]> git.openfabrics.org - ~shefty/rdma-dev.git/commitdiff
wimax i2400m: fix race condition while accessing rx_roq by using kref count
authorPrasanna S. Panchamukhi <prasannax.s.panchamukhi@intel.com>
Tue, 13 Apr 2010 23:35:58 +0000 (16:35 -0700)
committerInaky Perez-Gonzalez <inaky.perez-gonzalez@intel.com>
Tue, 11 May 2010 21:08:23 +0000 (14:08 -0700)
This patch fixes the race condition when one thread tries to destroy
the memory allocated for rx_roq, while another thread still happen
to access rx_roq.
Such a race condition occurs when i2400m-sdio kernel module gets
unloaded, destroying the memory allocated for rx_roq while rx_roq
is accessed by i2400m_rx_edata(), as explained below:
$thread1                                $thread2
$ void i2400m_rx_edata()                $
$Access rx_roq[]                        $
$roq = &i2400m->rx_roq[ro_cin]          $
$ i2400m_roq_[reset/queue/update_ws]    $
$                                       $ void i2400m_rx_release();
$                                       $kfree(rx->roq);
$                                       $rx->roq = NULL;
$Oops! rx_roq is NULL

This patch fixes the race condition using refcount approach.

Signed-off-by: Prasanna S. Panchamukhi <prasannax.s.panchamukhi@intel.com>
drivers/net/wimax/i2400m/i2400m.h
drivers/net/wimax/i2400m/rx.c

index 7a9c2c5b25cb92a9e5b271d0dbcf6758aecaf046..b8c7dbfead498a3b04b77165baf9c7ece39e7792 100644 (file)
@@ -412,7 +412,7 @@ struct i2400m_barker_db;
  *
  * @tx_size_max: biggest TX message sent.
  *
- * @rx_lock: spinlock to protect RX members
+ * @rx_lock: spinlock to protect RX members and rx_roq_refcount.
  *
  * @rx_pl_num: total number of payloads received
  *
@@ -436,6 +436,10 @@ struct i2400m_barker_db;
  *     delivered. Then the driver can release them to the host. See
  *     drivers/net/i2400m/rx.c for details.
  *
+ * @rx_roq_refcount: refcount rx_roq. This refcounts any access to
+ *     rx_roq thus preventing rx_roq being destroyed when rx_roq
+ *     is being accessed. rx_roq_refcount is protected by rx_lock.
+ *
  * @rx_reports: reports received from the device that couldn't be
  *     processed because the driver wasn't still ready; when ready,
  *     they are pulled from here and chewed.
@@ -597,10 +601,12 @@ struct i2400m {
                tx_num, tx_size_acc, tx_size_min, tx_size_max;
 
        /* RX stuff */
-       spinlock_t rx_lock;             /* protect RX state */
+       /* protect RX state and rx_roq_refcount */
+       spinlock_t rx_lock;
        unsigned rx_pl_num, rx_pl_max, rx_pl_min,
                rx_num, rx_size_acc, rx_size_min, rx_size_max;
-       struct i2400m_roq *rx_roq;      /* not under rx_lock! */
+       struct i2400m_roq *rx_roq;      /* access is refcounted */
+       struct kref rx_roq_refcount;    /* refcount access to rx_roq */
        u8 src_mac_addr[ETH_HLEN];
        struct list_head rx_reports;    /* under rx_lock! */
        struct work_struct rx_report_ws;
index fa2e11e5b4b9cb32718bb827f41ecbe1c3c580ef..71b697f3a68d0cc8d4e5f527b9484cc151ea5267 100644 (file)
@@ -916,6 +916,25 @@ void i2400m_roq_queue_update_ws(struct i2400m *i2400m, struct i2400m_roq *roq,
 }
 
