MIPS: Fix race on setting and getting cpu_online_mask

While testing cpu hoptlug (cpu down and up in loops) on kernel 4.4, it was
observed that occasionally check for cpu online will fail in kernel/cpu.c,
_cpu_up:

https://git.kernel.org/pub/scm/linux/kernel/git/stable/linux-stable.git/tree/kernel/cpu.c?h=v4.4.79#n485
 518        /* Arch-specific enabling code. */
 519        ret = __cpu_up(cpu, idle);
 520
 521        if (ret != 0)
 522                goto out_notify;
 523        BUG_ON(!cpu_online(cpu));

Reason is race between start_secondary and _cpu_up. cpu_callin_map is set
before cpu_online_mask. In __cpu_up, cpu_callin_map is waited for, but cpu
online mask is not, resulting in race in which secondary processor started
and set cpu_callin_map, but not yet set the online mask,resulting in above
BUG being hit.

Upstream differs in the area. cpu_online check is in bringup_wait_for_ap,
which is after cpu reached AP_ONLINE_IDLE,where secondary passed its start
function. Nonetheless, fix makes start_secondary safe and not depending on
other locks throughout the code. It protects as well against cpu_online
checks put in between sometimes in the future.

Fix this by moving completion after all flags are set.

Signed-off-by: Matija Glavinic Pecotic <matija.glavinic-pecotic.ext@nokia.com>
Cc: Alexander Sverdlin <alexander.sverdlin@nokia.com>
Cc: linux-mips@linux-mips.org
Patchwork: https://patchwork.linux-mips.org/patch/16925/
Signed-off-by: Ralf Baechle <ralf@linux-mips.org>
This commit is contained in:
Matija Glavinic Pecotic 2017-08-03 08:20:22 +02:00 committed by Ralf Baechle
parent 00e06297b3
commit 6f542ebeae

View file

@ -376,9 +376,6 @@ asmlinkage void start_secondary(void)
cpumask_set_cpu(cpu, &cpu_coherent_mask);
notify_cpu_starting(cpu);
complete(&cpu_running);
synchronise_count_slave(cpu);
set_cpu_online(cpu, true);
set_cpu_sibling_map(cpu);
@ -386,6 +383,9 @@ asmlinkage void start_secondary(void)
calculate_cpu_foreign_map();
complete(&cpu_running);
synchronise_count_slave(cpu);
/*
* irq will be enabled in ->smp_finish(), enabling it too early
* is dangerous.