Hi Joerg,
This does look like Andrew's commit (from 2008) is buggy (and this is a mainline bug, not an RT one). (top posting this so Andrew knows to look further ;-) On Thu, 28 May 2020 13:41:08 +0200 Joerg Vehlow <[email protected]> wrote: > Hi, > > I think I found a bug in the kernel with rt patches (or maybe even without). > This applies to all kernels propably starting at 2.6.27. > > When a kernel panic is triggered from an interrupt handler, the dump-capture > kernel is not started, instead the system acts as if it was not installed. > The reason for this is, that panic calls __crash_kexec, which is protected > by a mutex. On an rt kernel this mutex is an rt mutex and when trylock > is called > on an rt mutex, the first check is whether the current kthread is in an > nmi or > irq handler. If it is, the function just returns 0 -> locking failed. > > According to rt_mutex_trylock documentation, it is not allowed to call this > function from an irq handler, but panic can be called from everywhere > and thus > rt_mutex_trylock can be called from everywhere. Actually even > mutex_trylock has > the comment, that it is not supposed to be used from interrupt context, > but it > still locks the mutex. I guess this could also be a bug in the non-rt > kernel. > > I found this problem using a test module, that triggers the softlock > detection. > It is a pretty simple module, that creates a kthread, that disables > preemption, > spins 60 seconds in an endless loop and then reenables preemption and > terminates > the thread. This reliably triggers the softlock detection and if > kernel.softlockup_panic=0, the system resumes perfectly fine afterwards. If > kernel.softlockup_panic=1 I would expect the dump-capture kernel to be > executed, > but it is not due to the bug (without rt patches it works), instead the > panic > function is executed until the end to the endless loop. > > > A stacktrace captured at the trylock call inside kexec_code looks like this: > #0 __rt_mutex_trylock (lock=0xffffffff81701aa0 <kexec_mutex>) at > /usr/src/kernel/kernel/locking/rtmutex.c:2110 > #1 0xffffffff8087601a in _mutex_trylock (lock=<optimised out>) at > /usr/src/kernel/kernel/locking/mutex-rt.c:185 > #2 0xffffffff803022a0 in __crash_kexec (regs=0x0 <irq_stack_union>) at > /usr/src/kernel/kernel/kexec_core.c:941 > #3 0xffffffff8027af59 in panic (fmt=0xffffffff80fa3d66 "softlockup: > hung tasks") at /usr/src/kernel/kernel/panic.c:198 > #4 0xffffffff80325b6d in watchdog_timer_fn (hrtimer=<optimised out>) at > /usr/src/kernel/kernel/watchdog.c:464 > #5 0xffffffff802e6b90 in __run_hrtimer (flags=<optimised out>, > now=<optimised out>, timer=<optimised out>, base=<optimised out>, > cpu_base=<optimised out>) at /usr/src/kernel/kernel/time/hrtimer.c:1417 > #6 __hrtimer_run_queues (cpu_base=0xffff88807db1c000, now=<optimised > out>, flags=<optimised out>, active_mask=<optimised out>) at > /usr/src/kernel/kernel/time/hrtimer.c:1479 > #7 0xffffffff802e7704 in hrtimer_interrupt (dev=<optimised out>) at > /usr/src/kernel/kernel/time/hrtimer.c:1539 > #8 0xffffffff80a020f2 in local_apic_timer_interrupt () at > /usr/src/kernel/arch/x86/kernel/apic/apic.c:1067 > #9 smp_apic_timer_interrupt (regs=<optimised out>) at > /usr/src/kernel/arch/x86/kernel/apic/apic.c:1092 > #10 0xffffffff80a015df in apic_timer_interrupt () at > /usr/src/kernel/arch/x86/entry/entry_64.S:909 > > > Obviously and as expected the panic was triggered in the context of the apic > interrupt. So in_irq() is true and trylock fails. > > > About 12 years ago this was not implemented using a mutex, but using xchg. > See: 8c5a1cf0ad3ac5fcdf51314a63b16a440870f6a2 Yes, that commit is wrong, because mutex_trylock() is not to be taken in interrupt context, where crash_kexec() looks like it can be called. Unless back then crash_kexec() wasn't called in interrupt context, then the commit that calls it from that combined with this commit is the issue. -- Steve > > > Since my knowledege about mutexes inside the kernel is very limited, I > do not > know how this can be fixed and whether it should be fixed in the rt > patches or > if this really is a bug in mainline kernel (because trylock is also not > allowed > to be used in interrupt handlers. > > > Jörg
