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

rdma: clean up SGE definitions and properly set values

Browse Source 
We have historically conflated SPDK_NVMF_MAX_SGL_ENTRIES and the maximum
number of SGEs associated with a wr data object. For now these are the
same thing, but there should be nothing tying the number of NVMe request
SGL elements to the number of rdma request wr sgl elements.
Also, clarify the rx_sge and tx_sge enums to reflect the actual maximum
number of SGEs associated with either the send and receive queues.
This change doesn't actually modify these values, but sets us up to do
things like split the data in an NVMe SGE into multiple WR SGEs in case
the buffer associated with the NVMe SGE is not contained in a single
RDMA mr.
We also need to store these values in the qpair for later usage.

Change-Id: Iff3756fc72787a4b72a99b2bdf90bf486a8010fa
Signed-off-by: Seth Howell <seth.howell@intel.com>
Reviewed-on: https://review.gerrithub.io/433196
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Chandler-Test-Pool: SPDK Automated Test System <sys_sgsw@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
This commit is contained in:
Seth Howell 2018-11-13 17:49:54 -07:00 committed by Jim Harris
parent e0b669312a
commit 1180bf8343
1 changed files with 32 additions and 24 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

56
lib/nvmf/rdma.c
View File

@ -54,9 +54,9 @@
/*
RDMA Connection Resource Defaults
*/
#define NVMF_DEFAULT_TX_SGE 1
#define NVMF_DEFAULT_TX_SGE SPDK_NVMF_MAX_SGL_ENTRIES
#define NVMF_DEFAULT_RSP_SGE 1
#define NVMF_DEFAULT_RX_SGE 2
#define NVMF_DEFAULT_DATA_SGE 16
/* The RDMA completion queue size */
#define NVMF_RDMA_CQ_SIZE 4096
@ -229,14 +229,14 @@ struct spdk_nvmf_rdma_request {
struct {
struct spdk_nvmf_rdma_wr rdma_wr;
struct ibv_send_wr wr;
struct ibv_sge sgl[NVMF_DEFAULT_TX_SGE];
struct ibv_sge sgl[NVMF_DEFAULT_RSP_SGE];
} rsp;
struct {
struct spdk_nvmf_rdma_wr rdma_wr;
struct ibv_send_wr wr;
struct ibv_sge sgl[SPDK_NVMF_MAX_SGL_ENTRIES];
void *buffers[SPDK_NVMF_MAX_SGL_ENTRIES];
struct ibv_sge sgl[NVMF_DEFAULT_TX_SGE];
void *buffers[NVMF_DEFAULT_TX_SGE];
} data;
struct spdk_nvmf_rdma_wr rdma_wr;
@ -266,6 +266,12 @@ struct spdk_nvmf_rdma_qpair {
/* The maximum number of active RDMA READ and WRITE operations at one time */
uint16_t max_rw_depth;
/* The maximum number of SGEs per WR on the send queue */
uint32_t max_send_sge;
/* The maximum number of SGEs per WR on the recv queue */
uint32_t max_recv_sge;
/* Receives that are waiting for a request object */
TAILQ_HEAD(, spdk_nvmf_rdma_recv) incoming_queue;
@ -275,8 +281,6 @@ struct spdk_nvmf_rdma_qpair {
/* Number of requests in each state */
uint32_t state_cntr[RDMA_REQUEST_NUM_STATES];
int max_sge;
/* Array of size "max_queue_depth" containing RDMA requests. */
struct spdk_nvmf_rdma_request *reqs;
@ -310,7 +314,6 @@ struct spdk_nvmf_rdma_qpair {
/* IBV queue pair attributes: they are used to manage
* qp state and recover from errors.
*/
struct ibv_qp_init_attr ibv_init_attr;
struct ibv_qp_attr ibv_attr;
uint32_t disconnect_flags;
@ -408,6 +411,7 @@ spdk_nvmf_rdma_qpair_dec_refcnt(struct spdk_nvmf_rdma_qpair *rqpair)
static enum ibv_qp_state
spdk_nvmf_rdma_update_ibv_state(struct spdk_nvmf_rdma_qpair *rqpair) {
enum ibv_qp_state old_state, new_state;
struct ibv_qp_init_attr init_attr;
int rc;
/* All the attributes needed for recovery */
