From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1753472AbaFHIfq (ORCPT ); Sun, 8 Jun 2014 04:35:46 -0400 Received: from mail-wg0-f52.google.com ([74.125.82.52]:55559 "EHLO mail-wg0-f52.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1753414AbaFHIfm (ORCPT ); Sun, 8 Jun 2014 04:35:42 -0400 Message-ID: <1402216538.31630.7.camel@marge.simpson.net> Subject: [RFC PATCH] rt/aio: fix rcu garbage collection might_sleep() splat From: Mike Galbraith To: RT Cc: LKML , Sebastian Andrzej Siewior , Steven Rostedt , Thomas Gleixner , "Paul E. McKenney" Date: Sun, 08 Jun 2014 10:35:38 +0200 Content-Type: text/plain; charset="UTF-8" X-Mailer: Evolution 3.2.3 Content-Transfer-Encoding: 7bit Mime-Version: 1.0 Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org [ 172.743098] BUG: sleeping function called from invalid context at kernel/locking/rtmutex.c:768 [ 172.743116] in_atomic(): 1, irqs_disabled(): 0, pid: 26, name: rcuos/2 [ 172.743117] 2 locks held by rcuos/2/26: [ 172.743128] #0: (rcu_callback){.+.+..}, at: [] rcu_nocb_kthread+0x1e2/0x380 [ 172.743135] #1: (rcu_read_lock_sched){.+.+..}, at: [] percpu_ref_kill_rcu+0xa6/0x1c0 [ 172.743138] Preemption disabled at:[] rcu_nocb_kthread+0x263/0x380 [ 172.743138] [ 172.743142] CPU: 0 PID: 26 Comm: rcuos/2 Not tainted 3.14.4-rt5 #31 [ 172.743143] Hardware name: MEDIONPC MS-7502/MS-7502, BIOS 6.00 PG 12/26/2007 [ 172.743148] ffff8802231aa190 ffff8802231a5d08 ffffffff81582e9e 0000000000000000 [ 172.743151] ffff8802231a5d28 ffffffff81077aeb ffff880209f68140 ffff880209f681c0 [ 172.743154] ffff8802231a5d48 ffffffff81589304 ffff880209f68000 ffff880209f68000 [ 172.743155] Call Trace: [ 172.743160] [] dump_stack+0x4e/0x9c [ 172.743163] [] __might_sleep+0xfb/0x170 [ 172.743167] [] rt_spin_lock+0x24/0x70 [ 172.743171] [] free_ioctx_users+0x30/0x130 [ 172.743174] [] percpu_ref_kill_rcu+0x1b4/0x1c0 [ 172.743177] [] ? percpu_ref_kill_rcu+0xa6/0x1c0 [ 172.743180] [] ? percpu_ref_kill_and_confirm+0x70/0x70 [ 172.743183] [] rcu_nocb_kthread+0x263/0x380 [ 172.743185] [] ? rcu_nocb_kthread+0x1e2/0x380 [ 172.743189] [] ? rcu_report_exp_rnp.isra.52+0xc0/0xc0 [ 172.743192] [] kthread+0xd6/0xf0 [ 172.743194] [] ? _raw_spin_unlock_irq+0x2c/0x70 [ 172.743197] [] ? __kthread_parkme+0x70/0x70 [ 172.743200] [] ret_from_fork+0x7c/0xb0 [ 172.743203] [] ? __kthread_parkme+0x70/0x70 crash> gdb list *percpu_ref_kill_rcu+0x1b4 0xffffffff812ace34 is in percpu_ref_kill_rcu (include/linux/percpu-refcount.h:169). 