hw/rdma/rdma_rm.c

   1 /*
   2  * QEMU paravirtual RDMA - Resource Manager Implementation
   3  *
   4  * Copyright (C) 2018 Oracle
   5  * Copyright (C) 2018 Red Hat Inc
   6  *
   7  * Authors:
   8  *     Yuval Shaia <yuval.shaia@oracle.com>
   9  *     Marcel Apfelbaum <marcel@redhat.com>
  10  *
  11  * This work is licensed under the terms of the GNU GPL, version 2 or later.
  12  * See the COPYING file in the top-level directory.
  13  *
  14  */
  15
  16 #include "qemu/osdep.h"
  17 #include "qapi/error.h"
  18 #include "cpu.h"
  19
  20 #include "rdma_utils.h"
  21 #include "rdma_backend.h"
  22 #include "rdma_rm.h"
  23
  24 /* Page directory and page tables */
  25 #define PG_DIR_SZ { TARGET_PAGE_SIZE / sizeof(__u64) }
  26 #define PG_TBL_SZ { TARGET_PAGE_SIZE / sizeof(__u64) }
  27
  28 static inline void res_tbl_init(const char *name, RdmaRmResTbl *tbl,
  29                                 uint32_t tbl_sz, uint32_t res_sz)
  30 {
  31     tbl->tbl = g_malloc(tbl_sz * res_sz);
  32
  33     strncpy(tbl->name, name, MAX_RM_TBL_NAME);
  34     tbl->name[MAX_RM_TBL_NAME - 1] = 0;
  35
  36     tbl->bitmap = bitmap_new(tbl_sz);
  37     tbl->tbl_sz = tbl_sz;
  38     tbl->res_sz = res_sz;
  39     qemu_mutex_init(&tbl->lock);
  40 }
  41
  42 static inline void res_tbl_free(RdmaRmResTbl *tbl)
  43 {
  44     qemu_mutex_destroy(&tbl->lock);
  45     g_free(tbl->tbl);
  46     bitmap_zero_extend(tbl->bitmap, tbl->tbl_sz, 0);
  47 }
  48
  49 static inline void *res_tbl_get(RdmaRmResTbl *tbl, uint32_t handle)
  50 {
  51     pr_dbg("%s, handle=%d\n", tbl->name, handle);
  52
  53     if ((handle < tbl->tbl_sz) && (test_bit(handle, tbl->bitmap))) {
  54         return tbl->tbl + handle * tbl->res_sz;
  55     } else {
  56         pr_dbg("Invalid handle %d\n", handle);
  57         return NULL;
  58     }
  59 }
  60
  61 static inline void *res_tbl_alloc(RdmaRmResTbl *tbl, uint32_t *handle)
  62 {
  63     qemu_mutex_lock(&tbl->lock);
  64
  65     *handle = find_first_zero_bit(tbl->bitmap, tbl->tbl_sz);
  66     if (*handle > tbl->tbl_sz) {
  67         pr_dbg("Failed to alloc, bitmap is full\n");
  68         qemu_mutex_unlock(&tbl->lock);
  69         return NULL;
  70     }
  71
  72     set_bit(*handle, tbl->bitmap);
  73
  74     qemu_mutex_unlock(&tbl->lock);
  75
  76     memset(tbl->tbl + *handle * tbl->res_sz, 0, tbl->res_sz);
  77
  78     pr_dbg("%s, handle=%d\n", tbl->name, *handle);
  79
  80     return tbl->tbl + *handle * tbl->res_sz;
  81 }
  82
  83 static inline void res_tbl_dealloc(RdmaRmResTbl *tbl, uint32_t handle)
  84 {
  85     pr_dbg("%s, handle=%d\n", tbl->name, handle);
  86
  87     qemu_mutex_lock(&tbl->lock);
  88
  89     if (handle < tbl->tbl_sz) {
  90         clear_bit(handle, tbl->bitmap);
  91     }
  92
  93     qemu_mutex_unlock(&tbl->lock);
  94 }
  95
  96 int rdma_rm_alloc_pd(RdmaDeviceResources *dev_res, RdmaBackendDev *backend_dev,
  97                      uint32_t *pd_handle, uint32_t ctx_handle)
  98 {
  99     RdmaRmPD *pd;
 100     int ret = -ENOMEM;
 101
 102     pd = res_tbl_alloc(&dev_res->pd_tbl, pd_handle);
 103     if (!pd) {
 104         goto out;
 105     }
 106
 107     ret = rdma_backend_create_pd(backend_dev, &pd->backend_pd);
 108     if (ret) {
 109         ret = -EIO;
 110         goto out_tbl_dealloc;
 111     }
 112
 113     pd->ctx_handle = ctx_handle;
 114
