linux-kernel - Re: [PATCH v2 02/11] locking/ww_mutex: Re-check ww->ctx in the inner optimistic spin loop

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <e9d4cf0f-a2c5-e248-c6f9-3d8c447f9d06@redhat.com>
Date:   Tue, 6 Dec 2016 11:03:28 -0500
From:   Waiman Long <longman@...hat.com>
To:     Peter Zijlstra <peterz@...radead.org>,
        Nicolai Hähnle <nhaehnle@...il.com>
Cc:     linux-kernel@...r.kernel.org,
        Nicolai Hähnle <Nicolai.Haehnle@....com>,
        Ingo Molnar <mingo@...hat.com>,
        Maarten Lankhorst <dev@...ankhorst.nl>,
        Daniel Vetter <daniel@...ll.ch>,
        Chris Wilson <chris@...is-wilson.co.uk>,
        dri-devel@...ts.freedesktop.org
Subject: Re: [PATCH v2 02/11] locking/ww_mutex: Re-check ww->ctx in the inner
 optimistic spin loop

On 12/06/2016 10:06 AM, Peter Zijlstra wrote:
> On Thu, Dec 01, 2016 at 03:06:45PM +0100, Nicolai Hähnle wrote:
>> +++ b/kernel/locking/mutex.c
>> @@ -350,7 +350,8 @@ ww_mutex_set_context_slowpath(struct ww_mutex *lock,
>>   * access and not reliable.
>>   */
>>  static noinline
>> -bool mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
>> +bool mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner,
>> +			 bool use_ww_ctx, struct ww_acquire_ctx *ww_ctx)
>>  {
>>  	bool ret = true;
>>  
>> @@ -373,6 +374,28 @@ bool mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
>>  			break;
>>  		}
>>  
>> +		if (use_ww_ctx && ww_ctx->acquired > 0) {
>> +			struct ww_mutex *ww;
>> +
>> +			ww = container_of(lock, struct ww_mutex, base);
>> +
>> +			/*
>> +			 * If ww->ctx is set the contents are undefined, only
>> +			 * by acquiring wait_lock there is a guarantee that
>> +			 * they are not invalid when reading.
>> +			 *
>> +			 * As such, when deadlock detection needs to be
>> +			 * performed the optimistic spinning cannot be done.
>> +			 *
>> +			 * Check this in every inner iteration because we may
>> +			 * be racing against another thread's ww_mutex_lock.
>> +			 */
>> +			if (READ_ONCE(ww->ctx)) {
>> +				ret = false;
>> +				break;
>> +			}
>> +		}
>> +
>>  		cpu_relax();
>>  	}
>>  	rcu_read_unlock();
> Aside from the valid question about mutex_can_spin_on_owner() there's
> another 'problem' here, mutex_spin_on_owner() is marked noinline, so all
> the use_ww_ctx stuff doesn't 'work' here.

The mutex_spin_on_owner() function was originally marked noinline
because it could be a major consumer of CPU cycles in a contended lock.
Having it shown separately in the perf output will help the users have a
better understanding of what is consuming all the CPU cycles. So I would
still like to keep it this way.

If you have concern about additional latency for non-ww_mutex calls, one
alternative can be:

diff --git a/kernel/locking/mutex.c b/kernel/locking/mutex.c
index 0afa998..777338d 100644
--- a/kernel/locking/mutex.c
+++ b/kernel/locking/mutex.c
@@ -349,9 +349,9 @@ static __always_inline void ww_mutex_lock_acquired(struct ww
  * Look out! "owner" is an entirely speculative pointer
  * access and not reliable.
  */
-static noinline
-bool mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner,
-                        bool use_ww_ctx, struct ww_acquire_ctx *ww_ctx)
+static __always_inline
+bool __mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner,
+                          const bool use_ww_ctx, struct ww_acquire_ctx *ww_ctx)
 {
        bool ret = true;
 
@@ -403,6 +403,19 @@ bool mutex_spin_on_owner(struct mutex *lock, struct task_st
        return ret;
 }
 
+static noinline
+bool mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
+{
+       return __mutex_spin_on_owner(lock, owner, false, NULL);
+}
+
+static noinline
+bool ww_mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner,
+                           struct ww_acquire_ctx *ww_ctx)
+{
+       return __mutex_spin_on_owner(lock, owner, true, ww_ctx);
+}
+
 /*
  * Initial check for entering the mutex spinning loop
  */
@@ -489,13 +502,17 @@ static bool mutex_optimistic_spin(struct mutex *lock,
                 */
                owner = __mutex_owner(lock);
                if (owner) {
+                       bool spin_ok;
+
                        if (waiter && owner == task) {
                                smp_mb(); /* ACQUIRE */
                                break;
                        }
 
-                       if (!mutex_spin_on_owner(lock, owner, use_ww_ctx,
-                                                ww_ctx))
+                       spin_ok = use_ww_ctx
+                               ? ww_mutex_spin_on_owner(lock, owner, ww_ctx)
+                               : mutex_spin_on_owner(lock, owner);
+                       if (!spin_ok)
                                goto fail_unlock;
                }