Tao Ma
2010-Aug-16 07:39 UTC
[Ocfs2-devel] [PATCH] CMWQ: Set workqueue name before we process one work.
Now in CMWQ, workqueue threads are named as 'kworker/*'. So it is a little boring to see in the 'top'(below) and actually it isn't meaningful for the users. 12606 root 20 0 0 0 0 S 2 0.0 0:03.22 kworker/u:0 12607 root 20 0 0 0 0 S 1 0.0 0:03.69 kworker/u:4 So this patch just try to set the proper workqueue name if it does exist. Yes, workqueue now is a thread pool and the data may be not accurate but I think it is better(below) than the 'kworker*' stuff. And if there is something block the workqueue, we can find the caller easily. 12606 root 20 0 0 0 0 D 2 0.0 0:02.90 dlm_wq 12607 root 20 0 0 0 0 D 0 0.0 0:03.21 ocfs2_wq What's more, in ocfs2, we sometimes want to print some debug info in the system log and we use 'current->comm' to print the thread and this change also does help. The only thing I am not clear is that do we need [gs]et_task_comm? I guess not and this patch just try to use strlcpy without task_lock. Cc: Tejun Heo <tj at kernel.org> Signed-off-by: Tao Ma <tao.ma at oracle.com> --- kernel/workqueue.c | 10 ++++++++++ 1 files changed, 10 insertions(+), 0 deletions(-) diff --git a/kernel/workqueue.c b/kernel/workqueue.c index 2994a0e..69402f8 100644 --- a/kernel/workqueue.c +++ b/kernel/workqueue.c @@ -1730,6 +1730,7 @@ static void process_one_work(struct worker *worker, struct work_struct *work) work_func_t f = work->func; int work_color; struct worker *collision; + char comm_saved[TASK_COMM_LEN]; #ifdef CONFIG_LOCKDEP /* * It is permissible to free the struct work_struct from @@ -1790,7 +1791,16 @@ static void process_one_work(struct worker *worker, struct work_struct *work) work_clear_pending(work); lock_map_acquire(&cwq->wq->lockdep_map); lock_map_acquire(&lockdep_map); + if (cwq->wq->name) { + strlcpy(comm_saved, worker->task->comm, + sizeof(comm_saved)); + strlcpy(worker->task->comm, cwq->wq->name, + sizeof(worker->task->comm)); + } f(work); + if (cwq->wq->name) + strlcpy(worker->task->comm, comm_saved, + sizeof(worker->task->comm)); lock_map_release(&lockdep_map); lock_map_release(&cwq->wq->lockdep_map); -- 1.7.1.GIT
Tejun Heo
2010-Aug-16 07:44 UTC
[Ocfs2-devel] [PATCH] CMWQ: Set workqueue name before we process one work.
On 08/16/2010 09:39 AM, Tao Ma wrote:> Now in CMWQ, workqueue threads are named as 'kworker/*'. So it is > a little boring to see in the 'top'(below) and actually it isn't > meaningful for the users. > 12606 root 20 0 0 0 0 S 2 0.0 0:03.22 kworker/u:0 > 12607 root 20 0 0 0 0 S 1 0.0 0:03.69 kworker/u:4 > > So this patch just try to set the proper workqueue name if it does > exist. Yes, workqueue now is a thread pool and the data may be not > accurate but I think it is better(below) than the 'kworker*' stuff. > And if there is something block the workqueue, we can find the caller > easily. > 12606 root 20 0 0 0 0 D 2 0.0 0:02.90 dlm_wq > 12607 root 20 0 0 0 0 D 0 0.0 0:03.21 ocfs2_wq > > What's more, in ocfs2, we sometimes want to print some debug info in > the system log and we use 'current->comm' to print the thread and this > change also does help. > > The only thing I am not clear is that do we need [gs]et_task_comm? > I guess not and this patch just try to use strlcpy without task_lock. > > Cc: Tejun Heo <tj at kernel.org> > Signed-off-by: Tao Ma <tao.ma at oracle.com>I thought about the same thing but this doesn't provide any more information than stack traces for debugging and for run time tracking implementing tracing API (scheduled to be added in this devel cycle) is the right approach. Also, I don't think it's wise to constantly change program name. It will confuse more than help. So, unless there are other reasons for doing this, I don't think I'm gonna take this one. Thanks. -- tejun
Tao Ma
2010-Aug-16 08:03 UTC
[Ocfs2-devel] [PATCH] CMWQ: Set workqueue name before we process one work.
Hi Tejun, Thanks for the quick response. On 08/16/2010 03:44 PM, Tejun Heo wrote:> On 08/16/2010 09:39 AM, Tao Ma wrote: >> Now in CMWQ, workqueue threads are named as 'kworker/*'. So it is >> a little boring to see in the 'top'(below) and actually it isn't >> meaningful for the users. >> 12606 root 20 0 0 0 0 S 2 0.0 0:03.22 kworker/u:0 >> 12607 root 20 0 0 0 0 S 1 0.0 0:03.69 kworker/u:4 >> >> So this patch just try to set the proper workqueue name if it does >> exist. Yes, workqueue now is a thread pool and the data may be not >> accurate but I think it is better(below) than the 'kworker*' stuff. >> And if there is something block the workqueue, we can find the caller >> easily. >> 12606 root 20 0 0 0 0 D 2 0.0 0:02.90 dlm_wq >> 12607 root 20 0 0 0 0 D 0 0.0 0:03.21 ocfs2_wq >> >> What's more, in ocfs2, we sometimes want to print some debug info in >> the system log and we use 'current->comm' to print the thread and this >> change also does help. >> >> The only thing I am not clear is that do we need [gs]et_task_comm? >> I guess not and this patch just try to use strlcpy without task_lock. >> >> Cc: Tejun Heo<tj at kernel.org> >> Signed-off-by: Tao Ma<tao.ma at oracle.com> > > I thought about the same thing but this doesn't provide any more > information than stack traces for debugging and for run time tracking > implementing tracing API (scheduled to be added in this devel cycle) > is the right approach.Do you have the link for this or it will show up in the mail list soon?> change program name. It will confuse more than help. So, unless > there are other reasons for doing this, I don't think I'm gonna take > this one.OK, let me find other ways ocfs2 can use to trace the real callers of some functions in runtime. Regards, Tao