1/* 2 * Copyright (c) 2007 Cisco Systems, Inc. All rights reserved. 3 * Copyright (c) 2007, 2008 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 34#include <linux/slab.h> 35 36#include "mlx4_ib.h" 37 38static u32 convert_access(int acc) 39{ 40 return (acc & IB_ACCESS_REMOTE_ATOMIC ? MLX4_PERM_ATOMIC : 0) | 41 (acc & IB_ACCESS_REMOTE_WRITE ? MLX4_PERM_REMOTE_WRITE : 0) | 42 (acc & IB_ACCESS_REMOTE_READ ? MLX4_PERM_REMOTE_READ : 0) | 43 (acc & IB_ACCESS_LOCAL_WRITE ? MLX4_PERM_LOCAL_WRITE : 0) | 44 (acc & IB_ACCESS_MW_BIND ? MLX4_PERM_BIND_MW : 0) | 45 MLX4_PERM_LOCAL_READ; 46} 47 48static enum mlx4_mw_type to_mlx4_type(enum ib_mw_type type) 49{ 50 switch (type) { 51 case IB_MW_TYPE_1: return MLX4_MW_TYPE_1; 52 case IB_MW_TYPE_2: return MLX4_MW_TYPE_2; 53 default: return -1; 54 } 55} 56 57struct ib_mr *mlx4_ib_get_dma_mr(struct ib_pd *pd, int acc) 58{ 59 struct mlx4_ib_mr *mr; 60 int err; 61 62 mr = kmalloc(sizeof *mr, GFP_KERNEL); 63 if (!mr) 64 return ERR_PTR(-ENOMEM); 65 66 err = mlx4_mr_alloc(to_mdev(pd->device)->dev, to_mpd(pd)->pdn, 0, 67 ~0ull, convert_access(acc), 0, 0, &mr->mmr); 68 if (err) 69 goto err_free; 70 71 err = mlx4_mr_enable(to_mdev(pd->device)->dev, &mr->mmr); 72 if (err) 73 goto err_mr; 74 75 mr->ibmr.rkey = mr->ibmr.lkey = mr->mmr.key; 76 mr->umem = NULL; 77 78 return &mr->ibmr; 79 80err_mr: 81 (void) mlx4_mr_free(to_mdev(pd->device)->dev, &mr->mmr); 82 83err_free: 84 kfree(mr); 85 86 return ERR_PTR(err); 87} 88 89int mlx4_ib_umem_write_mtt(struct mlx4_ib_dev *dev, struct mlx4_mtt *mtt, 90 struct ib_umem *umem) 91{ 92 u64 *pages; 93 int i, k, entry; 94 int n; 95 int len; 96 int err = 0; 97 struct scatterlist *sg; 98 99 pages = (u64 *) __get_free_page(GFP_KERNEL); 100 if (!pages) 101 return -ENOMEM; 102 103 i = n = 0; 104 105 for_each_sg(umem->sg_head.sgl, sg, umem->nmap, entry) { 106 len = sg_dma_len(sg) >> mtt->page_shift; 107 for (k = 0; k < len; ++k) { 108 pages[i++] = sg_dma_address(sg) + 109 umem->page_size * k; 110 /* 111 * Be friendly to mlx4_write_mtt() and 112 * pass it chunks of appropriate size. 113 */ 114 if (i == PAGE_SIZE / sizeof (u64)) { 115 err = mlx4_write_mtt(dev->dev, mtt, n, 116 i, pages); 117 if (err) 118 goto out; 119 n += i; 120 i = 0; 121 } 122 } 123 } 124 125 if (i) 126 err = mlx4_write_mtt(dev->dev, mtt, n, i, pages); 127 128out: 129 free_page((unsigned long) pages); 130 return err; 131} 132 133struct ib_mr *mlx4_ib_reg_user_mr(struct ib_pd *pd, u64 start, u64 length, 134 u64 virt_addr, int access_flags, 135 struct ib_udata *udata) 136{ 137 struct mlx4_ib_dev *dev = to_mdev(pd->device); 138 struct mlx4_ib_mr *mr; 139 int shift; 140 int err; 141 int n; 142 143 mr = kmalloc(sizeof *mr, GFP_KERNEL); 144 if (!mr) 145 return ERR_PTR(-ENOMEM); 146 147 /* Force registering the memory as writable. */ 148 /* Used for memory re-registeration. HCA protects the access */ 149 mr->umem = ib_umem_get(pd->uobject->context, start, length, 150 access_flags | IB_ACCESS_LOCAL_WRITE, 0); 151 if (IS_ERR(mr->umem)) { 152 err = PTR_ERR(mr->umem); 