IB/ehca: Fix sync between completion handler and destroy cq
authorHoang-Nam Nguyen <hnguyen@linux.vnet.ibm.com>
Wed, 28 Feb 2007 17:01:02 +0000 (18:01 +0100)
committerRoland Dreier <rolandd@cisco.com>
Thu, 1 Mar 2007 21:04:05 +0000 (13:04 -0800)
This patch fixes two issues reported by Roland Dreier and Christoph Hellwig:

- Mismatched sync/locking between completion handler and destroy cq We
  introduced a counter nr_events per cq to track number of irq events
  seen. This counter is incremented when an event queue entry is seen
  and decremented after completion handler has been called regardless
  if scaling code is active or not. Note that nr_callbacks tracks
  number of events assigned to a cpu and both counters can potentially
  diverge.

  The sync between running completion handler and destroy cq is done
  by using the global spin lock ehca_cq_idr_lock.

- Replace yield by wait_event on the counter above to become zero.

Signed-off-by: Hoang-Nam Nguyen <hnguyen@de.ibm.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
drivers/infiniband/hw/ehca/ehca_classes.h
drivers/infiniband/hw/ehca/ehca_cq.c
drivers/infiniband/hw/ehca/ehca_irq.c
drivers/infiniband/hw/ehca/ehca_main.c

index 40404c9..82ded44 100644 (file)
@@ -52,6 +52,8 @@ struct ehca_mw;
 struct ehca_pd;
 struct ehca_av;
 
+#include <linux/wait.h>
+
 #include <rdma/ib_verbs.h>
 #include <rdma/ib_user_verbs.h>
 
@@ -153,7 +155,9 @@ struct ehca_cq {
        spinlock_t cb_lock;
        struct hlist_head qp_hashtab[QP_HASHTAB_LEN];
        struct list_head entry;
-       u32 nr_callbacks;
+       u32 nr_callbacks; /* #events assigned to cpu by scaling code */
+       u32 nr_events;    /* #events seen */
+       wait_queue_head_t wait_completion;
        spinlock_t task_lock;
        u32 ownpid;
        /* mmap counter for resources mapped into user space */
index 6ebfa27..e2cdc1a 100644 (file)
@@ -146,6 +146,7 @@ struct ib_cq *ehca_create_cq(struct ib_device *device, int cqe,
        spin_lock_init(&my_cq->spinlock);
        spin_lock_init(&my_cq->cb_lock);
        spin_lock_init(&my_cq->task_lock);
+       init_waitqueue_head(&my_cq->wait_completion);
        my_cq->ownpid = current->tgid;
 
        cq = &my_cq->ib_cq;
@@ -302,6 +303,16 @@ create_cq_exit1:
        return cq;
 }
 
+static int get_cq_nr_events(struct ehca_cq *my_cq)
+{
+       int ret;
+       unsigned long flags;
+       spin_lock_irqsave(&ehca_cq_idr_lock, flags);
+       ret = my_cq->nr_events;
+       spin_unlock_irqrestore(&ehca_cq_idr_lock, flags);
+       return ret;
+}
+
 int ehca_destroy_cq(struct ib_cq *cq)
 {
        u64 h_ret;
@@ -329,10 +340,11 @@ int ehca_destroy_cq(struct ib_cq *cq)
        }
 
        spin_lock_irqsave(&ehca_cq_idr_lock, flags);
-       while (my_cq->nr_callbacks) {
+       while (my_cq->nr_events) {
                spin_unlock_irqrestore(&ehca_cq_idr_lock, flags);
-               yield();
+               wait_event(my_cq->wait_completion, !get_cq_nr_events(my_cq));
                spin_lock_irqsave(&ehca_cq_idr_lock, flags);
+               /* recheck nr_events to assure no cqe has just arrived */
        }
 
