From mboxrd@z Thu Jan 1 00:00:00 1970 Received: from eggs.gnu.org ([2001:4830:134:3::10]:59367) by lists.gnu.org with esmtp (Exim 4.71) (envelope-from ) id 1elIX6-0008MQ-37 for qemu-devel@nongnu.org; Mon, 12 Feb 2018 13:08:59 -0500 Received: from Debian-exim by eggs.gnu.org with spam-scanned (Exim 4.71) (envelope-from ) id 1elIX1-0002gj-3V for qemu-devel@nongnu.org; Mon, 12 Feb 2018 13:08:52 -0500 Received: from mx3-rdu2.redhat.com ([66.187.233.73]:54846 helo=mx1.redhat.com) by eggs.gnu.org with esmtps (TLS1.0:DHE_RSA_AES_256_CBC_SHA1:32) (Exim 4.71) (envelope-from ) id 1elIX0-0002fx-JV for qemu-devel@nongnu.org; Mon, 12 Feb 2018 13:08:46 -0500 From: Marcel Apfelbaum Date: Mon, 12 Feb 2018 20:08:16 +0200 Message-Id: <20180212180819.82556-7-marcel@redhat.com> In-Reply-To: <20180212180819.82556-1-marcel@redhat.com> References: <20180212180819.82556-1-marcel@redhat.com> Subject: [Qemu-devel] [PATCH V10 6/9] hw/rdma: Implementation of generic rdma device layers List-Id: List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , To: qemu-devel@nongnu.org Cc: peter.maydell@linaro.org, ehabkost@redhat.com, yuval.shaia@oracle.com, marcel@redhat.com, mst@redhat.com, dotanb@mellanox.com, yanjun.zhu@oracle.com From: Yuval Shaia This layer is composed of two sub-modules, backend and resource manager. Backend sub-module is responsible for all the interaction with IB layers such as ibverbs and umad (external libraries). Resource manager is a collection of functions and structures to manage RDMA resources such as QPs, CQs and MRs. Reviewed-by: Dotan Barak Signed-off-by: Yuval Shaia Signed-off-by: Marcel Apfelbaum --- Makefile.objs | 1 + configure | 9 +- hw/rdma/Makefile.objs | 2 +- hw/rdma/rdma_backend.c | 818 +++++++++++++++++++++++++++++++++++++++++++++++++ hw/rdma/rdma_backend.h | 98 ++++++ hw/rdma/rdma_rm.c | 544 ++++++++++++++++++++++++++++++++ hw/rdma/rdma_rm.h | 69 +++++ hw/rdma/trace-events | 5 + 8 files changed, 1541 insertions(+), 5 deletions(-) create mode 100644 hw/rdma/rdma_backend.c create mode 100644 hw/rdma/rdma_backend.h create mode 100644 hw/rdma/rdma_rm.c create mode 100644 hw/rdma/rdma_rm.h create mode 100644 hw/rdma/trace-events diff --git a/Makefile.objs b/Makefile.objs index 2efba6d768..f3a3d28304 100644 --- a/Makefile.objs +++ b/Makefile.objs @@ -130,6 +130,7 @@ trace-events-subdirs += hw/block/dataplane trace-events-subdirs += hw/char trace-events-subdirs += hw/intc trace-events-subdirs += hw/net +trace-events-subdirs += hw/rdma trace-events-subdirs += hw/virtio trace-events-subdirs += hw/audio trace-events-subdirs += hw/misc diff --git a/configure b/configure index 62562f08cf..c0ecd1858f 100755 --- a/configure +++ b/configure @@ -1575,7 +1575,7 @@ disabled with --disable-FEATURE, default is enabled if available: hax HAX acceleration support hvf Hypervisor.framework acceleration support whpx Windows Hypervisor Platform acceleration support - rdma RDMA-based migration support + rdma Enable RDMA-based migration and PVRDMA support vde support for vde network netmap support for netmap network linux-aio Linux AIO support @@ -2926,15 +2926,16 @@ if test "$rdma" != "no" ; then #include int main(void) { return 0; } EOF - rdma_libs="-lrdmacm -libverbs" + rdma_libs="-lrdmacm -libverbs -libumad" if compile_prog "" "$rdma_libs" ; then rdma="yes" + libs_softmmu="$libs_softmmu $rdma_libs" else if test "$rdma" = "yes" ; then error_exit \ - " OpenFabrics librdmacm/libibverbs not present." \ + " OpenFabrics librdmacm/libibverbs/libibumad not present." \ " Your options:" \ - " (1) Fast: Install infiniband packages from your distro." \ + " (1) Fast: Install infiniband packages (devel) from your distro." \ " (2) Cleanest: Install libraries from www.openfabrics.org" \ " (3) Also: Install softiwarp if you don't have RDMA hardware" fi diff --git a/hw/rdma/Makefile.objs b/hw/rdma/Makefile.objs index cdffe4a9a3..6a59bf0d5b 100644 --- a/hw/rdma/Makefile.objs +++ b/hw/rdma/Makefile.objs @@ -1,3 +1,3 @@ ifeq ($(CONFIG_RDMA),y) -obj-$(CONFIG_PCI) += rdma_utils.o +obj-$(CONFIG_PCI) += rdma_utils.o rdma_backend.o rdma_rm.o endif diff --git a/hw/rdma/rdma_backend.c b/hw/rdma/rdma_backend.c new file mode 100644 index 0000000000..e306fba534 --- /dev/null +++ b/hw/rdma/rdma_backend.c @@ -0,0 +1,818 @@ +/* + * QEMU paravirtual RDMA - Generic RDMA backend + * + * Copyright (C) 2018 Oracle + * Copyright (C) 2018 Red Hat Inc + * + * Authors: + * Yuval Shaia + * Marcel Apfelbaum + * + * This work is licensed under the terms of the GNU GPL, version 2 or later. + * See the COPYING file in the top-level directory. + * + */ + +#include +#include +#include + +#include + +#include "trace.h" +#include "rdma_utils.h" +#include "rdma_rm.h" +#include "rdma_backend.h" + +/* Vendor Errors */ +#define VENDOR_ERR_FAIL_BACKEND 0x201 +#define VENDOR_ERR_TOO_MANY_SGES 0x202 +#define VENDOR_ERR_NOMEM 0x203 +#define VENDOR_ERR_QP0 0x204 +#define VENDOR_ERR_NO_SGE 0x205 +#define VENDOR_ERR_MAD_SEND 0x206 +#define VENDOR_ERR_INVLKEY 0x207 +#define VENDOR_ERR_MR_SMALL 0x208 + +#define THR_NAME_LEN 16 + +typedef struct BackendCtx { + uint64_t req_id; + void *up_ctx; + bool is_tx_req; +} BackendCtx; + +static void (*comp_handler)(int status, unsigned int vendor_err, void *ctx); + +static void dummy_comp_handler(int status, unsigned int vendor_err, void *ctx) +{ + pr_err("No completion handler is registered\n"); +} + +static void poll_cq(RdmaDeviceResources *rdma_dev_res, struct ibv_cq *ibcq) +{ + int i, ne; + BackendCtx *bctx; + struct ibv_wc wc[2]; + + pr_dbg("Entering poll_cq