1 /* 2 * Copyright (c) 2004, 2005, 2006 Voltaire, Inc. All rights reserved. 3 * Copyright (c) 2013-2014 Mellanox Technologies. All rights reserved. 4 * 5 * This software is available to you under a choice of one of two 6 * licenses. You may choose to be licensed under the terms of the GNU 7 * General Public License (GPL) Version 2, available from the file 8 * COPYING in the main directory of this source tree, or the 9 * OpenIB.org BSD license below: 10 * 11 * Redistribution and use in source and binary forms, with or 12 * without modification, are permitted provided that the following 13 * conditions are met: 14 * 15 * - Redistributions of source code must retain the above 16 * copyright notice, this list of conditions and the following 17 * disclaimer. 18 * 19 * - Redistributions in binary form must reproduce the above 20 * copyright notice, this list of conditions and the following 21 * disclaimer in the documentation and/or other materials 22 * provided with the distribution. 23 * 24 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, 25 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF 26 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND 27 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS 28 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN 29 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN 30 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 31 * SOFTWARE. 32 */ 33 #include <linux/module.h> 34 #include <linux/kernel.h> 35 #include <linux/slab.h> 36 #include <linux/mm.h> 37 #include <linux/highmem.h> 38 #include <linux/scatterlist.h> 39 40 #include "iscsi_iser.h" 41 static 42 int iser_fast_reg_fmr(struct iscsi_iser_task *iser_task, 43 struct iser_data_buf *mem, 44 struct iser_reg_resources *rsc, 45 struct iser_mem_reg *mem_reg); 46 static 47 int iser_fast_reg_mr(struct iscsi_iser_task *iser_task, 48 struct iser_data_buf *mem, 49 struct iser_reg_resources *rsc, 50 struct iser_mem_reg *mem_reg); 51 52 static const struct iser_reg_ops fastreg_ops = { 53 .alloc_reg_res = iser_alloc_fastreg_pool, 54 .free_reg_res = iser_free_fastreg_pool, 55 .reg_mem = iser_fast_reg_mr, 56 .unreg_mem = iser_unreg_mem_fastreg, 57 .reg_desc_get = iser_reg_desc_get_fr, 58 .reg_desc_put = iser_reg_desc_put_fr, 59 }; 60 61 static const struct iser_reg_ops fmr_ops = { 62 .alloc_reg_res = iser_alloc_fmr_pool, 63 .free_reg_res = iser_free_fmr_pool, 64 .reg_mem = iser_fast_reg_fmr, 65 .unreg_mem = iser_unreg_mem_fmr, 66 .reg_desc_get = iser_reg_desc_get_fmr, 67 .reg_desc_put = iser_reg_desc_put_fmr, 68 }; 69 70 void iser_reg_comp(struct ib_cq *cq, struct ib_wc *wc) 71 { 72 iser_err_comp(wc, "memreg"); 73 } 74 75 int iser_assign_reg_ops(struct iser_device *device) 76 { 77 struct ib_device *ib_dev = device->ib_device; 78 79 /* Assign function handles - based on FMR support */ 80 if (ib_dev->ops.alloc_fmr && ib_dev->ops.dealloc_fmr && 81 ib_dev->ops.map_phys_fmr && ib_dev->ops.unmap_fmr) { 82 iser_info("FMR supported, using FMR for registration\n"); 83 device->reg_ops = &fmr_ops; 84 } else if (ib_dev->attrs.device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) { 85 iser_info("FastReg supported, using FastReg for registration\n"); 86 device->reg_ops = &fastreg_ops; 87 device->remote_inv_sup = iser_always_reg; 88 } else { 89 iser_err("IB device does not support FMRs nor FastRegs, can't register memory\n"); 90 return -1; 91 } 92 93 return 0; 94 } 95 96 struct iser_fr_desc * 97 iser_reg_desc_get_fr(struct ib_conn *ib_conn) 98 { 99 struct iser_fr_pool *fr_pool = &ib_conn->fr_pool; 100 struct iser_fr_desc *desc; 101 unsigned long flags; 102 103 spin_lock_irqsave(&fr_pool->lock, flags); 104 desc = list_first_entry(&fr_pool->list, 105 struct iser_fr_desc, list); 106 list_del(&desc->list); 107 spin_unlock_irqrestore(&fr_pool->lock, flags); 108 109 return desc; 110 } 111 112 void 113 iser_reg_desc_put_fr(struct ib_conn *ib_conn, 114 struct iser_fr_desc *desc) 115 { 116 struct iser_fr_pool *fr_pool = &ib_conn->fr_pool; 117 unsigned long flags; 118 119 spin_lock_irqsave(&fr_pool->lock, flags); 120 list_add(&desc->list, &fr_pool->list); 121 spin_unlock_irqrestore(&fr_pool->lock, flags); 122 } 123 124 struct iser_fr_desc * 125 iser_reg_desc_get_fmr(struct ib_conn *ib_conn) 126 { 127 struct iser_fr_pool *fr_pool = &ib_conn->fr_pool; 128 129 return list_first_entry(&fr_pool->list, 130 struct iser_fr_desc, list); 131 } 132 133 void 134 iser_reg_desc_put_fmr(struct ib_conn *ib_conn, 135 struct iser_fr_desc *desc) 136 { 137 } 138 139 static void iser_data_buf_dump(struct iser_data_buf *data, 140 struct ib_device *ibdev) 141 { 142 struct scatterlist *sg; 143 int i; 144 145 for_each_sg(data->sg, sg, data->dma_nents, i) 146 iser_dbg("sg[%d] dma_addr:0x%lX page:0x%p " 147 "off:0x%x sz:0x%x dma_len:0x%x\n", 148 i, (unsigned long)sg_dma_address(sg), 149 sg_page(sg), sg->offset, sg->length, sg_dma_len(sg)); 150 } 151 152 static void iser_dump_page_vec(struct iser_page_vec *page_vec) 153 { 154 int i; 155 156 iser_err("page vec npages %d data length %lld\n", 157 page_vec->npages, page_vec->fake_mr.length); 158 for (i = 0; i < page_vec->npages; i++) 159 iser_err("vec[%d]: %llx\n", i, page_vec->pages[i]); 160 } 161 162 int iser_dma_map_task_data(struct iscsi_iser_task *iser_task, 163 struct iser_data_buf *data, 164 enum iser_data_dir iser_dir, 165 enum dma_data_direction dma_dir) 166 { 167 struct ib_device *dev; 168 169 iser_task->dir[iser_dir] = 1; 170 dev = iser_task->iser_conn->ib_conn.device->ib_device; 171 172 data->dma_nents = ib_dma_map_sg(dev, data->sg, data->size, dma_dir); 173 if (unlikely(data->dma_nents == 0)) { 174 iser_err("dma_map_sg failed!!!