Thread (26 messages) 26 messages, 5 authors, 2020-02-07

Re: [PATCH V5 0/6] blk-mq: improvement CPU hotplug

From: Thomas Gleixner <hidden>
Date: 2020-02-01 11:32:15

Marc Zyngier [off-list ref] writes:
On Sat, 1 Feb 2020 09:31:17 +0800
Ming Lei [off-list ref] wrote:
quoted
On Sat, Feb 1, 2020 at 2:02 AM John Garry [off-list ref] wrote:

gic_set_affinity shouldn't have switched out, so looks like one gic
issue.
Given that gic_set_affinity doesn't sleep, this looks pretty unlikely.
And __irq_set_affinity() holds a spinlock with irq disabled, so I can't
really explain how you'd get there. I've just booted a lockdep enabled
v5.5 on my D05, moved SPIs around (because that's the only way to reach
this code), and nothing caught fire.

Either the stack trace isn't reliable (when I read things like
"80d:00h:35m:42s" in the trace, I'm a bit suspicious), or CPU hotplug is
doing something really funky here.
The hotplug code cannot end up in schedule either and it holds desc lock
as normal affinity setting. The other backtrace is more complete,

[  728.741808] rcu: INFO: rcu_preempt detected stalls on CPUs/tasks:80d:00h:35m:42s]
[  728.747895] rcu: 48-...0: (0 ticks this GP) idle=b3e/1/0x4000000000000000 softirq=5548/5548 fqs=2626
[  728.757197] (detected by 63, t=5255 jiffies, g=40989, q=1890)
[  728.763018] Task dump for CPU 48:
[  728.766321] irqbalance      R  running task        0  1272      1 0x00000002
[  728.773358] Call trace:
[  728.775801]  __switch_to+0xbc/0x218
[  728.779283]  gic_set_affinity+0x16c/0x1d8
[  728.783282]  irq_do_set_affinity+0x30/0xd0
[  728.787365]  irq_set_affinity_locked+0xc8/0xf0
[  728.791796]  __irq_set_affinity+0x4c/0x80
[  728.795794]  write_irq_affinity.isra.7+0x104/0x120
[  728.800572]  irq_affinity_proc_write+0x1c/0x28
[  728.805008]  proc_reg_write+0x78/0xb8
[  728.808660]  __vfs_write+0x18/0x38
[  728.812050]  vfs_write+0xb4/0x1e0
[  728.815352]  ksys_write+0x68/0xf8
[  728.818655]  __arm64_sys_write+0x18/0x20
[  728.822567]  el0_svc_common.constprop.2+0x64/0x160
[  728.827345]  el0_svc_handler+0x20/0x80
[  728.831082]  el0_sync_handler+0xe4/0x188
[  728.834991]  el0_sync+0x140/0x180

But the __switch_to() there definitely does not make any sense at all.

Thanks,

        tglx
Keyboard shortcuts
hback out one level
jnext message in thread
kprevious message in thread
ldrill in
Escclose help / fold thread tree
?toggle this help