153 goto err_free; 154 } 155 156 n = ib_umem_page_count(mr->umem); 157 shift = ilog2(mr->umem->page_size); 158 159 err = mlx4_mr_alloc(dev->dev, to_mpd(pd)->pdn, virt_addr, length, 160 convert_access(access_flags), n, shift, &mr->mmr); 161 if (err) 162 goto err_umem; 163 164 err = mlx4_ib_umem_write_mtt(dev, &mr->mmr.mtt, mr->umem); 165 if (err) 166 goto err_mr; 167 168 err = mlx4_mr_enable(dev->dev, &mr->mmr); 169 if (err) 170 goto err_mr; 171 172 mr->ibmr.rkey = mr->ibmr.lkey = mr->mmr.key; 173 174 return &mr->ibmr; 175 176err_mr: 177 (void) mlx4_mr_free(to_mdev(pd->device)->dev, &mr->mmr); 178 179err_umem: 180 ib_umem_release(mr->umem); 181 182err_free: 183 kfree(mr); 184 185 return ERR_PTR(err); 186} 187 188int mlx4_ib_rereg_user_mr(struct ib_mr *mr, int flags, 189 u64 start, u64 length, u64 virt_addr, 190 int mr_access_flags, struct ib_pd *pd, 191 struct ib_udata *udata) 192{ 193 struct mlx4_ib_dev *dev = to_mdev(mr->device); 194 struct mlx4_ib_mr *mmr = to_mmr(mr); 195 struct mlx4_mpt_entry *mpt_entry; 196 struct mlx4_mpt_entry **pmpt_entry = &mpt_entry; 197 int err; 198 199 /* Since we synchronize this call and mlx4_ib_dereg_mr via uverbs, 200 * we assume that the calls can't run concurrently. Otherwise, a 201 * race exists. 202 */ 203 err = mlx4_mr_hw_get_mpt(dev->dev, &mmr->mmr, &pmpt_entry); 204 205 if (err) 206 return err; 207 208 if (flags & IB_MR_REREG_PD) { 209 err = mlx4_mr_hw_change_pd(dev->dev, *pmpt_entry, 210 to_mpd(pd)->pdn); 211 212 if (err) 213 goto release_mpt_entry; 214 } 215 216 if (flags & IB_MR_REREG_ACCESS) { 217 err = mlx4_mr_hw_change_access(dev->dev, *pmpt_entry, 218 convert_access(mr_access_flags)); 219 220 if (err) 221 goto release_mpt_entry; 222 } 223 224 if (flags & IB_MR_REREG_TRANS) { 225 int shift; 226 int err; 227 int n; 228 229 mlx4_mr_rereg_mem_cleanup(dev->dev, &mmr->mmr); 230 ib_umem_release(mmr->umem); 231 mmr->umem = ib_umem_get(mr->uobject->context, start, length, 232 mr_access_flags | 233 IB_ACCESS_LOCAL_WRITE, 234 0); 235 if (IS_ERR(mmr->umem)) { 236 err = PTR_ERR(mmr->umem); 237 /* Prevent mlx4_ib_dereg_mr from free'ing invalid pointer */ 238 mmr->umem = NULL; 239 goto release_mpt_entry; 240 } 241 n = ib_umem_page_count(mmr->umem); 242 shift = ilog2(mmr->umem->page_size); 243 244 err = mlx4_mr_rereg_mem_write(dev->dev, &mmr->mmr, 245 virt_addr, length, n, shift, 246 *pmpt_entry); 247 if (err) { 248 ib_umem_release(mmr->umem); 249 goto release_mpt_entry; 250 } 251 mmr->mmr.iova = virt_addr; 252 mmr->mmr.size = length; 253 254 err = mlx4_ib_umem_write_mtt(dev, &mmr->mmr.mtt, mmr->umem); 255 if (err) { 256 mlx4_mr_rereg_mem_cleanup(dev->dev, &mmr->mmr); 257 ib_umem_release(mmr->umem); 258 goto release_mpt_entry; 259 } 260 } 261 262 /* If we couldn't transfer the MR to the HCA, just remember to 263 * return a failure. But dereg_mr will free the resources. 