Skip to contentSkip to conversation

workqueue for v2.6.36

GIT PULL Tejun Heo 5 participants 8 messages Aug 4, 2010 – Aug 9, 2010

lore
Share & tools
lore

Tejun Heo

author
#0 lore raw

Hello, Linus.

Please consider pulling from

git://git.kernel.org/pub/scm/linux/kernel/git/tj/wq.git for-linus

to receive the concurrencey managed workqueue patches. The branch contains 32 patches to prepare for and implement cmwq and 23 patches fixing bugs and converting libata, async, fscache and other slow-work users to workqueue and remove slow-work.

The following overview section gives a brief overview. For more detailed information, please refer to the last posting of cmwq patchset.

http://thread.gmane.org/gmane.linux.kernel/1003710

Most objections have been addressed and all the contained conversions have been acked by respective subsystem maintainers.

One that wasn't addressed was Daniel Walker's objection on the ground that cmwq would make it impossible to adjust priorities of workqueue threads which can be useful as an ad-hoc optimization. I don't plan to address this concern (suggested solution is to add userland visible knobs to adjust workqueue priorities) at this point because it is an implementation detail that userspace shouldn't diddle with in the first place. For details, please read the following thread.

http://thread.gmane.org/gmane.linux.kernel/998652/focus=999232

Thanks.

OVERVIEW
========

The bulk of changes is concentrated on making all the different workqueues share per-cpu global worker pools, which greatly lessens up-front resource requirement per workqueue thus increasing scalability and reducing use case constraints.

One major restriction which is removed by the use of shared worker pool is the level of concurrency per workqueue. Normal workqueues only provide one execution context per cpu, single cpu workqueues one per each workqueue. This often introduces unnecessary and irregular latencies in work execution and easily creates deadlocks around execution resources. With shared worker pool, workqueues can easily provide high level of concurrency and most of the issues become marginal.

The 'concurreny-managed' part of name comes from how each per-cpu global worker pool manages its concurrency. It hooks into scheduler code and tracks the number of runnable workers and starts executing new works iff it reaches zero. This maintains just enough level of concurrency without depending on fragile heuristics which are usually needed for thread pools. In most cases, workqueues are used as a way to obtain a sleepable execution context (ie. they don't burn a lot of cpu cycles) and the minimal level of concurrency fits this usage model very well - it doesn't add to latency while maximizing batch execution and reuse of workers.

The basics of cmwq haven't changed much since its initial posting from about a year ago. Most of updates were regarding interaction w/ scheduler and features which were necessary to convert users which were using private pools. On macro level, the followings are notable.

* WQ_NON_REENTRANT ordering. By default, workqueues retain the same loose execution semantics where only non-reentrancy on the same CPU is guaranteed. WQ_NON_REENTRANT guarantees non-reetrancy across all CPUs. This is useful for single CPU workqueue users which don't really need full ordering.

* WQ_CPU_INTENSIVE. This is created to serve cpu-bound cpu intensive workloads. Works which may consume a lot of cpu cycles shouldn't participate in concurrency management as they may block other works for a long time.

* WQ_HIGHPRI for highpri workqueues. Works scheduled on highpri workqueues are queued at the head of global work queue.

* Unbound workqueue. Workqueues created with WQ_UNBOUND is not bound to any specific workqueue and basically behaves as simple thread pool which spawns and assigns workers on-demand. This is used for cases where there can be a lot of long running cpu intensive workers which can be better served by regular thread scheduling. It's also used to serve single cpu workqueues as managing concurrency isn't as useful for them and unbound workers are handled as if they all are on the same cpu making implementing the ordering requirement trivial.

CURRENT STATE AND TODOS
=======================

The core code has been mostly stable for some time and conversions of different types (libata taking advantage of the flexibility of cmwq, replacement of backend worker pool for async, replacement of slow-work mechanism) were successfully done and acked by respective maintainers. TODO items are...

* Currently, a lot of workqueues needlessly are single CPU and/or have WQ_RESCUER set through safe default conversion of create_workqueue() wrappers. Audit each workqueue users and convert them to use new alloc_workqueue() function w/ only necessary restrictions and features.

* Conversions of other private worker pools. Writeback worker pool is currently being worked on and SCSI EH pool would probably follow.

* Debug facilities using the tracing API.

* (maybe) Better lockdep annotation. The current lockdep annotation still assumes single execution context per cpu.

* Documentation (probably from previous patchset head messages).

MERGE CONFLICTS AND RESOLUSTIONS
================================

Merging with the current mainline results in the following three conflicts. All of them are under fs/cifs/.

1. fs/cifs/cifsfs.c

This is between cmwq conversion dropping slow-work clean up path and cifs updating DFL_UPCALL cleanup path. As there's no later failure path, just removing the updated function in the cleanup path is enough.