 115     return 0;
 116
 117 out_tbl_dealloc:
 118     res_tbl_dealloc(&dev_res->pd_tbl, *pd_handle);
 119
 120 out:
 121     return ret;
 122 }
 123
 124 RdmaRmPD *rdma_rm_get_pd(RdmaDeviceResources *dev_res, uint32_t pd_handle)
 125 {
 126     return res_tbl_get(&dev_res->pd_tbl, pd_handle);
 127 }
 128
 129 void rdma_rm_dealloc_pd(RdmaDeviceResources *dev_res, uint32_t pd_handle)
 130 {
 131     RdmaRmPD *pd = rdma_rm_get_pd(dev_res, pd_handle);
 132
 133     if (pd) {
 134         rdma_backend_destroy_pd(&pd->backend_pd);
 135         res_tbl_dealloc(&dev_res->pd_tbl, pd_handle);
 136     }
 137 }
 138
 139 int rdma_rm_alloc_mr(RdmaDeviceResources *dev_res, uint32_t pd_handle,
 140                      uint64_t guest_start, size_t guest_length, void *host_virt,
 141                      int access_flags, uint32_t *mr_handle, uint32_t *lkey,
 142                      uint32_t *rkey)
 143 {
 144     RdmaRmMR *mr;
 145     int ret = 0;
 146     RdmaRmPD *pd;
 147
 148     pd = rdma_rm_get_pd(dev_res, pd_handle);
 149     if (!pd) {
 150         pr_dbg("Invalid PD\n");
 151         return -EINVAL;
 152     }
 153
 154     mr = res_tbl_alloc(&dev_res->mr_tbl, mr_handle);
 155     if (!mr) {
 156         pr_dbg("Failed to allocate obj in table\n");
 157         return -ENOMEM;
 158     }
 159     pr_dbg("mr_handle=%d\n", *mr_handle);
 160
 161     pr_dbg("host_virt=0x%p\n", host_virt);
 162     pr_dbg("guest_start=0x%" PRIx64 "\n", guest_start);
 163     pr_dbg("length=%zu\n", guest_length);
 164
 165     if (host_virt) {
 166         mr->virt = host_virt;
 167         mr->start = guest_start;
 168         mr->length = guest_length;
 169         mr->virt += (mr->start & (TARGET_PAGE_SIZE - 1));
 170
 171         ret = rdma_backend_create_mr(&mr->backend_mr, &pd->backend_pd, mr->virt,
 172                                      mr->length, access_flags);
 173         if (ret) {
 174             pr_dbg("Fail in rdma_backend_create_mr, err=%d\n", ret);
 175             ret = -EIO;
 176             goto out_dealloc_mr;
 177         }
 178     }
 179
 180     /* We keep mr_handle in lkey so send and recv get get mr ptr */
 181     *lkey = *mr_handle;
 182     *rkey = -1;
 183
 184     mr->pd_handle = pd_handle;
 185
 186     return 0;
 187
 188 out_dealloc_mr:
 189     res_tbl_dealloc(&dev_res->mr_tbl, *mr_handle);
 190
 191     return ret;
 192 }
 193
 194 RdmaRmMR *rdma_rm_get_mr(RdmaDeviceResources *dev_res, uint32_t mr_handle)
 195 {
 196     return res_tbl_get(&dev_res->mr_tbl, mr_handle);
 197 }
 198
 199 void rdma_rm_dealloc_mr(RdmaDeviceResources *dev_res, uint32_t mr_handle)
 200 {
 201     RdmaRmMR *mr = rdma_rm_get_mr(dev_res, mr_handle);
 202
 203     if (mr) {
 204         rdma_backend_destroy_mr(&mr->backend_mr);
 205         pr_dbg("start=0x%" PRIx64 "\n", mr->start);
 206         if (mr->start) {
 207             mr->virt -= (mr->start & (TARGET_PAGE_SIZE - 1));
 208             munmap(mr->virt, mr->length);
 209         }
 210         res_tbl_dealloc(&dev_res->mr_tbl, mr_handle);
 211     }
 212 }
 213
 214 int rdma_rm_alloc_uc(RdmaDeviceResources *dev_res, uint32_t pfn,
 215                      uint32_t *uc_handle)
 216 {
 217     RdmaRmUC *uc;
 218
 219     /* TODO: Need to make sure pfn is between bar start address and
 220      * bsd+RDMA_BAR2_UAR_SIZE
 221     if (pfn > RDMA_BAR2_UAR_SIZE) {
