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     void *addr;
 148     size_t length;
 149
 150     pd = rdma_rm_get_pd(dev_res, pd_handle);
 151     if (!pd) {
 152         pr_dbg("Invalid PD\n");
 153         return -EINVAL;
 154     }
 155
 156     mr = res_tbl_alloc(&dev_res->mr_tbl, mr_handle);
 157     if (!mr) {
 158         pr_dbg("Failed to allocate obj in table\n");
 159         return -ENOMEM;
 160     }
 161
 162     if (!host_virt) {
 163         /* TODO: This is my guess but not so sure that this needs to be
 164          * done */
 165         length = TARGET_PAGE_SIZE;
 166         addr = g_malloc(length);
 167     } else {
 168         mr->user_mr.host_virt = host_virt;
 169         pr_dbg("host_virt=0x%p\n", mr->user_mr.host_virt);
 170         mr->user_mr.length = guest_length;
 171         pr_dbg("length=%zu\n", guest_length);
 172         mr->user_mr.guest_start = guest_start;
 173         pr_dbg("guest_start=0x%" PRIx64 "\n", mr->user_mr.guest_start);
 174
 175         length = mr->user_mr.length;
 176         addr = mr->user_mr.host_virt;
 177     }
 178
 179     ret = rdma_backend_create_mr(&mr->backend_mr, &pd->backend_pd, addr, length,
 180                                  access_flags);
 181     if (ret) {
 182         pr_dbg("Fail in rdma_backend_create_mr, err=%d\n", ret);
 183         ret = -EIO;
 184         goto out_dealloc_mr;
 185     }
 186
 187     if (!host_virt) {
 188         *lkey = mr->lkey = rdma_backend_mr_lkey(&mr->backend_mr);
 189         *rkey = mr->rkey = rdma_backend_mr_rkey(&mr->backend_mr);
 190     } else {
 191         /* We keep mr_handle in lkey so send and recv get get mr ptr */
 192         *lkey = *mr_handle;
 193         *rkey = -1;
 194     }
 195
 196     mr->pd_handle = pd_handle;
 197
 198     return 0;
 199
 200 out_dealloc_mr:
 201     res_tbl_dealloc(&dev_res->mr_tbl, *mr_handle);
 202
 203     return ret;
 204 }
 205
 206 RdmaRmMR *rdma_rm_get_mr(RdmaDeviceResources *dev_res, uint32_t mr_handle)
 207 {
 208     return res_tbl_get(&dev_res->mr_tbl, mr_handle);
 209 }
 210
 211 void rdma_rm_dealloc_mr(RdmaDeviceResources *dev_res, uint32_t mr_handle)
 212 {
 213     RdmaRmMR *mr = rdma_rm_get_mr(dev_res, mr_handle);
 214
 215     if (mr) {
 216         rdma_backend_destroy_mr(&mr->backend_mr);
 217         munmap(mr->user_mr.host_virt, mr->user_mr.length);
 218         res_tbl_dealloc(&dev_res->mr_tbl, mr_handle);
 219     }
 220 }
 221
 222 int rdma_rm_alloc_uc(RdmaDeviceResources *dev_res, uint32_t pfn,
 223                      uint32_t *uc_handle)
 224 {
 225     RdmaRmUC *uc;
 226
 227     /* TODO: Need to make sure pfn is between bar start address and
 228      * bsd+RDMA_BAR2_UAR_SIZE
 229     if (pfn > RDMA_BAR2_UAR_SIZE) {
 230         pr_err("pfn out of range (%d > %d)\n", pfn, RDMA_BAR2_UAR_SIZE);
 231         return -ENOMEM;
 232     }
 233     */
 234
 235     uc = res_tbl_alloc(&dev_res->uc_tbl, uc_handle);
 236     if (!uc) {
 237         return -ENOMEM;
 238     }
 239
 240     return 0;
 241 }
 242
 243 RdmaRmUC *rdma_rm_get_uc(RdmaDeviceResources *dev_res, uint32_t uc_handle)
 244 {
 245     return res_tbl_get(&dev_res->uc_tbl, uc_handle);
 246 }
 247
 248 void rdma_rm_dealloc_uc(RdmaDeviceResources *dev_res, uint32_t uc_handle)
