Using huge pages decreases the amount of internal resources in the HCA
needed for performing rdma, which in turn can improve the performance,
specifically the number of messages per second.
Change the rdma buffer pool to be allocated using huge pages with a
fallback to valloc on failure.
Using this patch provided improvement of 30% more IOPS for 4K reads.
Signed-off-by: Roi Dayan <roid@xxxxxxxxxxxx>
---
usr/iscsi/iser.c | 59 +++++++++++++++++++++++++++++++++++++++++++++++++++--
usr/iscsi/iser.h | 3 ++
2 files changed, 59 insertions(+), 3 deletions(-)
diff --git a/usr/iscsi/iser.c b/usr/iscsi/iser.c
index b60ce74..b58b13d 100644
--- a/usr/iscsi/iser.c
+++ b/usr/iscsi/iser.c
@@ -32,6 +32,8 @@
#include <sys/epoll.h>
#include <infiniband/verbs.h>
#include <rdma/rdma_cma.h>
+#include <sys/ipc.h>
+#include <sys/shm.h>
#include "util.h"
#include "iscsid.h"
@@ -559,16 +561,66 @@ static inline void iser_set_rsp_stat_sn(struct iscsi_session *session,
}
}
+static uint8_t* iser_alloc_pool(size_t pool_size, int *shmid)
+{
+ int shmemid;
+ uint8_t *buf;
+
+ /* allocate memory */
+ shmemid = shmget(IPC_PRIVATE, pool_size,
+ SHM_HUGETLB | IPC_CREAT | SHM_R | SHM_W);
+
+ if (shmemid < 0) {
+ eprintf("shmget rdma pool sz:%zu failed\n", pool_size);
+ goto failed_huge_page;
+ }
+
+ /* get pointer to allocated memory */
+ buf = shmat(shmemid, NULL, 0);
+
+ if (buf == (void*)-1) {
+ eprintf("Shared memory attach failure (errno=%d %m)", errno);
+ shmctl(shmemid, IPC_RMID, NULL);
+ goto failed_huge_page;
+ }
+
+ /* mark 'to be destroyed' when process detaches from shmem segment
+ this will clear the HugePage resources even if process if killed not nicely.
+ From checking shmctl man page it is unlikely that it will fail here. */
+ if (shmctl(shmemid, IPC_RMID, NULL)) {
+ eprintf("Shared memory contrl mark 'to be destroyed' failed (errno=%d %m)", errno);
+ }
+
+ dprintf("Allocated huge page sz:%zu\n", pool_size);
+ *shmid = shmemid;
+ return buf;
+
+ failed_huge_page:
+ *shmid = -1;
+ return valloc(pool_size);
+}
+
+static void iser_free_pool(uint8_t *pool_buf, int shmid) {
+ if (shmid >= 0) {
+ if (shmdt(pool_buf) != 0) {
+ eprintf("shmem detach failure (errno=%d %m)", errno);
+ }
+ } else {
+ free(pool_buf);
+ }
+}
+
static int iser_init_rdma_buf_pool(struct iser_device *dev)
{
uint8_t *pool_buf, *list_buf;
size_t pool_size, list_size;
struct iser_membuf *rdma_buf;
+ int shmid;
int i;
membuf_size = roundup(membuf_size, pagesize);
pool_size = membuf_num * membuf_size;
- pool_buf = valloc(pool_size);
+ pool_buf = iser_alloc_pool(pool_size, &shmid);
if (!pool_buf) {
eprintf("malloc rdma pool sz:%zu failed\n", pool_size);
return -ENOMEM;
@@ -578,7 +630,7 @@ static int iser_init_rdma_buf_pool(struct iser_device *dev)
list_buf = malloc(list_size);
if (!list_buf) {
eprintf("malloc list_buf sz:%zu failed\n", list_size);
- free(pool_buf);
+ iser_free_pool(pool_buf, shmid);
return -ENOMEM;
}
@@ -587,13 +639,14 @@ static int iser_init_rdma_buf_pool(struct iser_device *dev)
IBV_ACCESS_LOCAL_WRITE);
if (!dev->membuf_mr) {
eprintf("ibv_reg_mr failed, %m\n");
- free(pool_buf);
+ iser_free_pool(pool_buf, shmid);
free(list_buf);
return -1;
}
dprintf("pool buf:%p list:%p mr:%p lkey:0x%x\n",
pool_buf, list_buf, dev->membuf_mr, dev->membuf_mr->lkey);
+ dev->rdma_hugetbl_shmid = shmid;
dev->membuf_regbuf = pool_buf;
dev->membuf_listbuf = list_buf;
INIT_LIST_HEAD(&dev->membuf_free);
diff --git a/usr/iscsi/iser.h b/usr/iscsi/iser.h
index d06fc6e..0749f50 100644
--- a/usr/iscsi/iser.h
+++ b/usr/iscsi/iser.h
@@ -252,6 +252,9 @@ struct iser_device {
struct ibv_mr *membuf_mr;
int waiting_for_mem;
+ /* shared memory identifier */
+ int rdma_hugetbl_shmid;
+
struct event_data poll_sched;
/* free and allocated membuf entries */
--
1.7.8.2
--
To unsubscribe from this list: send the line "unsubscribe stgt" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