 222         pr_err("pfn out of range (%d > %d)\n", pfn, RDMA_BAR2_UAR_SIZE);
 223         return -ENOMEM;
 224     }
 225     */
 226
 227     uc = res_tbl_alloc(&dev_res->uc_tbl, uc_handle);
 228     if (!uc) {
 229         return -ENOMEM;
 230     }
 231
 232     return 0;
 233 }
 234
 235 RdmaRmUC *rdma_rm_get_uc(RdmaDeviceResources *dev_res, uint32_t uc_handle)
 236 {
 237     return res_tbl_get(&dev_res->uc_tbl, uc_handle);
 238 }
 239
 240 void rdma_rm_dealloc_uc(RdmaDeviceResources *dev_res, uint32_t uc_handle)
 241 {
 242     RdmaRmUC *uc = rdma_rm_get_uc(dev_res, uc_handle);
 243
 244     if (uc) {
 245         res_tbl_dealloc(&dev_res->uc_tbl, uc_handle);
 246     }
 247 }
 248
 249 RdmaRmCQ *rdma_rm_get_cq(RdmaDeviceResources *dev_res, uint32_t cq_handle)
 250 {
 251     return res_tbl_get(&dev_res->cq_tbl, cq_handle);
 252 }
 253
 254 int rdma_rm_alloc_cq(RdmaDeviceResources *dev_res, RdmaBackendDev *backend_dev,
 255                      uint32_t cqe, uint32_t *cq_handle, void *opaque)
 256 {
 257     int rc;
 258     RdmaRmCQ *cq;
 259
 260     cq = res_tbl_alloc(&dev_res->cq_tbl, cq_handle);
 261     if (!cq) {
 262         return -ENOMEM;
 263     }
 264
 265     cq->opaque = opaque;
 266     cq->notify = false;
 267
 268     rc = rdma_backend_create_cq(backend_dev, &cq->backend_cq, cqe);
 269     if (rc) {
 270         rc = -EIO;
 271         goto out_dealloc_cq;
 272     }
 273
 274     return 0;
 275
 276 out_dealloc_cq:
 277     rdma_rm_dealloc_cq(dev_res, *cq_handle);
 278
 279     return rc;
 280 }
 281
 282 void rdma_rm_req_notify_cq(RdmaDeviceResources *dev_res, uint32_t cq_handle,
 283                            bool notify)
 284 {
 285     RdmaRmCQ *cq;
 286
 287     pr_dbg("cq_handle=%d, notify=0x%x\n", cq_handle, notify);
 288
 289     cq = rdma_rm_get_cq(dev_res, cq_handle);
 290     if (!cq) {
 291         return;
 292     }
 293
 294     cq->notify = notify;
 295     pr_dbg("notify=%d\n", cq->notify);
 296 }
 297
 298 void rdma_rm_dealloc_cq(RdmaDeviceResources *dev_res, uint32_t cq_handle)
 299 {
 300     RdmaRmCQ *cq;
 301
 302     cq = rdma_rm_get_cq(dev_res, cq_handle);
 303     if (!cq) {
 304         return;
 305     }
 306
 307     rdma_backend_destroy_cq(&cq->backend_cq);
 308
 309     res_tbl_dealloc(&dev_res->cq_tbl, cq_handle);
 310 }
 311
 312 RdmaRmQP *rdma_rm_get_qp(RdmaDeviceResources *dev_res, uint32_t qpn)
 313 {
 314     GBytes *key = g_bytes_new(&qpn, sizeof(qpn));
 315
 316     RdmaRmQP *qp = g_hash_table_lookup(dev_res->qp_hash, key);
 317
 318     g_bytes_unref(key);
 319
 320     return qp;
 321 }
 322
 323 int rdma_rm_alloc_qp(RdmaDeviceResources *dev_res, uint32_t pd_handle,
 324                      uint8_t qp_type, uint32_t max_send_wr,
 325                      uint32_t max_send_sge, uint32_t send_cq_handle,
 326                      uint32_t max_recv_wr, uint32_t max_recv_sge,
 327                      uint32_t recv_cq_handle, void *opaque, uint32_t *qpn)
 328 {
 329     int rc;
 330     RdmaRmQP *qp;
 331     RdmaRmCQ *scq, *rcq;
 332     RdmaRmPD *pd;
 333     uint32_t rm_qpn;
 334
 335     pr_dbg("qp_type=%d\n", qp_type);
 336
 337     pd = rdma_rm_get_pd(dev_res, pd_handle);
 338     if (!pd) {
 339         pr_err("Invalid pd handle (%d)\n", pd_handle);
 340         return -EINVAL;
 341     }
 342
 343     scq = rdma_rm_get_cq(dev_res, send_cq_handle);
 344     rcq = rdma_rm_get_cq(dev_res, recv_cq_handle);