\n"); 175 return -EINVAL; 176 } 177 return 0; 178 } 179 180 void iser_dma_unmap_task_data(struct iscsi_iser_task *iser_task, 181 struct iser_data_buf *data, 182 enum dma_data_direction dir) 183 { 184 struct ib_device *dev; 185 186 dev = iser_task->iser_conn->ib_conn.device->ib_device; 187 ib_dma_unmap_sg(dev, data->sg, data->size, dir); 188 } 189 190 static int 191 iser_reg_dma(struct iser_device *device, struct iser_data_buf *mem, 192 struct iser_mem_reg *reg) 193 { 194 struct scatterlist *sg = mem->sg; 195 196 reg->sge.lkey = device->pd->local_dma_lkey; 197 /* 198 * FIXME: rework the registration code path to differentiate 199 * rkey/lkey use cases 200 */ 201 202 if (device->pd->flags & IB_PD_UNSAFE_GLOBAL_RKEY) 203 reg->rkey = device->pd->unsafe_global_rkey; 204 else 205 reg->rkey = 0; 206 reg->sge.addr = sg_dma_address(&sg[0]); 207 reg->sge.length = sg_dma_len(&sg[0]); 208 209 iser_dbg("Single DMA entry: lkey=0x%x, rkey=0x%x, addr=0x%llx," 210 " length=0x%x\n", reg->sge.lkey, reg->rkey, 211 reg->sge.addr, reg->sge.length); 212 213 return 0; 214 } 215 216 static int iser_set_page(struct ib_mr *mr, u64 addr) 217 { 218 struct iser_page_vec *page_vec = 219 container_of(mr, struct iser_page_vec, fake_mr); 220 221 page_vec->pages[page_vec->npages++] = addr; 222 223 return 0; 224 } 225 226 static 227 int iser_fast_reg_fmr(struct iscsi_iser_task *iser_task, 228 struct iser_data_buf *mem, 229 struct iser_reg_resources *rsc, 230 struct iser_mem_reg *reg) 231 { 232 struct ib_conn *ib_conn = &iser_task->iser_conn->ib_conn; 233 struct iser_device *device = ib_conn->device; 234 struct iser_page_vec *page_vec = rsc->page_vec; 235 struct ib_fmr_pool *fmr_pool = rsc->fmr_pool; 236 struct ib_pool_fmr *fmr; 237 int ret, plen; 238 239 page_vec->npages = 0; 240 page_vec->fake_mr.page_size = SZ_4K; 241 plen = ib_sg_to_pages(&page_vec->fake_mr, mem->sg, 242 mem->dma_nents, NULL, iser_set_page); 243 if (unlikely(plen < mem->dma_nents)) { 244 iser_err("page vec too short to hold this SG\n"); 245 iser_data_buf_dump(mem, device->ib_device); 246 iser_dump_page_vec(page_vec); 247 return -EINVAL; 248 } 249 250 fmr = ib_fmr_pool_map_phys(fmr_pool, page_vec->pages, 251 page_vec->npages, page_vec->pages[0]); 252 if (IS_ERR(fmr)) { 253 ret = PTR_ERR(fmr); 254 iser_err("ib_fmr_pool_map_phys failed: %d\n", ret); 255 return ret; 256 } 257 258 reg->sge.lkey = fmr->fmr->lkey; 259 reg->rkey = fmr->fmr->rkey; 260 reg->sge.addr = page_vec->fake_mr.iova; 261 reg->sge.length = page_vec->fake_mr.length; 262 reg->mem_h = fmr; 263 264 iser_dbg("fmr reg: lkey=0x%x, rkey=0x%x, addr=0x%llx," 265 " length=0x%x\n", reg->sge.lkey, reg->rkey, 266 reg->sge.addr, reg->sge.length); 267 268 return 0; 269 } 270 271 /** 272 * Unregister (previosuly registered using FMR) memory. 273 * If memory is non-FMR does nothing. 