#ifdef CONFIG_CIFS_DFS_UPCALL
<<<<<<< HEAD
=======

	   cifs_exit_dns_resolver();
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

3a09b1be53d23df780a0cd0e4087a05e2ca4a00c

out_unregister_key_type:
#endif

Resolution

   #ifdef CONFIG_CIFS_DFS_UPCALL
    out_unregister_key_type:
   #endif

2. fs/cifs/file.c

This is simple context conflict.

<<<<<<< HEAD
void cifs_oplock_break(struct work_struct *work)
=======
static int cifs_release_page(struct page *page, gfp_t gfp)

   {
	   if (PagePrivate(page))
		   return 0;

	   return cifs_fscache_release_page(page, gfp);

}

static void cifs_invalidate_page(struct page *page, unsigned long offset)

   {
	   struct cifsInodeInfo *cifsi = CIFS_I(page->mapping->host);

	   if (offset == 0)
		   cifs_fscache_invalidate_page(page, &cifsi->vfs_inode);

}

static void
cifs_oplock_break(struct slow_work *work)

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

3a09b1be53d23df780a0cd0e4087a05e2ca4a00c

Resolution

static int cifs_release_page(struct page *page, gfp_t gfp)

 {
	 if (PagePrivate(page))
		 return 0;

	 return cifs_fscache_release_page(page, gfp);

}

static void cifs_invalidate_page(struct page *page, unsigned long offset)

 {
	 struct cifsInodeInfo *cifsi = CIFS_I(page->mapping->host);

	 if (offset == 0)
		 cifs_fscache_invalidate_page(page, &cifsi->vfs_inode);

}

void cifs_oplock_break(struct work_struct *work)

3. fs/cifs/cifsglob.h

Another context conflict.

    <<<<<<< HEAD
    void cifs_oplock_break(struct work_struct *work);
    void cifs_oplock_break_get(struct cifsFileInfo *cfile);
    void cifs_oplock_break_put(struct cifsFileInfo *cfile);
    =======
    extern const struct slow_work_ops cifs_oplock_break_ops;

    #endif	/* _CIFS_GLOB_H */
    >>>>>>> 3a09b1be53d23df780a0cd0e4087a05e2ca4a00c

Resolution

    void cifs_oplock_break(struct work_struct *work);
    void cifs_oplock_break_get(struct cifsFileInfo *cfile);
    void cifs_oplock_break_put(struct cifsFileInfo *cfile);

    extern const struct slow_work_ops cifs_oplock_break_ops;

    #endif	/* _CIFS_GLOB_H */

COMMITS AND CHANGES
===================

Suresh Siddha (1):
workqueue: mark init_workqueues() as early_initcall()

Tejun Heo (54):
kthread: implement kthread_worker
ivtv: use kthread_worker instead of workqueue
kthread: implement kthread_data()
acpi: use queue_work_on() instead of binding workqueue worker to cpu0
workqueue: kill RT workqueue
workqueue: misc/cosmetic updates
workqueue: merge feature parameters into flags
workqueue: define masks for work flags and conditionalize STATIC flags
workqueue: separate out process_one_work()
workqueue: temporarily remove workqueue tracing
workqueue: kill cpu_populated_map
workqueue: update cwq alignement
workqueue: reimplement workqueue flushing using color coded works
workqueue: introduce worker
workqueue: reimplement work flushing using linked works
workqueue: implement per-cwq active work limit
workqueue: reimplement workqueue freeze using max_active
workqueue: introduce global cwq and unify cwq locks
workqueue: implement worker states
workqueue: reimplement CPU hotplugging support using trustee
workqueue: make single thread workqueue shared worker pool friendly
workqueue: add find_worker_executing_work() and track current_cwq
workqueue: carry cpu number in work data once execution starts
workqueue: implement WQ_NON_REENTRANT
workqueue: use shared worklist and pool all workers per cpu
workqueue: implement worker_{set|clr}_flags()
workqueue: implement concurrency managed dynamic worker pool
workqueue: increase max_active of keventd and kill current_is_keventd()
workqueue: s/__create_workqueue()/alloc_workqueue()/, and add system workqueues
workqueue: implement several utility APIs
workqueue: implement high priority workqueue
workqueue: implement cpu intensive workqueue
workqueue: use worker_set/clr_flags() only from worker itself
workqueue: fix race condition in flush_workqueue()
workqueue: fix incorrect cpu number BUG_ON() in get_work_gcwq()
workqueue: fix worker management invocation without pending works
libata: take advantage of cmwq and remove concurrency limitations
workqueue: prepare for WQ_UNBOUND implementation
workqueue: implement unbound workqueue
workqueue: remove WQ_SINGLE_CPU and use WQ_UNBOUND instead
async: use workqueue for worker pool
workqueue: fix locking in retry path of maybe_create_worker()
workqueue: fix build problem on !CONFIG_SMP
workqueue: fix mayday_mask handling on UP
workqueue: fix how cpu number is stored in work->data
fscache: convert object to use workqueue instead of slow-work
fscache: convert operation to use workqueue instead of slow-work
fscache: drop references to slow-work
cifs: use workqueue instead of slow-work
drm: use workqueue instead of slow-work
gfs2: use workqueue instead of slow-work
slow-work: kill it
fscache: fix build on !CONFIG_SYSCTL
workqueue: explain for_each_*cwq_cpu() iterators