loop on cq %p\n", ibcq); + do { + ne = ibv_poll_cq(ibcq, ARRAY_SIZE(wc), wc); + + pr_dbg("Got %d completion(s) from cq %p\n", ne, ibcq); + + for (i = 0; i < ne; i++) { + pr_dbg("wr_id=0x%lx\n", wc[i].wr_id); + pr_dbg("status=%d\n", wc[i].status); + + bctx = rdma_rm_get_cqe_ctx(rdma_dev_res, wc[i].wr_id); + if (unlikely(!bctx)) { + pr_dbg("Error: Failed to find ctx for req %ld\n", wc[i].wr_id); + continue; + } + pr_dbg("Processing %s CQE\n", bctx->is_tx_req ? "send" : "recv"); + + comp_handler(wc[i].status, wc[i].vendor_err, bctx->up_ctx); + + rdma_rm_dealloc_cqe_ctx(rdma_dev_res, wc[i].wr_id); + g_free(bctx); + } + } while (ne > 0); + + if (ne < 0) { + pr_dbg("Got error %d from ibv_poll_cq\n", ne); + } +} + +static void *comp_handler_thread(void *arg) +{ + RdmaBackendDev *backend_dev = (RdmaBackendDev *)arg; + int rc; + struct ibv_cq *ev_cq; + void *ev_ctx; + + pr_dbg("Starting\n"); + + while (backend_dev->comp_thread.run) { + pr_dbg("Waiting for completion on channel %p\n", backend_dev->channel); + rc = ibv_get_cq_event(backend_dev->channel, &ev_cq, &ev_ctx); + pr_dbg("ibv_get_cq_event=%d\n", rc); + if (unlikely(rc)) { + pr_dbg("---> ibv_get_cq_event (%d)\n", rc); + continue; + } + + rc = ibv_req_notify_cq(ev_cq, 0); + if (unlikely(rc)) { + pr_dbg("Error %d from ibv_req_notify_cq\n", rc); + } + + poll_cq(backend_dev->rdma_dev_res, ev_cq); + + ibv_ack_cq_events(ev_cq, 1); + } + + pr_dbg("Going down\n"); + + /* TODO: Post cqe for all remaining buffs that were posted */ + + return NULL; +} + +void rdma_backend_register_comp_handler(void (*handler)(int status, + unsigned int vendor_err, void *ctx)) +{ + comp_handler = handler; +} + +void rdma_backend_unregister_comp_handler(void) +{ + rdma_backend_register_comp_handler(dummy_comp_handler); +} + +int rdma_backend_query_port(RdmaBackendDev *backend_dev, + struct ibv_port_attr *port_attr) +{ + int rc; + + rc = ibv_query_port(backend_dev->context, backend_dev->port_num, port_attr); + if (rc) { + pr_dbg("Error %d from ibv_query_port\n", rc); + return -EIO; + } + + return 0; +} + +void rdma_backend_poll_cq(RdmaDeviceResources *rdma_dev_res, RdmaBackendCQ *cq) +{ + poll_cq(rdma_dev_res, cq->ibcq); +} + +static GHashTable *ah_hash; + +static struct ibv_ah *create_ah(RdmaBackendDev *backend_dev, struct ibv_pd *pd, + uint8_t sgid_idx, union ibv_gid *dgid) +{ + GBytes *ah_key = g_bytes_new(dgid, sizeof(*dgid)); + struct ibv_ah *ah = g_hash_table_lookup(ah_hash, ah_key); + + if (ah) { + trace_create_ah_cache_hit(be64_to_cpu(dgid->global.subnet_prefix), + be64_to_cpu(dgid->global.interface_id)); + g_bytes_unref(ah_key); + } else { + struct ibv_ah_attr ah_attr = { + .is_global = 1, + .port_num = backend_dev->port_num, + .grh.hop_limit = 1, + }; + + ah_attr.grh.dgid = *dgid; + ah_attr.grh.sgid_index = sgid_idx; + + ah = ibv_create_ah(pd, &ah_attr); + if (ah) { + g_hash_table_insert(ah_hash, ah_key, ah); + } else { + g_bytes_unref(ah_key); + pr_dbg("ibv_create_ah failed for gid <%lx %lx>\n", + be64_to_cpu(dgid->global.subnet_prefix), + be64_to_cpu(dgid->global.interface_id)); + } + + trace_create_ah_cache_miss(be64_to_cpu(dgid->global.subnet_prefix), + be64_to_cpu(dgid->global.interface_id)); + } + + return ah; +} + +static void destroy_ah_hash_key(gpointer data) +{ + g_bytes_unref(data); +} + +static void destroy_ah_hast_data(gpointer data) +{ + struct ibv_ah *ah = data; + + ibv_destroy_ah(ah); +} + +static void ah_cache_init(void) +{ + ah_hash = g_hash_table_new_full(g_bytes_hash, g_bytes_equal, + destroy_ah_hash_key, destroy_ah_hast_data); +} + +static int build_host_sge_array(RdmaDeviceResources *rdma_dev_res, + struct ibv_sge *dsge, struct ibv_sge *ssge, + uint8_t num_sge) +{ + RdmaRmMR *mr; + int ssge_idx; + + pr_dbg("num_sge=%d\n", num_sge); + + for (ssge_idx = 0; ssge_idx < num_sge; ssge_idx++) { + mr = rdma_rm_get_mr(rdma_dev_res, ssge[ssge_idx].lkey); + if (unlikely(!mr)) { + pr_dbg("Invalid lkey 0x%x\n", ssge[ssge_idx].lkey); + return VENDOR_ERR_INVLKEY | ssge[ssge_idx].lkey; + } + + dsge->addr = mr->user_mr.host_virt + ssge[ssge_idx].addr - + mr->user_mr.guest_start; + dsge->length = ssge[ssge_idx].length; + dsge->lkey = rdma_backend_mr_lkey(&mr->backend_mr); + + pr_dbg("ssge->addr=0x%lx\n", (uint64_t)ssge[ssge_idx].addr); + pr_dbg("dsge->addr=0x%lx\n", dsge->addr); + pr_dbg("dsge->length=%d\n", dsge->length); + pr_dbg("dsge->lkey=0x%x\n", dsge->lkey); + + dsge++; + } + + return 0; +} + +void rdma_backend_post_send(RdmaBackendDev *backend_dev, + RdmaBackendQP *qp, uint8_t qp_type, + struct ibv_sge *sge, uint32_t num_sge, + union ibv_gid *dgid, uint32_t dqpn, + uint32_t dqkey, void *ctx) +{ + BackendCtx *bctx; + struct ibv_sge new_sge[MAX_SGE]; + uint32_t bctx_id; + int rc; + struct ibv_send_wr wr = {0}, *bad_wr; + + if (!qp->ibqp) { /* This field does not get initialized for QP0 and QP1 */ + if (qp_type == IBV_QPT_SMI) { + pr_dbg("QP0 unsupported\n"); + comp_handler(IBV_WC_GENERAL_ERR, VENDOR_ERR_QP0, ctx); + } else if (qp_type == IBV_QPT_GSI) { + pr_dbg("QP1\n"); + comp_handler(IBV_WC_GENERAL_ERR, VENDOR_ERR_MAD_SEND, ctx); + } + pr_dbg("qp->ibqp is NULL for qp_type %d!!!