 345
 346     if (!scq || !rcq) {
 347         pr_err("Invalid send_cqn or recv_cqn (%d, %d)\n",
 348                send_cq_handle, recv_cq_handle);
 349         return -EINVAL;
 350     }
 351
 352     qp = res_tbl_alloc(&dev_res->qp_tbl, &rm_qpn);
 353     if (!qp) {
 354         return -ENOMEM;
 355     }
 356     pr_dbg("rm_qpn=%d\n", rm_qpn);
 357
 358     qp->qpn = rm_qpn;
 359     qp->qp_state = IBV_QPS_RESET;
 360     qp->qp_type = qp_type;
 361     qp->send_cq_handle = send_cq_handle;
 362     qp->recv_cq_handle = recv_cq_handle;
 363     qp->opaque = opaque;
 364
 365     rc = rdma_backend_create_qp(&qp->backend_qp, qp_type, &pd->backend_pd,
 366                                 &scq->backend_cq, &rcq->backend_cq, max_send_wr,
 367                                 max_recv_wr, max_send_sge, max_recv_sge);
 368     if (rc) {
 369         rc = -EIO;
 370         goto out_dealloc_qp;
 371     }
 372
 373     *qpn = rdma_backend_qpn(&qp->backend_qp);
 374     pr_dbg("rm_qpn=%d, backend_qpn=0x%x\n", rm_qpn, *qpn);
 375     g_hash_table_insert(dev_res->qp_hash, g_bytes_new(qpn, sizeof(*qpn)), qp);
 376
 377     return 0;
 378
 379 out_dealloc_qp:
 380     res_tbl_dealloc(&dev_res->qp_tbl, qp->qpn);
 381
 382     return rc;
 383 }
 384
 385 int rdma_rm_modify_qp(RdmaDeviceResources *dev_res, RdmaBackendDev *backend_dev,
 386                       uint32_t qp_handle, uint32_t attr_mask,
 387                       union ibv_gid *dgid, uint32_t dqpn,
 388                       enum ibv_qp_state qp_state, uint32_t qkey,
 389                       uint32_t rq_psn, uint32_t sq_psn)
 390 {
 391     RdmaRmQP *qp;
 392     int ret;
 393
 394     pr_dbg("qpn=0x%x\n", qp_handle);
 395
 396     qp = rdma_rm_get_qp(dev_res, qp_handle);
 397     if (!qp) {
 398         return -EINVAL;
 399     }
 400
 401     pr_dbg("qp_type=%d\n", qp->qp_type);
 402     pr_dbg("attr_mask=0x%x\n", attr_mask);
 403
 404     if (qp->qp_type == IBV_QPT_SMI) {
 405         pr_dbg("QP0 unsupported\n");
 406         return -EPERM;
 407     } else if (qp->qp_type == IBV_QPT_GSI) {
 408         pr_dbg("QP1\n");
 409         return 0;
 410     }
 411
 412     if (attr_mask & IBV_QP_STATE) {
 413         qp->qp_state = qp_state;
 414         pr_dbg("qp_state=%d\n", qp->qp_state);
 415
 416         if (qp->qp_state == IBV_QPS_INIT) {
 417             ret = rdma_backend_qp_state_init(backend_dev, &qp->backend_qp,
 418                                              qp->qp_type, qkey);
 419             if (ret) {
 420                 return -EIO;
 421             }
 422         }
 423
 424         if (qp->qp_state == IBV_QPS_RTR) {
 425             ret = rdma_backend_qp_state_rtr(backend_dev, &qp->backend_qp,
 426                                             qp->qp_type, dgid, dqpn, rq_psn,
 427                                             qkey, attr_mask & IBV_QP_QKEY);
 428             if (ret) {
 429                 return -EIO;
 430             }
 431         }
 432
 433         if (qp->qp_state == IBV_QPS_RTS) {
 434             ret = rdma_backend_qp_state_rts(&qp->backend_qp, qp->qp_type,
 435                                             sq_psn, qkey,
 436                                             attr_mask & IBV_QP_QKEY);
 437             if (ret) {
 438                 return -EIO;
 439             }
 440         }
 441     }
 442
 443     return 0;
 444 }
 445
