[SCSI] ibmvfc: Fix dropped interrupts
authorBrian King <brking@linux.vnet.ibm.com>
Fri, 20 Mar 2009 20:44:35 +0000 (15:44 -0500)
committerJames Bottomley <James.Bottomley@HansenPartnership.com>
Fri, 3 Apr 2009 14:22:42 +0000 (09:22 -0500)
This patch fixes a problem of possible dropped interrupts. Currently,
the ibmvfc driver has a race condition where after ibmvfc_interrupt
gets run, the platform code clears the interrupt. This can result in
lost interrupts and, in worst case scenarios, result in command
timeouts. Fix this by implementing a tasklet similar to what the
ibmvscsi driver does so that interrupt processing is no longer done in
the actual interrupt handler, which eliminates the race.

Signed-off-by: Brian King <brking@linux.vnet.ibm.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
drivers/scsi/ibmvscsi/ibmvfc.c
drivers/scsi/ibmvscsi/ibmvfc.h

index 93d1fbe4ee5d2a3ef40eea6458e6f2cc8e3a0c44..229e360f02202dac3b08d9c7801cf1c3212c870a 100644 (file)
@@ -640,6 +640,7 @@ static void ibmvfc_release_crq_queue(struct ibmvfc_host *vhost)
 
        ibmvfc_dbg(vhost, "Releasing CRQ\n");
        free_irq(vdev->irq, vhost);
+       tasklet_kill(&vhost->tasklet);
        do {
                rc = plpar_hcall_norets(H_FREE_CRQ, vdev->unit_address);
        } while (rc == H_BUSY || H_IS_LONG_BUSY(rc));
@@ -2699,6 +2700,25 @@ static struct ibmvfc_crq *ibmvfc_next_crq(struct ibmvfc_host *vhost)
 static irqreturn_t ibmvfc_interrupt(int irq, void *dev_instance)
 {
        struct ibmvfc_host *vhost = (struct ibmvfc_host *)dev_instance;
+       unsigned long flags;
+
+       spin_lock_irqsave(vhost->host->host_lock, flags);
+       vio_disable_interrupts(to_vio_dev(vhost->dev));
+       tasklet_schedule(&vhost->tasklet);
+       spin_unlock_irqrestore(vhost->host->host_lock, flags);
+       return IRQ_HANDLED;
+}
+
+/**
+ * ibmvfc_tasklet - Interrupt handler tasklet
+ * @data:              ibmvfc host struct
+ *
+ * Returns:
+ *     Nothing
+ **/
+static void ibmvfc_tasklet(void *data)
+{
+       struct ibmvfc_host *vhost = data;
        struct vio_dev *vdev = to_vio_dev(vhost->dev);
        struct ibmvfc_crq *crq;
        struct ibmvfc_async_crq *async;
@@ -2706,7 +2726,6 @@ static irqreturn_t ibmvfc_interrupt(int irq, void *dev_instance)
        int done = 0;
 
        spin_lock_irqsave(vhost->host->host_lock, flags);
-       vio_disable_interrupts(to_vio_dev(vhost->dev));
        while (!done) {
                /* Pull all the valid messages off the CRQ */
                while ((crq = ibmvfc_next_crq(vhost)) != NULL) {
@@ -2734,7 +2753,6 @@ static irqreturn_t ibmvfc_interrupt(int irq, void *dev_instance)
        }
 
        spin_unlock_irqrestore(vhost->host->host_lock, flags);
-       return IRQ_HANDLED;
 }
 
 /**
@@ -3859,6 +3877,8 @@ static int ibmvfc_init_crq(struct ibmvfc_host *vhost)
 
        retrc = 0;
 
+       tasklet_init(&vhost->tasklet, (void *)ibmvfc_tasklet, (unsigned long)vhost);
+
        if ((rc = request_irq(vdev->irq, ibmvfc_interrupt, 0, IBMVFC_NAME, vhost))) {
                dev_err(dev, "Couldn't register irq 0x%x. rc=%d\n", vdev->irq, rc);
                goto req_irq_failed;
@@ -3874,6 +3894,7 @@ static int ibmvfc_init_crq(struct ibmvfc_host *vhost)
        return retrc;
 
 req_irq_failed:
+       tasklet_kill(&vhost->tasklet);
        do {
                rc = plpar_hcall_norets(H_FREE_CRQ, vdev->unit_address);
        } while (rc == H_BUSY || H_IS_LONG_BUSY(rc));
index b21e071b986214875074ea35748e775a4ec51469..70107522e3a91d782a7b6b6fe88710276fcac01b 100644 (file)
@@ -684,6 +684,7 @@ struct ibmvfc_host {
        char partition_name[97];
        void (*job_step) (struct ibmvfc_host *);
        struct task_struct *work_thread;
+       struct tasklet_struct tasklet;
        wait_queue_head_t init_wait_q;
        wait_queue_head_t work_wait_q;
 };