[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20190925093153.GC4553@hirez.programming.kicks-ass.net>
Date: Wed, 25 Sep 2019 11:31:53 +0200
From: Peter Zijlstra <peterz@...radead.org>
To: Qian Cai <cai@....pw>
Cc: akpm@...ux-foundation.org, bigeasy@...utronix.de,
tglx@...utronix.de, thgarnie@...gle.com, tytso@....edu,
cl@...ux.com, penberg@...nel.org, rientjes@...gle.com,
mingo@...hat.com, will@...nel.org, linux-mm@...ck.org,
linux-kernel@...r.kernel.org, keescook@...omium.org
Subject: Re: [PATCH] mm/slub: fix a deadlock in shuffle_freelist()
On Fri, Sep 13, 2019 at 12:27:44PM -0400, Qian Cai wrote:
> The commit b7d5dc21072c ("random: add a spinlock_t to struct
> batched_entropy") insists on acquiring "batched_entropy_u32.lock" in
> get_random_u32() which introduced the lock chain,
>
> "&rq->lock --> batched_entropy_u32.lock"
>
> even after crng init. As the result, it could result in deadlock below.
> Fix it by using get_random_bytes() in shuffle_freelist() which does not
> need to take on the batched_entropy locks.
>
> WARNING: possible circular locking dependency detected
> 5.3.0-rc7-mm1+ #3 Tainted: G L
> ------------------------------------------------------
> make/7937 is trying to acquire lock:
> ffff900012f225f8 (random_write_wait.lock){....}, at:
> __wake_up_common_lock+0xa8/0x11c
>
> but task is already holding lock:
> ffff0096b9429c00 (batched_entropy_u32.lock){-.-.}, at:
> get_random_u32+0x6c/0x1dc
>
> which lock already depends on the new lock.
>
> the existing dependency chain (in reverse order) is:
>
> -> #3 (batched_entropy_u32.lock){-.-.}:
> lock_acquire+0x31c/0x360
> _raw_spin_lock_irqsave+0x7c/0x9c
> get_random_u32+0x6c/0x1dc
> new_slab+0x234/0x6c0
> ___slab_alloc+0x3c8/0x650
> kmem_cache_alloc+0x4b0/0x590
> __debug_object_init+0x778/0x8b4
> debug_object_init+0x40/0x50
> debug_init+0x30/0x29c
> hrtimer_init+0x30/0x50
> init_dl_task_timer+0x24/0x44
> __sched_fork+0xc0/0x168
> init_idle+0x78/0x26c
> fork_idle+0x12c/0x178
> idle_threads_init+0x108/0x178
> smp_init+0x20/0x1bc
> kernel_init_freeable+0x198/0x26c
> kernel_init+0x18/0x334
> ret_from_fork+0x10/0x18
>
> -> #2 (&rq->lock){-.-.}:
This relation is silly..
I suspect the below 'works'...
diff --git a/kernel/sched/core.c b/kernel/sched/core.c
index 63900ca029e0..ec1d72f18b34 100644
--- a/kernel/sched/core.c
+++ b/kernel/sched/core.c
@@ -6027,10 +6027,11 @@ void init_idle(struct task_struct *idle, int cpu)
struct rq *rq = cpu_rq(cpu);
unsigned long flags;
+ __sched_fork(0, idle);
+
raw_spin_lock_irqsave(&idle->pi_lock, flags);
raw_spin_lock(&rq->lock);
- __sched_fork(0, idle);
idle->state = TASK_RUNNING;
idle->se.exec_start = sched_clock();
idle->flags |= PF_IDLE;
Powered by blists - more mailing lists