1 /* 2 * Copyright (c) 2004, 2005 Topspin Communications. All rights reserved. 3 * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved. 4 * Copyright (c) 2005, 2006 Cisco Systems. All rights reserved. 5 * Copyright (c) 2005 Mellanox Technologies. All rights reserved. 6 * Copyright (c) 2004 Voltaire, Inc. All rights reserved. 7 * 8 * This software is available to you under a choice of one of two 9 * licenses. You may choose to be licensed under the terms of the GNU 10 * General Public License (GPL) Version 2, available from the file 11 * COPYING in the main directory of this source tree, or the 12 * OpenIB.org BSD license below: 13 * 14 * Redistribution and use in source and binary forms, with or 15 * without modification, are permitted provided that the following 16 * conditions are met: 17 * 18 * - Redistributions of source code must retain the above 19 * copyright notice, this list of conditions and the following 20 * disclaimer. 21 * 22 * - Redistributions in binary form must reproduce the above 23 * copyright notice, this list of conditions and the following 24 * disclaimer in the documentation and/or other materials 25 * provided with the distribution. 26 * 27 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, 28 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF 29 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND 30 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS 31 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN 32 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN 33 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 34 * SOFTWARE. 35 */ 36 37 #include <rdma/ib_smi.h> 38 #include <rdma/ib_umem.h> 39 #include <rdma/ib_user_verbs.h> 40 #include <rdma/uverbs_ioctl.h> 41 42 #include <linux/sched.h> 43 #include <linux/slab.h> 44 #include <linux/stat.h> 45 #include <linux/mm.h> 46 #include <linux/export.h> 47 48 #include "mthca_dev.h" 49 #include "mthca_cmd.h" 50 #include <rdma/mthca-abi.h> 51 #include "mthca_memfree.h" 52 53 static void init_query_mad(struct ib_smp *mad) 54 { 55 mad->base_version = 1; 56 mad->mgmt_class = IB_MGMT_CLASS_SUBN_LID_ROUTED; 57 mad->class_version = 1; 58 mad->method = IB_MGMT_METHOD_GET; 59 } 60 61 static int mthca_query_device(struct ib_device *ibdev, struct ib_device_attr *props, 62 struct ib_udata *uhw) 63 { 64 struct ib_smp *in_mad = NULL; 65 struct ib_smp *out_mad = NULL; 66 int err = -ENOMEM; 67 struct mthca_dev *mdev = to_mdev(ibdev); 68 69 if (uhw->inlen || uhw->outlen) 70 return -EINVAL; 71 72 in_mad = kzalloc(sizeof *in_mad, GFP_KERNEL); 73 out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL); 74 if (!in_mad || !out_mad) 75 goto out; 76 77 memset(props, 0, sizeof *props); 78 79 props->fw_ver = mdev->fw_ver; 80 81 init_query_mad(in_mad); 82 in_mad->attr_id = IB_SMP_ATTR_NODE_INFO; 83 84 err = mthca_MAD_IFC(mdev, 1, 1, 85 1, NULL, NULL, in_mad, out_mad); 86 if (err) 87 goto out; 88 89 props->device_cap_flags = mdev->device_cap_flags; 90 props->vendor_id = be32_to_cpup((__be32 *) (out_mad->data + 36)) & 91 0xffffff; 92 props->vendor_part_id = be16_to_cpup((__be16 *) (out_mad->data + 30)); 93 props->hw_ver = be32_to_cpup((__be32 *) (out_mad->data + 32)); 94 memcpy(&props->sys_image_guid, out_mad->data + 4, 8); 95 96 props->max_mr_size = ~0ull; 97 props->page_size_cap = mdev->limits.page_size_cap; 98 props->max_qp = mdev->limits.num_qps - mdev->limits.reserved_qps; 99 props->max_qp_wr = mdev->limits.max_wqes; 100 props->max_send_sge = mdev->limits.max_sg; 101 props->max_recv_sge = mdev->limits.max_sg; 102 props->max_sge_rd = mdev->limits.max_sg; 103 props->max_cq = mdev->limits.num_cqs - mdev->limits.reserved_cqs; 104 props->max_cqe = mdev->limits.max_cqes; 105 props->max_mr = mdev->limits.num_mpts - mdev->limits.reserved_mrws; 106 props->max_pd = mdev->limits.num_pds - mdev->limits.reserved_pds; 107 props->max_qp_rd_atom = 1 << mdev->qp_table.rdb_shift; 108 props->max_qp_init_rd_atom = mdev->limits.max_qp_init_rdma; 109 props->max_res_rd_atom = props->max_qp_rd_atom * props->max_qp; 110 props->max_srq = mdev->limits.num_srqs - mdev->limits.reserved_srqs; 111 props->max_srq_wr = mdev->limits.max_srq_wqes; 112 props->max_srq_sge = mdev->limits.max_srq_sge; 113 props->local_ca_ack_delay = mdev->limits.local_ca_ack_delay; 114 props->atomic_cap = mdev->limits.flags & DEV_LIM_FLAG_ATOMIC ? 