IB/rdmavt: Avoid queuing work into a destroyed cq kthread worker
authorPetr Mladek <pmladek@suse.com>
Wed, 19 Oct 2016 12:07:19 +0000 (14:07 +0200)
committerDoug Ledford <dledford@redhat.com>
Wed, 14 Dec 2016 17:16:11 +0000 (12:16 -0500)
The memory barrier is not enough to protect queuing works into
a destroyed cq kthread. Just imagine the following situation:

CPU1 CPU2

rvt_cq_enter()
  worker =  cq->rdi->worker;

rvt_cq_exit()
  rdi->worker = NULL;
  smp_wmb();
  kthread_flush_worker(worker);
  kthread_stop(worker->task);
  kfree(worker);

  // nothing queued yet =>
  // nothing flushed and
  // happily stopped and freed

  if (likely(worker)) {
     // true => read before CPU2 acted
     cq->notify = RVT_CQ_NONE;
     cq->triggered++;
     kthread_queue_work(worker, &cq->comptask);

  BANG: worker has been flushed/stopped/freed in the meantime.

This patch solves this by protecting the critical sections by
rdi->n_cqs_lock. It seems that this lock is not much contended
and looks reasonable for this purpose.

One catch is that rvt_cq_enter() might be called from IRQ context.
Therefore we must always take the lock with IRQs disabled to avoid
a possible deadlock.

Signed-off-by: Petr Mladek <pmladek@suse.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
drivers/infiniband/sw/rdmavt/cq.c

index 6d9904a4a0abe7efb9e882942cd76f96dc920806..223ec4589fc7ad25deec99b37bed272bd01746cb 100644 (file)
@@ -119,18 +119,17 @@ void rvt_cq_enter(struct rvt_cq *cq, struct ib_wc *entry, bool solicited)
        if (cq->notify == IB_CQ_NEXT_COMP ||
            (cq->notify == IB_CQ_SOLICITED &&
             (solicited || entry->status != IB_WC_SUCCESS))) {
-               struct kthread_worker *worker;
                /*
                 * This will cause send_complete() to be called in
                 * another thread.
                 */
-               smp_read_barrier_depends(); /* see rvt_cq_exit */
-               worker = cq->rdi->worker;
-               if (likely(worker)) {
+               spin_lock(&cq->rdi->n_cqs_lock);
+               if (likely(cq->rdi->worker)) {
                        cq->notify = RVT_CQ_NONE;
                        cq->triggered++;
-                       kthread_queue_work(worker, &cq->comptask);
+                       kthread_queue_work(cq->rdi->worker, &cq->comptask);
                }
+               spin_unlock(&cq->rdi->n_cqs_lock);
        }
 
        spin_unlock_irqrestore(&cq->lock, flags);
@@ -240,15 +239,15 @@ struct ib_cq *rvt_create_cq(struct ib_device *ibdev,
                }
        }
 
-       spin_lock(&rdi->n_cqs_lock);
+       spin_lock_irq(&rdi->n_cqs_lock);
        if (rdi->n_cqs_allocated == rdi->dparms.props.max_cq) {
-               spin_unlock(&rdi->n_cqs_lock);
+               spin_unlock_irq(&rdi->n_cqs_lock);
                ret = ERR_PTR(-ENOMEM);
                goto bail_ip;
        }
 
        rdi->n_cqs_allocated++;
-       spin_unlock(&rdi->n_cqs_lock);
+       spin_unlock_irq(&rdi->n_cqs_lock);
 
        if (cq->ip) {
                spin_lock_irq(&rdi->pending_lock);
@@ -296,9 +295,9 @@ int rvt_destroy_cq(struct ib_cq *ibcq)
        struct rvt_dev_info *rdi = cq->rdi;
 
        kthread_flush_work(&cq->comptask);
-       spin_lock(&rdi->n_cqs_lock);
+       spin_lock_irq(&rdi->n_cqs_lock);
        rdi->n_cqs_allocated--;
-       spin_unlock(&rdi->n_cqs_lock);
+       spin_unlock_irq(&rdi->n_cqs_lock);
        if (cq->ip)
                kref_put(&cq->ip->ref, rvt_release_mmap_info);
        else
@@ -541,12 +540,15 @@ void rvt_cq_exit(struct rvt_dev_info *rdi)
 {
        struct kthread_worker *worker;
 
-       worker = rdi->worker;
-       if (!worker)
+       /* block future queuing from send_complete() */
+       spin_lock_irq(&rdi->n_cqs_lock);
+       if (!rdi->worker) {
+               spin_unlock_irq(&rdi->n_cqs_lock);
                return;
-       /* blocks future queuing from send_complete() */
+       }
        rdi->worker = NULL;
-       smp_wmb(); /* See rdi_cq_enter */
+       spin_unlock_irq(&rdi->n_cqs_lock);
+
        kthread_flush_worker(worker);
        kthread_stop(worker->task);
        kfree(worker);