Re: WITNESS observes 2 LORs on Boot of Release 10.1
On 11/22/2014 03:51 PM, Benjamin Kaduk wrote: On Fri, 21 Nov 2014, Ellis H. Wilson III wrote: Before I start, and this is mainly geared to my responder Benjamin Kaduk, based on your response, are you suggesting that the cnputc WITNESS panic you expected to happen is now completely unavoidable in FreeBSD 10? I.E., is this a spinlock that WITNESS falls over each time but that is provably deadlock free that the developers have decided cannot be BLESSED for some reason? https://lists.freebsd.org/pipermail/freebsd-current/2012-January/031316.html looks to be a better explanation than the previous link I sent ... in short, console output is hard. I guess I just can't wrap my head around why we would ever move to a regime where SKIPSPIN is the default for testing... That just seems like an open invitation for introducing spinlock regressions. I don't think anyone made the conscious decision to do that, it just happened by default as no one spent the time to fix the aforementioned issue. Moving onto the LORs I'm seeing, a question I have as a newbie to WITNESS debugging is how exactly to interpret the output if I see a stacktrace and then a LOR output like the following: lock order reversal: 1st 0x81633d88 entropy harvest mutex (entropy harvest mutex) @ /usr/src/sys/dev/random/random_harvestq.c:198 2nd 0x813b6208 scrlock (scrlock) @ /usr/src/sys/dev/syscons/syscons.c:2682 Does this mean WITNESS has already stored an ordering of #1 harvest_mtx then #2 scp-scr_lock, and somewhere somebody tried to lock scp-scr_lock without first getting harvest_mtx? Or the reverse (WITNESS previously recorded scrlock and then harvest and the lines it spit out were the offenders?) I believe it is the latter (the ordering being printed is the bad one which caused WITNESS to complain). Thanks so much for the additional info Ben. This fleshes out the history of this issue for me significantly. I have filed a bug on this at: https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=195262 Xin Li was able to identify the ordering that caused the problem and proposed a possible patch to fix it. I can confirm that now I'm booting with solely WITNESS (i.e., not WITNESS_SKIPSPIN) without panic. Thanks! ellis ___ freebsd-current@freebsd.org mailing list http://lists.freebsd.org/mailman/listinfo/freebsd-current To unsubscribe, send any mail to freebsd-current-unsubscr...@freebsd.org
Re: WITNESS observes 2 LORs on Boot of Release 10.1
On Fri, 21 Nov 2014, Ellis H. Wilson III wrote: Before I start, and this is mainly geared to my responder Benjamin Kaduk, based on your response, are you suggesting that the cnputc WITNESS panic you expected to happen is now completely unavoidable in FreeBSD 10? I.E., is this a spinlock that WITNESS falls over each time but that is provably deadlock free that the developers have decided cannot be BLESSED for some reason? https://lists.freebsd.org/pipermail/freebsd-current/2012-January/031316.html looks to be a better explanation than the previous link I sent ... in short, console output is hard. I guess I just can't wrap my head around why we would ever move to a regime where SKIPSPIN is the default for testing... That just seems like an open invitation for introducing spinlock regressions. I don't think anyone made the conscious decision to do that, it just happened by default as no one spent the time to fix the aforementioned issue. Moving onto the LORs I'm seeing, a question I have as a newbie to WITNESS debugging is how exactly to interpret the output if I see a stacktrace and then a LOR output like the following: lock order reversal: 1st 0x81633d88 entropy harvest mutex (entropy harvest mutex) @ /usr/src/sys/dev/random/random_harvestq.c:198 2nd 0x813b6208 scrlock (scrlock) @ /usr/src/sys/dev/syscons/syscons.c:2682 Does this mean WITNESS has already stored an ordering of #1 harvest_mtx then #2 scp-scr_lock, and somewhere somebody tried to lock scp-scr_lock without first getting harvest_mtx? Or the reverse (WITNESS previously recorded scrlock and then harvest and the lines it spit out were the offenders?) I believe it is the latter (the ordering being printed is the bad one which caused WITNESS to complain). -Ben ___ freebsd-current@freebsd.org mailing list http://lists.freebsd.org/mailman/listinfo/freebsd-current To unsubscribe, send any mail to freebsd-current-unsubscr...@freebsd.org
Re: WITNESS observes 2 LORs on Boot of Release 10.1
On 11/18/2014 08:35 PM, Ellis H. Wilson III wrote: If nobody has seen these before, I'll try and put together fixes for them. Please somebody speak up if you have seen them or have useful information for me to go on in my patches. I've started to dig into the relevant code in sys/dev/random/random_harvestq.c, sys/dev/syscons/syscons.c, and sys/kern/subr_sleepqueue.c. Before I start, and this is mainly geared to my responder Benjamin Kaduk, based on your response, are you suggesting that the cnputc WITNESS panic you expected to happen is now completely unavoidable in FreeBSD 10? I.E., is this a spinlock that WITNESS falls over each time but that is provably deadlock free that the developers have decided cannot be BLESSED for some reason? I guess I just can't wrap my head around why we would ever move to a regime where SKIPSPIN is the default for testing... That just seems like an open invitation for introducing spinlock regressions. Moving onto the LORs I'm seeing, a question I have as a newbie to WITNESS debugging is how exactly to interpret the output if I see a stacktrace and then a LOR output like the following: lock order reversal: 1st 0x81633d88 entropy harvest mutex (entropy harvest mutex) @ /usr/src/sys/dev/random/random_harvestq.c:198 2nd 