115 IB_ATOMIC_HCA : IB_ATOMIC_NONE; 116 props->max_pkeys = mdev->limits.pkey_table_len; 117 props->max_mcast_grp = mdev->limits.num_mgms + mdev->limits.num_amgms; 118 props->max_mcast_qp_attach = MTHCA_QP_PER_MGM; 119 props->max_total_mcast_qp_attach = props->max_mcast_qp_attach * 120 props->max_mcast_grp; 121 122 err = 0; 123 out: 124 kfree(in_mad); 125 kfree(out_mad); 126 return err; 127 } 128 129 static int mthca_query_port(struct ib_device *ibdev, 130 u8 port, struct ib_port_attr *props) 131 { 132 struct ib_smp *in_mad = NULL; 133 struct ib_smp *out_mad = NULL; 134 int err = -ENOMEM; 135 136 in_mad = kzalloc(sizeof *in_mad, GFP_KERNEL); 137 out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL); 138 if (!in_mad || !out_mad) 139 goto out; 140 141 /* props being zeroed by the caller, avoid zeroing it here */ 142 143 init_query_mad(in_mad); 144 in_mad->attr_id = IB_SMP_ATTR_PORT_INFO; 145 in_mad->attr_mod = cpu_to_be32(port); 146 147 err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1, 148 port, NULL, NULL, in_mad, out_mad); 149 if (err) 150 goto out; 151 152 props->lid = be16_to_cpup((__be16 *) (out_mad->data + 16)); 153 props->lmc = out_mad->data[34] & 0x7; 154 props->sm_lid = be16_to_cpup((__be16 *) (out_mad->data + 18)); 155 props->sm_sl = out_mad->data[36] & 0xf; 156 props->state = out_mad->data[32] & 0xf; 157 props->phys_state = out_mad->data[33] >> 4; 158 props->port_cap_flags = be32_to_cpup((__be32 *) (out_mad->data + 20)); 159 props->gid_tbl_len = to_mdev(ibdev)->limits.gid_table_len; 160 props->max_msg_sz = 0x80000000; 161 props->pkey_tbl_len = to_mdev(ibdev)->limits.pkey_table_len; 162 props->bad_pkey_cntr = be16_to_cpup((__be16 *) (out_mad->data + 46)); 163 props->qkey_viol_cntr = be16_to_cpup((__be16 *) (out_mad->data + 48)); 164 props->active_width = out_mad->data[31] & 0xf; 165 props->active_speed = out_mad->data[35] >> 4; 166 props->max_mtu = out_mad->data[41] & 0xf; 167 props->active_mtu = out_mad->data[36] >> 4; 168 props->subnet_timeout = out_mad->data[51] & 0x1f; 169 props->max_vl_num = out_mad->data[37] >> 4; 170 props->init_type_reply = out_mad->data[41] >> 4; 171 172 out: 173 kfree(in_mad); 174 kfree(out_mad); 175 return err; 176 } 177 178 static int mthca_modify_device(struct ib_device *ibdev, 179 int mask, 180 struct ib_device_modify *props) 181 { 182 if (mask & ~IB_DEVICE_MODIFY_NODE_DESC) 183 return -EOPNOTSUPP; 184 185 if (mask & IB_DEVICE_MODIFY_NODE_DESC) { 186 if (mutex_lock_interruptible(&to_mdev(ibdev)->cap_mask_mutex)) 187 return -ERESTARTSYS; 188 memcpy(ibdev->node_desc, props->node_desc, 189 IB_DEVICE_NODE_DESC_MAX); 190 mutex_unlock(&to_mdev(ibdev)->cap_mask_mutex); 191 } 192 193 return 0; 194 } 195 196 static int mthca_modify_port(struct ib_device *ibdev, 197 u8 port, int port_modify_mask, 198 struct ib_port_modify *props) 199 { 200 struct mthca_set_ib_param set_ib; 201 struct ib_port_attr attr; 202 int err; 203 204 if (mutex_lock_interruptible(&to_mdev(ibdev)->cap_mask_mutex)) 205 return -ERESTARTSYS; 206 207 err = ib_query_port(ibdev, port, &attr); 208 if (err) 209 goto out; 210 211 set_ib.set_si_guid = 0; 212 set_ib.reset_qkey_viol = !!(port_modify_mask & IB_PORT_RESET_QKEY_CNTR); 213 214 set_ib.cap_mask = (attr.port_cap_flags | props->set_port_cap_mask) & 215 ~props->clr_port_cap_mask; 216 217 err = mthca_SET_IB(to_mdev(ibdev), &set_ib, port); 218 if (err) 219 goto out; 220 out: 221 mutex_unlock(&to_mdev(ibdev)->cap_mask_mutex); 222 return err; 223 } 224 225 static int mthca_query_pkey(struct ib_device *ibdev, 226 u8 port, u16 index, u16 *pkey) 227 { 228 struct ib_smp *in_mad = NULL; 229 struct ib_smp *out_mad = NULL; 230 int err = -ENOMEM; 231 232 in_mad = kzalloc(sizeof *in_mad, GFP_KERNEL); 233 out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL); 234 if (!in_mad || !out_mad) 235 goto out; 236 237 init_query_mad(in_mad); 238 in_mad->attr_id = IB_SMP_ATTR_PKEY_TABLE; 239 in_mad->attr_mod = cpu_to_be32(index / 32); 240 241 err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1, 242 port, NULL, NULL, in_mad, out_mad); 243 if (err) 244 goto out; 245 246 *pkey = be16_to_cpu(((__be16 *) out_mad->data)[index % 32]); 247 248 out: 249 kfree(in_mad); 250 kfree(out_mad); 251 return