274 */ 275 void iser_unreg_mem_fmr(struct iscsi_iser_task *iser_task, 276 enum iser_data_dir cmd_dir) 277 { 278 struct iser_mem_reg *reg = &iser_task->rdma_reg[cmd_dir]; 279 280 if (!reg->mem_h) 281 return; 282 283 iser_dbg("PHYSICAL Mem.Unregister mem_h %p\n", reg->mem_h); 284 285 ib_fmr_pool_unmap((struct ib_pool_fmr *)reg->mem_h); 286 287 reg->mem_h = NULL; 288 } 289 290 void iser_unreg_mem_fastreg(struct iscsi_iser_task *iser_task, 291 enum iser_data_dir cmd_dir) 292 { 293 struct iser_device *device = iser_task->iser_conn->ib_conn.device; 294 struct iser_mem_reg *reg = &iser_task->rdma_reg[cmd_dir]; 295 struct iser_fr_desc *desc; 296 struct ib_mr_status mr_status; 297 298 desc = reg->mem_h; 299 if (!desc) 300 return; 301 302 /* 303 * The signature MR cannot be invalidated and reused without checking. 304 * libiscsi calls the check_protection transport handler only if 305 * SCSI-Response is received. And the signature MR is not checked if 306 * the task is completed for some other reason like a timeout or error 307 * handling. That's why we must check the signature MR here before 308 * putting it to the free pool. 309 */ 310 if (unlikely(desc->sig_protected)) { 311 desc->sig_protected = false; 312 ib_check_mr_status(desc->rsc.sig_mr, IB_MR_CHECK_SIG_STATUS, 313 &mr_status); 314 } 315 device->reg_ops->reg_desc_put(&iser_task->iser_conn->ib_conn, desc); 316 reg->mem_h = NULL; 317 } 318 319 static void 320 iser_set_dif_domain(struct scsi_cmnd *sc, struct ib_sig_domain *domain) 321 { 322 domain->sig_type = IB_SIG_TYPE_T10_DIF; 323 domain->sig.dif.pi_interval = scsi_prot_interval(sc); 324 domain->sig.dif.ref_tag = t10_pi_ref_tag(sc->request); 325 /* 326 * At the moment we hard code those, but in the future 327 * we will take them from sc. 328 */ 329 domain->sig.dif.apptag_check_mask = 0xffff; 330 domain->sig.dif.app_escape = true; 331 domain->sig.dif.ref_escape = true; 332 if (sc->prot_flags & SCSI_PROT_REF_INCREMENT) 333 domain->sig.dif.ref_remap = true; 334 }; 335 336 static int 337 iser_set_sig_attrs(struct scsi_cmnd *sc, struct ib_sig_attrs *sig_attrs) 338 { 339 switch (scsi_get_prot_op(sc)) { 340 case SCSI_PROT_WRITE_INSERT: 341 case SCSI_PROT_READ_STRIP: 342 sig_attrs->mem.sig_type = IB_SIG_TYPE_NONE; 343 iser_set_dif_domain(sc, &sig_attrs->wire); 344 sig_attrs->wire.sig.dif.bg_type = IB_T10DIF_CRC; 345 break; 346 case SCSI_PROT_READ_INSERT: 347 case SCSI_PROT_WRITE_STRIP: 348 sig_attrs->wire.sig_type = IB_SIG_TYPE_NONE; 349 iser_set_dif_domain(sc, &sig_attrs->mem); 350 sig_attrs->mem.sig.dif.bg_type = sc->prot_flags & SCSI_PROT_IP_CHECKSUM ? 351 IB_T10DIF_CSUM : IB_T10DIF_CRC; 352 break; 353 case SCSI_PROT_READ_PASS: 354 case SCSI_PROT_WRITE_PASS: 355 iser_set_dif_domain(sc, &sig_attrs->wire); 356 sig_attrs->wire.sig.dif.bg_type = IB_T10DIF_CRC; 357 iser_set_dif_domain(sc, &sig_attrs->mem); 358 sig_attrs->mem.sig.dif.bg_type = sc->prot_flags & SCSI_PROT_IP_CHECKSUM ? 