Commit 5e86c809 authored by Daniel Verkamp's avatar Daniel Verkamp
Browse files

nvme/pcie: limit max completions based on queue size 



For a given hardware queue size, only allow a quarter of the queue size
to be returned as completions in a single call to
spdk_nvme_qpair_process_completions(), and adjust num_trackers to match
so that num_trackers + max_completions_cap doesn't exceed the hardware
completion queue size.  This ensures that there is room in the
completion queue if new I/O is issued in response to completions before
we ring the completion queue doorbell.

The choice of 1/4 queue size is arbitrary; this seems to be a good
compromise between completion batch size and number of trackers.

Change-Id: I2c5aad7b98bfc8b33e53242240b2c9254fa05b4e
Signed-off-by: default avatarDaniel Verkamp <daniel.verkamp@intel.com>
Reviewed-on: https://review.gerrithub.io/393529


Tested-by: default avatarSPDK Automated Test System <sys_sgsw@intel.com>
Reviewed-by: default avatarBen Walker <benjamin.walker@intel.com>
Reviewed-by: default avatarJim Harris <james.r.harris@intel.com>
Reviewed-by: default avatarChangpeng Liu <changpeng.liu@intel.com>
parent 1ba93e61

lib/nvme/nvme_pcie.c

+25 −15
Original line number Diff line number Diff line
@@ -42,8 +42,14 @@ 
#include "nvme_internal.h"

#include "nvme_uevent.h"



/*

 * Number of completion queue entries to process before ringing the

 *  completion queue doorbell.

 */

#define NVME_MIN_COMPLETIONS	(1)

#define NVME_MAX_COMPLETIONS	(128)



#define NVME_ADMIN_ENTRIES	(128)

#define NVME_ADMIN_TRACKERS	(64)



/*

 * NVME_MAX_SGL_DESCRIPTORS defines the maximum number of descriptors in one SGL
@@ -158,6 +164,8 @@ struct nvme_pcie_qpair { 


	uint16_t num_entries;



	uint16_t max_completions_cap;



	uint16_t sq_tail;

	uint16_t cq_head;

	uint16_t sq_head;
@@ -871,15 +879,18 @@ nvme_pcie_qpair_construct(struct spdk_nvme_qpair *qpair) 
	uint16_t		num_trackers;

	size_t 			page_size = sysconf(_SC_PAGESIZE);



	if (qpair->id == 0) {

		num_trackers = NVME_ADMIN_TRACKERS;

	} else {

	/*

		 *  Note that for a queue size of N, we can only have (N-1)

		 *  commands outstanding, hence the "-1" here.

	 * Limit the maximum number of completions to return per call to prevent wraparound,

	 * and calculate how many trackers can be submitted at once without overflowing the

	 * completion queue.

	 */

		num_trackers = pqpair->num_entries - 1;

	}

	pqpair->max_completions_cap = pqpair->num_entries / 4;

	pqpair->max_completions_cap = spdk_max(pqpair->max_completions_cap, NVME_MIN_COMPLETIONS);

	pqpair->max_completions_cap = spdk_min(pqpair->max_completions_cap, NVME_MAX_COMPLETIONS);

	num_trackers = pqpair->num_entries - pqpair->max_completions_cap;



	SPDK_INFOLOG(SPDK_LOG_NVME, "max_completions_cap = %" PRIu16 " num_trackers = %" PRIu16 "\n",

		     pqpair->max_completions_cap, num_trackers);



	assert(num_trackers != 0);
@@ -1971,14 +1982,13 @@ nvme_pcie_qpair_process_completions(struct spdk_nvme_qpair *qpair, uint32_t max_ 
		nvme_robust_mutex_lock(&ctrlr->ctrlr_lock);

	}



	if (max_completions == 0 || (max_completions > (pqpair->num_entries - 1U))) {



	if (max_completions == 0 || max_completions > pqpair->max_completions_cap) {

		/*

		 * max_completions == 0 means unlimited, but complete at most one

		 * queue depth batch of I/O at a time so that the completion

		 * max_completions == 0 means unlimited, but complete at most

		 * max_completions_cap batch of I/O at a time so that the completion

		 * queue doorbells don't wrap around.

		 */

		max_completions = pqpair->num_entries - 1;

		max_completions = pqpair->max_completions_cap;

	}



	while (1) {