 Documentation/filesystems/caching/fscache.txt |   10 +-
 Documentation/slow-work.txt                   |  322 ---
 arch/ia64/kernel/smpboot.c                    |    2 +-
 arch/x86/kernel/smpboot.c                     |    2 +-
 drivers/acpi/osl.c                            |   40 +-
 drivers/ata/libata-core.c                     |   20 +-
 drivers/ata/libata-eh.c                       |    4 +-
 drivers/ata/libata-scsi.c                     |   10 +-
 drivers/ata/libata-sff.c                      |    9 +-
 drivers/ata/libata.h                          |    1 -
 drivers/gpu/drm/drm_crtc_helper.c             |   29 +-
 drivers/media/video/ivtv/ivtv-driver.c        |   26 +-
 drivers/media/video/ivtv/ivtv-driver.h        |    8 +-
 drivers/media/video/ivtv/ivtv-irq.c           |   15 +-
 drivers/media/video/ivtv/ivtv-irq.h           |    2 +-
 fs/cachefiles/namei.c                         |   13 +-
 fs/cachefiles/rdwr.c                          |    4 +-
 fs/cifs/Kconfig                               |    1 -
 fs/cifs/cifsfs.c                              |    5 -
 fs/cifs/cifsglob.h                            |    8 +-
 fs/cifs/dir.c                                 |    2 +-
 fs/cifs/file.c                                |   30 +-
 fs/cifs/misc.c                                |   20 +-
 fs/fscache/Kconfig                            |    1 -
 fs/fscache/internal.h                         |    8 +
 fs/fscache/main.c                             |  106 +-
 fs/fscache/object-list.c                      |   11 +-
 fs/fscache/object.c                           |  106 +-
 fs/fscache/operation.c                        |   67 +-
 fs/fscache/page.c                             |   36 +-
 fs/gfs2/Kconfig                               |    1 -
 fs/gfs2/incore.h                              |    3 +-
 fs/gfs2/main.c                                |   14 +-
 fs/gfs2/ops_fstype.c                          |    8 +-
 fs/gfs2/recovery.c                            |   54 +-
 fs/gfs2/recovery.h                            |    6 +-
 fs/gfs2/sys.c                                 |    3 +-
 include/drm/drm_crtc.h                        |    3 +-
 include/linux/cpu.h                           |    2 +
 include/linux/fscache-cache.h                 |   47 +-
 include/linux/kthread.h                       |   65 +
 include/linux/libata.h                        |    1 +
 include/linux/slow-work.h                     |  163 --
 include/linux/workqueue.h                     |  154 +-
 include/trace/events/workqueue.h              |   92 -
 init/Kconfig                                  |   24 -
 init/main.c                                   |    2 -
 kernel/Makefile                               |    2 -
 kernel/async.c                                |  141 +-
 kernel/kthread.c                              |  164 ++
 kernel/power/process.c                        |   21 +-
 kernel/slow-work-debugfs.c                    |  227 --
 kernel/slow-work.c                            | 1068 ---------
 kernel/slow-work.h                            |   72 -
 kernel/sysctl.c                               |    8 -
 kernel/trace/Kconfig                          |   11 -
 kernel/workqueue.c                            | 3160 +++++++++++++++++++++----
 kernel/workqueue_sched.h                      |   13 +-
 58 files changed, 3505 insertions(+), 2942 deletions(-)
 delete mode 100644 Documentation/slow-work.txt
 delete mode 100644 include/linux/slow-work.h
 delete mode 100644 include/trace/events/workqueue.h
 delete mode 100644 kernel/slow-work-debugfs.c
 delete mode 100644 kernel/slow-work.c
 delete mode 100644 kernel/slow-work.h

--
tejun

#1 lore raw
Quoted 6 lines · Tejun Heo

Hello, Linus.

Please consider pulling from

git://git.kernel.org/pub/scm/linux/kernel/git/tj/wq.git for-linus

Quoted 8 lines

to receive the concurrencey managed workqueue patches. The branch contains 32 patches to prepare for and implement cmwq and 23 patches fixing bugs and converting libata, async, fscache and other slow-work users to workqueue and remove slow-work.

The following overview section gives a brief overview. For more detailed information, please refer to the last posting of cmwq patchset.