 
+/*
+ * This routine destroys the memory allocated for rx_roq, when no
+ * other thread is accessing it. Access to rx_roq is refcounted by
+ * rx_roq_refcount, hence memory allocated must be destroyed when
+ * rx_roq_refcount becomes zero. This routine gets executed when
+ * rx_roq_refcount becomes zero.
+ */
+void i2400m_rx_roq_destroy(struct kref *ref)
+{
+       unsigned itr;
+       struct i2400m *i2400m
+                       = container_of(ref, struct i2400m, rx_roq_refcount);
+       for (itr = 0; itr < I2400M_RO_CIN + 1; itr++)
+               __skb_queue_purge(&i2400m->rx_roq[itr].queue);
+       kfree(i2400m->rx_roq[0].log);
+       kfree(i2400m->rx_roq);
+       i2400m->rx_roq = NULL;
+}
+
 /*
  * Receive and send up an extended data packet
  *
@@ -969,6 +988,7 @@ void i2400m_rx_edata(struct i2400m *i2400m, struct sk_buff *skb_rx,
        unsigned ro_needed, ro_type, ro_cin, ro_sn;
        struct i2400m_roq *roq;
        struct i2400m_roq_data *roq_data;
+       unsigned long flags;
 
        BUILD_BUG_ON(ETH_HLEN > sizeof(*hdr));
 
@@ -1007,7 +1027,16 @@ void i2400m_rx_edata(struct i2400m *i2400m, struct sk_buff *skb_rx,
                ro_cin = (reorder >> I2400M_RO_CIN_SHIFT) & I2400M_RO_CIN;
                ro_sn = (reorder >> I2400M_RO_SN_SHIFT) & I2400M_RO_SN;
 
+               spin_lock_irqsave(&i2400m->rx_lock, flags);
                roq = &i2400m->rx_roq[ro_cin];
+               if (roq == NULL) {
+                       kfree_skb(skb); /* rx_roq is already destroyed */
+                       spin_unlock_irqrestore(&i2400m->rx_lock, flags);
+                       goto error;
+               }
+               kref_get(&i2400m->rx_roq_refcount);
+               spin_unlock_irqrestore(&i2400m->rx_lock, flags);
+
                roq_data = (struct i2400m_roq_data *) &skb->cb;
                roq_data->sn = ro_sn;
                roq_data->cs = cs;
@@ -1034,6 +1063,10 @@ void i2400m_rx_edata(struct i2400m *i2400m, struct sk_buff *skb_rx,
                default:
                        dev_err(dev, "HW BUG? unknown reorder type %u\n", ro_type);
                }
+
+               spin_lock_irqsave(&i2400m->rx_lock, flags);
+               kref_put(&i2400m->rx_roq_refcount, i2400m_rx_roq_destroy);
+               spin_unlock_irqrestore(&i2400m->rx_lock, flags);
        }
        else
                i2400m_net_erx(i2400m, skb, cs);
@@ -1344,6 +1377,7 @@ int i2400m_rx_setup(struct i2400m *i2400m)
                        __i2400m_roq_init(&i2400m->rx_roq[itr]);
                        i2400m->rx_roq[itr].log = &rd[itr];
                }
+               kref_init(&i2400m->rx_roq_refcount);
        }
        return 0;
 
@@ -1357,12 +1391,12 @@ error_roq_alloc:
 /* Tear down the RX queue and infrastructure */
 void i2400m_rx_release(struct i2400m *i2400m)
 {
+       unsigned long flags;
+
        if (i2400m->rx_reorder) {
-               unsigned itr;
-               for(itr = 0; itr < I2400M_RO_CIN + 1; itr++)
-                       __skb_queue_purge(&i2400m->rx_roq[itr].queue);
-               kfree(i2400m->rx_roq[0].log);
-               kfree(i2400m->rx_roq);
+               spin_lock_irqsave(&i2400m->rx_lock, flags);
+               kref_put(&i2400m->rx_roq_refcount, i2400m_rx_roq_destroy);
+               spin_unlock_irqrestore(&i2400m->rx_lock, flags);
        }
        /* at this point, nothing can be received... */
        i2400m_report_hook_flush(i2400m);