@ -430,7 +434,7 @@ spdk_nvmf_rdma_update_ibv_state(struct spdk_nvmf_rdma_qpair *rqpair) {
old_state = rqpair->ibv_attr.qp_state;
rc = ibv_query_qp(rqpair->cm_id->qp, &rqpair->ibv_attr,
spdk_nvmf_ibv_attr_mask, &rqpair->ibv_init_attr);
spdk_nvmf_ibv_attr_mask, &init_attr);
if (rc)
{
@ -637,25 +641,28 @@ spdk_nvmf_rdma_qpair_initialize(struct spdk_nvmf_qpair *qpair)
int rc, i;
struct spdk_nvmf_rdma_recv *rdma_recv;
struct spdk_nvmf_rdma_request *rdma_req;
struct spdk_nvmf_transport *transport;
struct spdk_nvmf_transport *transport;
struct spdk_nvmf_rdma_device *device;
struct ibv_qp_init_attr ibv_init_attr;
rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport);
transport = &rtransport->transport;
device = rqpair->port->device;
memset(&rqpair->ibv_init_attr, 0, sizeof(struct ibv_qp_init_attr));
rqpair->ibv_init_attr.qp_context = rqpair;
rqpair->ibv_init_attr.qp_type = IBV_QPT_RC;
rqpair->ibv_init_attr.send_cq = rqpair->poller->cq;
rqpair->ibv_init_attr.recv_cq = rqpair->poller->cq;
rqpair->ibv_init_attr.cap.max_send_wr = rqpair->max_queue_depth *
2 + 1; /* SEND, READ, and WRITE operations + dummy drain WR */
rqpair->ibv_init_attr.cap.max_recv_wr = rqpair->max_queue_depth +
1; /* RECV operations + dummy drain WR */
rqpair->ibv_init_attr.cap.max_send_sge = rqpair->max_sge;
rqpair->ibv_init_attr.cap.max_recv_sge = NVMF_DEFAULT_RX_SGE;
memset(&ibv_init_attr, 0, sizeof(struct ibv_qp_init_attr));
ibv_init_attr.qp_context = rqpair;
ibv_init_attr.qp_type = IBV_QPT_RC;
ibv_init_attr.send_cq = rqpair->poller->cq;
ibv_init_attr.recv_cq = rqpair->poller->cq;
ibv_init_attr.cap.max_send_wr = rqpair->max_queue_depth *
2 + 1; /* SEND, READ, and WRITE operations + dummy drain WR */
ibv_init_attr.cap.max_recv_wr = rqpair->max_queue_depth +
1; /* RECV operations + dummy drain WR */
ibv_init_attr.cap.max_send_sge = spdk_min(device->attr.max_sge, NVMF_DEFAULT_TX_SGE);
ibv_init_attr.cap.max_recv_sge = spdk_min(device->attr.max_sge, NVMF_DEFAULT_RX_SGE);
rc = rdma_create_qp(rqpair->cm_id, rqpair->port->device->pd, &rqpair->ibv_init_attr);
rc = rdma_create_qp(rqpair->cm_id, rqpair->port->device->pd, &ibv_init_attr);
if (rc) {
SPDK_ERRLOG("rdma_create_qp failed: errno %d: %s\n", errno, spdk_strerror(errno));
rdma_destroy_id(rqpair->cm_id);
@ -664,6 +671,8 @@ spdk_nvmf_rdma_qpair_initialize(struct spdk_nvmf_qpair *qpair)
return -1;
}
rqpair->max_send_sge = spdk_min(NVMF_DEFAULT_TX_SGE, ibv_init_attr.cap.max_send_sge);
rqpair->max_recv_sge = spdk_min(NVMF_DEFAULT_RX_SGE, ibv_init_attr.cap.max_recv_sge);
spdk_trace_record(TRACE_RDMA_QP_CREATE, 0, 0, (uintptr_t)rqpair->cm_id, 0);
SPDK_DEBUGLOG(SPDK_LOG_RDMA, "New RDMA Connection: %p\n", qpair);
@ -1020,7 +1029,6 @@ nvmf_rdma_connect(struct spdk_nvmf_transport *transport, struct rdma_cm_event *e
rqpair->cm_id = event->id;
rqpair->listen_id = event->listen_id;
rqpair->qpair.transport = transport;
rqpair->max_sge = spdk_min(port->device->attr.max_sge, SPDK_NVMF_MAX_SGL_ENTRIES);
TAILQ_INIT(&rqpair->incoming_queue);
event->id->context = &rqpair->qpair;
@ -1570,7 +1578,7 @@ spdk_nvmf_rdma_create(struct spdk_nvmf_transport_opts *opts)
}
sge_count = opts->max_io_size / opts->io_unit_size;
if (sge_count > SPDK_NVMF_MAX_SGL_ENTRIES) {
if (sge_count > NVMF_DEFAULT_TX_SGE) {
SPDK_ERRLOG("Unsupported IO Unit size specified, %d bytes\n", opts->io_unit_size);
spdk_nvmf_rdma_destroy(&rtransport->transport);
return NULL;