264 */ 265 err = mlx4_mr_hw_write_mpt(dev->dev, &mmr->mmr, pmpt_entry); 266 if (!err && flags & IB_MR_REREG_ACCESS) 267 mmr->mmr.access = mr_access_flags; 268 269release_mpt_entry: 270 mlx4_mr_hw_put_mpt(dev->dev, pmpt_entry); 271 272 return err; 273} 274 275int mlx4_ib_dereg_mr(struct ib_mr *ibmr) 276{ 277 struct mlx4_ib_mr *mr = to_mmr(ibmr); 278 int ret; 279 280 ret = mlx4_mr_free(to_mdev(ibmr->device)->dev, &mr->mmr); 281 if (ret) 282 return ret; 283 if (mr->umem) 284 ib_umem_release(mr->umem); 285 kfree(mr); 286 287 return 0; 288} 289 290struct ib_mw *mlx4_ib_alloc_mw(struct ib_pd *pd, enum ib_mw_type type) 291{ 292 struct mlx4_ib_dev *dev = to_mdev(pd->device); 293 struct mlx4_ib_mw *mw; 294 int err; 295 296 mw = kmalloc(sizeof(*mw), GFP_KERNEL); 297 if (!mw) 298 return ERR_PTR(-ENOMEM); 299 300 err = mlx4_mw_alloc(dev->dev, to_mpd(pd)->pdn, 301 to_mlx4_type(type), &mw->mmw); 302 if (err) 303 goto err_free; 304 305 err = mlx4_mw_enable(dev->dev, &mw->mmw); 306 if (err) 307 goto err_mw; 308 309 mw->ibmw.rkey = mw->mmw.key; 310 311 return &mw->ibmw; 312 313err_mw: 314 mlx4_mw_free(dev->dev, &mw->mmw); 315 316err_free: 317 kfree(mw); 318 319 return ERR_PTR(err); 320} 321 322int mlx4_ib_bind_mw(struct ib_qp *qp, struct ib_mw *mw, 323 struct ib_mw_bind *mw_bind) 324{ 325 struct ib_send_wr wr; 326 struct ib_send_wr *bad_wr; 327 int ret; 328 329 memset(&wr, 0, sizeof(wr)); 330 wr.opcode = IB_WR_BIND_MW; 331 wr.wr_id = mw_bind->wr_id; 332 wr.send_flags = mw_bind->send_flags; 333 wr.wr.bind_mw.mw = mw; 334 wr.wr.bind_mw.bind_info = mw_bind->bind_info; 335 wr.wr.bind_mw.rkey = ib_inc_rkey(mw->rkey); 336 337 ret = mlx4_ib_post_send(qp, &wr, &bad_wr); 338 if (!ret) 339 mw->rkey = wr.wr.bind_mw.rkey; 340 341 return ret; 342} 343 344int mlx4_ib_dealloc_mw(struct ib_mw *ibmw) 345{ 346 struct mlx4_ib_mw *mw = to_mmw(ibmw); 347 348 mlx4_mw_free(to_mdev(ibmw->device)->dev, &mw->mmw); 349 kfree(mw); 350 351 return 0; 352} 353 354struct ib_mr *mlx4_ib_alloc_fast_reg_mr(struct ib_pd *pd, 355 int max_page_list_len) 356{ 357 struct mlx4_ib_dev *dev = to_mdev(pd->device); 358 struct mlx4_ib_mr *mr; 359 int err; 360 361 mr = kmalloc(sizeof *mr, GFP_KERNEL); 362 if (!mr) 363 return ERR_PTR(-ENOMEM); 364 365 err = mlx4_mr_alloc(dev->dev, to_mpd(pd)->pdn, 0, 0, 0, 366 max_page_list_len, 0, &mr->mmr); 367 if (err) 368 goto err_free; 369 370 err = mlx4_mr_enable(dev->dev, &mr->mmr); 371 if (err) 372 goto err_mr; 373 374 mr->ibmr.rkey = mr->ibmr.lkey = mr->mmr.key; 375 mr->umem = NULL; 376 377 return &mr->ibmr; 378 379err_mr: 380 (void) mlx4_mr_free(dev->dev, &mr->mmr); 381 382err_free: 383 kfree(mr); 384 return ERR_PTR(err); 385} 386 387struct ib_fast_reg_page_list *mlx4_ib_alloc_fast_reg_page_list(struct ib_device *ibdev, 388 int page_list_len) 389{ 390 struct mlx4_ib_dev *dev = to_mdev(ibdev); 391 struct mlx4_ib_fast_reg_page_list *mfrpl; 392 int size = page_list_len * sizeof (u64); 393 394 if (page_list_len > MLX4_MAX_FAST_REG_PAGES) 395 return ERR_PTR(-EINVAL); 396 397 mfrpl = kmalloc(sizeof *mfrpl, GFP_KERNEL); 398 if (!mfrpl) 399 return ERR_PTR(-ENOMEM); 400 401 mfrpl->ibfrpl.page_list = kmalloc(size, GFP_KERNEL); 402 if (!mfrpl->ibfrpl.page_list) 403 goto