Re: [Nfs-ganesha-devel] svc_ioq_callback() issues

2015-08-11 Thread Malahal Naineni
> Evidently you're doing something illegal here, but perhaps you're just > exposing a bug that was already present. We certainly want to rule that > out. Thank you Matt for your time. I just found out that I was releasing mutex twice! I wish we had PTHREAD_MUTEX_{loc,unlock} macros in tirpc code.

Re: [Nfs-ganesha-devel] svc_ioq_callback() issues

2015-08-11 Thread Malahal Naineni
William Allen Simpson [william.allen.simp...@gmail.com] wrote: > On 8/11/15 3:26 AM, Malahal Naineni wrote: > > ... We see work pool mutex (svc_work_pool.pqh.qmutex) > > at the top now, so I am trying to fix this. I made a small patch so > > that a work pool thread waits for 1 full second before wo

Re: [Nfs-ganesha-devel] svc_ioq_callback() issues

2015-08-11 Thread Matt Benjamin
Hi Malahal, - Original Message - > From: "Malahal Naineni" > To: "Matt Benjamin" > Cc: Nfs-ganesha-devel@lists.sourceforge.net, "william allen simpson" > , "Daniel > Gryniewicz" > Sent: Tuesday, August 11, 2015 11:04:04 AM > Subject: Re: svc_ioq_callback() issues > > > I have some co

Re: [Nfs-ganesha-devel] svc_ioq_callback() issues

2015-08-11 Thread Malahal Naineni
Thank you Matt and Bill. I am actually not proposing this as a patch that needs to be merged upstream. We are just experimenting and mutrace shows that work pool lock has the most contention time. Not sure how that reports and whether this is really an issue is still unknown. One thing we noticed i

Re: [Nfs-ganesha-devel] svc_ioq_callback() issues

2015-08-11 Thread Matt Benjamin
Hi Malahal, - Original Message - > From: "Malahal Naineni" > To: "Matt Benjamin" > Cc: Nfs-ganesha-devel@lists.sourceforge.net, "william allen simpson" > , "Daniel > Gryniewicz" > Sent: Tuesday, August 11, 2015 3:26:23 AM > Subject: svc_ioq_callback() issues > > Thank you Matt. > > "

Re: [Nfs-ganesha-devel] svc_ioq_callback() issues

2015-08-11 Thread William Allen Simpson
On 8/11/15 3:26 AM, Malahal Naineni wrote: > ... We see work pool mutex (svc_work_pool.pqh.qmutex) > at the top now, so I am trying to fix this. I made a small patch so > that a work pool thread waits for 1 full second before working on a > different task or waiting in the global pool. > DO NOT do

[Nfs-ganesha-devel] svc_ioq_callback() issues

2015-08-11 Thread Malahal Naineni
Thank you Matt. "git blame" does indicate that you experimented with it but it was disabled due to not making much difference with throughput. Makes sense. I didn't go deeper into the git history. I enabled it and the decoder fridge mutex has gone out from the contention list. We see work pool mu