 249 {
 250     RdmaRmUC *uc = rdma_rm_get_uc(dev_res, uc_handle);
 251
 252     if (uc) {
 253         res_tbl_dealloc(&dev_res->uc_tbl, uc_handle);
 254     }
 255 }
 256
 257 RdmaRmCQ *rdma_rm_get_cq(RdmaDeviceResources *dev_res, uint32_t cq_handle)
 258 {
 259     return res_tbl_get(&dev_res->cq_tbl, cq_handle);
 260 }
 261
 262 int rdma_rm_alloc_cq(RdmaDeviceResources *dev_res, RdmaBackendDev *backend_dev,
 263                      uint32_t cqe, uint32_t *cq_handle, void *opaque)
 264 {
 265     int rc;
 266     RdmaRmCQ *cq;
 267
 268     cq = res_tbl_alloc(&dev_res->cq_tbl, cq_handle);
 269     if (!cq) {
 270         return -ENOMEM;
 271     }
 272
 273     cq->opaque = opaque;
 274     cq->notify = false;
 275
 276     rc = rdma_backend_create_cq(backend_dev, &cq->backend_cq, cqe);
 277     if (rc) {
 278         rc = -EIO;
 279         goto out_dealloc_cq;
 280     }
 281
 282     return 0;
 283
 284 out_dealloc_cq:
 285     rdma_rm_dealloc_cq(dev_res, *cq_handle);
 286
 287     return rc;
 288 }
 289
 290 void rdma_rm_req_notify_cq(RdmaDeviceResources *dev_res, uint32_t cq_handle,
 291                            bool notify)
 292 {
 293     RdmaRmCQ *cq;
 294
 295     pr_dbg("cq_handle=%d, notify=0x%x\n", cq_handle, notify);
 296
 297     cq = rdma_rm_get_cq(dev_res, cq_handle);
 298     if (!cq) {
 299         return;
 300     }
 301
 302     cq->notify = notify;
 303     pr_dbg("notify=%d\n", cq->notify);
 304 }
 305
 306 void rdma_rm_dealloc_cq(RdmaDeviceResources *dev_res, uint32_t cq_handle)
 307 {
 308     RdmaRmCQ *cq;
 309
 310     cq = rdma_rm_get_cq(dev_res, cq_handle);
 311     if (!cq) {
 312         return;
 313     }
 314
 315     rdma_backend_destroy_cq(&cq->backend_cq);
 316
 317     res_tbl_dealloc(&dev_res->cq_tbl, cq_handle);
 318 }
 319
 320 RdmaRmQP *rdma_rm_get_qp(RdmaDeviceResources *dev_res, uint32_t qpn)
 321 {
 322     GBytes *key = g_bytes_new(&qpn, sizeof(qpn));
 323
 324     RdmaRmQP *qp = g_hash_table_lookup(dev_res->qp_hash, key);
 325
 326     g_bytes_unref(key);
 327
 328     return qp;
 329 }
 330
 331 int rdma_rm_alloc_qp(RdmaDeviceResources *dev_res, uint32_t pd_handle,
 332                      uint8_t qp_type, uint32_t max_send_wr,
 333                      uint32_t max_send_sge, uint32_t send_cq_handle,
 334                      uint32_t max_recv_wr, uint32_t max_recv_sge,
 335                      uint32_t recv_cq_handle, void *opaque, uint32_t *qpn)
 336 {
 337     int rc;
 338     RdmaRmQP *qp;
 339     RdmaRmCQ *scq, *rcq;
 340     RdmaRmPD *pd;
 341     uint32_t rm_qpn;
 342
 343     pr_dbg("qp_type=%d\n", qp_type);
 344
 345     pd = rdma_rm_get_pd(dev_res, pd_handle);
 346     if (!pd) {
 347         pr_err("Invalid pd handle (%d)\n", pd_handle);
 348         return -EINVAL;
 349     }
 350
 351     scq = rdma_rm_get_cq(dev_res, send_cq_handle);
 352     rcq = rdma_rm_get_cq(dev_res, recv_cq_handle);
 353
 354     if (!scq || !rcq) {
 355         pr_err("Invalid send_cqn or recv_cqn (%d, %d)\n",
 356                send_cq_handle, recv_cq_handle);
 357         return -EINVAL;
 358     }
 359
 360     qp = res_tbl_alloc(&dev_res->qp_tbl, &rm_qpn);
 361     if (!qp) {
 362         return -ENOMEM;
