On Tue, Apr 05, 2016 at 01:57:42PM +0200, Sebastian Andrzej Siewior wrote: > On 04/05/2016 01:51 PM, rcoch...@linutronix.de wrote: > > On Tue, Apr 05, 2016 at 01:36:38PM +0200, Heiko Carstens wrote: > >> On Tue, Apr 05, 2016 at 01:23:36PM +0200, Heiko Carstens wrote: > >>> Subsequently, in this case, the setup_pmc_cpu() call will be executed on > >>> the wrong cpu. > >> > >> .. or to illustrate this behaviour: the following patch (white space > >> damaged due to copy-paste) results in the following: > > > > I guess you are missing the following commit? > … > > cpu/hotplug: Move online calls to hotplugged cpu > > No, Heiko is right here. If one of the "CPU_DOWN_PREPARE" fails then > the following CPU_DOWN_FAILED will be invoked on the correct CPU. > > However if we are further down the road and the final ARCH specific > "die" failed (just before CPU_DYING) are invoked then we get this done > on the wrong CPU.
I think there is more broken: if I willingly let __cpu_disable() fail and try to offline e.g. cpu 2 for the second time chcpu will never return. Plus the console contains several "NOHZ: local_softirq_pending 01" messages. # cat /proc/1619/stack [<000000000013e460>] cpuhp_kick_ap_work+0x78/0x1b8 [<00000000008a1972>] _cpu_down+0xca/0x1c0 [<000000000013f362>] do_cpu_down+0x5a/0x88 [<0000000000682308>] device_offline+0xb8/0xe0 [<000000000068244e>] online_store+0x5e/0x98 [<000000000037ecea>] kernfs_fop_write+0x13a/0x190 [<00000000002ee26e>] __vfs_write+0x36/0x108 [<00000000002ef3e4>] vfs_write+0x94/0x1a0 [<00000000002f0ace>] SyS_write+0x66/0xd8 [<00000000008aa944>] system_call+0x244/0x264 [<ffffffffffffffff>] 0xffffffffffffffff (1619 is the pid of chcpu) All of this works without problems on vanilla 4.5 kernel. I think you can reproduce this on any architecture :)
