From mboxrd@z Thu Jan 1 00:00:00 1970 From: "Paul E. McKenney" Subject: Re: linux-next: Tree for March 25 (Call trace: RCU|workqueues|block|VFS|ext4 related?) Date: Mon, 28 Mar 2011 06:24:36 -0700 Message-ID: <20110328132436.GL2322@linux.vnet.ibm.com> References: <20110327000900.GB2322@linux.vnet.ibm.com> <20110327032529.GC2322@linux.vnet.ibm.com> <20110327050720.GA31424@linux.vnet.ibm.com> <20110327213220.GE2322@linux.vnet.ibm.com> <20110328040823.GG2322@linux.vnet.ibm.com> Reply-To: paulmck@linux.vnet.ibm.com Mime-Version: 1.0 Content-Type: text/plain; charset=iso-8859-1 Content-Transfer-Encoding: QUOTED-PRINTABLE Return-path: Received: from e3.ny.us.ibm.com ([32.97.182.143]:33745 "EHLO e3.ny.us.ibm.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1752443Ab1C1NYz (ORCPT ); Mon, 28 Mar 2011 09:24:55 -0400 Content-Disposition: inline In-Reply-To: Sender: linux-next-owner@vger.kernel.org List-ID: To: sedat.dilek@gmail.com Cc: Josh Triplett , linux-next , LKML , Stephen Rothwell , Randy Dunlap , Theodore Ts'o , Jens Axboe , Tejun Heo , Al Viro , Nick Piggin On Mon, Mar 28, 2011 at 02:33:36PM +0200, Sedat Dilek wrote: > On Mon, Mar 28, 2011 at 6:08 AM, Paul E. McKenney > wrote: > > On Sun, Mar 27, 2011 at 11:48:30PM +0200, Sedat Dilek wrote: > >> On Sun, Mar 27, 2011 at 11:32 PM, Paul E. McKenney > >> wrote: > >> > On Sun, Mar 27, 2011 at 02:26:15PM +0200, Sedat Dilek wrote: > >> >> On Sun, Mar 27, 2011 at 7:07 AM, Paul E. McKenney > >> >> wrote: > >> >> > On Sat, Mar 26, 2011 at 08:25:29PM -0700, Paul E. McKenney wr= ote: > >> >> >> On Sun, Mar 27, 2011 at 03:30:34AM +0200, Sedat Dilek wrote: > >> >> >> > On Sun, Mar 27, 2011 at 1:09 AM, Paul E. McKenney > >> >> >> > wrote: > >> >> >> > > On Sat, Mar 26, 2011 at 11:15:22PM +0100, Sedat Dilek wr= ote: > >> >> > > >> >> > [ . . . ] > >> >> > > >> >> >> > >> But then came RCU :-(. > >> >> >> > > > >> >> >> > > Well, if it turns out to be a problem in RCU I will cert= ainly apologize. > >> >> >> > > > >> >> >> > > >> >> >> > No, that's not so dramatic. > >> >> >> > Dealing with this RCU issue has nice side-effects: I remem= bered (and > >> >> >> > finally did) to use a reduced kernel-config set. > >> >> >> > The base for it I created with 'make localmodconfig' and d= id some > >> >> >> > manual fine-tuning afterwards (throw out media, rc, dvd, u= nneeded FSs, > >> >> >> > etc.). > >> >> >> > Also, I can use fresh gcc-4.6 (4.6.0-1) from the official = Debian repos. > >> >> >> > > >> >> >> > So, I started building with > >> >> >> > "revert-rcu-patches/0001-Revert-rcu-introduce-kfree_rcu.pa= tch". > >> >> >> > I will let you know. > >> >> >> > >> >> >> And please also check for tasks consuming all available CPU. > >> >> > > >> >> > And I still cannot reproduce with the full RCU stack (but bas= ed off of > >> >> > 2.6.38 rather than -next). =A0Nevertheless, if you would like= to try a > >> >> > speculative patch, here you go. > >> >> > >> >> You are right and my strategy on handling the (possible RCU?) i= ssue is wrong. > >> >> Surely, you tested your RCU stuff in your own repo and everythi= ng > >> >> might be OK on top of stable 2.6.38. > >> >> Linux-next gets daily updates from a lot of different trees, so= there > >> >> might be interferences with other stuff. > >> >> Please, understand I am interested in finding out what is the c= ause > >> >> for my issues, my aim is not to blame you. > >> > > >> > I am not worried about