 363     }
 364     pr_dbg("rm_qpn=%d\n", rm_qpn);
 365
 366     qp->qpn = rm_qpn;
 367     qp->qp_state = IBV_QPS_RESET;
 368     qp->qp_type = qp_type;
 369     qp->send_cq_handle = send_cq_handle;
 370     qp->recv_cq_handle = recv_cq_handle;
 371     qp->opaque = opaque;
 372
 373     rc = rdma_backend_create_qp(&qp->backend_qp, qp_type, &pd->backend_pd,
 374                                 &scq->backend_cq, &rcq->backend_cq, max_send_wr,
 375                                 max_recv_wr, max_send_sge, max_recv_sge);
 376     if (rc) {
 377         rc = -EIO;
 378         goto out_dealloc_qp;
 379     }
 380
 381     *qpn = rdma_backend_qpn(&qp->backend_qp);
 382     pr_dbg("rm_qpn=%d, backend_qpn=0x%x\n", rm_qpn, *qpn);
 383     g_hash_table_insert(dev_res->qp_hash, g_bytes_new(qpn, sizeof(*qpn)), qp);
 384
 385     return 0;
 386
 387 out_dealloc_qp:
 388     res_tbl_dealloc(&dev_res->qp_tbl, qp->qpn);
 389
 390     return rc;
 391 }
 392
 393 int rdma_rm_modify_qp(RdmaDeviceResources *dev_res, RdmaBackendDev *backend_dev,
 394                       uint32_t qp_handle, uint32_t attr_mask,
 395                       union ibv_gid *dgid, uint32_t dqpn,
 396                       enum ibv_qp_state qp_state, uint32_t qkey,
 397                       uint32_t rq_psn, uint32_t sq_psn)
 398 {
 399     RdmaRmQP *qp;
 400     int ret;
 401
 402     pr_dbg("qpn=%d\n", qp_handle);
 403
 404     qp = rdma_rm_get_qp(dev_res, qp_handle);
 405     if (!qp) {
 406         return -EINVAL;
 407     }
 408
 409     pr_dbg("qp_type=%d\n", qp->qp_type);
 410     pr_dbg("attr_mask=0x%x\n", attr_mask);
 411
 412     if (qp->qp_type == IBV_QPT_SMI) {
 413         pr_dbg("QP0 unsupported\n");
 414         return -EPERM;
 415     } else if (qp->qp_type == IBV_QPT_GSI) {
 416         pr_dbg("QP1\n");
 417         return 0;
 418     }
 419
 420     if (attr_mask & IBV_QP_STATE) {
 421         qp->qp_state = qp_state;
 422         pr_dbg("qp_state=%d\n", qp->qp_state);
 423
 424         if (qp->qp_state == IBV_QPS_INIT) {
 425             ret = rdma_backend_qp_state_init(backend_dev, &qp->backend_qp,
 426                                              qp->qp_type, qkey);
 427             if (ret) {
 428                 return -EIO;
 429             }
 430         }
 431
 432         if (qp->qp_state == IBV_QPS_RTR) {
 433             ret = rdma_backend_qp_state_rtr(backend_dev, &qp->backend_qp,
 434                                             qp->qp_type, dgid, dqpn, rq_psn,
 435                                             qkey, attr_mask & IBV_QP_QKEY);
 436             if (ret) {
 437                 return -EIO;
 438             }
 439         }
 440
 441         if (qp->qp_state == IBV_QPS_RTS) {
 442             ret = rdma_backend_qp_state_rts(&qp->backend_qp, qp->qp_type,
 443                                             sq_psn, qkey,
 444                                             attr_mask & IBV_QP_QKEY);
 445             if (ret) {
 446                 return -EIO;
 447             }
 448         }
 449     }
 450
 451     return 0;
 452 }
 453
 454 int rdma_rm_query_qp(RdmaDeviceResources *dev_res, RdmaBackendDev *backend_dev,
 455                      uint32_t qp_handle, struct ibv_qp_attr *attr,
