d/numam-spdk
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

nvme/pcie: reduce physically contiguous memory for CQ/SQ

Browse Source 
Following patch made sure that CQ/SQ are allocated in
physically contiguous manner:
(64db67) nvme/pcie: make sure sq and cq are physically contiguous

Using MAX_IO_QUEUE_ENTRIES is enough to make sure that either
queue does not span multiple hugepages.

Yet the patch made sure that whole page is occupied only
by the queue. Which unnecessarily increases memory consumption
up to two hugepages per each qpair.

This patch changes it so that each queue alignment is limited
up to its size.

Changes in hugepages consumed when allocating io_qpair in hello_world
application:
io_queue_size		Without patch	With patch
256			8MiB		0MiB
1024			12MiB		4MiB
4096			24MiB		16MiB
Note: 0MiB means no new hugepages were required and qpair fits into
previously allocated hugepages (see all steps before io_qpair
allocation in hello_world).

Intersting result of this patch is that since we required alignment
up to the hugepage size this resulted in reserving even two 2MiB
hugepages to account for DPDK internal malloc trailing element.
See alloc_sz in try_expand_heap_primary() within malloc_heap.c

This patch not only reduces overall memory reserved for the
queues, but decreases increase in heap consumption on DPDK side.

Signed-off-by: Tomasz Zawadzki <tomasz.zawadzki@intel.com>
Change-Id: I75bf86e93674b4822d8204df3fb99458dec61e9c
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/2244
Community-CI: Mellanox Build Bot
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
This commit is contained in:
Tomasz Zawadzki 2020-05-07 11:35:14 -04:00
parent 2623c4c4d8
commit d3cf561199
1 changed files with 9 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

17
lib/nvme/nvme_pcie.c
View File

@ -1032,7 +1032,8 @@ nvme_pcie_qpair_construct(struct spdk_nvme_qpair *qpair,
uint16_t i;
volatile uint32_t *doorbell_base;
uint16_t num_trackers;
size_t page_align = VALUE_2MB;
size_t page_align = sysconf(_SC_PAGESIZE);
size_t queue_align, queue_len;
uint32_t flags = SPDK_MALLOC_DMA;
uint64_t sq_paddr = 0;
uint64_t cq_paddr = 0;
@ -1070,7 +1071,7 @@ nvme_pcie_qpair_construct(struct spdk_nvme_qpair *qpair,
/* cmd and cpl rings must be aligned on page size boundaries. */
if (ctrlr->opts.use_cmb_sqs) {
pqpair->cmd = nvme_pcie_ctrlr_alloc_cmb(ctrlr, pqpair->num_entries * sizeof(struct spdk_nvme_cmd),
sysconf(_SC_PAGESIZE), &pqpair->cmd_bus_addr);
page_align, &pqpair->cmd_bus_addr);
if (pqpair->cmd != NULL) {
pqpair->sq_in_cmb = true;
}
@ -1083,9 +1084,9 @@ nvme_pcie_qpair_construct(struct spdk_nvme_qpair *qpair,
/* To ensure physical address contiguity we make each ring occupy
* a single hugepage only. See MAX_IO_QUEUE_ENTRIES.
*/
pqpair->cmd = spdk_zmalloc(pqpair->num_entries * sizeof(struct spdk_nvme_cmd),
page_align, NULL,
SPDK_ENV_SOCKET_ID_ANY, flags);
queue_len = pqpair->num_entries * sizeof(struct spdk_nvme_cmd);
queue_align = spdk_max(spdk_align32pow2(queue_len), page_align);
pqpair->cmd = spdk_zmalloc(queue_len, queue_align, NULL, SPDK_ENV_SOCKET_ID_ANY, flags);
if (pqpair->cmd == NULL) {
SPDK_ERRLOG("alloc qpair_cmd failed\n");
return -ENOMEM;
@ -1106,9 +1107,9 @@ nvme_pcie_qpair_construct(struct spdk_nvme_qpair *qpair,
if (pqpair->cq_vaddr) {
pqpair->cpl = pqpair->cq_vaddr;
} else {
pqpair->cpl = spdk_zmalloc(pqpair->num_entries * sizeof(struct spdk_nvme_cpl),
page_align, NULL,
SPDK_ENV_SOCKET_ID_ANY, flags);
queue_len = pqpair->num_entries * sizeof(struct spdk_nvme_cpl);
queue_align = spdk_max(spdk_align32pow2(queue_len), page_align);
pqpair->cpl = spdk_zmalloc(queue_len, queue_align, NULL, SPDK_ENV_SOCKET_ID_ANY, flags);
if (pqpair->cpl == NULL) {
SPDK_ERRLOG("alloc qpair_cpl failed\n");
return -ENOMEM;