|| ||Tejun Heo <email@example.com> |
|| ||firstname.lastname@example.org |
|| ||[PATCHSET] workqueue: remove gcwq and make worker_pool the only backend abstraction |
|| ||Wed, 16 Jan 2013 17:42:32 -0800|
|| ||Lai Jiangshan <email@example.com>|
|| ||Article, Thread
Currently, on the backend side, there are two layers of abstraction.
For each CPU and the special unbound wq-specific CPU, there's one
global_cwq. gcwq in turn hosts two worker_pools - one for normal
priority, the other for highpri - each of which actually serves the
worker_pool is the later addition to support separate pool of workers
for highpri workqueues. Stuff was moved to worker_pool on as-needed
basis and, as a result, the two pools belonging to the same CPU share
some stuff in the gcwq - most notably the lock and the hash table for
work items currently being executed.
It seems like we'll need to support worker pools with custom
attributes, which is planned to be implemented as extra worker_pools
for the unbound CPU. Removing gcwq and having worker_pool as the top
level abstraction makes things much simpler for such designs. Also,
there's scalability benefit to not sharing locking and busy hash among
different worker pools as worker pools w/ custom attributes are likely
to have widely different memory / cpu locality characteristics.
In retrospect, it might have been less churn if we just converted to
have multiple gcwqs per CPU when we were adding highpri pool support.
Oh well, such is life and the name worker_pool fits the role much
better anyway at this point.
This patchset moves the remaining stuff in gcwq to worker_pool and
then removes gcwq entirely making worker_pool the top level and the
only backend abstraction. In the process, this patchset also prepares
for later addition of worker_pools with custom attributes.
This patchset shouldn't introduce any visible differences outside of
workqueue proper and contains the following 17 patches.
0001-0002 are misc preps.
0003-0004 move flags from gcwq to pool.
0005-0007 make work->data off-queue backlink point to worker_pools
instead of CPUs, which is necessary to move busy_hash to pool.
0008-0010 move busy_hash, cpu and locking to pool.
0011-0014 make operations per-pool and remove gcwq usages.
0015-0017 remove gcwq and cleanup afterwards.
This patchset is on top of wq/for-3.9 023f27d3d6f ("workqueue: fix
find_worker_executing_work() brekage from hashtable conversion") and
available in the following git branch.
include/linux/workqueue.h | 17
include/trace/events/workqueue.h | 2
kernel/workqueue.c | 897 +++++++++++++++++++--------------------
3 files changed, 461 insertions(+), 455 deletions(-)
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to firstname.lastname@example.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/