\n", qp_type); + return; + } + + pr_dbg("num_sge=%d\n", num_sge); + if (!num_sge) { + pr_dbg("num_sge=0\n"); + comp_handler(IBV_WC_GENERAL_ERR, VENDOR_ERR_NO_SGE, ctx); + return; + } + + bctx = g_malloc0(sizeof(*bctx)); + bctx->up_ctx = ctx; + bctx->is_tx_req = 1; + + rc = rdma_rm_alloc_cqe_ctx(backend_dev->rdma_dev_res, &bctx_id, bctx); + if (unlikely(rc)) { + pr_dbg("Failed to allocate cqe_ctx\n"); + comp_handler(IBV_WC_GENERAL_ERR, VENDOR_ERR_NOMEM, ctx); + goto out_free_bctx; + } + + rc = build_host_sge_array(backend_dev->rdma_dev_res, new_sge, sge, num_sge); + if (rc) { + pr_dbg("Error: Failed to build host SGE array\n"); + comp_handler(IBV_WC_GENERAL_ERR, rc, ctx); + goto out_dealloc_cqe_ctx; + } + + if (qp_type == IBV_QPT_UD) { + wr.wr.ud.ah = create_ah(backend_dev, qp->ibpd, + backend_dev->backend_gid_idx, dgid); + wr.wr.ud.remote_qpn = dqpn; + wr.wr.ud.remote_qkey = dqkey; + } + + wr.num_sge = num_sge; + wr.opcode = IBV_WR_SEND; + wr.send_flags = IBV_SEND_SIGNALED; + wr.sg_list = new_sge; + wr.wr_id = bctx_id; + + rc = ibv_post_send(qp->ibqp, &wr, &bad_wr); + pr_dbg("ibv_post_send=%d\n", rc); + if (rc) { + pr_dbg("Fail (%d, %d) to post send WQE to qpn %d\n", rc, errno, + qp->ibqp->qp_num); + comp_handler(IBV_WC_GENERAL_ERR, VENDOR_ERR_FAIL_BACKEND, ctx); + goto out_dealloc_cqe_ctx; + } + + return; + +out_dealloc_cqe_ctx: + rdma_rm_dealloc_cqe_ctx(backend_dev->rdma_dev_res, bctx_id); + +out_free_bctx: + g_free(bctx); +} + +void rdma_backend_post_recv(RdmaBackendDev *backend_dev, + RdmaDeviceResources *rdma_dev_res, + RdmaBackendQP *qp, uint8_t qp_type, + struct ibv_sge *sge, uint32_t num_sge, void *ctx) +{ + BackendCtx *bctx; + struct ibv_sge new_sge[MAX_SGE]; + uint32_t bctx_id; + int rc; + struct ibv_recv_wr wr = {0}, *bad_wr; + + if (!qp->ibqp) { /* This field does not get initialized for QP0 and QP1 */ + if (qp_type == IBV_QPT_SMI) { + pr_dbg("QP0 unsupported\n"); + comp_handler(IBV_WC_GENERAL_ERR, VENDOR_ERR_QP0, ctx); + } + if (qp_type == IBV_QPT_GSI) { + pr_dbg("QP1\n"); + comp_handler(IBV_WC_GENERAL_ERR, VENDOR_ERR_MAD_SEND, ctx); + } + return; + } + + pr_dbg("num_sge=%d\n", num_sge); + if (!num_sge) { + pr_dbg("num_sge=0\n"); + comp_handler(IBV_WC_GENERAL_ERR, VENDOR_ERR_NO_SGE, ctx); + return; + } + + bctx = g_malloc0(sizeof(*bctx)); + bctx->up_ctx = ctx; + bctx->is_tx_req = 0; + + rc = rdma_rm_alloc_cqe_ctx(rdma_dev_res, &bctx_id, bctx); + if (unlikely(rc)) { + pr_dbg("Failed to allocate cqe_ctx\n"); + comp_handler(IBV_WC_GENERAL_ERR, VENDOR_ERR_NOMEM, ctx); + goto out_free_bctx; + } + + rc = build_host_sge_array(rdma_dev_res, new_sge, sge, num_sge); + if (rc) { + pr_dbg("Error: Failed to build host SGE array\n"); + comp_handler(IBV_WC_GENERAL_ERR, rc, ctx); + goto out_dealloc_cqe_ctx; + } + + wr.num_sge = num_sge; + wr.sg_list = new_sge; + wr.wr_id = bctx_id; + rc = ibv_post_recv(qp->ibqp, &wr, &bad_wr); + pr_dbg("ibv_post_recv=%d\n", rc); + if (rc) { + pr_dbg("Fail (%d, %d) to post recv WQE to qpn %d\n", rc, errno, + qp->ibqp->qp_num); + comp_handler(IBV_WC_GENERAL_ERR, VENDOR_ERR_FAIL_BACKEND, ctx); + goto out_dealloc_cqe_ctx; + } + + return; + +out_dealloc_cqe_ctx: + rdma_rm_dealloc_cqe_ctx(rdma_dev_res, bctx_id); + +out_free_bctx: + g_free(bctx); +} + +int rdma_backend_create_pd(RdmaBackendDev *backend_dev, RdmaBackendPD *pd) +{ + pd->ibpd = ibv_alloc_pd(backend_dev->context); + + return pd->ibpd ? 0 : -EIO; +} + +void rdma_backend_destroy_pd(RdmaBackendPD *pd) +{ + if (pd->ibpd) { + ibv_dealloc_pd(pd->ibpd); + } +} + +int rdma_backend_create_mr(RdmaBackendMR *mr, RdmaBackendPD *pd, uint64_t addr, + size_t length, int access) +{ + pr_dbg("addr=0x%lx\n", addr); + pr_dbg("len=%ld\n", length); + mr->ibmr = ibv_reg_mr(pd->ibpd, (void *)addr, length, access); + if (mr->ibmr) { + pr_dbg("lkey=0x%x\n", mr->ibmr->lkey); + pr_dbg("rkey=0x%x\n", mr->ibmr->rkey); + mr->ibpd = pd->ibpd; + } + + return mr->ibmr ? 0 : -EIO; +} + +void rdma_backend_destroy_mr(RdmaBackendMR *mr) +{ + if (mr->ibmr) { + ibv_dereg_mr(mr->ibmr); + } +} + +int rdma_backend_create_cq(RdmaBackendDev *backend_dev, RdmaBackendCQ *cq, + int cqe) +{ + int rc; + + pr_dbg("cqe=%d\n", cqe); + + pr_dbg("dev->channel=%p\n", backend_dev->channel); + cq->ibcq = ibv_create_cq(backend_dev->context, cqe + 1, NULL, + backend_dev->channel, 0); + + if (cq->ibcq) { + rc = ibv_req_notify_cq(cq->ibcq, 0); + if (rc) { + pr_dbg("Error %d from ibv_req_notify_cq\n", rc); + } + cq->backend_dev = backend_dev; + } + + return cq->ibcq ? 0 : -EIO; +} + +void rdma_backend_destroy_cq(RdmaBackendCQ *cq) +{ + if (cq->ibcq) { + ibv_destroy_cq(cq->ibcq); + } +} + +int rdma_backend_create_qp(RdmaBackendQP *qp, uint8_t qp_type, + RdmaBackendPD *pd, RdmaBackendCQ *scq, + RdmaBackendCQ *rcq, uint32_t max_send_wr, + uint32_t max_recv_wr, uint32_t max_send_sge, + uint32_t max_recv_sge) +{ + struct ibv_qp_init_attr attr = {0}; + + qp->ibqp = 0; + pr_dbg("qp_type=%d\n", qp_type); + + switch (qp_type) { + case IBV_QPT_GSI: + pr_dbg("QP1 unsupported\n"); + return 0; + + case IBV_QPT_RC: + /* fall through */ + case IBV_QPT_UD: + /* do nothing */ + break; + + default: + pr_dbg("Unsupported QP type %d\n", qp_type); + return -EIO; + } + + attr.qp_type = qp_type; + attr.send_cq = scq->ibcq; + attr.recv_cq = rcq->ibcq; + attr.cap.max_send_wr = max_send_wr; + attr.cap.max_recv_wr = max_recv_wr; + attr.cap.max_send_sge = max_send_sge; + attr.cap.max_recv_sge = max_recv_sge; + + pr_dbg("max_send_wr=%d\n", max_send_wr); + pr_dbg("max_recv_wr=%d\n", max_recv_wr); + pr_dbg("max_send_sge=%d\n", max_send_sge); + pr_dbg("max_recv_sge=%d\n", max_recv_sge); + + qp->ibqp = ibv_create_qp(pd->ibpd, &attr); + if (likely(!qp->ibqp)) { + pr_dbg("Error from ibv_create_qp\n"); + return -EIO; + } + + qp->ibpd = pd->ibpd; + + /* TODO: Query QP to get max_inline_data and save it to be used in send */ + + pr_dbg("qpn=0x%x\n", qp->ibqp->qp_num); + + return 0; +} + +int rdma_backend_qp_state_init(RdmaBackendDev *backend_dev, RdmaBackendQP *qp, + uint8_t qp_type, uint32_t qkey) +{ + struct ibv_qp_attr attr = {0}; + int rc, attr_mask; + + pr_dbg("qpn=0x%x\n", qp->ibqp->qp_num); + pr_dbg("sport_num=%d\n", backend_dev->port_num); + + attr_mask = IBV_QP_STATE | IBV_QP_PKEY_INDEX | IBV_QP_PORT; + attr.qp_state = IBV_QPS_INIT; + attr.pkey_index = 0; + attr.port_num = backend_dev->port_num; + + switch (qp_type) { + case IBV_QPT_RC: + attr_mask |= IBV_QP_ACCESS_FLAGS; + break; + + case IBV_QPT_UD: + attr.qkey = qkey; + attr_mask |= IBV_QP_QKEY; + break; + + default: + pr_dbg("Unsupported QP type %d\n", qp_type); + return -EIO; + } + + rc = ibv_modify_qp(qp->ibqp, &attr, attr_mask); + if (rc) { + pr_dbg("Error %d from ibv_modify_qp\n", rc); + return -EIO; + } + + return 0; +} + +int rdma_backend_qp_state_rtr(RdmaBackendDev *backend_dev, RdmaBackendQP *qp, + uint8_t qp_type, union ibv_gid *dgid, + uint32_t dqpn, uint32_t rq_psn, uint32_t qkey, + bool use_qkey) +{ + struct ibv_qp_attr attr = {0}; + union ibv_gid ibv_gid = { + .global.interface_id = dgid->global.interface_id, + .global.subnet_prefix = dgid->global.subnet_prefix + }; + int rc, attr_mask; + + attr.qp_state = IBV_QPS_RTR; + attr_mask = IBV_QP_STATE; + + switch (qp_type) { + case IBV_QPT_RC: + pr_dbg("dgid=0x%lx,%lx\n", + be64_to_cpu(ibv_gid.global.subnet_prefix), + be64_to_cpu(ibv_gid.global.interface_id)); + pr_dbg("dqpn=0x%x\n", dqpn); + pr_dbg("sgid_idx=%d\n", backend_dev->backend_gid_idx); + pr_dbg("sport_num=%d\n", backend_dev->port_num); + pr_dbg("rq_psn=0x%x\n", rq_psn); + + attr.path_mtu = IBV_MTU_1024; + attr.dest_qp_num = dqpn; + attr.max_dest_rd_atomic = 1; + attr.min_rnr_timer = 12; + attr.ah_attr.port_num = backend_dev->port_num; + attr.ah_attr.is_global = 1; + attr.ah_attr.grh.hop_limit = 1; + attr.ah_attr.grh.dgid = ibv_gid; + attr.ah_attr.grh.sgid_index = backend_dev->backend_gid_idx; + attr.rq_psn = rq_psn; + + attr_mask |= IBV_QP_AV | IBV_QP_PATH_MTU | IBV_QP_DEST_QPN | + IBV_QP_RQ_PSN | IBV_QP_MAX_DEST_RD_ATOMIC | + IBV_QP_MIN_RNR_TIMER; + break; + + case IBV_QPT_UD: + if (use_qkey) { + pr_dbg("qkey=0x%x\n", qkey); + attr.qkey = qkey; + attr_mask |= IBV_QP_QKEY; + } + break; + } + + rc = ibv_modify_qp(qp->ibqp, &attr, attr_mask); + if (rc) { + pr_dbg("Error %d from ibv_modify_qp\n", rc); + return -EIO; + } + + return 0; +} + +int rdma_backend_qp_state_rts(RdmaBackendQP *qp, uint8_t qp_type, + uint32_t sq_psn, uint32_t qkey, bool use_qkey) +{ + struct ibv_qp_attr attr = {0}; + int rc, attr_mask; + + pr_dbg("qpn=0x%x\n", qp->ibqp->qp_num); + pr_dbg("sq_psn=0x%x\n", sq_psn); + + attr.qp_state = IBV_QPS_RTS; + attr.sq_psn = sq_psn; + attr_mask = IBV_QP_STATE | IBV_QP_SQ_PSN; + + switch (qp_type) { + case IBV_QPT_RC: + attr.timeout = 14; + attr.retry_cnt = 7; + attr.rnr_retry = 7; + attr.max_rd_atomic = 1; + + attr_mask |= IBV_QP_TIMEOUT | IBV_QP_RETRY_CNT | IBV_QP_RNR_RETRY | + IBV_QP_MAX_QP_RD_ATOMIC; + break; + + case IBV_QPT_UD: + if (use_qkey) { + pr_dbg("qkey=0x%x\n", qkey); + attr.qkey = qkey; + attr_mask |= IBV_QP_QKEY; + } + break; + } + + rc = ibv_modify_qp(qp->ibqp, &attr, attr_mask); + if (rc) { + pr_dbg("Error %d from ibv_modify_qp\n", rc); + return -EIO; + } + + return 0; +} + +void rdma_backend_destroy_qp(RdmaBackendQP *qp) +{ + if (qp->ibqp) { + ibv_destroy_qp(qp->ibqp); + } +} + +#define CHK_ATTR(req, dev, member, fmt) ({ \ + pr_dbg("%s="fmt","fmt"\n", #member, dev.member, req->member); \ + if (req->member > dev.member) { \ + warn_report("%s = 0x%lx is higher than host device capability 0x%lx", \ + #member, (uint64_t)req->member, (uint64_t)dev.member); \ + req->member = dev.member; \ + } \ + pr_dbg("%s="fmt"\n", #member, req->member); }) + +static int init_device_caps(RdmaBackendDev *backend_dev, + struct ibv_device_attr *dev_attr) +{ + if (ibv_query_device(backend_dev->context, &backend_dev->dev_attr)) { + return -EIO; + } + + CHK_ATTR(dev_attr, backend_dev->dev_attr, max_mr_size, "%ld"); + CHK_ATTR(dev_attr, backend_dev->dev_attr, max_qp, "%d"); + CHK_ATTR(dev_attr, backend_dev->dev_attr, max_sge, "%d"); + CHK_ATTR(dev_attr, backend_dev->dev_attr, max_qp_wr, "%d"); + CHK_ATTR(dev_attr, backend_dev->dev_attr, max_cq, "%d"); + CHK_ATTR(dev_attr, backend_dev->dev_attr, max_cqe, "%d"); + CHK_ATTR(dev_attr, backend_dev->dev_attr, max_mr, "%d"); + CHK_ATTR(dev_attr, backend_dev->dev_attr, max_pd, "%d"); + CHK_ATTR(dev_attr, backend_dev->dev_attr, max_qp_rd_atom, "%d"); + CHK_ATTR(dev_attr, backend_dev->dev_attr, max_qp_init_rd_atom, "%d"); + CHK_ATTR(dev_attr, backend_dev->dev_attr, max_ah, "%d"); + + return 0; +} + +int rdma_backend_init(RdmaBackendDev *backend_dev, + RdmaDeviceResources *rdma_dev_res, + const char *backend_device_name, uint8_t port_num, + uint8_t backend_gid_idx, struct ibv_device_attr *dev_attr, + Error **errp) +{ + int i; + int ret = 0; + int num_ibv_devices; + char thread_name[THR_NAME_LEN] = {0}; + struct ibv_device **dev_list; + struct ibv_port_attr port_attr; + + backend_dev->backend_gid_idx = backend_gid_idx; + backend_dev->port_num = port_num; + backend_dev->rdma_dev_res = rdma_dev_res; + + rdma_backend_register_comp_handler(dummy_comp_handler); + + dev_list = ibv_get_device_list(&num_ibv_devices); + if (!dev_list) { + error_setg(errp, "Failed to get IB devices list"); + return -EIO; + } + + if (num_ibv_devices == 0) { + error_setg(errp, "No IB devices were found"); + ret = -ENXIO; + goto out_free_dev_list; + } + + if (backend_device_name) { + for (i = 0; dev_list[i]; ++i) { + if (!strcmp(ibv_get_device_name(dev_list[i]), + backend_device_name)) { + break; + } + } + + backend_dev->ib_dev = dev_list[i]; + if (!backend_dev->ib_dev) { + error_setg(errp, "Failed to find IB device %s", + backend_device_name); + ret = -EIO; + goto out_free_dev_list; + } + } else { + backend_dev->ib_dev = *dev_list; + } + + pr_dbg("Using backend device %s, port %d, gid_idx %d\n", + ibv_get_device_name(backend_dev->ib_dev), + backend_dev->port_num, backend_dev->backend_gid_idx); + + backend_dev->context = ibv_open_device(backend_dev->ib_dev); + if (!backend_dev->context) { + error_setg(errp, "Failed to open IB device"); + ret = -EIO; + goto out; + } + + backend_dev->channel = ibv_create_comp_channel(backend_dev->context); + if (!backend_dev->channel) { + error_setg(errp, "Failed to create IB communication channel"); + ret = -EIO; + goto out_close_device; + } + pr_dbg("dev->backend_dev.channel=%p\n", backend_dev->channel); + + ret = ibv_query_port(backend_dev->context, backend_dev->port_num, + &port_attr); + if (ret) { + error_setg(errp, "Error %d from ibv_query_port", ret); + ret = -EIO; + goto out_destroy_comm_channel; + } + + if (backend_dev->backend_gid_idx > port_attr.gid_tbl_len) { + error_setg(errp, "Invalid backend_gid_idx, should be less than %d", + port_attr.gid_tbl_len); + goto out_destroy_comm_channel; + } + + ret = init_device_caps(backend_dev, dev_attr); + if (ret) { + error_setg(errp, "Failed to initialize device capabilities"); + ret = -EIO; + goto out_destroy_comm_channel; + } + + ret = ibv_query_gid(backend_dev->context, backend_dev->port_num, + backend_dev->backend_gid_idx, &backend_dev->gid); + if (ret) { + error_setg(errp, "Failed to query gid %d", + backend_dev->backend_gid_idx); + ret = -EIO; + goto out_destroy_comm_channel; + } + pr_dbg("subnet_prefix=0x%lx\n", + be64_to_cpu(backend_dev->gid.global.subnet_prefix)); + pr_dbg("interface_id=0x%lx\n", + be64_to_cpu(backend_dev->gid.global.interface_id)); + + snprintf(thread_name, sizeof(thread_name), "rdma_comp_%s", + ibv_get_device_name(backend_dev->ib_dev)); + backend_dev->comp_thread.run = true; + qemu_thread_create(&backend_dev->comp_thread.thread, thread_name, + comp_handler_thread, backend_dev, QEMU_THREAD_DETACHED); + + ah_cache_init(); + + goto out_free_dev_list; + +out_destroy_comm_channel: + ibv_destroy_comp_channel(backend_dev->channel); + +out_close_device: + ibv_close_device(backend_dev->context); + +out_free_dev_list: + ibv_free_device_list(dev_list); + +out: + return ret; +} + +void rdma_backend_fini(RdmaBackendDev *backend_dev) +{ + g_hash_table_destroy(ah_hash); + ibv_destroy_comp_channel(backend_dev->channel); + ibv_close_device(backend_dev->context); +} diff --git a/hw/rdma/rdma_backend.h b/hw/rdma/rdma_backend.h new file mode 100644 index 0000000000..68f2b05ca7 --- /dev/null +++ b/hw/rdma/rdma_backend.h @@ -0,0 +1,98 @@ +/* + * RDMA device: Definitions of Backend Device functions + * + * Copyright (C) 2018 Oracle + * Copyright (C) 2018 Red Hat Inc + * + * Authors: + * Yuval Shaia + * Marcel Apfelbaum + * + * This work is licensed under the terms of the GNU GPL, version 2 or later. + * See the COPYING file in the top-level directory. + * + */ + +#ifndef RDMA_BACKEND_H +#define RDMA_BACKEND_H + +#include +#include "rdma_rm_defs.h" +#include "rdma_backend_defs.h" + +/* Add definition for QP0 and QP1 as there is no userspace enums for them */ +enum ibv_special_qp_type { + IBV_QPT_SMI = 0, + IBV_QPT_GSI = 1, +}; + +static inline union ibv_gid *rdma_backend_gid(RdmaBackendDev *dev) +{ + return &dev->gid; +} + +static inline uint32_t rdma_backend_qpn(const RdmaBackendQP *qp) +{ + return qp->ibqp ? qp->ibqp->qp_num : 0; +} + +static inline uint32_t rdma_backend_mr_lkey(const RdmaBackendMR *mr) +{ + return mr->ibmr ? mr->ibmr->lkey : 0; +} + +static inline uint32_t rdma_backend_mr_rkey(const RdmaBackendMR *mr) +{ + return mr->ibmr ? mr->ibmr->rkey : 0; +} + +int rdma_backend_init(RdmaBackendDev *backend_dev, + RdmaDeviceResources *rdma_dev_res, + const char *backend_device_name, uint8_t port_num, + uint8_t backend_gid_idx, struct ibv_device_attr *dev_attr, + Error **errp); +void rdma_backend_fini(RdmaBackendDev *backend_dev); +void rdma_backend_register_comp_handler(void (*handler)(int status, + unsigned int vendor_err, void *ctx)); +void rdma_backend_unregister_comp_handler(void); + +int rdma_backend_query_port(RdmaBackendDev *backend_dev, + struct ibv_port_attr *port_attr); +int rdma_backend_create_pd(RdmaBackendDev *backend_dev, RdmaBackendPD *pd); +void rdma_backend_destroy_pd(RdmaBackendPD *pd); + +int rdma_backend_create_mr(RdmaBackendMR *mr, RdmaBackendPD *pd, uint64_t addr, + size_t length, int access); +void rdma_backend_destroy_mr(RdmaBackendMR *mr); + +int rdma_backend_create_cq(RdmaBackendDev *backend_dev, RdmaBackendCQ *cq, + int cqe); +void rdma_backend_destroy_cq(RdmaBackendCQ *cq); +void rdma_backend_poll_cq(RdmaDeviceResources *rdma_dev_res, RdmaBackendCQ *cq); + +int rdma_backend_create_qp(RdmaBackendQP *qp, uint8_t qp_type, + RdmaBackendPD *pd, RdmaBackendCQ *scq, + RdmaBackendCQ *rcq, uint32_t max_send_wr, + uint32_t max_recv_wr, uint32_t max_send_sge, + uint32_t max_recv_sge); +int rdma_backend_qp_state_init(RdmaBackendDev *backend_dev, RdmaBackendQP *qp, + uint8_t qp_type, uint32_t qkey); +int rdma_backend_qp_state_rtr(RdmaBackendDev *backend_dev, RdmaBackendQP *qp, + uint8_t qp_type, union ibv_gid *dgid, + uint32_t dqpn, uint32_t rq_psn, uint32_t qkey, + bool use_qkey); +int rdma_backend_qp_state_rts(RdmaBackendQP *qp, uint8_t qp_type, + uint32_t sq_psn, uint32_t qkey, bool