err; 252 } 253 254 static int mthca_query_gid(struct ib_device *ibdev, u8 port, 255 int index, union ib_gid *gid) 256 { 257 struct ib_smp *in_mad = NULL; 258 struct ib_smp *out_mad = NULL; 259 int err = -ENOMEM; 260 261 in_mad = kzalloc(sizeof *in_mad, GFP_KERNEL); 262 out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL); 263 if (!in_mad || !out_mad) 264 goto out; 265 266 init_query_mad(in_mad); 267 in_mad->attr_id = IB_SMP_ATTR_PORT_INFO; 268 in_mad->attr_mod = cpu_to_be32(port); 269 270 err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1, 271 port, NULL, NULL, in_mad, out_mad); 272 if (err) 273 goto out; 274 275 memcpy(gid->raw, out_mad->data + 8, 8); 276 277 init_query_mad(in_mad); 278 in_mad->attr_id = IB_SMP_ATTR_GUID_INFO; 279 in_mad->attr_mod = cpu_to_be32(index / 8); 280 281 err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1, 282 port, NULL, NULL, in_mad, out_mad); 283 if (err) 284 goto out; 285 286 memcpy(gid->raw + 8, out_mad->data + (index % 8) * 8, 8); 287 288 out: 289 kfree(in_mad); 290 kfree(out_mad); 291 return err; 292 } 293 294 static int mthca_alloc_ucontext(struct ib_ucontext *uctx, 295 struct ib_udata *udata) 296 { 297 struct ib_device *ibdev = uctx->device; 298 struct mthca_alloc_ucontext_resp uresp = {}; 299 struct mthca_ucontext *context = to_mucontext(uctx); 300 int err; 301 302 if (!(to_mdev(ibdev)->active)) 303 return -EAGAIN; 304 305 uresp.qp_tab_size = to_mdev(ibdev)->limits.num_qps; 306 if (mthca_is_memfree(to_mdev(ibdev))) 307 uresp.uarc_size = to_mdev(ibdev)->uar_table.uarc_size; 308 else 309 uresp.uarc_size = 0; 310 311 err = mthca_uar_alloc(to_mdev(ibdev), &context->uar); 312 if (err) 313 return err; 314 315 context->db_tab = mthca_init_user_db_tab(to_mdev(ibdev)); 316 if (IS_ERR(context->db_tab)) { 317 err = PTR_ERR(context->db_tab); 318 mthca_uar_free(to_mdev(ibdev), &context->uar); 319 return err; 320 } 321 322 if (ib_copy_to_udata(udata, &uresp, sizeof(uresp))) { 323 mthca_cleanup_user_db_tab(to_mdev(ibdev), &context->uar, context->db_tab); 324 mthca_uar_free(to_mdev(ibdev), &context->uar); 325 return -EFAULT; 326 } 327 328 context->reg_mr_warned = 0; 329 330 return 0; 331 } 332 333 static void mthca_dealloc_ucontext(struct ib_ucontext *context) 334 { 335 mthca_cleanup_user_db_tab(to_mdev(context->device), &to_mucontext(context)->uar, 336 to_mucontext(context)->db_tab); 337 mthca_uar_free(to_mdev(context->device), &to_mucontext(context)->uar); 338 } 339 340 static int mthca_mmap_uar(struct ib_ucontext *context, 341 struct vm_area_struct *vma) 342 { 343 if (vma->vm_end - vma->vm_start != PAGE_SIZE) 344 return -EINVAL; 345 346 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot); 347 348 if (io_remap_pfn_range(vma, vma->vm_start, 349 to_mucontext(context)->uar.pfn, 350 PAGE_SIZE, vma->vm_page_prot)) 351 return -EAGAIN; 352 353 return 0; 354 } 355 356 static int mthca_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata) 357 { 358 struct ib_device *ibdev = ibpd->device; 359 struct mthca_pd *pd = to_mpd(ibpd); 360 int err; 361 362 err = mthca_pd_alloc(to_mdev(ibdev), !udata, pd); 363 if (err) 364 return err; 365 366 if (udata) { 367 if (ib_copy_to_udata(udata, &pd->pd_num, sizeof (__u32))) { 368 mthca_pd_free(to_mdev(ibdev), pd); 369 return -EFAULT; 370 } 371 } 372 373 return 0; 374 } 375 376 static void mthca_dealloc_pd(struct ib_pd *pd, struct ib_udata *udata) 377 { 378 mthca_pd_free(to_mdev(pd->device), to_mpd(pd)); 379 } 380 381 static int mthca_ah_create(struct ib_ah *ibah, 382 struct rdma_ah_init_attr *init_attr, 383 struct ib_udata *udata) 384 385 { 386 struct mthca_ah *ah = to_mah(ibah); 387 388 return mthca_create_ah(to_mdev(ibah->device), to_mpd(ibah->pd), 389 init_attr->ah_attr, ah); 390 } 391 392 static void mthca_ah_destroy(struct ib_ah *ah, u32 flags) 393 { 394 mthca_destroy_ah(to_mdev(ah->device), to_mah(ah)); 395 } 396 397 static int mthca_create_srq(struct ib_srq *ibsrq, 398 struct ib_srq_init_attr *init_attr, 399 struct ib_udata *udata) 400 { 401 struct mthca_create_srq ucmd; 402 struct mthca_ucontext *context = rdma_udata_to_drv_context( 403 udata, struct mthca_ucontext, ibucontext); 404 struct mthca_srq *srq = to_msrq(ibsrq); 405 int err; 406 407 if (init_attr->srq_type != IB_SRQT_BASIC) 408 return -EOPNOTSUPP; 409 410 