x86 on dovetail: Stress-ng gets my system down

Bezdeka, Florian florian.bezdeka at siemens.com
Mon May 3 15:28:48 CEST 2021


On Mon, 2021-05-03 at 15:09 +0200, Philippe Gerum wrote:
> Bezdeka, Florian <florian.bezdeka at siemens.com> writes:
> 
> > Hi,
> > 
> > while trying to debug one of the Xenomai 3.2 issues listed at [1] I run
> > into the situation described below on my x86 system. The problem (or at
> > least the "system hang" is reproducible on real hardware and on
> > qemu/kvm.
> > 
> > Once the system is frozen, attaching GDB to the qemu process shows me:
> > 
> > (gdb) info threads
> >   Id   Target Id                  Frame 
> > * 1    Thread 1 (CPU#0 [running]) csd_lock_wait (csd=0xffff88803e92ea00) at kernel/smp.c:228
> >   2    Thread 2 (CPU#1 [running]) 0x0000564f0b05d36d in ?? ()
> >   3    Thread 3 (CPU#2 [running]) csd_lock_wait (csd=0xffff88803e82f1e0) at kernel/smp.c:228
> >   4    Thread 4 (CPU#3 [running]) csd_lock_wait (csd=0xffff88803e82f200) at kernel/smp.c:228
> > 
> > So three of my CPUs are waiting for other CPUs to complete a function
> > call IPI. It looks like CPU1 is not responding anymore. The system is
> > completely unusable at this point.
> > 
> > (gdb) bt
> > #0  csd_lock_wait (csd=0xffff88803e92ea00) at kernel/smp.c:228
> > #1  smp_call_function_many_cond (mask=mask at entry=0xffff88800448c340, func=func at entry=0xffffffff81055bb0 <flush_tlb_func_remote>, info=info at entry=0xffffffff8200acc0 <full_flush_tlb_info>, wait=wait at entry=true, 
> >     cond_func=cond_func at entry=0xffffffff810550b0 <tlb_is_not_lazy>) at kernel/smp.c:693
> > #2  0xffffffff810f56f5 in on_each_cpu_cond_mask (cond_func=cond_func at entry=0xffffffff810550b0 <tlb_is_not_lazy>, func=func at entry=0xffffffff81055bb0 <flush_tlb_func_remote>, info=info at entry=0xffffffff8200acc0 <full_flush_tlb_info>, wait=wait at entry=true, 
> >     mask=mask at entry=0xffff88800448c340) at kernel/smp.c:904
> > #3  0xffffffff81055538 in native_flush_tlb_others (cpumask=cpumask at entry=0xffff88800448c340, info=info at entry=0xffffffff8200acc0 <full_flush_tlb_info>) at arch/x86/mm/tlb.c:840
> > #4  0xffffffff81055fac in flush_tlb_others (info=0xffffffff8200acc0 <full_flush_tlb_info>, cpumask=0xffff88800448c340) at arch/x86/mm/tlb.c:1170
> > #5  arch_tlbbatch_flush (batch=batch at entry=0xffff88800448c340) at arch/x86/mm/tlb.c:1170
> > #6  0xffffffff811ae3e1 in try_to_unmap_flush () at mm/rmap.c:602
> > #7  0xffffffff8117d9d3 in shrink_page_list (page_list=page_list at entry=0xffffc9000306f910, pgdat=pgdat at entry=0xffff88803ffdb000, sc=sc at entry=0xffffc9000306fb18, stat=stat at entry=0xffffc9000306f924, ignore_references=ignore_references at entry=false) at mm/vmscan.c:1487
> > #8  0xffffffff8117f79c in shrink_inactive_list (nr_to_scan=<optimized out>, lruvec=lruvec at entry=0xffff88803ffde508, sc=sc at entry=0xffffc9000306fb18, lru=lru at entry=LRU_INACTIVE_FILE) at mm/vmscan.c:1962
> > #9  0xffffffff811800dc in shrink_list (sc=0xffffc9000306fb18, lruvec=0xffff88803ffde508, nr_to_scan=<optimized out>, lru=<optimized out>) at mm/vmscan.c:2169
> > #10 shrink_lruvec (lruvec=lruvec at entry=0xffff88803ffde508, sc=sc at entry=0xffffc9000306fb18) at mm/vmscan.c:2464
> > #11 0xffffffff81180374 in shrink_node_memcgs (sc=0xffffc9000306fb18, pgdat=0xffff88803ffdb000) at mm/vmscan.c:2652
> > #12 shrink_node (pgdat=pgdat at entry=0xffff88803ffdb000, sc=sc at entry=0xffffc9000306fb18) at mm/vmscan.c:2769
> > #13 0xffffffff811806c8 in shrink_zones (sc=0xffffc9000306fb18, zonelist=0xffff88803ffdc400) at mm/vmscan.c:2972
> > #14 do_try_to_free_pages (zonelist=zonelist at entry=0xffff88803ffdc400, sc=sc at entry=0xffffc9000306fb18) at mm/vmscan.c:3027
> > #15 0xffffffff811817f6 in try_to_free_pages (zonelist=0xffff88803ffdc400, order=order at entry=1, gfp_mask=gfp_mask at entry=4197824, nodemask=<optimized out>) at mm/vmscan.c:3266
> > #16 0xffffffff811ba411 in __perform_reclaim (ac=0xffffc9000306fc90, ac=0xffffc9000306fc90, order=1, gfp_mask=4197824) at mm/page_alloc.c:4335
> > #17 __alloc_pages_direct_reclaim (did_some_progress=<synthetic pointer>, ac=0xffffc9000306fc90, alloc_flags=2112, order=1, gfp_mask=4197824) at mm/page_alloc.c:4356
> > #18 __alloc_pages_slowpath (gfp_mask=<optimized out>, gfp_mask at entry=4197824, order=order at entry=1, ac=ac at entry=0xffffc9000306fc90) at mm/page_alloc.c:4760
