Re: [PATCH] RDMA/core: reduce IB_POLL_BATCH constant

From: Sagi Grimberg
Date: Wed Feb 21 2018 - 08:44:50 EST



On Tue, 2018-02-20 at 21:59 +0100, Arnd Bergmann wrote:
/* # of WCs to poll for with a single call to ib_poll_cq */
-#define IB_POLL_BATCH 16
+#define IB_POLL_BATCH 8

The purpose of batch polling is to minimize contention on the cq spinlock.
Reducing the IB_POLL_BATCH constant may affect performance negatively. Has
the performance impact of this change been verified for all affected drivers
(ib_srp, ib_srpt, ib_iser, ib_isert, NVMeOF, NVMeOF target, SMB Direct, NFS
over RDMA, ...)?

Only the users of the DIRECT polling method use an on-stack
array of ib_wc's. This is only the SRP drivers.

The other two modes have use of a dynamically allocated array
of ib_wc's that hangs off the ib_cq. These shouldn't need any
reduction in the size of this array, and they are the common
case.

IMO a better solution would be to change ib_process_cq_direct
to use a smaller on-stack array, and leave IB_POLL_BATCH alone.

The only reason why I added this array on-stack was to allow consumers
that did not use ib_alloc_cq api to call it, but that seems like a
wrong decision when thinking it over again (as probably these users
did not set the wr_cqe correctly).

How about we make ib_process_cq_direct use the cq wc array and add
a WARN_ON statement (and fail it gracefully) if the caller used this
API without calling ib_alloc_cq?

--
diff --git a/drivers/infiniband/core/cq.c b/drivers/infiniband/core/cq.c
index bc79ca8215d7..cd3e9e124834 100644
--- a/drivers/infiniband/core/cq.c
+++ b/drivers/infiniband/core/cq.c
@@ -25,10 +25,10 @@
#define IB_POLL_FLAGS \
(IB_CQ_NEXT_COMP | IB_CQ_REPORT_MISSED_EVENTS)

-static int __ib_process_cq(struct ib_cq *cq, int budget, struct ib_wc *poll_wc)
+static int __ib_process_cq(struct ib_cq *cq, int budget)
{
int i, n, completed = 0;
- struct ib_wc *wcs = poll_wc ? : cq->wc;
+ struct ib_wc *wcs = cq->wc;

/*
* budget might be (-1) if the caller does not
@@ -72,9 +72,9 @@ static int __ib_process_cq(struct ib_cq *cq, int budget, struct ib_wc *poll_wc)
*/
int ib_process_cq_direct(struct ib_cq *cq, int budget)
{
- struct ib_wc wcs[IB_POLL_BATCH];
-
- return __ib_process_cq(cq, budget, wcs);
+ if (unlikely(WARN_ON_ONCE(!cq->wc)))
+ return 0;
+ return __ib_process_cq(cq, budget);
}
EXPORT_SYMBOL(ib_process_cq_direct);

@@ -88,7 +88,7 @@ static int ib_poll_handler(struct irq_poll *iop, int budget)
struct ib_cq *cq = container_of(iop, struct ib_cq, iop);
int completed;

- completed = __ib_process_cq(cq, budget, NULL);
+ completed = __ib_process_cq(cq, budget);
if (completed < budget) {
irq_poll_complete(&cq->iop);
if (ib_req_notify_cq(cq, IB_POLL_FLAGS) > 0)
@@ -108,7 +108,7 @@ static void ib_cq_poll_work(struct work_struct *work)
struct ib_cq *cq = container_of(work, struct ib_cq, work);
int completed;

- completed = __ib_process_cq(cq, IB_POLL_BUDGET_WORKQUEUE, NULL);
+ completed = __ib_process_cq(cq, IB_POLL_BUDGET_WORKQUEUE);
if (completed >= IB_POLL_BUDGET_WORKQUEUE ||
ib_req_notify_cq(cq, IB_POLL_FLAGS) > 0)
queue_work(ib_comp_wq, &cq->work);
--