1) lockdep will complain when run_workqueue() performs recursion.

2) The recursive implementation of run_workqueue() means that
   flush_workqueue() and its documentation are inconsistent.  This may
   hide deadlocks and other bugs.

3) The recursion in run_workqueue() will poison cwq->current_work, but
   flush_work() and __cancel_work_timer(), etcetera need a reliable
   cwq->current_work.
Signed-off-by: NLai Jiangshan <laijs@cn.fujitsu.com>
Acked-by: NOleg Nesterov <oleg@redhat.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Eric Dumazet <dada1@cosmosbay.com>
Cc: Rusty Russell <rusty@rustcorp.com.au>
Signed-off-by: NAndrew Morton <akpm@linux-foundation.org>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

Impact: cleanup

Time to clean up remaining laggards using the old cpu_ functions.
Signed-off-by: NRusty Russell <rusty@rustcorp.com.au>
Cc: Greg Kroah-Hartman <gregkh@suse.de>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Trond.Myklebust@netapp.com

Impact: remove potential clashes with generic kevent workqueue

Annoyingly, some places we want to use work_on_cpu are already in
workqueues.  As per Ingo's suggestion, we create a different workqueue
for work_on_cpu.
Signed-off-by: NRusty Russell <rusty@rustcorp.com.au>
Signed-off-by: NMike Travis <travis@sgi.com>
Signed-off-by: NIngo Molnar <mingo@elte.hu>

Impact: remove potential circular lock dependency with cpu hotplug lock

This has caused more problems than it solved, with a pile of cpu
hotplug locking issues.

Followup patches will get_online_cpus() in callers that need it, but
if they don't do it they're no worse than before when they were using
set_cpus_allowed without locking.
Signed-off-by: NRusty Russell <rusty@rustcorp.com.au>
Signed-off-by: NMike Travis <travis@sgi.com>
Signed-off-by: NIngo Molnar <mingo@elte.hu>