 456                      int attr_mask, struct ibv_qp_init_attr *init_attr)
 457 {
 458     RdmaRmQP *qp;
 459
 460     pr_dbg("qpn=%d\n", qp_handle);
 461
 462     qp = rdma_rm_get_qp(dev_res, qp_handle);
 463     if (!qp) {
 464         return -EINVAL;
 465     }
 466
 467     pr_dbg("qp_type=%d\n", qp->qp_type);
 468
 469     return rdma_backend_query_qp(&qp->backend_qp, attr, attr_mask, init_attr);
 470 }
 471
 472 void rdma_rm_dealloc_qp(RdmaDeviceResources *dev_res, uint32_t qp_handle)
 473 {
 474     RdmaRmQP *qp;
 475     GBytes *key;
 476
 477     key = g_bytes_new(&qp_handle, sizeof(qp_handle));
 478     qp = g_hash_table_lookup(dev_res->qp_hash, key);
 479     g_hash_table_remove(dev_res->qp_hash, key);
 480     g_bytes_unref(key);
 481
 482     if (!qp) {
 483         return;
 484     }
 485
 486     rdma_backend_destroy_qp(&qp->backend_qp);
 487
 488     res_tbl_dealloc(&dev_res->qp_tbl, qp->qpn);
 489 }
 490
 491 void *rdma_rm_get_cqe_ctx(RdmaDeviceResources *dev_res, uint32_t cqe_ctx_id)
 492 {
 493     void **cqe_ctx;
 494
 495     cqe_ctx = res_tbl_get(&dev_res->cqe_ctx_tbl, cqe_ctx_id);
 496     if (!cqe_ctx) {
 497         return NULL;
 498     }
 499
 500     pr_dbg("ctx=%p\n", *cqe_ctx);
 501
 502     return *cqe_ctx;
 503 }
 504
 505 int rdma_rm_alloc_cqe_ctx(RdmaDeviceResources *dev_res, uint32_t *cqe_ctx_id,
 506                           void *ctx)
 507 {
 508     void **cqe_ctx;
 509
 510     cqe_ctx = res_tbl_alloc(&dev_res->cqe_ctx_tbl, cqe_ctx_id);
 511     if (!cqe_ctx) {
 512         return -ENOMEM;
 513     }
 514
 515     pr_dbg("ctx=%p\n", ctx);
 516     *cqe_ctx = ctx;
 517
 518     return 0;
 519 }
 520
 521 void rdma_rm_dealloc_cqe_ctx(RdmaDeviceResources *dev_res, uint32_t cqe_ctx_id)
 522 {
 523     res_tbl_dealloc(&dev_res->cqe_ctx_tbl, cqe_ctx_id);
 524 }
 525
 526 static void destroy_qp_hash_key(gpointer data)
 527 {
 528     g_bytes_unref(data);
 529 }
 530
 531 int rdma_rm_init(RdmaDeviceResources *dev_res, struct ibv_device_attr *dev_attr,
 532                  Error **errp)
 533 {
 534     dev_res->qp_hash = g_hash_table_new_full(g_bytes_hash, g_bytes_equal,
 535                                              destroy_qp_hash_key, NULL);
 536     if (!dev_res->qp_hash) {
 537         return -ENOMEM;
 538     }
 539
 540     res_tbl_init("PD", &dev_res->pd_tbl, dev_attr->max_pd, sizeof(RdmaRmPD));
 541     res_tbl_init("CQ", &dev_res->cq_tbl, dev_attr->max_cq, sizeof(RdmaRmCQ));
 542     res_tbl_init("MR", &dev_res->mr_tbl, dev_attr->max_mr, sizeof(RdmaRmMR));
 543     res_tbl_init("QP", &dev_res->qp_tbl, dev_attr->max_qp, sizeof(RdmaRmQP));
 544     res_tbl_init("CQE_CTX", &dev_res->cqe_ctx_tbl, dev_attr->max_qp *
 545                        dev_attr->max_qp_wr, sizeof(void *));
 546     res_tbl_init("UC", &dev_res->uc_tbl, MAX_UCS, sizeof(RdmaRmUC));
 547
 548     return 0;
 549 }
 550
 551 void rdma_rm_fini(RdmaDeviceResources *dev_res)
 552 {
 553     res_tbl_free(&dev_res->uc_tbl);
 554     res_tbl_free(&dev_res->cqe_ctx_tbl);
 555     res_tbl_free(&dev_res->qp_tbl);
 556     res_tbl_free(&dev_res->cq_tbl);
 557     res_tbl_free(&dev_res->mr_tbl);
 558     res_tbl_free(&dev_res->pd_tbl);
 559     g_hash_table_destroy(dev_res->qp_hash);
 560 }