remove yield() and use wait_for_completion() in order to wait for running
completion handlers finished before destroying associated completion queue
Signed-off-by: Hoang-Nam Nguyen <[email protected]>
---
ehca_classes.h | 3 +++
ehca_cq.c | 5 +++--
ehca_irq.c | 6 +++++-
3 files changed, 11 insertions(+), 3 deletions(-)
diff --git a/drivers/infiniband/hw/ehca/ehca_classes.h b/drivers/infiniband/hw/ehca/ehca_classes.h
index 40404c9..d8ce0c8 100644
--- a/drivers/infiniband/hw/ehca/ehca_classes.h
+++ b/drivers/infiniband/hw/ehca/ehca_classes.h
@@ -52,6 +52,8 @@ struct ehca_mw;
struct ehca_pd;
struct ehca_av;
+#include <linux/completion.h>
+
#include <rdma/ib_verbs.h>
#include <rdma/ib_user_verbs.h>
@@ -154,6 +156,7 @@ struct ehca_cq {
struct hlist_head qp_hashtab[QP_HASHTAB_LEN];
struct list_head entry;
u32 nr_callbacks;
+ struct completion zero_callbacks;
spinlock_t task_lock;
u32 ownpid;
/* mmap counter for resources mapped into user space */
diff --git a/drivers/infiniband/hw/ehca/ehca_cq.c b/drivers/infiniband/hw/ehca/ehca_cq.c
index 9291a86..906bd5b 100644
--- a/drivers/infiniband/hw/ehca/ehca_cq.c
+++ b/drivers/infiniband/hw/ehca/ehca_cq.c
@@ -147,6 +147,7 @@ struct ib_cq *ehca_create_cq(struct ib_d
spin_lock_init(&my_cq->spinlock);
spin_lock_init(&my_cq->cb_lock);
spin_lock_init(&my_cq->task_lock);
+ init_completion(&my_cq->zero_callbacks);
my_cq->ownpid = current->tgid;
cq = &my_cq->ib_cq;
@@ -330,9 +331,9 @@ int ehca_destroy_cq(struct ib_cq *cq)
}
spin_lock_irqsave(&ehca_cq_idr_lock, flags);
- while (my_cq->nr_callbacks) {
+ if (my_cq->nr_callbacks) {
spin_unlock_irqrestore(&ehca_cq_idr_lock, flags);
- yield();
+ wait_for_completion(&my_cq->zero_callbacks);
spin_lock_irqsave(&ehca_cq_idr_lock, flags);
}
diff --git a/drivers/infiniband/hw/ehca/ehca_irq.c b/drivers/infiniband/hw/ehca/ehca_irq.c
index 3ec53c6..7db39b7 100644
--- a/drivers/infiniband/hw/ehca/ehca_irq.c
+++ b/drivers/infiniband/hw/ehca/ehca_irq.c
@@ -605,6 +605,7 @@ static void run_comp_task(struct ehca_cp
spin_lock_irqsave(&cct->task_lock, flags);
while (!list_empty(&cct->cq_list)) {
+ int is_complete = 0;
cq = list_entry(cct->cq_list.next, struct ehca_cq, entry);
spin_unlock_irqrestore(&cct->task_lock, flags);
comp_event_callback(cq);
@@ -612,11 +613,14 @@ static void run_comp_task(struct ehca_cp
spin_lock(&cq->task_lock);
cq->nr_callbacks--;
- if (cq->nr_callbacks == 0) {
+ is_complete = (cq->nr_callbacks == 0);
+ if (is_complete) {
list_del_init(cct->cq_list.next);
cct->cq_jobs--;
}
spin_unlock(&cq->task_lock);
+ if (is_complete) /* wake up waiting destroy_cq() */
+ complete(&cq->zero_callbacks);
}
spin_unlock_irqrestore(&cct->task_lock, flags);
Looking at this one more time, I think it actually may be buggy:
> @@ -147,6 +147,7 @@ struct ib_cq *ehca_create_cq(struct ib_d
> spin_lock_init(&my_cq->spinlock);
> spin_lock_init(&my_cq->cb_lock);
> spin_lock_init(&my_cq->task_lock);
> + init_completion(&my_cq->zero_callbacks);
So you initialize the zero_callbacks completion once, at
ehca_create_cq().
But then
> @@ -612,11 +613,14 @@ static void run_comp_task(struct ehca_cp
>
> spin_lock(&cq->task_lock);
> cq->nr_callbacks--;
> - if (cq->nr_callbacks == 0) {
> + is_complete = (cq->nr_callbacks == 0);
> + if (is_complete) {
> list_del_init(cct->cq_list.next);
> cct->cq_jobs--;
> }
> spin_unlock(&cq->task_lock);
> + if (is_complete) /* wake up waiting destroy_cq() */
> + complete(&cq->zero_callbacks);
> }
every time nr_callbacks drops to 0, you complete the zero_callbacks
completion. So the first time a callback runs, you will complete
zero_callbacks, which will let wait_for_completion() finish even if
you later increment nr_callbacks again.
Also this
> - while (my_cq->nr_callbacks) {
> + if (my_cq->nr_callbacks) {
> spin_unlock_irqrestore(&ehca_cq_idr_lock, flags);
> - yield();
> + wait_for_completion(&my_cq->zero_callbacks);
> spin_lock_irqsave(&ehca_cq_idr_lock, flags);
> }
looks rather unsafe -- I don't see any common locking protecting both
this test of nr_callbacks and the setting of nr_callbacks in the ehca
irq handling... so I don't see anything protecting you from seeing
nr_callbacks==0 and not going into the if() (or while() -- the old
code has the same problem I think) but then doing ++nr_callbacks
somewhere else. In fact since you do the idr_remove() and
hipz_h_destroy_cq() *after* you make sure no callbacks are running,
this seems like it could happen easily.
So I'm holding off on applying this for now. Please think it over and
either tell me the current patch is OK, or fix it up. There's not
really too much urgency because a change like this is something I
would be comfortable merging between 2.6.21-rc1 and -rc2.
- R.
Hi,
> So I'm holding off on applying this for now. Please think it over and
> either tell me the current patch is OK, or fix it up. There's not
> really too much urgency because a change like this is something I
> would be comfortable merging between 2.6.21-rc1 and -rc2.
You're absolutely right. Let's target for rc2.
Thanks for this good catch!
Nam