        idr_remove(&ehca_cq_idr, my_cq->token);
index 3ec53c6..20f36bf 100644 (file)
@@ -404,10 +404,11 @@ static inline void process_eqe(struct ehca_shca *shca, struct ehca_eqe *eqe)
        u32 token;
        unsigned long flags;
        struct ehca_cq *cq;
+
        eqe_value = eqe->entry;
        ehca_dbg(&shca->ib_device, "eqe_value=%lx", eqe_value);
        if (EHCA_BMASK_GET(EQE_COMPLETION_EVENT, eqe_value)) {
-               ehca_dbg(&shca->ib_device, "... completion event");
+               ehca_dbg(&shca->ib_device, "Got completion event");
                token = EHCA_BMASK_GET(EQE_CQ_TOKEN, eqe_value);
                spin_lock_irqsave(&ehca_cq_idr_lock, flags);
                cq = idr_find(&ehca_cq_idr, token);
@@ -419,16 +420,20 @@ static inline void process_eqe(struct ehca_shca *shca, struct ehca_eqe *eqe)
                        return;
                }
                reset_eq_pending(cq);
-               if (ehca_scaling_code) {
+               cq->nr_events++;
+               spin_unlock_irqrestore(&ehca_cq_idr_lock, flags);
+               if (ehca_scaling_code)
                        queue_comp_task(cq);
-                       spin_unlock_irqrestore(&ehca_cq_idr_lock, flags);
-               } else {
-                       spin_unlock_irqrestore(&ehca_cq_idr_lock, flags);
+               else {
                        comp_event_callback(cq);
+                       spin_lock_irqsave(&ehca_cq_idr_lock, flags);
+                       cq->nr_events--;
+                       if (!cq->nr_events)
+                               wake_up(&cq->wait_completion);
+                       spin_unlock_irqrestore(&ehca_cq_idr_lock, flags);
                }
        } else {
-               ehca_dbg(&shca->ib_device,
-                        "Got non completion event");
+               ehca_dbg(&shca->ib_device, "Got non completion event");
                parse_identifier(shca, eqe_value);
        }
 }
@@ -478,6 +483,7 @@ void ehca_process_eq(struct ehca_shca *shca, int is_irq)
                                         "token=%x", token);
                                continue;
                        }
+                       eqe_cache[eqe_cnt].cq->nr_events++;
                        spin_unlock(&ehca_cq_idr_lock);
                } else
                        eqe_cache[eqe_cnt].cq = NULL;
@@ -504,12 +510,18 @@ void ehca_process_eq(struct ehca_shca *shca, int is_irq)
        /* call completion handler for cached eqes */
        for (i = 0; i < eqe_cnt; i++)
                if (eq->eqe_cache[i].cq) {
-                       if (ehca_scaling_code) {
-                               spin_lock(&ehca_cq_idr_lock);
+                       if (ehca_scaling_code)
                                queue_comp_task(eq->eqe_cache[i].cq);
-                               spin_unlock(&ehca_cq_idr_lock);
-                       } else
-                               comp_event_callback(eq->eqe_cache[i].cq);
+                       else {
+                               struct ehca_cq *cq = eq->eqe_cache[i].cq;
+                               comp_event_callback(cq);
+                               spin_lock_irqsave(&ehca_cq_idr_lock, flags);
+                               cq->nr_events--;
+                               if (!cq->nr_events)
+                                       wake_up(&cq->wait_completion);
+                               spin_unlock_irqrestore(&ehca_cq_idr_lock,
+                                                      flags);
+                       }
                } else {
                        ehca_dbg(&shca->ib_device, "Got non completion event");
                        parse_identifier(shca, eq->eqe_cache[i].eqe->entry);
@@ -523,7 +535,6 @@ void ehca_process_eq(struct ehca_shca *shca, int is_irq)
                if (!eqe)
                        break;
                process_eqe(shca, eqe);
-               eqe_cnt++;
        } while (1);
 
 unlock_irq_spinlock:
@@ -567,8 +578,7 @@ static void __queue_comp_task(struct ehca_cq *__cq,
                list_add_tail(&__cq->entry, &cct->cq_list);
                cct->cq_jobs++;
                wake_up(&cct->wait_queue);
-       }
-       else
+       } else
                __cq->nr_callbacks++;
 
        spin_unlock(&__cq->task_lock);
@@ -577,18 +587,21 @@ static void __queue_comp_task(struct ehca_cq *__cq,
 
 static void queue_comp_task(struct ehca_cq *__cq)
 {
-       int cpu;
        int cpu_id;
        struct ehca_cpu_comp_task *cct;
+       int cq_jobs;
+       unsigned long flags;
 
-       cpu = get_cpu();
        cpu_id = find_next_online_cpu(pool);
        BUG_ON(!cpu_online(cpu_id));
 
        cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu_id);
        BUG_ON(!cct);
 
-       if (cct->cq_jobs > 0) {
+       spin_lock_irqsave(&cct->task_lock, flags);
+       cq_jobs = cct->cq_jobs;
+       spin_unlock_irqrestore(&cct->task_lock, flags);
+       if (cq_jobs > 0) {
                cpu_id = find_next_online_cpu(pool);
                cct = per_cpu_ptr(pool->cpu_comp_tasks, cpu_id);
                BUG_ON(!cct);
@@ -608,11 +621,17 @@ static void run_comp_task(struct ehca_cpu_comp_task* cct)
                cq = list_entry(cct->cq_list.next, struct ehca_cq, entry);
                spin_unlock_irqrestore(&cct->task_lock, flags);
                comp_event_callback(cq);
-               spin_lock_irqsave(&cct->task_lock, flags);
 
+               spin_lock_irqsave(&ehca_cq_idr_lock, flags);
+               cq->nr_events--;
+               if (!cq->nr_events)
+                       wake_up(&cq->wait_completion);
+               spin_unlock_irqrestore(&ehca_cq_idr_lock, flags);
+
+               spin_lock_irqsave(&cct->task_lock, flags);
                spin_lock(&cq->task_lock);
                cq->nr_callbacks--;
-               if (cq->nr_callbacks == 0) {
+               if (!cq->nr_callbacks) {
                        list_del_init(cct->cq_list.next);
                        cct->cq_jobs--;
                }
index c183512..059da96 100644 (file)
@@ -52,7 +52,7 @@
 MODULE_LICENSE("Dual BSD/GPL");
 MODULE_AUTHOR("Christoph Raisch <raisch@de.ibm.com>");
 MODULE_DESCRIPTION("IBM eServer HCA InfiniBand Device Driver");
-MODULE_VERSION("SVNEHCA_0021");
+MODULE_VERSION("SVNEHCA_0022");
 
 int ehca_open_aqp1     = 0;
 int ehca_debug_level   = 0;
@@ -810,7 +810,7 @@ int __init ehca_module_init(void)
        int ret;
 
        printk(KERN_INFO "eHCA Infiniband Device Driver "
-              "(Rel.: SVNEHCA_0021)\n");
+              "(Rel.: SVNEHCA_0022)\n");
        idr_init(&ehca_qp_idr);
        idr_init(&ehca_cq_idr);
        spin_lock_init(&ehca_qp_idr_lock);