Messages in this thread | | | Date | Sat, 19 May 2012 19:32:59 -0700 (PDT) | From | David Rientjes <> | Subject | Re: [PATCH] pidmap: Use GFP_ATOMIC to allocate page (was: Re: [ 00/54] 3.0.32-stable review) |
| |
On Sat, 19 May 2012, Steven Rostedt wrote:
> Ran 18 tests against 3.0.32-rc, 2 failed (16 passed!). One failed by a > lock up (still looking into it) and the other is this one: > > Calibrating delay loop (skipped), value calculated using timer frequency.. 4799.59 BogoMIPS (lpj=2399799) > pid_max: default: 32768 minimum: 301 > BUG: scheduling while atomic: swapper/0/0x10000002 > Modules linked in: > Pid: 0, comm: swapper Not tainted 3.1.0-rc9-test+ #1 > Call Trace: > [<ffffffff81435b3e>] __schedule_bug+0x57/0x5b > [<ffffffff8143c44f>] __schedule+0x90/0x672 > [<ffffffff810636dd>] ? kzalloc.constprop.2+0x24/0x26 > [<ffffffff810460d1>] __cond_resched+0x23/0x2f > [<ffffffff8143ca83>] _cond_resched+0x14/0x1d^M > [<ffffffff810f7cd3>] slab_pre_alloc_hook.isra.52+0x28/0x2c > [<ffffffff810f8b1a>] kmem_cache_alloc_trace+0x29/0xbd > [<ffffffff810636dd>] kzalloc.constprop.2+0x24/0x26 > [<ffffffff81b7a434>] pidmap_init+0x81/0xc0 > [<ffffffff81b61b03>] start_kernel+0x330/0x3c2 > [<ffffffff81b612c4>] x86_64_start_reservations+0xaf/0xb3 > [<ffffffff81b61140>] ? early_idt_handlers+0x140/0x140 > [<ffffffff81b613ca>] x86_64_start_kernel+0x102/0x111 > Security Framework initialized > > Had CONFIG_SLUB, IRQ_FORCED_THREADING and voluntary preempt enabled. > Seems the might_resched() check from might_sleep_if() in > pre_alloc_hook() caused a schedule in early boot (pidmap_init). Not sure > why it scheduled that early, perhaps a threaded interrupt was added? I > haven't been able to reproduce it again. > > Perhaps we should not be using GFP_KERNEL when allocating memory for > pidmap_init()? As we have preemption disabled at this point, we should > be using GFP_ATOMIC. Also add a BUG_ON if the page fails to get > allocated. > > Signed-off-by: Steven Rostedt <rostedt@goodmis.org> > > This patch is against 3.4-rc2 (just happened to be a kernel checkout I > had). > > diff --git a/kernel/pid.c b/kernel/pid.c > index 9f08dfa..32dab20 100644 > --- a/kernel/pid.c > +++ b/kernel/pid.c > @@ -563,7 +563,9 @@ void __init pidmap_init(void) > PIDS_PER_CPU_MIN * num_possible_cpus()); > pr_info("pid_max: default: %u minimum: %u\n", pid_max, pid_max_min); > > - init_pid_ns.pidmap[0].page = kzalloc(PAGE_SIZE, GFP_KERNEL); > + /* Preemption is still disabled at this point */ > + init_pid_ns.pidmap[0].page = kzalloc(PAGE_SIZE, GFP_ATOMIC); > + BUG_ON(!init_pid_ns.pidmap[0].page); > /* Reserve PID 0. We never call free_pidmap(0) */ > set_bit(0, init_pid_ns.pidmap[0].page); > atomic_dec(&init_pid_ns.pidmap[0].nr_free);
Why wasn't this caught by gfp_allowed_mask in slab_pre_alloc_hook()? GFP_KERNEL should be allowed in this context.
| |