err_free; 404 405 mfrpl->mapped_page_list = dma_alloc_coherent(&dev->dev->pdev->dev, 406 size, &mfrpl->map, 407 GFP_KERNEL); 408 if (!mfrpl->mapped_page_list) 409 goto err_free; 410 411 WARN_ON(mfrpl->map & 0x3f); 412 413 return &mfrpl->ibfrpl; 414 415err_free: 416 kfree(mfrpl->ibfrpl.page_list); 417 kfree(mfrpl); 418 return ERR_PTR(-ENOMEM); 419} 420 421void mlx4_ib_free_fast_reg_page_list(struct ib_fast_reg_page_list *page_list) 422{ 423 struct mlx4_ib_dev *dev = to_mdev(page_list->device); 424 struct mlx4_ib_fast_reg_page_list *mfrpl = to_mfrpl(page_list); 425 int size = page_list->max_page_list_len * sizeof (u64); 426 427 dma_free_coherent(&dev->dev->pdev->dev, size, mfrpl->mapped_page_list, 428 mfrpl->map); 429 kfree(mfrpl->ibfrpl.page_list); 430 kfree(mfrpl); 431} 432 433struct ib_fmr *mlx4_ib_fmr_alloc(struct ib_pd *pd, int acc, 434 struct ib_fmr_attr *fmr_attr) 435{ 436 struct mlx4_ib_dev *dev = to_mdev(pd->device); 437 struct mlx4_ib_fmr *fmr; 438 int err = -ENOMEM; 439 440 fmr = kmalloc(sizeof *fmr, GFP_KERNEL); 441 if (!fmr) 442 return ERR_PTR(-ENOMEM); 443 444 err = mlx4_fmr_alloc(dev->dev, to_mpd(pd)->pdn, convert_access(acc), 445 fmr_attr->max_pages, fmr_attr->max_maps, 446 fmr_attr->page_shift, &fmr->mfmr); 447 if (err) 448 goto err_free; 449 450 err = mlx4_fmr_enable(to_mdev(pd->device)->dev, &fmr->mfmr); 451 if (err) 452 goto err_mr; 453 454 fmr->ibfmr.rkey = fmr->ibfmr.lkey = fmr->mfmr.mr.key; 455 456 return &fmr->ibfmr; 457 458err_mr: 459 (void) mlx4_mr_free(to_mdev(pd->device)->dev, &fmr->mfmr.mr); 460 461err_free: 462 kfree(fmr); 463 464 return ERR_PTR(err); 465} 466 467int mlx4_ib_map_phys_fmr(struct ib_fmr *ibfmr, u64 *page_list, 468 int npages, u64 iova) 469{ 470 struct mlx4_ib_fmr *ifmr = to_mfmr(ibfmr); 471 struct mlx4_ib_dev *dev = to_mdev(ifmr->ibfmr.device); 472 473 return mlx4_map_phys_fmr(dev->dev, &ifmr->mfmr, page_list, npages, iova, 474 &ifmr->ibfmr.lkey, &ifmr->ibfmr.rkey); 475} 476 477int mlx4_ib_unmap_fmr(struct list_head *fmr_list) 478{ 479 struct ib_fmr *ibfmr; 480 int err; 481 struct mlx4_dev *mdev = NULL; 482 483 list_for_each_entry(ibfmr, fmr_list, list) { 484 if (mdev && to_mdev(ibfmr->device)->dev != mdev) 485 return -EINVAL; 486 mdev = to_mdev(ibfmr->device)->dev; 487 } 488 489 if (!mdev) 490 return 0; 491 492 list_for_each_entry(ibfmr, fmr_list, list) { 493 struct mlx4_ib_fmr *ifmr = to_mfmr(ibfmr); 494 495 mlx4_fmr_unmap(mdev, &ifmr->mfmr, &ifmr->ibfmr.lkey, &ifmr->ibfmr.rkey); 496 } 497 498 /* 499 * Make sure all MPT status updates are visible before issuing 500 * SYNC_TPT firmware command. 501 */ 502 wmb(); 503 504 err = mlx4_SYNC_TPT(mdev); 505 if (err) 506 pr_warn("SYNC_TPT error %d when " 507 "unmapping FMRs\n", err); 508 509 return 0; 510} 511 512int mlx4_ib_fmr_dealloc(struct ib_fmr *ibfmr) 513{ 514 struct mlx4_ib_fmr *ifmr = to_mfmr(ibfmr); 515 struct mlx4_ib_dev *dev = to_mdev(ibfmr->device); 516 int err; 517 518 err = mlx4_fmr_free(dev->dev, &ifmr->mfmr); 519 520 if (!err) 521 kfree(ifmr); 522 523 return err; 524} 525