 446 int rdma_rm_query_qp(RdmaDeviceResources *dev_res, RdmaBackendDev *backend_dev,
 447                      uint32_t qp_handle, struct ibv_qp_attr *attr,
 448                      int attr_mask, struct ibv_qp_init_attr *init_attr)
 449 {
 450     RdmaRmQP *qp;
 451
 452     pr_dbg("qpn=0x%x\n", qp_handle);
 453
 454     qp = rdma_rm_get_qp(dev_res, qp_handle);
 455     if (!qp) {
 456         return -EINVAL;
 457     }
 458
 459     pr_dbg("qp_type=%d\n", qp->qp_type);
 460
 461     return rdma_backend_query_qp(&qp->backend_qp, attr, attr_mask, init_attr);
 462 }
 463
 464 void rdma_rm_dealloc_qp(RdmaDeviceResources *dev_res, uint32_t qp_handle)
 465 {
 466     RdmaRmQP *qp;
 467     GBytes *key;
 468
 469     key = g_bytes_new(&qp_handle, sizeof(qp_handle));
 470     qp = g_hash_table_lookup(dev_res->qp_hash, key);
 471     g_hash_table_remove(dev_res->qp_hash, key);
 472     g_bytes_unref(key);
 473
 474     if (!qp) {
 475         return;
 476     }
 477
 478     rdma_backend_destroy_qp(&qp->backend_qp);
 479
 480     res_tbl_dealloc(&dev_res->qp_tbl, qp->qpn);
 481 }
 482
 483 void *rdma_rm_get_cqe_ctx(RdmaDeviceResources *dev_res, uint32_t cqe_ctx_id)
 484 {
 485     void **cqe_ctx;
 486
 487     cqe_ctx = res_tbl_get(&dev_res->cqe_ctx_tbl, cqe_ctx_id);
 488     if (!cqe_ctx) {
 489         return NULL;
 490     }
 491
 492     pr_dbg("ctx=%p\n", *cqe_ctx);
 493
 494     return *cqe_ctx;
 495 }
 496
 497 int rdma_rm_alloc_cqe_ctx(RdmaDeviceResources *dev_res, uint32_t *cqe_ctx_id,
 498                           void *ctx)
 499 {
 500     void **cqe_ctx;
 501
 502     cqe_ctx = res_tbl_alloc(&dev_res->cqe_ctx_tbl, cqe_ctx_id);
 503     if (!cqe_ctx) {
 504         return -ENOMEM;
 505     }
 506
 507     pr_dbg("ctx=%p\n", ctx);
 508     *cqe_ctx = ctx;
 509
 510     return 0;
 511 }
 512
 513 void rdma_rm_dealloc_cqe_ctx(RdmaDeviceResources *dev_res, uint32_t cqe_ctx_id)
 514 {
 515     res_tbl_dealloc(&dev_res->cqe_ctx_tbl, cqe_ctx_id);
 516 }
 517
 518 static void destroy_qp_hash_key(gpointer data)
 519 {
 520     g_bytes_unref(data);
 521 }
 522
 523 int rdma_rm_init(RdmaDeviceResources *dev_res, struct ibv_device_attr *dev_attr,
 524                  Error **errp)
 525 {
 526     dev_res->qp_hash = g_hash_table_new_full(g_bytes_hash, g_bytes_equal,
 527                                              destroy_qp_hash_key, NULL);
 528     if (!dev_res->qp_hash) {
 529         return -ENOMEM;
 530     }
 531
 532     res_tbl_init("PD", &dev_res->pd_tbl, dev_attr->max_pd, sizeof(RdmaRmPD));
 533     res_tbl_init("CQ", &dev_res->cq_tbl, dev_attr->max_cq, sizeof(RdmaRmCQ));
 534     res_tbl_init("MR", &dev_res->mr_tbl, dev_attr->max_mr, sizeof(RdmaRmMR));
 535     res_tbl_init("QP", &dev_res->qp_tbl, dev_attr->max_qp, sizeof(RdmaRmQP));
 536     res_tbl_init("CQE_CTX", &dev_res->cqe_ctx_tbl, dev_attr->max_qp *
 537                        dev_attr->max_qp_wr, sizeof(void *));
 538     res_tbl_init("UC", &dev_res->uc_tbl, MAX_UCS, sizeof(RdmaRmUC));
 539
 540     return 0;
 541 }
 542
 543 void rdma_rm_fini(RdmaDeviceResources *dev_res)
 544 {
 545     res_tbl_free(&dev_res->uc_tbl);
 546     res_tbl_free(&dev_res->cqe_ctx_tbl);
 547     res_tbl_free(&dev_res->qp_tbl);
 548     res_tbl_free(&dev_res->mr_tbl);
 549     res_tbl_free(&dev_res->cq_tbl);
 550     res_tbl_free(&dev_res->pd_tbl);
 551
 552     g_hash_table_destroy(dev_res->qp_hash);
 553 }