I haven't seen anything that shows your adding back the same expressiveness that your removing .. So I still don't think this should be merged.

Daniel

Signature · 2 lines
Sent by an consultant of the Qualcomm Innovation Center, Inc.
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum.
#2 lore raw
Quoted 3 lines · Daniel Walker

I haven't seen anything that shows your adding back the same expressiveness that your removing .. So I still don't think this should be merged.

Do you mean by expressiveness the ability to hack around a suboptimally working driver in userland (by requiring the administrator to play with kernel thread priorities of dedicated worker threads)? Is it known whether this driver/ these drivers still require this hack after Tejun's patch set is applied? If yes, how about finding someone to fix this driver for good. Meanwhile, let's gets rid of the problem of having both too few and too many workers in countless present usages of the workqueue API please.

Signature · 3 lines
Stefan Richter
-=====-==-=- =--- --=--
http://arcgraph.de/sr/
#3 lore raw
Quoted 5 lines · Stefan Richter
Quoted 3 lines · Daniel Walker

I haven't seen anything that shows your adding back the same expressiveness that your removing .. So I still don't think this should be merged.

Quoted 8 lines

Do you mean by expressiveness the ability to hack around a suboptimally working driver in userland (by requiring the administrator to play with kernel thread priorities of dedicated worker threads)? Is it known whether this driver/ these drivers still require this hack after Tejun's patch set is applied? If yes, how about finding someone to fix this driver for good. Meanwhile, let's gets rid of the problem of having both too few and too many workers in countless present usages of the workqueue API please.

I mean by expressiveness the fact that a user can prioritize their system more fully in the current system.

Daniel

Signature · 2 lines
Sent by an consultant of the Qualcomm Innovation Center, Inc.
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum.
#4 lore raw
In reply to Tejun Heo
Quoted 4 lines · Tejun Heo

Please consider pulling from

 git://git.kernel.org/pub/scm/linux/kernel/git/tj/wq.git for-linus

Hmm. So I pulled, because I think it is the right thing in the long run. But right now I'm bisecting a non-booting machine, and it seems to be due to the workqueue pull.

It's not fully bisected yet, but the only remaining thing I actually enable is workqueues. The symptom is a machine that stops at

...
dracut: Starting plymouth daemon
dracut: Scanning devices sda2 for LVM volume groups
dracut: Reading all physical volumes. This may take a while...
dracut: Found volume group "VolGroup" using metadata type lvm2

and the next thing I'd expect to see is the mounting of the filesystem and the SElinux notifications. And I never do.

I'll bisect it all the way, but thought I'd let you know.

Linus

#5 lore raw
Quoted 19 lines · Linus Torvalds
Quoted 4 lines · Tejun Heo

Please consider pulling from

 git://git.kernel.org/pub/scm/linux/kernel/git/tj/wq.git for-linus

Hmm. So I pulled, because I think it is the right thing in the long run. But right now I'm bisecting a non-booting machine, and it seems to be due to the workqueue pull.

It's not fully bisected yet, but the only remaining thing I actually enable is workqueues. The symptom is a machine that stops at

...
dracut: Starting plymouth daemon
dracut: Scanning devices sda2 for LVM volume groups
dracut: Reading all physical volumes. This may take a while...
dracut: Found volume group "VolGroup" using metadata type lvm2

Quoted 4 lines

and the next thing I'd expect to see is the mounting of the filesystem and the SElinux notifications. And I never do.

I'll bisect it all the way, but thought I'd let you know.

No need to, it's already fixed:
http://lkml.org/lkml/2010/8/9/94

Signature · 2 lines
»A man who doesn't know he is in prison can never escape.«
William S. Burroughs
#6 lore raw
Quoted 6 lines · Markus Trippelsdorf
Quoted 2 lines · Linus Torvalds

I'll bisect it all the way, but thought I'd let you know.

No need to, it's already fixed:
http://lkml.org/lkml/2010/8/9/94

Right you are. That patch fixes it for me too.

Tejun, please send me a pull request for the fixes (or patches). I hate picking up patches that haven't been explicitly sent to me.

Linus

author
#7 lore raw

Hello, Linus.

Quoted 6 lines · Linus Torvalds
Quoted 4 lines · Markus Trippelsdorf
Quoted 2 lines · Linus Torvalds

I'll bisect it all the way, but thought I'd let you know.

Quoted 7 lines
Quoted 2 lines

No need to, it's already fixed:
http://lkml.org/lkml/2010/8/9/94

Right you are. That patch fixes it for me too.

Tejun, please send me a pull request for the fixes (or patches). I hate picking up patches that haven't been explicitly sent to me.

Oh, sorry about that. I was gonna send a pull request tomorrow. Sending it right now.

Thanks.

Signature · 1 line
tejun