164 pcpu_count = ACCESS_ONCE(ref->pcpu_count); 165 166 if (likely(REF_STATUS(pcpu_count) == PCPU_REF_PTR)) 167 __this_cpu_dec(*pcpu_count); 168 else if (unlikely(atomic_dec_and_test(&ref->count))) 169 ref->release(ref); 170 171 rcu_read_unlock_sched(); 172 } 173 Ok, so ->release() can't do anything where it may meet a sleeping lock, but in an -rt kernel, it does that. Convert struct kioctx ctx_lock/completion_lock to raw_spinlock_t, and defer final free to a time when we're not under rcu_read_lock_sched(). runltp -f ltp-aio-stress.part1 runs kernel/ltp gripe free. INFO: ltp-pan reported all tests PASS LTP Version: 20140422 ############################################################### Done executing testcases. LTP Version: 20140422 ############################################################### Signed-off-by: Mike Galbraith --- fs/aio.c | 61 ++++++++++++++++++++++++++++++++++++++++++++----------------- 1 file changed, 44 insertions(+), 17 deletions(-) --- a/fs/aio.c +++ b/fs/aio.c @@ -125,7 +125,7 @@ struct kioctx { } ____cacheline_aligned_in_smp; struct { - spinlock_t ctx_lock; + raw_spinlock_t ctx_lock; struct list_head active_reqs; /* used for cancellation */ } ____cacheline_aligned_in_smp; @@ -136,13 +136,16 @@ struct kioctx { struct { unsigned tail; - spinlock_t completion_lock; + raw_spinlock_t completion_lock; } ____cacheline_aligned_in_smp; struct page *internal_pages[AIO_RING_PAGES]; struct file *aio_ring_file; unsigned id; +#ifdef CONFIG_PREEMPT_RT_BASE + struct rcu_head rcu; +#endif }; /*------ sysctl variables----*/ @@ -334,11 +337,11 @@ static int aio_migratepage(struct addres * while the old page is copied to the new. This prevents new * events from being lost. */ - spin_lock_irqsave(&ctx->completion_lock, flags); + raw_spin_lock_irqsave(&ctx->completion_lock, flags); migrate_page_copy(new, old); BUG_ON(ctx->ring_pages[idx] != old); ctx->ring_pages[idx] = new; - spin_unlock_irqrestore(&ctx->completion_lock, flags); + raw_spin_unlock_irqrestore(&ctx->completion_lock, flags); /* The old page is no longer accessible. */ put_page(old); @@ -461,14 +464,14 @@ void kiocb_set_cancel_fn(struct kiocb *r struct kioctx *ctx = req->ki_ctx; unsigned long flags; - spin_lock_irqsave(&ctx->ctx_lock, flags); + raw_spin_lock_irqsave(&ctx->ctx_lock, flags); if (!req->ki_list.next) list_add(&req->ki_list, &ctx->active_reqs); req->ki_cancel = cancel; - spin_unlock_irqrestore(&ctx->ctx_lock, flags); + raw_spin_unlock_irqrestore(&ctx->ctx_lock, flags); } EXPORT_SYMBOL(kiocb_set_cancel_fn); @@ -493,6 +496,7 @@ static int kiocb_cancel(struct kioctx *c return cancel(kiocb); } +#ifndef CONFIG_PREEMPT_RT_BASE static void free_ioctx(struct work_struct *work) { struct kioctx *ctx = container_of(work, struct kioctx, free_work); @@ -503,13 +507,36 @@ static void free_ioctx(struct work_struc free_percpu(ctx->cpu); kmem_cache_free(kioctx_cachep, ctx); } +#else +static void free_ioctx_rcu(struct rcu_head *rcu) +{ + struct kioctx *ctx = container_of(rcu, struct kioctx, rcu); + + pr_debug("freeing %p\n", ctx); + + aio_free_ring(ctx); + free_percpu(ctx->cpu); + kmem_cache_free(kioctx_cachep, ctx); +} +#endif static void free_ioctx_reqs(struct percpu_ref *ref) { struct kioctx *ctx = container_of(ref, struct kioctx, reqs); +#ifndef CONFIG_PREEMPT_RT_BASE INIT_WORK(&ctx->free_work, free_ioctx); schedule_work(&ctx->free_work); +#else + /* + * We're in ->release() under rcu_read_lock_sched(), and can't do + * anything that requires taking a sleeping lock, so ->release() + * becomes a two stage rcu process for -rt. We've now done the + * rcu work that we can under locks made raw to get us this far. + * Defer the freeing bit until we're again allowed to schedule(). + */ + call_rcu(&ctx->rcu, free_ioctx_rcu); +#endif } /* @@ -522,7 +549,7 @@ static void free_ioctx_users(struct perc struct kioctx *ctx = container_of(ref, struct kioctx, users); struct kiocb *req; - spin_lock_irq(&ctx->ctx_lock); + raw_spin_lock_irq(&ctx->ctx_lock); while (!list_empty(&ctx->active_reqs)) { req = list_first_entry(&ctx->active_reqs, @@ -532,7 +559,7 @@ static void free_ioctx_users(struct perc kiocb_cancel(ctx, req); } - spin_unlock_irq(&ctx->ctx_lock); + raw_spin_unlock_irq(&ctx->ctx_lock); percpu_ref_kill(&ctx->reqs); percpu_ref_put(&ctx->reqs); @@ -645,8 +672,8 @@ static struct kioctx *ioctx_alloc(unsign ctx->max_reqs = nr_events; - spin_lock_init(&ctx->ctx_lock); - spin_lock_init(&ctx->completion_lock); + raw_spin_lock_init(&ctx->ctx_lock); + raw_spin_lock_init(&ctx->completion_lock); mutex_init(&ctx->ring_lock); /* Protect against page migration throughout kiotx setup by keeping * the ring_lock mutex held until setup is complete. */ @@ -948,9 +975,9 @@ void aio_complete(struct kiocb *iocb, lo if (iocb->ki_list.next) { unsigned long flags; - spin_lock_irqsave(&ctx->ctx_lock, flags); + raw_spin_lock_irqsave(&ctx->ctx_lock, flags); list_del(&iocb->ki_list); - spin_unlock_irqrestore(&ctx->ctx_lock, flags); + raw_spin_unlock_irqrestore(&ctx->ctx_lock, flags); } /* @@ -958,7 +985,7 @@ void aio_complete(struct kiocb *iocb, lo * ctx->completion_lock to prevent other code from messing with the tail * pointer since we might be called from irq context. */ - spin_lock_irqsave(&ctx->completion_lock, flags); + raw_spin_lock_irqsave(&ctx->completion_lock, flags); tail = ctx->tail; pos = tail + AIO_EVENTS_OFFSET; @@ -993,7 +1020,7 @@ void aio_complete(struct kiocb *iocb, lo kunmap_atomic(ring); flush_dcache_page(ctx->ring_pages[0]); - spin_unlock_irqrestore(&ctx->completion_lock, flags); + raw_spin_unlock_irqrestore(&ctx->completion_lock, flags); pr_debug("added to ring %p at [%u]\n", iocb, tail); @@ -1515,7 +1542,7 @@ static struct kiocb *lookup_kiocb(struct { struct list_head *pos; - assert_spin_locked(&ctx->ctx_lock); + assert_raw_spin_locked(&ctx->ctx_lock); if (key != KIOCB_KEY) return NULL; @@ -1555,7 +1582,7 @@ SYSCALL_DEFINE3(io_cancel, aio_context_t if (unlikely(!ctx)) return -EINVAL; - spin_lock_irq(&ctx->ctx_lock); + raw_spin_lock_irq(&ctx->ctx_lock); kiocb = lookup_kiocb(ctx, iocb, key); if (kiocb) @@ -1563,7 +1590,7 @@ SYSCALL_DEFINE3(io_cancel, aio_context_t else ret = -EINVAL; - spin_unlock_irq(&ctx->ctx_lock); + raw_spin_unlock_irq(&ctx->ctx_lock); if (!ret) { /*