> > #19 0xffffffff811baf44 in __alloc_pages_nodemask (gfp_mask=<optimized out>, gfp_mask at entry=4197824, order=order at entry=1, preferred_nid=<optimized out>, nodemask=0x0 <fixed_percpu_data>) at mm/page_alloc.c:4970
> > #20 0xffffffff811ce039 in alloc_pages_current (gfp=gfp at entry=4197824, order=order at entry=1) at ./include/linux/topology.h:88
> > #21 0xffffffff811b6248 in alloc_pages (order=order at entry=1, gfp_mask=4197824) at ./include/linux/gfp.h:547
> > #22 __get_free_pages (gfp_mask=gfp_mask at entry=4197824, order=order at entry=1) at mm/page_alloc.c:4994
> > #23 0xffffffff8105482c in _pgd_alloc () at arch/x86/mm/pgtable.c:430
> > #24 pgd_alloc (mm=mm at entry=0xffff88800315e400) at arch/x86/mm/pgtable.c:430
> > #25 0xffffffff8105efae in mm_alloc_pgd (mm=0xffff88800315e400) at kernel/fork.c:1054
> > #26 mm_init (mm=mm at entry=0xffff88800315e400, user_ns=<optimized out>, p=0xffff888002bbc880) at kernel/fork.c:1054
> > #27 0xffffffff8105f624 in dup_mm (oldmm=0xffff888004efa800, tsk=0xffff888002bbc880) at kernel/fork.c:1369
> > #28 0xffffffff810616a5 in copy_mm (tsk=0xffff888002bbc880, clone_flags=0) at ./arch/x86/include/asm/current.h:15
> > #29 copy_process (pid=pid at entry=0x0 <fixed_percpu_data>, trace=trace at entry=0, node=node at entry=-1, args=args at entry=0xffffc9000306fed0) at kernel/fork.c:2110
> > #30 0xffffffff81061934 in kernel_clone (args=args at entry=0xffffc9000306fed0) at kernel/fork.c:2471
> > #31 0xffffffff81061c8f in __do_sys_fork (__unused=<optimized out>) at kernel/fork.c:2534
> > #32 0xffffffff81b41693 in do_syscall_64 (nr=<optimized out>, regs=0xffffc9000306ff58) at arch/x86/entry/common.c:55
> > #33 0xffffffff81c0007c in entry_SYSCALL_64 () at arch/x86/entry/entry_64.S:120
> > #34 0x00000000000526aa in ?? ()
> > #35 0x0000564f0b0bc2b0 in ?? ()
> > #36 0x0000000000000001 in fixed_percpu_data ()
> > #37 0x00007ffe0ee549f0 in ?? ()
> > #38 0x0000564f0b657260 in ?? ()
> > #39 0x0000000000000000 in ?? ()
> > 
> > 
> > Kernel-Config: Attached. It's a x86_64 defconfig with the following
> > modifications:
> >  - CONFIG_XENOMAI disabled
> >  - CONFIG_DOVETAIL disabled
> >  - CONFIG_MIGRATION disabled
> >  - CONFIG_DEBUG_INFO enabled (to be able to debug)
> >  - CONFIG_GDB_SCRIPTS enabled (debugging...)
> > 
> > I disabled Xenomai and Dovetail to limit the search scope. The problem
> > remains reproducible without them.
> > 
> > 
> > Workload:
> > Stressing the system with stress-ng. After 45 to 60 minutes the system
> > is frozen.
> > cmdline: stress-ng --cpu 4 --io 2 --vm 2 --vm-bytes 128M --fork 4 --timeout 0
> > 
> > 
> > IRQ flag:
> > All CPUs (or gdb threads) waiting at kernel/smp.c:228 have the IF flag
> > (part of eflag register) unset, while other CPUs have it set:
> > 
> > (gdb) info register
> > eflags         0x2                 [ ]
> > 
> > vs
> > 
> > eflags         0x202               [ IF ]
> > 
> > 
> > smp_call_function_many_cond() has some notes about deadlocks that might
> > appear when being called with IRQs disabled, but I actually never saw
> > one of the warnings that should come up. As IF flag is unset, someone
> > has to turn off IRQs later (while waiting) and that might be the reason
> > for the deadlock.
> 
> You should be able to prove it that way:
> 
> diff --git a/kernel/smp.c b/kernel/smp.c
> index 25240fb2df94922..35c439d1011b8fc 100644
> --- a/kernel/smp.c
> +++ b/kernel/smp.c
> @@ -211,6 +211,8 @@ static __always_inline void csd_lock_wait(call_single_data_t *csd)
>   	ts1 = ts0 = sched_clock();
> 
>  	for (;;) {
> +		WARN_ON_ONCE(hard_irqs_disabled());
> +		WARN_ON_ONCE(irqs_disabled());
>  		if (csd_lock_wait_toolong(csd, ts0, &ts1, &bug_id))
>  			break;
>  		cpu_relax();
> 
> It did happen that an issue in trap/irq mgmt caused the (virtual)
> interrupt state to become suddenly broken inside a loop. That snippet
> would detect it, without pinpointing the offender precisely though.

I will give it a try and report back.

> 
> A couple of questions:
> 
> - how long does it take to lockup a KVM instance with stress-ng that
>   way?

In my case about 45 minutes.

> 
> - are you running this kernel?
> https://git.evlproject.org/linux-evl.git/log/?h=dovetail/v5.10

Yes. Sorry for not mentioning that.

> 
> Thanks,
> 



More information about the Xenomai mailing list