use_qkey); +void rdma_backend_destroy_qp(RdmaBackendQP *qp); + +void rdma_backend_post_send(RdmaBackendDev *backend_dev, + RdmaBackendQP *qp, uint8_t qp_type, + struct ibv_sge *sge, uint32_t num_sge, + union ibv_gid *dgid, uint32_t dqpn, uint32_t dqkey, + void *ctx); +void rdma_backend_post_recv(RdmaBackendDev *backend_dev, + RdmaDeviceResources *rdma_dev_res, + RdmaBackendQP *qp, uint8_t qp_type, + struct ibv_sge *sge, uint32_t num_sge, void *ctx); + +#endif diff --git a/hw/rdma/rdma_rm.c b/hw/rdma/rdma_rm.c new file mode 100644 index 0000000000..b5fc45ddab --- /dev/null +++ b/hw/rdma/rdma_rm.c @@ -0,0 +1,544 @@ +/* + * QEMU paravirtual RDMA - Resource Manager Implementation + * + * Copyright (C) 2018 Oracle + * Copyright (C) 2018 Red Hat Inc + * + * Authors: + * Yuval Shaia + * Marcel Apfelbaum + * + * This work is licensed under the terms of the GNU GPL, version 2 or later. + * See the COPYING file in the top-level directory. + * + */ + +#include +#include +#include + +#include "rdma_utils.h" +#include "rdma_backend.h" +#include "rdma_rm.h" + +#define MAX_RM_TBL_NAME 16 + +/* Page directory and page tables */ +#define PG_DIR_SZ { TARGET_PAGE_SIZE / sizeof(__u64) } +#define PG_TBL_SZ { TARGET_PAGE_SIZE / sizeof(__u64) } + +static inline void res_tbl_init(const char *name, RdmaRmResTbl *tbl, + uint32_t tbl_sz, uint32_t res_sz) +{ + tbl->tbl = g_malloc(tbl_sz * res_sz); + + strncpy(tbl->name, name, MAX_RM_TBL_NAME); + tbl->name[MAX_RM_TBL_NAME - 1] = 0; + + tbl->bitmap = bitmap_new(tbl_sz); + tbl->tbl_sz = tbl_sz; + tbl->res_sz = res_sz; + qemu_mutex_init(&tbl->lock); +} + +static inline void res_tbl_free(RdmaRmResTbl *tbl) +{ + qemu_mutex_destroy(&tbl->lock); + g_free(tbl->tbl); + bitmap_zero_extend(tbl->bitmap, tbl->tbl_sz, 0); +} + +static inline void *res_tbl_get(RdmaRmResTbl *tbl, uint32_t handle) +{ + pr_dbg("%s, handle=%d\n", tbl->name, handle); + + if ((handle < tbl->tbl_sz) && (test_bit(handle, tbl->bitmap))) { + return tbl->tbl + handle * tbl->res_sz; + } else { + pr_dbg("Invalid handle %d\n", handle); + return NULL; + } +} + +static inline void *res_tbl_alloc(RdmaRmResTbl *tbl, uint32_t *handle) +{ + qemu_mutex_lock(&tbl->lock); + + *handle = find_first_zero_bit(tbl->bitmap, tbl->tbl_sz); + if (*handle > tbl->tbl_sz) { + pr_dbg("Failed to alloc, bitmap is full\n"); + qemu_mutex_unlock(&tbl->lock); + return NULL; + } + + set_bit(*handle, tbl->bitmap); + + qemu_mutex_unlock(&tbl->lock); + + memset(tbl->tbl + *handle * tbl->res_sz, 0, tbl->res_sz); + + pr_dbg("%s, handle=%d\n", tbl->name, *handle); + + return tbl->tbl + *handle * tbl->res_sz; +} + +static inline void res_tbl_dealloc(RdmaRmResTbl *tbl, uint32_t handle) +{ + pr_dbg("%s, handle=%d\n", tbl->name, handle); + + qemu_mutex_lock(&tbl->lock); + + if (handle < tbl->tbl_sz) { + clear_bit(handle, tbl->bitmap); + } + + qemu_mutex_unlock(&tbl->lock); +} + +int rdma_rm_alloc_pd(RdmaDeviceResources *dev_res, RdmaBackendDev *backend_dev, + uint32_t *pd_handle, uint32_t ctx_handle) +{ + RdmaRmPD *pd; + int ret = -ENOMEM; + + pd = res_tbl_alloc(&dev_res->pd_tbl, pd_handle); + if (!pd) { + goto out; + } + + ret = rdma_backend_create_pd(backend_dev, &pd->backend_pd); + if (ret) { + ret = -EIO; + goto out_tbl_dealloc; + } + + pd->ctx_handle = ctx_handle; + + return 0; + +out_tbl_dealloc: + res_tbl_dealloc(&dev_res->pd_tbl, *pd_handle); + +out: + return ret; +} + +RdmaRmPD *rdma_rm_get_pd(RdmaDeviceResources *dev_res, uint32_t pd_handle) +{ + return res_tbl_get(&dev_res->pd_tbl, pd_handle); +} + +void rdma_rm_dealloc_pd(RdmaDeviceResources *dev_res, uint32_t pd_handle) +{ + RdmaRmPD *pd = rdma_rm_get_pd(dev_res, pd_handle); + + if (pd) { + rdma_backend_destroy_pd(&pd->backend_pd); + res_tbl_dealloc(&dev_res->pd_tbl, pd_handle); + } +} + +int rdma_rm_alloc_mr(RdmaDeviceResources *dev_res, uint32_t pd_handle, + uint64_t guest_start, size_t guest_length, void *host_virt, + int access_flags, uint32_t *mr_handle, uint32_t *lkey, + uint32_t *rkey) +{ + RdmaRmMR *mr; + int ret = 0; + RdmaRmPD *pd; + uint64_t addr; + size_t length; + + pd = rdma_rm_get_pd(dev_res, pd_handle); + if (!pd) { + pr_dbg("Invalid PD\n"); + return -EINVAL; + } + + mr = res_tbl_alloc(&dev_res->mr_tbl, mr_handle); + if (!mr) { + pr_dbg("Failed to allocate obj in table\n"); + return -ENOMEM; + } + + if (!host_virt) { + /* TODO: This is my guess but not so sure that this needs to be + * done */ + length = TARGET_PAGE_SIZE; + addr = (uint64_t)g_malloc(length); + } else { + mr->user_mr.host_virt = (uint64_t) host_virt; + pr_dbg("host_virt=0x%lx\n", mr->user_mr.host_virt); + mr->user_mr.length = guest_length; + pr_dbg("length=0x%lx\n", guest_length); + mr->user_mr.guest_start = guest_start; + pr_dbg("guest_start=0x%lx\n", mr->user_mr.guest_start); + + length = mr->user_mr.length; + addr = mr->user_mr.host_virt; + } + + ret = rdma_backend_create_mr(&mr->backend_mr, &pd->backend_pd, addr, length, + access_flags); + if (ret) { + pr_dbg("Fail in rdma_backend_create_mr, err=%d\n", ret); + ret = -EIO; + goto out_dealloc_mr; + } + + if (!host_virt) { + *lkey = mr->lkey = rdma_backend_mr_lkey(&mr->backend_mr); + *rkey = mr->rkey = rdma_backend_mr_rkey(&mr->backend_mr); + } else { + /* We keep mr_handle in lkey so send and recv get get mr ptr */ + *lkey = *mr_handle; + *rkey = -1; + } + + mr->pd_handle = pd_handle; + + return 0; + +out_dealloc_mr: + res_tbl_dealloc(&dev_res->mr_tbl, *mr_handle); + + return ret; +} + +RdmaRmMR *rdma_rm_get_mr(RdmaDeviceResources *dev_res, uint32_t mr_handle) +{ + return