if (udata) { 411 if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd))) 412 return -EFAULT; 413 414 err = mthca_map_user_db(to_mdev(ibsrq->device), &context->uar, 415 context->db_tab, ucmd.db_index, 416 ucmd.db_page); 417 418 if (err) 419 return err; 420 421 srq->mr.ibmr.lkey = ucmd.lkey; 422 srq->db_index = ucmd.db_index; 423 } 424 425 err = mthca_alloc_srq(to_mdev(ibsrq->device), to_mpd(ibsrq->pd), 426 &init_attr->attr, srq, udata); 427 428 if (err && udata) 429 mthca_unmap_user_db(to_mdev(ibsrq->device), &context->uar, 430 context->db_tab, ucmd.db_index); 431 432 if (err) 433 return err; 434 435 if (context && ib_copy_to_udata(udata, &srq->srqn, sizeof(__u32))) { 436 mthca_free_srq(to_mdev(ibsrq->device), srq); 437 return -EFAULT; 438 } 439 440 return 0; 441 } 442 443 static void mthca_destroy_srq(struct ib_srq *srq, struct ib_udata *udata) 444 { 445 if (udata) { 446 struct mthca_ucontext *context = 447 rdma_udata_to_drv_context( 448 udata, 449 struct mthca_ucontext, 450 ibucontext); 451 452 mthca_unmap_user_db(to_mdev(srq->device), &context->uar, 453 context->db_tab, to_msrq(srq)->db_index); 454 } 455 456 mthca_free_srq(to_mdev(srq->device), to_msrq(srq)); 457 } 458 459 static struct ib_qp *mthca_create_qp(struct ib_pd *pd, 460 struct ib_qp_init_attr *init_attr, 461 struct ib_udata *udata) 462 { 463 struct mthca_ucontext *context = rdma_udata_to_drv_context( 464 udata, struct mthca_ucontext, ibucontext); 465 struct mthca_create_qp ucmd; 466 struct mthca_qp *qp; 467 int err; 468 469 if (init_attr->create_flags) 470 return ERR_PTR(-EINVAL); 471 472 switch (init_attr->qp_type) { 473 case IB_QPT_RC: 474 case IB_QPT_UC: 475 case IB_QPT_UD: 476 { 477 qp = kzalloc(sizeof(*qp), GFP_KERNEL); 478 if (!qp) 479 return ERR_PTR(-ENOMEM); 480 481 if (udata) { 482 if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) { 483 kfree(qp); 484 return ERR_PTR(-EFAULT); 485 } 486 487 err = mthca_map_user_db(to_mdev(pd->device), &context->uar, 488 context->db_tab, 489 ucmd.sq_db_index, ucmd.sq_db_page); 490 if (err) { 491 kfree(qp); 492 return ERR_PTR(err); 493 } 494 495 err = mthca_map_user_db(to_mdev(pd->device), &context->uar, 496 context->db_tab, 497 ucmd.rq_db_index, ucmd.rq_db_page); 498 if (err) { 499 mthca_unmap_user_db(to_mdev(pd->device), 500 &context->uar, 501 context->db_tab, 502 ucmd.sq_db_index); 503 kfree(qp); 504 return ERR_PTR(err); 505 } 506 507 qp->mr.ibmr.lkey = ucmd.lkey; 508 qp->sq.db_index = ucmd.sq_db_index; 509 qp->rq.db_index = ucmd.rq_db_index; 510 } 511 512 err = mthca_alloc_qp(to_mdev(pd->device), to_mpd(pd), 513 to_mcq(init_attr->send_cq), 514 to_mcq(init_attr->recv_cq), 515 init_attr->qp_type, init_attr->sq_sig_type, 516 &init_attr->cap, qp, udata); 517 518 if (err && udata) { 519 mthca_unmap_user_db(to_mdev(pd->device), 520 &context->uar, 521 context->db_tab, 522 ucmd.sq_db_index); 523 mthca_unmap_user_db(to_mdev(pd->device), 524 &context->uar, 525 context->db_tab, 526 ucmd.rq_db_index); 527 } 528 529 qp->ibqp.qp_num = qp->qpn; 530 break; 531 } 532 case IB_QPT_SMI: 533 case IB_QPT_GSI: 534 { 535 /* Don't allow userspace to create special QPs */ 536 if (udata) 537 return ERR_PTR(-EINVAL); 538 539 qp = kzalloc(sizeof(struct mthca_sqp), GFP_KERNEL); 540 if (!qp) 541 return ERR_PTR(-ENOMEM); 542 543 qp->ibqp.qp_num = init_attr->qp_type == IB_QPT_SMI ? 0 : 1; 544 545 err = mthca_alloc_sqp(to_mdev(pd->device), to_mpd(pd), 546 to_mcq(init_attr->send_cq), 547 to_mcq(init_attr->recv_cq), 548 init_attr->sq_sig_type, &init_attr->cap, 549 qp->ibqp.qp_num, init_attr->port_num, 550 to_msqp(qp), udata); 551 break; 552 } 553 default: 554 /* Don't support raw QPs */ 555 return ERR_PTR(-EOPNOTSUPP); 556 } 557 558 if (err) { 559 kfree(qp); 560 return ERR_PTR(err); 561 } 562 563 init_attr->cap.max_send_wr = qp->sq.max; 564 init_attr->cap.max_recv_wr = qp->rq.max; 565 init_attr->cap.max_send_sge = qp->sq.max_gs; 566 init_attr->cap.max_recv_sge = qp->rq.max_gs; 567 init_attr->cap.max_inline_data = qp->max_inline_data; 568 569 return &qp->ibqp; 570 } 571 572 static int mthca_destroy_qp(struct ib_qp *qp, struct ib_udata *udata) 573 { 574 if (udata) { 575 struct mthca_ucontext *context = 576 rdma_udata_to_drv_context( 577 udata, 578 struct mthca_ucontext, 579 ibucontext); 580 581 mthca_unmap_user_db(to_mdev(qp->device), 