blame, but rather getting the bug fixed. = =A0The > >> > bug might be in RCU, it might be elsewhere, or it might be a com= bination > >> > of problems in RCU and elsewhere. > >> > > >> > So the first priority is locating the bug. > >> > > >> > And that is why I have been asking you over and over to PLEASE t= ake > >> > a look at what tasks are consuming CPU while the problem is occu= ring. > >> > The reason that I have been asking over and over is that the sym= ptoms > >> > you describe are likely caused by a loop in some kernel code. =A0= Yes, > >> > there might be other causes, but this is the most likely. =A0Giv= en that > >> > TREE_PREEMPT_RCU behaves better than TREE_RCU, it is likely that= this > >> > loop is in preemptible code with irqs enabled. =A0Therefore, the= process > >> > accounting code is likely to be able to see the CPU consumption,= and > >> > you should be able to see it via the "top" or "ps" commands -- o= r via > >> > any number of other tools. > >> > > >> > For example, if the problem is confined to RCU, you would likely= see > >> > the "rcuc0" or "rcun0" tasks consuming lots of CPU. =A0This woul= d narrow > >> > the problem down to a few tens of lines of code. =A0If the probl= em was > >> > in some other kthread, then identifying the kthread would very l= ikely > >> > narrow things down as well. > >> > > >> > So, please do take a look to see what taks consuming CPU. > >> > > >> >> As I was wrong and want to be 99.9% sure it is RCU stuff, I rev= erted > >> >> all (18) RCU patches from linux-next (next-20110325) by keeping= the > >> >> RCU|PREEMPT|HZ settings from last working next-20110323. > >> > > >> > Makes sense. > >> > > >> >> $ egrep 'RCU|PREEMPT|_HZ' /boot/config-2.6.38-next20110325-7-68= 6-iniza > >> >> # RCU Subsystem > >> >> CONFIG_TREE_RCU=3Dy > >> >> # CONFIG_PREEMPT_RCU is not set > >> >> # CONFIG_RCU_TRACE is not set > >> >> CONFIG_RCU_FANOUT=3D32 > >> >> # CONFIG_RCU_FANOUT_EXACT is not set > >> >> CONFIG_RCU_FAST_NO_HZ=3Dy > >> >> # CONFIG_TREE_RCU_TRACE is not set > >> >> CONFIG_PREEMPT_NOTIFIERS=3Dy > >> >> CONFIG_NO_HZ=3Dy > >> >> # CONFIG_PREEMPT_NONE is not set > >> >> CONFIG_PREEMPT_VOLUNTARY=3Dy > >> >> # CONFIG_PREEMPT is not set > >> >> # CONFIG_HZ_100 is not set > >> >> CONFIG_HZ_250=3Dy > >> >> # CONFIG_HZ_300 is not set > >> >> # CONFIG_HZ_1000 is not set > >> >> CONFIG_HZ=3D250 > >> >> # CONFIG_SPARSE_RCU_POINTER is not set > >> >> # CONFIG_RCU_TORTURE_TEST is not set > >> >> # CONFIG_RCU_CPU_STALL_DETECTOR is not set > >> >> > >> >> I will work and stress this kernel before doing any step-by-ste= p > >> >> revert of RCU stuff. > >> >> > >> >> Thanks for your patch, I applied it on top of "naked" next-2011= 0325, > >> >> but I still see call-traces. > >> > > >> > Thank you very much for testing it! > >> > > >> > I intend to keep that patch, as it should increase robustness in= other > >> > situations. > >> > > >> > =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 = =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0Thanx, Paul > >> > > >> >> - Sedat - > >> >> > >> >> > >> >> > >> >> > =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0= =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0 =A0Thanx, Paul > >> >> > > >> >> > -------------------------------------------------------------= ----------- > >> >> > > >> >> > rcu: further lower priority in rcu_yield() > >> >> > > >> >> > Although rcu_yield() dropped from