res_tbl_get(&dev_res->mr_tbl, mr_handle); +} + +void rdma_rm_dealloc_mr(RdmaDeviceResources *dev_res, uint32_t mr_handle) +{ + RdmaRmMR *mr = rdma_rm_get_mr(dev_res, mr_handle); + + if (mr) { + rdma_backend_destroy_mr(&mr->backend_mr); + munmap((void *)mr->user_mr.host_virt, mr->user_mr.length); + res_tbl_dealloc(&dev_res->mr_tbl, mr_handle); + } +} + +int rdma_rm_alloc_uc(RdmaDeviceResources *dev_res, uint32_t pfn, + uint32_t *uc_handle) +{ + RdmaRmUC *uc; + + /* TODO: Need to make sure pfn is between bar start address and + * bsd+RDMA_BAR2_UAR_SIZE + if (pfn > RDMA_BAR2_UAR_SIZE) { + pr_err("pfn out of range (%d > %d)\n", pfn, RDMA_BAR2_UAR_SIZE); + return -ENOMEM; + } + */ + + uc = res_tbl_alloc(&dev_res->uc_tbl, uc_handle); + if (!uc) { + return -ENOMEM; + } + + return 0; +} + +RdmaRmUC *rdma_rm_get_uc(RdmaDeviceResources *dev_res, uint32_t uc_handle) +{ + return res_tbl_get(&dev_res->uc_tbl, uc_handle); +} + +void rdma_rm_dealloc_uc(RdmaDeviceResources *dev_res, uint32_t uc_handle) +{ + RdmaRmUC *uc = rdma_rm_get_uc(dev_res, uc_handle); + + if (uc) { + res_tbl_dealloc(&dev_res->uc_tbl, uc_handle); + } +} + +RdmaRmCQ *rdma_rm_get_cq(RdmaDeviceResources *dev_res, uint32_t cq_handle) +{ + return res_tbl_get(&dev_res->cq_tbl, cq_handle); +} + +int rdma_rm_alloc_cq(RdmaDeviceResources *dev_res, RdmaBackendDev *backend_dev, + uint32_t cqe, uint32_t *cq_handle, void *opaque) +{ + int rc; + RdmaRmCQ *cq; + + cq = res_tbl_alloc(&dev_res->cq_tbl, cq_handle); + if (!cq) { + return -ENOMEM; + } + + cq->opaque = opaque; + cq->notify = false; + + rc = rdma_backend_create_cq(backend_dev, &cq->backend_cq, cqe); + if (rc) { + rc = -EIO; + goto out_dealloc_cq; + } + + return 0; + +out_dealloc_cq: + rdma_rm_dealloc_cq(dev_res, *cq_handle); + + return rc; +} + +void rdma_rm_req_notify_cq(RdmaDeviceResources *dev_res, uint32_t cq_handle, + bool notify) +{ + RdmaRmCQ *cq; + + pr_dbg("cq_handle=%d, notify=0x%x\n", cq_handle, notify); + + cq = rdma_rm_get_cq(dev_res, cq_handle); + if (!cq) { + return; + } + + cq->notify = notify; + pr_dbg("notify=%d\n", cq->notify); +} + +void rdma_rm_dealloc_cq(RdmaDeviceResources *dev_res, uint32_t cq_handle) +{ + RdmaRmCQ *cq; + + cq = rdma_rm_get_cq(dev_res, cq_handle); + if (!cq) { + return; + } + + rdma_backend_destroy_cq(&cq->backend_cq); + + res_tbl_dealloc(&dev_res->cq_tbl, cq_handle); +} + +RdmaRmQP *rdma_rm_get_qp(RdmaDeviceResources *dev_res, uint32_t qpn) +{ + GBytes *key = g_bytes_new(&qpn, sizeof(qpn)); + + RdmaRmQP *qp = g_hash_table_lookup(dev_res->qp_hash, key); + + g_bytes_unref(key); + + return qp; +} + +int rdma_rm_alloc_qp(RdmaDeviceResources *dev_res, uint32_t pd_handle, + uint8_t qp_type, uint32_t max_send_wr, + uint32_t max_send_sge, uint32_t send_cq_handle, + uint32_t max_recv_wr, uint32_t max_recv_sge, + uint32_t recv_cq_handle, void *opaque, uint32_t *qpn) +{ + int rc; + RdmaRmQP *qp; + RdmaRmCQ *scq, *rcq; + RdmaRmPD *pd; + uint32_t rm_qpn; + + pr_dbg("qp_type=%d\n", qp_type); + + pd = rdma_rm_get_pd(dev_res, pd_handle); + if (!pd) { + pr_err("Invalid pd handle (%d)\n", pd_handle); + return -EINVAL; + } + + scq = rdma_rm_get_cq(dev_res, send_cq_handle); + rcq = rdma_rm_get_cq(dev_res, recv_cq_handle); + + if (!scq || !rcq) { + pr_err("Invalid send_cqn or recv_cqn (%d, %d)\n", + send_cq_handle, recv_cq_handle); + return -EINVAL; + } + + qp = res_tbl_alloc(&dev_res->qp_tbl, &rm_qpn); + if (!qp) { + return -ENOMEM; + } + pr_dbg("rm_qpn=%d\n", rm_qpn); + + qp->qpn = rm_qpn; + qp->qp_state = IBV_QPS_RESET; + qp->qp_type = qp_type; + qp->send_cq_handle = send_cq_handle; + qp->recv_cq_handle = recv_cq_handle; + qp->opaque = opaque; + + rc = rdma_backend_create_qp(&qp->backend_qp, qp_type, &pd->backend_pd, + &scq->backend_cq, &rcq->backend_cq, max_send_wr, + max_recv_wr, max_send_sge, max_recv_sge); + if (rc) { + rc = -EIO; + goto out_dealloc_qp; + } + + *qpn = rdma_backend_qpn(&qp->backend_qp); + pr_dbg("rm_qpn=%d, backend_qpn=0x%x\n", rm_qpn, *qpn); + g_hash_table_insert(dev_res->qp_hash, g_bytes_new(qpn, sizeof(*qpn)), qp); + + return 0; + +out_dealloc_qp: + res_tbl_dealloc(&dev_res->qp_tbl, qp->qpn); + + return rc; +} + +int rdma_rm_modify_qp(RdmaDeviceResources *dev_res, RdmaBackendDev *backend_dev, + uint32_t qp_handle, uint32_t attr_mask, + union ibv_gid *dgid, uint32_t dqpn, + enum ibv_qp_state qp_state, uint32_t qkey, + uint32_t rq_psn, uint32_t sq_psn) +{ + RdmaRmQP *qp; + int ret; + + pr_dbg("qpn=%d\n", qp_handle); + + qp = rdma_rm_get_qp(dev_res, qp_handle); + if (!qp) { + return -EINVAL; + } + + pr_dbg("qp_type=%d\n", qp->qp_type); + pr_dbg("attr_mask=0x%x\n", attr_mask); + + if (qp->qp_type == IBV_QPT_SMI) { + pr_dbg("QP0 unsupported\n"); + return -EPERM; + } else if (qp->qp_type == IBV_QPT_GSI) { + pr_dbg("QP1\n"); + return 0; + } + + if (attr_mask & IBV_QP_STATE) { + qp->qp_state = qp_state; + pr_dbg("qp_state=%d\n", qp->qp_state); + + if (qp->qp_state == IBV_QPS_INIT) { + ret = rdma_backend_qp_state_init(backend_dev, &qp->backend_qp, + qp->qp_type, qkey); + if (ret) { + return -EIO; + } + } + + if (qp->qp_state == IBV_QPS_RTR) { + ret = rdma_backend_qp_state_rtr(backend_dev, &qp->backend_qp, + qp->qp_type, dgid, dqpn, rq_psn, + qkey, attr_mask & IBV_QP_QKEY); + if (ret) { + return -EIO; + } + } + + if (qp->qp_state == IBV_QPS_RTS) { + ret = rdma_backend_qp_state_rts(&qp->backend_qp, qp->qp_type, + sq_psn, qkey, + attr_mask & IBV_QP_QKEY); + if (ret) { + return -EIO; + } + } + } + + return 0; +} + +void rdma_rm_dealloc_qp(RdmaDeviceResources *dev_res, uint32_t qp_handle) +{ + RdmaRmQP *qp; + GBytes *key; + + key = g_bytes_new(&qp_handle, sizeof(qp_handle)); + qp = g_hash_table_lookup(dev_res->qp_hash, key); + g_hash_table_remove(dev_res->qp_hash, key); + g_bytes_unref(key); + + if (!qp) { + return; + } + + rdma_backend_destroy_qp(&qp->backend_qp); + + res_tbl_dealloc(&dev_res->qp_tbl, qp->qpn); +} + +void *rdma_rm_get_cqe_ctx(RdmaDeviceResources *dev_res, uint32_t cqe_ctx_id) +{ + void **cqe_ctx; + + cqe_ctx = res_tbl_get(&dev_res->cqe_ctx_tbl, cqe_ctx_id); + if (!cqe_ctx) { + return NULL; + } + + pr_dbg("ctx=%p\n", *cqe_ctx); + + return *cqe_ctx; +} + +int rdma_rm_alloc_cqe_ctx(RdmaDeviceResources *dev_res, uint32_t *cqe_ctx_id, + void *ctx) +{ + void **cqe_ctx; + + cqe_ctx = res_tbl_alloc(&dev_res->cqe_ctx_tbl, cqe_ctx_id); + if (!cqe_ctx) { + return -ENOMEM; + } + + pr_dbg("ctx=%p\n", ctx); + *cqe_ctx = ctx; + + return 0; +} + +void rdma_rm_dealloc_cqe_ctx(RdmaDeviceResources *dev_res, uint32_t cqe_ctx_id) +{ + res_tbl_dealloc(&dev_res->cqe_ctx_tbl, cqe_ctx_id); +} + +static void destroy_qp_hash_key(gpointer data) +{ + g_bytes_unref(data); +} + +int rdma_rm_init(RdmaDeviceResources *dev_res, struct ibv_device_attr *dev_attr, + Error **errp) +{ + dev_res->qp_hash = g_hash_table_new_full(g_bytes_hash, g_bytes_equal, + destroy_qp_hash_key, NULL); + if (!dev_res->qp_hash) { + return -ENOMEM; + } + + res_tbl_init("PD", &dev_res->pd_tbl, dev_attr->max_pd, sizeof(RdmaRmPD)); + res_tbl_init("CQ", &dev_res->cq_tbl, dev_attr->max_cq, sizeof(RdmaRmCQ)); + res_tbl_init("MR", &dev_res->mr_tbl, dev_attr->max_mr, sizeof(RdmaRmMR)); + res_tbl_init("QP", &dev_res->qp_tbl, dev_attr->max_qp, sizeof(RdmaRmQP)); + res_tbl_init("CQE_CTX", &dev_res->cqe_ctx_tbl, dev_attr->max_qp * + dev_attr->max_qp_wr, sizeof(void *)); + res_tbl_init("UC", &dev_res->uc_tbl, MAX_UCS, sizeof(RdmaRmUC)); + + return 0; +} + +void rdma_rm_fini(RdmaDeviceResources *dev_res) +{ + res_tbl_free(&dev_res->uc_tbl); + res_tbl_free(&dev_res->cqe_ctx_tbl); + res_tbl_free(&dev_res->qp_tbl); + res_tbl_free(&dev_res->cq_tbl); + res_tbl_free(&dev_res->mr_tbl); + res_tbl_free(&dev_res->pd_tbl); + g_hash_table_destroy(dev_res->qp_hash); +} diff --git a/hw/rdma/rdma_rm.h b/hw/rdma/rdma_rm.h new file mode 100644 index 0000000000..be95c1b0f4 --- /dev/null +++ b/hw/rdma/rdma_rm.h @@ -0,0 +1,69 @@ +/* + * RDMA device: Definitions of Resource Manager functions + * + * Copyright (C) 2018 Oracle + * Copyright (C) 2018 Red Hat Inc + * + * Authors: + * Yuval Shaia + * Marcel Apfelbaum + * + * This work is licensed under the terms of the GNU GPL, version 2 or later. + * See the COPYING file in the top-level directory. + * + */ + +#ifndef RDMA_RM_H +#define RDMA_RM_H + +#include +#include "rdma_backend_defs.h" +#include "rdma_rm_defs.h" + +int rdma_rm_init(RdmaDeviceResources *dev_res, struct ibv_device_attr *dev_attr, + Error **errp); +void rdma_rm_fini(RdmaDeviceResources *dev_res); + +int rdma_rm_alloc_pd(RdmaDeviceResources *dev_res, RdmaBackendDev *backend_dev, + uint32_t *pd_handle, uint32_t ctx_handle); +RdmaRmPD *rdma_rm_get_pd(RdmaDeviceResources *dev_res, uint32_t pd_handle); +void rdma_rm_dealloc_pd(RdmaDeviceResources *dev_res, uint32_t pd_handle); + +int rdma_rm_alloc_mr(RdmaDeviceResources *dev_res, uint32_t pd_handle, + uint64_t guest_start, size_t guest_length, void *host_virt, + int access_flags, uint32_t *mr_handle, uint32_t *lkey, + uint32_t *rkey); +RdmaRmMR *rdma_rm_get_mr(RdmaDeviceResources *dev_res, uint32_t mr_handle); +void rdma_rm_dealloc_mr(RdmaDeviceResources *dev_res, uint32_t mr_handle); + +int rdma_rm_alloc_uc(RdmaDeviceResources *dev_res, uint32_t pfn, + uint32_t *uc_handle); +RdmaRmUC *rdma_rm_get_uc(RdmaDeviceResources *dev_res, uint32_t uc_handle); +void rdma_rm_dealloc_uc(RdmaDeviceResources *dev_res, uint32_t uc_handle); + +int rdma_rm_alloc_cq(RdmaDeviceResources *dev_res, RdmaBackendDev *backend_dev, + uint32_t cqe, uint32_t *cq_handle, void *opaque); +RdmaRmCQ *rdma_rm_get_cq(RdmaDeviceResources *dev_res, uint32_t cq_handle); +void rdma_rm_req_notify_cq(RdmaDeviceResources *dev_res, uint32_t cq_handle, + bool notify); +void rdma_rm_dealloc_cq(RdmaDeviceResources *dev_res, uint32_t cq_handle); + +int rdma_rm_alloc_qp(RdmaDeviceResources *dev_res, uint32_t pd_handle, + uint8_t qp_type, uint32_t max_send_wr, + uint32_t max_send_sge, uint32_t send_cq_handle, + uint32_t max_recv_wr, uint32_t max_recv_sge, + uint32_t recv_cq_handle, void *opaque, uint32_t *qpn); +RdmaRmQP *rdma_rm_get_qp(RdmaDeviceResources *dev_res, uint32_t qpn); +int rdma_rm_modify_qp(RdmaDeviceResources *dev_res, RdmaBackendDev *backend_dev, + uint32_t qp_handle, uint32_t attr_mask, + union ibv_gid *dgid, uint32_t dqpn, + enum ibv_qp_state qp_state, uint32_t qkey, + uint32_t rq_psn, uint32_t sq_psn); +void rdma_rm_dealloc_qp(RdmaDeviceResources *dev_res, uint32_t qp_handle); + +int rdma_rm_alloc_cqe_ctx(RdmaDeviceResources *dev_res, uint32_t *cqe_ctx_id, + void *ctx); +void *rdma_rm_get_cqe_ctx(RdmaDeviceResources *dev_res, uint32_t cqe_ctx_id); +void rdma_rm_dealloc_cqe_ctx(RdmaDeviceResources *dev_res, uint32_t cqe_ctx_id); + +#endif diff --git a/hw/rdma/trace-events b/hw/rdma/trace-events new file mode 100644 index 0000000000..c4c202e647 --- /dev/null +++ b/hw/rdma/trace-events @@ -0,0 +1,5 @@ +# See docs/tracing.txt for syntax documentation. + +#hw/rdma/rdma_backend.c +create_ah_cache_hit(uint64_t subnet, uint64_t net_id) "subnet = 0x%"PRIx64" net_id = 0x%"PRIx64 +create_ah_cache_miss(uint64_t subnet, uint64_t net_id) "subnet = 0x%"PRIx64" net_id = 0x%"PRIx64 -- 2.13.5