582 &context->uar, 583 context->db_tab, 584 to_mqp(qp)->sq.db_index); 585 mthca_unmap_user_db(to_mdev(qp->device), 586 &context->uar, 587 context->db_tab, 588 to_mqp(qp)->rq.db_index); 589 } 590 mthca_free_qp(to_mdev(qp->device), to_mqp(qp)); 591 kfree(qp); 592 return 0; 593 } 594 595 static int mthca_create_cq(struct ib_cq *ibcq, 596 const struct ib_cq_init_attr *attr, 597 struct ib_udata *udata) 598 { 599 struct ib_device *ibdev = ibcq->device; 600 int entries = attr->cqe; 601 struct mthca_create_cq ucmd; 602 struct mthca_cq *cq; 603 int nent; 604 int err; 605 struct mthca_ucontext *context = rdma_udata_to_drv_context( 606 udata, struct mthca_ucontext, ibucontext); 607 608 if (attr->flags) 609 return -EINVAL; 610 611 if (entries < 1 || entries > to_mdev(ibdev)->limits.max_cqes) 612 return -EINVAL; 613 614 if (udata) { 615 if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd))) 616 return -EFAULT; 617 618 err = mthca_map_user_db(to_mdev(ibdev), &context->uar, 619 context->db_tab, ucmd.set_db_index, 620 ucmd.set_db_page); 621 if (err) 622 return err; 623 624 err = mthca_map_user_db(to_mdev(ibdev), &context->uar, 625 context->db_tab, ucmd.arm_db_index, 626 ucmd.arm_db_page); 627 if (err) 628 goto err_unmap_set; 629 } 630 631 cq = to_mcq(ibcq); 632 633 if (udata) { 634 cq->buf.mr.ibmr.lkey = ucmd.lkey; 635 cq->set_ci_db_index = ucmd.set_db_index; 636 cq->arm_db_index = ucmd.arm_db_index; 637 } 638 639 for (nent = 1; nent <= entries; nent <<= 1) 640 ; /* nothing */ 641 642 err = mthca_init_cq(to_mdev(ibdev), nent, context, 643 udata ? ucmd.pdn : to_mdev(ibdev)->driver_pd.pd_num, 644 cq); 645 if (err) 646 goto err_unmap_arm; 647 648 if (udata && ib_copy_to_udata(udata, &cq->cqn, sizeof(__u32))) { 649 mthca_free_cq(to_mdev(ibdev), cq); 650 err = -EFAULT; 651 goto err_unmap_arm; 652 } 653 654 cq->resize_buf = NULL; 655 656 return 0; 657 658 err_unmap_arm: 659 if (udata) 660 mthca_unmap_user_db(to_mdev(ibdev), &context->uar, 661 context->db_tab, ucmd.arm_db_index); 662 663 err_unmap_set: 664 if (udata) 665 mthca_unmap_user_db(to_mdev(ibdev), &context->uar, 666 context->db_tab, ucmd.set_db_index); 667 668 return err; 669 } 670 671 static int mthca_alloc_resize_buf(struct mthca_dev *dev, struct mthca_cq *cq, 672 int entries) 673 { 674 int ret; 675 676 spin_lock_irq(&cq->lock); 677 if (cq->resize_buf) { 678 ret = -EBUSY; 679 goto unlock; 680 } 681 682 cq->resize_buf = kmalloc(sizeof *cq->resize_buf, GFP_ATOMIC); 683 if (!cq->resize_buf) { 684 ret = -ENOMEM; 685 goto unlock; 686 } 687 688 cq->resize_buf->state = CQ_RESIZE_ALLOC; 689 690 ret = 0; 691 692 unlock: 693 spin_unlock_irq(&cq->lock); 694 695 if (ret) 696 return ret; 697 698 ret = mthca_alloc_cq_buf(dev, &cq->resize_buf->buf, entries); 699 if (ret) { 700 spin_lock_irq(&cq->lock); 701 kfree(cq->resize_buf); 702 cq->resize_buf = NULL; 703 spin_unlock_irq(&cq->lock); 704 return ret; 705 } 706 707 cq->resize_buf->cqe = entries - 1; 708 709 spin_lock_irq(&cq->lock); 710 cq->resize_buf->state = CQ_RESIZE_READY; 711 spin_unlock_irq(&cq->lock); 712 713 return 0; 714 } 715 716 static int mthca_resize_cq(struct ib_cq *ibcq, int entries, struct ib_udata *udata) 717 { 718 struct mthca_dev *dev = to_mdev(ibcq->device); 719 struct mthca_cq *cq = to_mcq(ibcq); 720 struct mthca_resize_cq ucmd; 721 u32 lkey; 722 int ret; 723 724 if (entries < 1 || entries > dev->limits.max_cqes) 725 return -EINVAL; 726 727 mutex_lock(&cq->mutex); 728 729 entries = roundup_pow_of_two(entries + 1); 730 if (entries == ibcq->cqe + 1) { 731 ret = 0; 732 goto out; 733 } 734 735 if (cq->is_kernel) { 736 ret = mthca_alloc_resize_buf(dev, cq, entries); 737 if (ret) 738 goto out; 739 lkey = cq->resize_buf->buf.mr.ibmr.lkey; 740 } else { 741 if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) { 742 ret = -EFAULT; 743 goto out; 744 } 745 lkey = ucmd.lkey; 746 } 747 748 ret = mthca_RESIZE_CQ(dev, cq->cqn, lkey, ilog2(entries)); 749 750 if (ret) { 751 if (cq->resize_buf) { 752 mthca_free_cq_buf(dev, &cq->resize_buf->buf, 753 cq->resize_buf->cqe); 754 kfree(cq->resize_buf); 755 spin_lock_irq(&cq->lock); 756 cq->resize_buf = NULL; 757 spin_unlock_irq(&cq->lock); 758 } 759 goto out; 760 } 761 762 if (cq->is_kernel) { 763 struct mthca_cq_buf tbuf; 764 int tcqe; 765 766 