real-time to normal priorit= y, there > >> >> > is always the possibility that the competing tasks have been = niced. > >> >> > So nice to 19 in rcu_yield() to help ensure that other tasks = have a > >> >> > better chance of running. > >> >> > > >> >> > =A0 =A0Signed-off-by: Paul E. McKenney > >> >> > > >> >> > diff --git a/kernel/rcutree.c b/kernel/rcutree.c > >> >> > index 759f54b..5477764 100644 > >> >> > --- a/kernel/rcutree.c > >> >> > +++ b/kernel/rcutree.c > >> >> > @@ -1492,6 +1492,7 @@ static void rcu_yield(void (*f)(unsigne= d long), unsigned long arg) > >> >> > =A0 =A0 =A0 =A0mod_timer(&yield_timer, jiffies + 2); > >> >> > =A0 =A0 =A0 =A0sp.sched_priority =3D 0; > >> >> > =A0 =A0 =A0 =A0sched_setscheduler_nocheck(current, SCHED_NORM= AL, &sp); > >> >> > + =A0 =A0 =A0 set_user_nice(current, 19); > >> >> > =A0 =A0 =A0 =A0schedule(); > >> >> > =A0 =A0 =A0 =A0sp.sched_priority =3D RCU_KTHREAD_PRIO; > >> >> > =A0 =A0 =A0 =A0sched_setscheduler_nocheck(current, SCHED_FIFO= , &sp); > >> > >> Sorry, my attempt was to identify and isolate the culprit commit. > >> > >> Reverting all RCU patches resulted in a stable system, the followi= ng 8 > >> kernels with reduced k-config setup where all built using this ker= nel. > >> > >> All kernels used TREE_RCU (see above), I did not change it (no > >> mixing/switching to PREEMPT and TREE_PREEMPT_RCU). > >> ( I doubt that TREE_PREEMPT_RCU was some kind of more stable here.= ) > >> > >> The culprit commit is bc56163ebd4580199ac7e63f5e160bf139ba0dd6 (fr= om > >> rcu/next GIT tree): > >> "rcu: move TREE_RCU from softirq to kthread" > > >=20 > Hi Paul, >=20 > > OK, please accept my apologies for your lost weekend. =A0And thank = you for > > testing this. > > >=20 > No worries, it was mostly a rainy day. > The only thing I did @ 16:30 was to go to regional election (the new > (regional) prime minister will be the 1st from The German Green > party). ;-) > But back to RCU :-): > The reduced kernel-config setup decreased the build-time from approx. > 2hrs (full, generic build) down to approx. 35mins. Very good! > >> I can do parallelly a tar job, open 20 tabs in firefox and run a f= lash > >> video in one of them (I did this several times). > > > > How many files in the tar job? =A0Is this creating a tar archive, e= xpanding > > it, or both? >=20 > I am doing a simple tar (filesize: 1.6G for full and 1.0G for reduced= build): >=20 > $ tar -cf $archivedir-on-external-usbhdd/$tarfile $kernel-build-dir OK, I was extracting a tarball and then removing the resulting source tree. I will try this. Though it does seem strange -- I can understan= d how removing a file tree would stress RCU, but not creating a tarball. Ah, well, if I fully understood it, there would not be a bug. > ...plus parallelly opening 20 tabs in firefox. > That's normally enough to get my system freaky and see RCU related > messages in the logs. Hmmm... My normal test systems don't have X -- I will need to set this up. > > Do you have a script for this? =A0Are all of these running at norma= l > > priority, or are some of them running at real-time priority? > > >=20 > Nothing special. OK. > >> [ setup.log ] > >> ... > >> =A0 (+) OK =A0 revert-rcu-patches/0001-Revert-rcu-introduce-kfree_= rcu.patch > >> =A0 (+) OK =A0 revert-rcu-patches/0002-Revert-rcu-fix-spelling.pat= ch > >> =A0 (+) OK =A0 revert-rcu-patches/0003-Revert-rcu-fix-rcu_cpu_kthr= ead_task-synchronization.patch > >> =A0 (+) OK =A0 revert-rcu-patches/0004-Revert-rcu-call-__rcu_read_= unlock-in-exit_rcu-for-tr.patch > >> =A0 (+) OK =A0 revert-rcu-patches/0005-Revert-rcu-Converge-TINY_RC= U-expedited-and-normal-bo.patch > >> =A0 (+) OK =A0 revert-rcu-patches/0006-Revert-rcu-remove-useless-b= oosted_this_gp-field.patch > >> =A0 (+) OK =A0 revert-rcu-patches/0007-Revert-rcu-code-cleanups-in= -TINY_RCU-priority-boosti.patch > >> =A0 (+) OK =A0 revert-rcu-patches/0008-Revert-rcu-Switch-to-this_c= pu-primitives.patch > >> =A0 (+) OK =A0 revert-rcu-patches/0009-Revert-rcu-Use-WARN_ON_ONCE= -for-DEBUG_OBJECTS_RCU_HE.patch > >> =A0 (+) OK =A0 revert-rcu-patches/0010-Revert-rcu-Enable-DEBUG_OBJ= ECTS_RCU_HEAD-from-PREEMP.patch > >> =A0 (+) OK =A0 revert-rcu-patches/0011-Revert-rcu-Add-boosting-to-= TREE_PREEMPT_RCU-tracing.patch > >> =A0 (+) OK =A0 revert-rcu-patches/0012-Revert-rcu-eliminate-unused= -boosting-statistics.patch > >> =A0 (+) OK =A0 revert-rcu-patches/0013-Revert-rcu-priority-boostin= g-for-TREE_PREEMPT_RCU.patch > >> =A0 (+) OK =A0 revert-rcu-patches/0014-Revert-rcu-move-TREE_RCU-fr= om-softirq-to-kthread.patch > >> ... > >> > >> Hope this helps to narrow down the problem. > >> > >> As I kept all kernels I can have a look at the tasks consuming hig= h > >> CPU usage tomorrow. > > > > Could you please? >=20 > I recalled (as you say I requested over and over again from you :-)) = I > looked with top, htop and 'ps axu', but there was nothing special. > Sometimes the system got frozen - at this point (or short before) I > did not see anything suspicious with top. OK, thank you for the info. > > Also, could you please mount debugfs and list out the files in the > > "rcu" directory? =A0The "ql=3D" value from the "rcu/rcudata" file i= s of > > particular interest. > > >=20 > Ah, before I forget... >=20 > I used TREE_RCU (was the default before noticing RCU issue) for > finding the culprit commit. > If it is from your POV more helpful to switch to PREEMPT + PREEMPT_RC= U > + RCU_BOOST, please let me *now* know. > ( Both RCU setups freaks up the system. ) If TREE_RCU hits problems faster, it is probably best to stay with TREE_RCU. > I think top & Co. are not enough to track the problem down. > I have seen tracing and debugging facililities for RCU. >=20 > Some questions to debug and trace setup: >=20 > Case #1: TREE_RCU >=20 > CONFIG_RCU_TRACE=3Dy > CONFIG_TREE_RCU_TRACE=3Dy Yep. > Case #2: PREEMPT + PREEMPT_RCU + RCU_BOOST >=20 > CONFIG_RCU_TRACE=3Dy > CONFIG_TREE_RCU_TRACE=3Dy > CONFIG_DEBUG_PREEMPT=3Dy <--- Helpful? > CONFIG_PREEMPT_TRACER=3Dy <--- Helpful? >=20 > Any other recommends for useful/helpful trace and/or debug options? >=20 > Any other intructions for debugging/tracing? Not at the moment. I will be looking at diagnostics will going through the code, so might have something later. > BTW, today's linux-next (next-20110328) is still freaky, I applied th= e > revert-rcu-patches patchset and all is fine. I reverted back to the commit preceding the one you pointed out last ni= ght my time, so the upcoming -next should be less freaky. > - Sedat - >=20 > P.S.: Note to myself >=20 > # mount -t debugfs none /sys/kernel/debug/ > # ln -s /sys/kernel/debug /debug >=20 > # find /debug -name rcu Or: # cd /debug/rcu then dump out everything except for the .csv file (which is the same as the non-.csv equivalent, but in spreadsheet format -- intended for systems with 100s or 1000s of CPUs). Thanx, Paul