0x813b6208 scrlock (scrlock) @ /usr/src/sys/dev/syscons/syscons.c:2682 Does this mean WITNESS has already stored an ordering of #1 harvest_mtx then #2 scp-scr_lock, and somewhere somebody tried to lock scp-scr_lock without first getting harvest_mtx? Or the reverse (WITNESS previously recorded scrlock and then harvest and the lines it spit out were the offenders?) Along those lines, in 10.0 and 10.1 releases I get two LORs showing up almost on-top of each other, with the other LOR showing up as: lock order reversal: 1st 0x81633d88 entropy harvest mutex (entropy harvest mutex) @ /usr/src/sys/dev/random/random_harvestq.c:198 2nd 0x81424bb8 sleepq chain (sleepq chain) @ /usr/src/sys/kern/subr_sleepqueue.c:240 This seems like maybe two LORs are detected at the same time, which perhaps suggests that the harvest_mtx should have been taken /after/ both of the other locks mentioned (scrlock and sleepq). I'm happy to do the legwork implementing, testing, and submitting a patch for this, but I would really appreciate a pointer in the right direction from somebody who already has handled some LORs before. Thanks! ellis ___ freebsd-current@freebsd.org mailing list http://lists.freebsd.org/mailman/listinfo/freebsd-current To unsubscribe, send any mail to freebsd-current-unsubscr...@freebsd.org
Re: WITNESS observes 2 LORs on Boot of Release 10.1
On Tue, 18 Nov 2014, Ellis H. Wilson III wrote: This is on a virtual machine via VirtualBox, but we have reproduced it on real hardware. Regarding the latter reversal, I have dug in the code and see the calls go about in the following order: In sys/dev/random/random_harvestq.c: 198 mtx_lock_spin(harvest_mtx); 209 msleep_spin_sbt(random_kthread_control, harvest_mtx, Now in sys/kern/kern_synch.c: 297 sleepq_lock(ident); Now in sys/kern/subr_sleepqueue.c 237 mtx_lock_spin(sc-sc_lock); The above line numbers might be a hair off from head since I'm looking at code synced last week. I'm happy to open a bug on this if that's the desirable course of action, or to even assist in fixing it, but I wanted to first see if anybody knew about these already (they didn't show up on the known LORs list on quick perusal) or if this was simply a case of WITNESS being overly conservative and throwing false positives. If this belongs on a different list just let me know. I don't believe it is known, and this list is fine. I'm observing the following two WITNESS LORs being thrown upon boot-up of 10.0 and I was tracking current, hoping they would go away by 10.1, but it seems they persist as shown below. I suspect this is because current is being built with WITNESS on but also with SKIPSPIN on. So these issues are unlikely to show up for any devs but those who specifically enable WITNESS and disable SKIPSPIN like myself. At my work we would greatly like to do our debugging with checking of spin-locking order included and panicing upon LOR detection. That's not possible with these in existence. However, I was under the impression that a kernel built with WITNESS and without WITNESS_SKIPSPIN would panic on boot on the cnputs_mutx (see, e.g., https://lists.freebsd.org/pipermail/freebsd-stable/2014-January/076864.html). So, (1) I'm surprised you can boot it, and (2) that would explain why no one else has been using it. -Ben ___ freebsd-current@freebsd.org mailing list http://lists.freebsd.org/mailman/listinfo/freebsd-current To unsubscribe, send any mail to freebsd-current-unsubscr...@freebsd.org
Re: WITNESS observes 2 LORs on Boot of Release 10.1
On 11/18/2014 05:39 PM, Benjamin Kaduk wrote: On Tue, 18 Nov 2014, Ellis H. Wilson III wrote: I'm observing the following two WITNESS LORs being thrown upon boot-up of 10.0 and I was tracking current, hoping they would go away by 10.1, but it seems they persist as shown below. I suspect this is because current is being built with WITNESS on but also with SKIPSPIN on. So these issues are unlikely to show up for any devs but those who specifically enable WITNESS and disable SKIPSPIN like myself. At my work we would greatly like to do our debugging with checking of spin-locking order included and panicing upon LOR detection. That's not possible with these in existence. However, I was under the impression that a kernel built with WITNESS and without WITNESS_SKIPSPIN would panic on boot on the cnputs_mutx (see, e.g., https://lists.freebsd.org/pipermail/freebsd-stable/2014-January/076864.html). So, (1) I'm surprised you can boot it, and (2) that would explain why no one else has been using it. That's a very interesting thread. I've seen another where a fellow developer suggested just throwing on the WITNESS_SKIPSPIN flag to solve the issue. I can't say that I agree with the approach, but I understand. I'd be willing to tackle a bit of WITNESS massaging to help it be instructed about known false positives better, if that's desirable. Why I'm able to boot however, is simple: I haven't enabled a full suite of debugging flags, KDB/DDB being the key ones that cause a panic to occur on a failure like the one I've seen. We originally were seeing the panic so WITNESS in its entirety was shut off. I was asked to try and get that back on-track, so to start I at least wanted to see how many LORs we were dealing with on boot. 2 is apparently that magic number, and maybe 3 if the cputs issue you refer to is still possible to be hit in the 10.1 wild. If nobody has seen these before, I'll try and put together fixes for them. Please somebody speak up if you have seen them or have useful information for me to go on in my patches. Thanks, ellis ___ freebsd-current@freebsd.org mailing list http://lists.freebsd.org/mailman/listinfo/freebsd-current To unsubscribe, send any mail to freebsd-current-unsubscr...@freebsd.org