spin_lock_irq(&cq->lock); 767 if (cq->resize_buf->state == CQ_RESIZE_READY) { 768 mthca_cq_resize_copy_cqes(cq); 769 tbuf = cq->buf; 770 tcqe = cq->ibcq.cqe; 771 cq->buf = cq->resize_buf->buf; 772 cq->ibcq.cqe = cq->resize_buf->cqe; 773 } else { 774 tbuf = cq->resize_buf->buf; 775 tcqe = cq->resize_buf->cqe; 776 } 777 778 kfree(cq->resize_buf); 779 cq->resize_buf = NULL; 780 spin_unlock_irq(&cq->lock); 781 782 mthca_free_cq_buf(dev, &tbuf, tcqe); 783 } else 784 ibcq->cqe = entries - 1; 785 786 out: 787 mutex_unlock(&cq->mutex); 788 789 return ret; 790 } 791 792 static void mthca_destroy_cq(struct ib_cq *cq, struct ib_udata *udata) 793 { 794 if (udata) { 795 struct mthca_ucontext *context = 796 rdma_udata_to_drv_context( 797 udata, 798 struct mthca_ucontext, 799 ibucontext); 800 801 mthca_unmap_user_db(to_mdev(cq->device), 802 &context->uar, 803 context->db_tab, 804 to_mcq(cq)->arm_db_index); 805 mthca_unmap_user_db(to_mdev(cq->device), 806 &context->uar, 807 context->db_tab, 808 to_mcq(cq)->set_ci_db_index); 809 } 810 mthca_free_cq(to_mdev(cq->device), to_mcq(cq)); 811 } 812 813 static inline u32 convert_access(int acc) 814 { 815 return (acc & IB_ACCESS_REMOTE_ATOMIC ? MTHCA_MPT_FLAG_ATOMIC : 0) | 816 (acc & IB_ACCESS_REMOTE_WRITE ? MTHCA_MPT_FLAG_REMOTE_WRITE : 0) | 817 (acc & IB_ACCESS_REMOTE_READ ? MTHCA_MPT_FLAG_REMOTE_READ : 0) | 818 (acc & IB_ACCESS_LOCAL_WRITE ? MTHCA_MPT_FLAG_LOCAL_WRITE : 0) | 819 MTHCA_MPT_FLAG_LOCAL_READ; 820 } 821 822 static struct ib_mr *mthca_get_dma_mr(struct ib_pd *pd, int acc) 823 { 824 struct mthca_mr *mr; 825 int err; 826 827 mr = kmalloc(sizeof *mr, GFP_KERNEL); 828 if (!mr) 829 return ERR_PTR(-ENOMEM); 830 831 err = mthca_mr_alloc_notrans(to_mdev(pd->device), 832 to_mpd(pd)->pd_num, 833 convert_access(acc), mr); 834 835 if (err) { 836 kfree(mr); 837 return ERR_PTR(err); 838 } 839 840 mr->umem = NULL; 841 842 return &mr->ibmr; 843 } 844 845 static struct ib_mr *mthca_reg_user_mr(struct ib_pd *pd, u64 start, u64 length, 846 u64 virt, int acc, struct ib_udata *udata) 847 { 848 struct mthca_dev *dev = to_mdev(pd->device); 849 struct sg_dma_page_iter sg_iter; 850 struct mthca_ucontext *context = rdma_udata_to_drv_context( 851 udata, struct mthca_ucontext, ibucontext); 852 struct mthca_mr *mr; 853 struct mthca_reg_mr ucmd; 854 u64 *pages; 855 int n, i; 856 int err = 0; 857 int write_mtt_size; 858 859 if (udata->inlen < sizeof ucmd) { 860 if (!context->reg_mr_warned) { 861 mthca_warn(dev, "Process '%s' did not pass in MR attrs.\n", 862 current->comm); 863 mthca_warn(dev, " Update libmthca to fix this.\n"); 864 } 865 ++context->reg_mr_warned; 866 ucmd.mr_attrs = 0; 867 } else if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) 868 return ERR_PTR(-EFAULT); 869 870 mr = kmalloc(sizeof *mr, GFP_KERNEL); 871 if (!mr) 872 return ERR_PTR(-ENOMEM); 873 874 mr->umem = ib_umem_get(pd->device, start, length, acc); 875 if (IS_ERR(mr->umem)) { 876 err = PTR_ERR(mr->umem); 877 goto err; 878 } 879 880 n = ib_umem_num_pages(mr->umem); 881 882 mr->mtt = mthca_alloc_mtt(dev, n); 883 if (IS_ERR(mr->mtt)) { 884 err = PTR_ERR(mr->mtt); 885 goto err_umem; 886 } 887 888 pages = (u64 *) __get_free_page(GFP_KERNEL); 889 if (!pages) { 890 err = -ENOMEM; 891 goto err_mtt; 892 } 893 894 i = n = 0; 895 896 write_mtt_size = min(mthca_write_mtt_size(dev), (int) (PAGE_SIZE / sizeof *pages)); 897 898 for_each_sg_dma_page(mr->umem->sg_head.sgl, &sg_iter, mr->umem->nmap, 0) { 899 pages[i++] = sg_page_iter_dma_address(&sg_iter); 900 901 /* 902 * Be friendly to write_mtt and pass it chunks 903 * of appropriate size. 904 */ 905 if (i == write_mtt_size) { 906 err = mthca_write_mtt(dev, mr->mtt, n, pages, i); 907 if (err) 908 goto mtt_done; 909 n += i; 910 i = 0; 911 } 912 } 913 914 if (i) 915 err = mthca_write_mtt(dev, mr->mtt, n, pages, i); 916 mtt_done: 917 free_page((unsigned long) pages); 918 if (err) 919 goto err_mtt; 920 921 err = mthca_mr_alloc(dev, to_mpd(pd)->pd_num, PAGE_SHIFT, virt, length, 922 convert_access(acc), mr); 923 924 if (err) 925 goto err_mtt; 926 927 return &mr->ibmr; 928 929 err_mtt: 930 mthca_free_mtt(dev, mr->mtt); 931 932 err_umem: 933 ib_umem_release(mr->umem); 934 935 err: 936 kfree(mr); 937 return ERR_PTR(err); 938 } 939 940 static int mthca_dereg_mr(struct ib_mr *mr, struct ib_udata *udata) 941 { 942 struct mthca_mr *mmr = to_mmr(mr); 943 944 mthca_free_mr(to_mdev(mr->device), mmr); 945 ib_umem_release(mmr->umem); 946 kfree(mmr); 947 948 return 0; 949 } 950 951 static ssize_t hw_rev_show(struct device *device, 952 struct device_attribute *attr, char *buf) 953 { 954 struct mthca_dev *dev = 955 rdma_device_to_drv_device(device, struct mthca_dev, ib_dev); 956 957 return sprintf(buf, "%x\n", dev->rev_id); 958 } 959 static DEVICE_ATTR_RO(hw_rev); 960 961 static ssize_t hca_type_show(struct device *device, 962 struct device_attribute *attr, char *buf) 963 { 964 struct mthca_dev *dev = 965 rdma_device_to_drv_device(device, struct mthca_dev, ib_dev); 966 967 switch (dev->pdev->device) { 968 case PCI_DEVICE_ID_MELLANOX_TAVOR: 969 return sprintf(buf, "MT23108\n"); 970 case PCI_DEVICE_ID_MELLANOX_ARBEL_COMPAT: 971 return sprintf(buf, "MT25208 (MT23108 compat mode)\n"); 972 case PCI_DEVICE_ID_MELLANOX_ARBEL: 973 return sprintf(buf, "MT25208\n"); 974 case PCI_DEVICE_ID_MELLANOX_SINAI: 975 case PCI_DEVICE_ID_MELLANOX_SINAI_OLD: 976 return sprintf(buf, "MT25204\n"); 977 default: 978 return sprintf(buf, "unknown\n"); 979 } 980 } 981 static DEVICE_ATTR_RO(hca_type); 982 983 static ssize_t board_id_show(struct device *device, 984 struct device_attribute *attr, char *buf) 985 { 986 struct mthca_dev *dev = 987 rdma_device_to_drv_device(device, struct mthca_dev, ib_dev); 988 989 return sprintf(buf, "%.*s\n", MTHCA_BOARD_ID_LEN, dev->board_id); 990 } 991 static DEVICE_ATTR_RO(board_id); 992 993 static struct attribute *mthca_dev_attributes[] = { 994 &dev_attr_hw_rev.attr, 995 &dev_attr_hca_type.attr, 996 &dev_attr_board_id.attr, 997 NULL 998 }; 999 1000 static const struct attribute_group mthca_attr_group = { 1001 .attrs = mthca_dev_attributes, 1002 }; 1003 1004 static int mthca_init_node_data(struct mthca_dev *dev) 1005 { 1006 struct ib_smp *in_mad = NULL; 1007 struct ib_smp *out_mad = NULL; 1008 int err = -ENOMEM; 1009 1010 in_mad = kzalloc(sizeof *in_mad, GFP_KERNEL); 1011 out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL); 1012 if (!in_mad || !out_mad) 1013 goto out; 1014 1015 init_query_mad(in_mad); 1016 in_mad->attr_id = IB_SMP_ATTR_NODE_DESC; 1017 1018 err = mthca_MAD_IFC(dev, 1, 1, 1019 1, NULL, NULL, in_mad, out_mad); 1020 if (err) 1021 goto out; 1022 1023 memcpy(dev->ib_dev.node_desc, out_mad->data, IB_DEVICE_NODE_DESC_MAX); 1024 1025 in_mad->attr_id = IB_SMP_ATTR_NODE_INFO; 1026 1027 err = mthca_MAD_IFC(dev, 1, 1, 1028 1, NULL, NULL, in_mad, out_mad); 1029 if (err) 1030 goto out; 1031 1032 if (mthca_is_memfree(dev)) 1033 dev->rev_id = be32_to_cpup((__be32 *) (out_mad->data + 32)); 1034 memcpy(&dev->ib_dev.node_guid, out_mad->data + 12, 8); 1035 1036 out: 1037 kfree(in_mad); 1038 kfree(out_mad); 1039 return err; 1040 } 1041 1042 static int mthca_port_immutable(struct ib_device *ibdev, u8 port_num, 1043 struct ib_port_immutable *immutable) 1044 { 1045 struct ib_port_attr attr; 1046 int err; 1047 1048 immutable->core_cap_flags = RDMA_CORE_PORT_IBA_IB; 1049 1050 err = ib_query_port(ibdev, port_num, &attr); 1051 if (err) 1052 return err; 1053 1054 immutable->pkey_tbl_len = attr.pkey_tbl_len; 1055 immutable->gid_tbl_len = attr.gid_tbl_len; 1056 immutable->max_mad_size = IB_MGMT_MAD_SIZE; 1057 1058 return 0; 1059 } 1060 1061 static void get_dev_fw_str(struct ib_device *device, char *str) 1062 { 1063 struct mthca_dev *dev = 1064 container_of(device, struct mthca_dev, ib_dev); 1065 snprintf(str, IB_FW_VERSION_NAME_MAX, "%d.%d.%d", 1066 (int) (dev->fw_ver >> 32), 1067 (int) (dev->fw_ver >> 16) & 0xffff, 1068 (int) dev->fw_ver & 0xffff); 1069 } 1070 1071 static const struct ib_device_ops mthca_dev_ops = { 1072 .owner = THIS_MODULE, 1073 .driver_id = RDMA_DRIVER_MTHCA, 1074 .uverbs_abi_ver = MTHCA_UVERBS_ABI_VERSION, 1075 .uverbs_no_driver_id_binding = 1, 1076 1077 .alloc_pd = mthca_alloc_pd, 1078 .alloc_ucontext = mthca_alloc_ucontext, 1079 .attach_mcast = mthca_multicast_attach, 1080 .create_ah = mthca_ah_create, 1081 .create_cq = mthca_create_cq, 1082 .create_qp = mthca_create_qp, 1083 .dealloc_pd = mthca_dealloc_pd, 1084 .dealloc_ucontext = mthca_dealloc_ucontext, 1085 .dereg_mr = mthca_dereg_mr, 1086 .destroy_ah = mthca_ah_destroy, 1087 .destroy_cq = mthca_destroy_cq, 1088 .destroy_qp = mthca_destroy_qp, 1089 .detach_mcast = mthca_multicast_detach, 1090 .get_dev_fw_str = get_dev_fw_str, 1091 .get_dma_mr = mthca_get_dma_mr, 1092 .get_port_immutable = mthca_port_immutable, 1093 .mmap = mthca_mmap_uar, 1094 .modify_device = mthca_modify_device, 1095 .modify_port = mthca_modify_port, 1096 .modify_qp = mthca_modify_qp, 1097 .poll_cq = mthca_poll_cq, 1098 .process_mad = mthca_process_mad, 1099 .query_ah = mthca_ah_query, 1100 .query_device = mthca_query_device, 1101 .query_gid = mthca_query_gid, 1102 .query_pkey = mthca_query_pkey, 1103 .query_port = mthca_query_port, 1104 .query_qp = mthca_query_qp, 1105 .reg_user_mr = mthca_reg_user_mr, 1106 .resize_cq = mthca_resize_cq, 1107 1108 INIT_RDMA_OBJ_SIZE(ib_ah, mthca_ah, ibah), 1109 INIT_RDMA_OBJ_SIZE(ib_cq, mthca_cq, ibcq), 1110 INIT_RDMA_OBJ_SIZE(ib_pd, mthca_pd, ibpd), 1111 INIT_RDMA_OBJ_SIZE(ib_ucontext, mthca_ucontext, ibucontext), 1112 }; 1113 1114 static const struct ib_device_ops mthca_dev_arbel_srq_ops = { 1115 .create_srq = mthca_create_srq, 1116 .destroy_srq = mthca_destroy_srq, 1117 .modify_srq = mthca_modify_srq, 1118 .post_srq_recv = mthca_arbel_post_srq_recv, 1119 .query_srq = mthca_query_srq, 1120 1121 INIT_RDMA_OBJ_SIZE(ib_srq, mthca_srq, ibsrq), 1122 }; 1123 1124 static const struct ib_device_ops mthca_dev_tavor_srq_ops = { 1125 .create_srq = mthca_create_srq, 1126 .destroy_srq = mthca_destroy_srq, 1127 .modify_srq = mthca_modify_srq, 1128 .post_srq_recv = mthca_tavor_post_srq_recv, 1129 .query_srq = mthca_query_srq, 1130 1131 INIT_RDMA_OBJ_SIZE(ib_srq, mthca_srq, ibsrq), 1132 }; 1133 1134 static const struct ib_device_ops mthca_dev_arbel_ops = { 1135 .post_recv = mthca_arbel_post_receive, 1136 .post_send = mthca_arbel_post_send, 1137 .req_notify_cq = mthca_arbel_arm_cq, 1138 }; 1139 1140 static const struct ib_device_ops mthca_dev_tavor_ops = { 1141 .post_recv = mthca_tavor_post_receive, 1142 .post_send = mthca_tavor_post_send, 1143 .req_notify_cq = mthca_tavor_arm_cq, 1144 }; 1145 1146 int mthca_register_device(struct mthca_dev *dev) 1147 { 1148 int ret; 1149 1150 ret = mthca_init_node_data(dev); 1151 if (ret) 1152 return ret; 1153 1154 dev->ib_dev.uverbs_cmd_mask = 1155 (1ull << IB_USER_VERBS_CMD_GET_CONTEXT) | 1156 (1ull << IB_USER_VERBS_CMD_QUERY_DEVICE) | 1157 (1ull << IB_USER_VERBS_CMD_QUERY_PORT) | 1158 (1ull << IB_USER_VERBS_CMD_ALLOC_PD) | 1159 (1ull << IB_USER_VERBS_CMD_DEALLOC_PD) | 1160 (1ull << IB_USER_VERBS_CMD_REG_MR) | 1161 (1ull << IB_USER_VERBS_CMD_DEREG_MR) | 1162 (1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL) | 1163 (1ull << IB_USER_VERBS_CMD_CREATE_CQ) | 1164 (1ull << IB_USER_VERBS_CMD_RESIZE_CQ) | 1165 (1ull << IB_USER_VERBS_CMD_DESTROY_CQ) | 1166 (1ull << IB_USER_VERBS_CMD_CREATE_QP) | 1167 (1ull << IB_USER_VERBS_CMD_QUERY_QP) | 1168 (1ull << IB_USER_VERBS_CMD_MODIFY_QP) | 1169 (1ull << IB_USER_VERBS_CMD_DESTROY_QP) | 1170 (1ull << IB_USER_VERBS_CMD_ATTACH_MCAST) | 1171 (1ull << IB_USER_VERBS_CMD_DETACH_MCAST); 1172 dev->ib_dev.node_type = RDMA_NODE_IB_CA; 1173 dev->ib_dev.phys_port_cnt = dev->limits.num_ports; 1174 dev->ib_dev.num_comp_vectors = 1; 1175 dev->ib_dev.dev.parent = &dev->pdev->dev; 1176 1177 if (dev->mthca_flags & MTHCA_FLAG_SRQ) { 1178 dev->ib_dev.uverbs_cmd_mask |= 1179 (1ull << IB_USER_VERBS_CMD_CREATE_SRQ) | 1180 (1ull << IB_USER_VERBS_CMD_MODIFY_SRQ) | 1181 (1ull << IB_USER_VERBS_CMD_QUERY_SRQ) | 1182 (1ull << IB_USER_VERBS_CMD_DESTROY_SRQ); 1183 1184 if (mthca_is_memfree(dev)) 1185 ib_set_device_ops(&dev->ib_dev, 1186 &mthca_dev_arbel_srq_ops); 1187 else 1188 ib_set_device_ops(&dev->ib_dev, 1189 &mthca_dev_tavor_srq_ops); 1190 } 1191 1192 ib_set_device_ops(&dev->ib_dev, &mthca_dev_ops); 1193 1194 if (mthca_is_memfree(dev)) 1195 ib_set_device_ops(&dev->ib_dev, &mthca_dev_arbel_ops); 1196 else 1197 ib_set_device_ops(&dev->ib_dev, &mthca_dev_tavor_ops); 1198 1199 mutex_init(&dev->cap_mask_mutex); 1200 1201 rdma_set_device_sysfs_group(&dev->ib_dev, &mthca_attr_group); 1202 ret = ib_register_device(&dev->ib_dev, "mthca%d"); 1203 if (ret) 1204 return ret; 1205 1206 mthca_start_catas_poll(dev); 1207 1208 return 0; 1209 } 1210 1211 void mthca_unregister_device(struct mthca_dev *dev) 1212 { 1213 mthca_stop_catas_poll(dev); 1214 ib_unregister_device(&dev->ib_dev); 1215 } 1216