359 IB_T10DIF_CSUM : IB_T10DIF_CRC; 360 break; 361 default: 362 iser_err("Unsupported PI operation %d\n", 363 scsi_get_prot_op(sc)); 364 return -EINVAL; 365 } 366 367 return 0; 368 } 369 370 static inline void 371 iser_set_prot_checks(struct scsi_cmnd *sc, u8 *mask) 372 { 373 *mask = 0; 374 if (sc->prot_flags & SCSI_PROT_REF_CHECK) 375 *mask |= IB_SIG_CHECK_REFTAG; 376 if (sc->prot_flags & SCSI_PROT_GUARD_CHECK) 377 *mask |= IB_SIG_CHECK_GUARD; 378 } 379 380 static inline void 381 iser_inv_rkey(struct ib_send_wr *inv_wr, 382 struct ib_mr *mr, 383 struct ib_cqe *cqe, 384 struct ib_send_wr *next_wr) 385 { 386 inv_wr->opcode = IB_WR_LOCAL_INV; 387 inv_wr->wr_cqe = cqe; 388 inv_wr->ex.invalidate_rkey = mr->rkey; 389 inv_wr->send_flags = 0; 390 inv_wr->num_sge = 0; 391 inv_wr->next = next_wr; 392 } 393 394 static int 395 iser_reg_sig_mr(struct iscsi_iser_task *iser_task, 396 struct iser_data_buf *mem, 397 struct iser_data_buf *sig_mem, 398 struct iser_reg_resources *rsc, 399 struct iser_mem_reg *sig_reg) 400 { 401 struct iser_tx_desc *tx_desc = &iser_task->desc; 402 struct ib_cqe *cqe = &iser_task->iser_conn->ib_conn.reg_cqe; 403 struct ib_mr *mr = rsc->sig_mr; 404 struct ib_sig_attrs *sig_attrs = mr->sig_attrs; 405 struct ib_reg_wr *wr = &tx_desc->reg_wr; 406 int ret; 407 408 memset(sig_attrs, 0, sizeof(*sig_attrs)); 409 ret = iser_set_sig_attrs(iser_task->sc, sig_attrs); 410 if (ret) 411 goto err; 412 413 iser_set_prot_checks(iser_task->sc, &sig_attrs->check_mask); 414 415 if (rsc->mr_valid) 416 iser_inv_rkey(&tx_desc->inv_wr, mr, cqe, &wr->wr); 417 418 ib_update_fast_reg_key(mr, ib_inc_rkey(mr->rkey)); 419 420 ret = ib_map_mr_sg_pi(mr, mem->sg, mem->dma_nents, NULL, 421 sig_mem->sg, sig_mem->dma_nents, NULL, SZ_4K); 422 if (unlikely(ret)) { 423 iser_err("failed to map PI sg (%d)\n", 424 mem->dma_nents + sig_mem->dma_nents); 425 goto err; 426 } 427 428 memset(wr, 0, sizeof(*wr)); 429 wr->wr.next = &tx_desc->send_wr; 430 wr->wr.opcode = IB_WR_REG_MR_INTEGRITY; 431 wr->wr.wr_cqe = cqe; 432 wr->wr.num_sge = 0; 433 wr->wr.send_flags = 0; 434 wr->mr = mr; 435 wr->key = mr->rkey; 436 wr->access = IB_ACCESS_LOCAL_WRITE | 437 IB_ACCESS_REMOTE_READ | 438 IB_ACCESS_REMOTE_WRITE; 439 rsc->mr_valid = 1; 440 441 sig_reg->sge.lkey = mr->lkey; 442 sig_reg->rkey = mr->rkey; 443 sig_reg->sge.addr = mr->iova; 444 sig_reg->sge.length = mr->length; 445 446 iser_dbg("lkey=0x%x rkey=0x%x addr=0x%llx length=%u\n", 447 sig_reg->sge.lkey, sig_reg->rkey, sig_reg->sge.addr, 448 sig_reg->sge.length); 449 err: 450 return ret; 451 } 452 453 static int iser_fast_reg_mr(struct iscsi_iser_task *iser_task, 454 struct iser_data_buf *mem, 455 struct iser_reg_resources *rsc, 456 struct iser_mem_reg *reg) 457 { 458 struct iser_tx_desc *tx_desc = &iser_task->desc; 459 struct ib_cqe *cqe = &iser_task->iser_conn->ib_conn.reg_cqe; 460 struct ib_mr *mr = rsc->mr; 461 struct ib_reg_wr *wr = &tx_desc->reg_wr; 462 int n; 463 464 if (rsc->mr_valid) 465 iser_inv_rkey(&tx_desc->inv_wr, mr, cqe, &wr->wr); 466 467 ib_update_fast_reg_key(mr, ib_inc_rkey(mr->rkey)); 468 469 n = ib_map_mr_sg(mr, mem->sg, mem->dma_nents, NULL, SZ_4K); 470 if (unlikely(n != mem->dma_nents)) { 471 iser_err("failed to map sg (%d/%d)\n", 472 n, mem->dma_nents); 473 return n < 0 ? n : -EINVAL; 474 } 475 476 wr->wr.next = &tx_desc->send_wr; 477 wr->wr.opcode = IB_WR_REG_MR; 478 wr->wr.wr_cqe = cqe; 479 wr->wr.send_flags = 0; 480 wr->wr.num_sge = 0; 481 wr->mr = mr; 482 wr->key = mr->rkey; 483 wr->access = IB_ACCESS_LOCAL_WRITE | 484 IB_ACCESS_REMOTE_WRITE | 485 IB_ACCESS_REMOTE_READ; 486 487 rsc->mr_valid = 1; 488 489 reg->sge.lkey = mr->lkey; 490 reg->rkey = mr->rkey; 491 reg->sge.addr = mr->iova; 492 reg->sge.length = mr->length; 493 494 iser_dbg("lkey=0x%x rkey=0x%x addr=0x%llx length=0x%x\n", 495 reg->sge.lkey, reg->rkey, reg->sge.addr, reg->sge.length); 496 497 return 0; 498 } 499 500 static int 501 iser_reg_data_sg(struct iscsi_iser_task *task, 502 struct iser_data_buf *mem, 503 struct iser_fr_desc *desc, 504 bool use_dma_key, 505 struct iser_mem_reg *reg) 506 { 507 struct iser_device *device = task->iser_conn->ib_conn.device; 508 509 if (use_dma_key) 510 return iser_reg_dma(device, mem, reg); 511 512 return device->reg_ops->reg_mem(task, mem, &desc->rsc, reg); 513 } 514 515 int iser_reg_rdma_mem(struct iscsi_iser_task *task, 516 enum iser_data_dir dir, 517 bool all_imm) 518 { 519 struct ib_conn *ib_conn = &task->iser_conn->ib_conn; 520 struct iser_device *device = ib_conn->device; 521 struct iser_data_buf *mem = &task->data[dir]; 522 struct iser_mem_reg *reg = &task->rdma_reg[dir]; 523 struct iser_fr_desc *desc = NULL; 524 bool use_dma_key; 525 int err; 526 527 use_dma_key = mem->dma_nents == 1 && (all_imm || !iser_always_reg) && 528 scsi_get_prot_op(task->sc) == SCSI_PROT_NORMAL; 529 530 if (!use_dma_key) { 531 desc = device->reg_ops->reg_desc_get(ib_conn); 532 reg->mem_h = desc; 533 } 534 535 if (scsi_get_prot_op(task->sc) == SCSI_PROT_NORMAL) { 536 err = iser_reg_data_sg(task, mem, desc, use_dma_key, reg); 537 if (unlikely(err)) 538 goto err_reg; 539 } else { 540 err = iser_reg_sig_mr(task, mem, &task->prot[dir], 541 &desc->rsc, reg); 542 if (unlikely(err)) 543 goto err_reg; 544 545 desc->sig_protected = true; 546 } 547 548 return 0; 549 550 err_reg: 551 if (desc) 552 device->reg_ops->reg_desc_put(ib_conn, desc); 553 554 return err; 555 } 556 557 void iser_unreg_rdma_mem(struct iscsi_iser_task *task, 558 enum iser_data_dir dir) 559 { 560 struct iser_device *device = task->iser_conn->ib_conn.device; 561 562 device->reg_ops->unreg_mem(task, dir); 563 } 564