1/* 2 * Copyright (c) 2006, 2007, 2008, 2009 QLogic Corporation. All rights reserved. 3 * Copyright (c) 2005, 2006 PathScale, Inc. 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 <rdma/ib_umem.h> 35#include <rdma/ib_smi.h> 36 37#include "qib.h" 38 39/* Fast memory region */ 40struct qib_fmr { 41 struct ib_fmr ibfmr; 42 u8 page_shift; 43 struct qib_mregion mr; /* must be last */ 44}; 45 46static inline struct qib_fmr *to_ifmr(struct ib_fmr *ibfmr) 47{ 48 return container_of(ibfmr, struct qib_fmr, ibfmr); 49} 50 51/** 52 * qib_get_dma_mr - get a DMA memory region 53 * @pd: protection domain for this memory region 54 * @acc: access flags 55 * 56 * Returns the memory region on success, otherwise returns an errno. 57 * Note that all DMA addresses should be created via the 58 * struct ib_dma_mapping_ops functions (see qib_dma.c). 59 */ 60struct ib_mr *qib_get_dma_mr(struct ib_pd *pd, int acc) 61{ 62 struct qib_ibdev *dev = to_idev(pd->device); 63 struct qib_mr *mr; 64 struct ib_mr *ret; 65 unsigned long flags; 66 67 if (to_ipd(pd)->user) { 68 ret = ERR_PTR(-EPERM); 69 goto bail; 70 } 71 72 mr = kzalloc(sizeof *mr, GFP_KERNEL); 73 if (!mr) { 74 ret = ERR_PTR(-ENOMEM); 75 goto bail; 76 } 77 78 mr->mr.access_flags = acc; 79 atomic_set(&mr->mr.refcount, 0); 80 81 spin_lock_irqsave(&dev->lk_table.lock, flags); 82 if (!dev->dma_mr) 83 dev->dma_mr = &mr->mr; 84 spin_unlock_irqrestore(&dev->lk_table.lock, flags); 85 86 ret = &mr->ibmr; 87 88bail: 89 return ret; 90} 91 92static struct qib_mr *alloc_mr(int count, struct qib_lkey_table *lk_table) 93{ 94 struct qib_mr *mr; 95 int m, i = 0; 96 97 /* Allocate struct plus pointers to first level page tables. */ 98 m = (count + QIB_SEGSZ - 1) / QIB_SEGSZ; 99 mr = kmalloc(sizeof *mr + m * sizeof mr->mr.map[0], GFP_KERNEL); 100 if (!mr) 101 goto done; 102 103 /* Allocate first level page tables. */ 104 for (; i < m; i++) { 105 mr->mr.map[i] = kmalloc(sizeof *mr->mr.map[0], GFP_KERNEL); 106 if (!mr->mr.map[i]) 107 goto bail; 108 } 109 mr->mr.mapsz = m; 110 mr->mr.max_segs = count; 111 112 /* 113 * ib_reg_phys_mr() will initialize mr->ibmr except for 114 * lkey and rkey. 115 */ 116 if (!qib_alloc_lkey(lk_table, &mr->mr)) 117 goto bail; 118 mr->ibmr.lkey = mr->mr.lkey; 119 mr->ibmr.rkey = mr->mr.lkey; 120 121 atomic_set(&mr->mr.refcount, 0); 122 goto done; 123 124bail: 125 while (i) 126 kfree(mr->mr.map[--i]); 127 kfree(mr); 128 mr = NULL; 129 130done: 131 return mr; 132} 133 134/** 135 * qib_reg_phys_mr - register a physical memory region 136 * @pd: protection domain for this memory region 137 * @buffer_list: pointer to the list of physical buffers to register 138 * @num_phys_buf: the number of physical buffers to register 139 * @iova_start: the starting address passed over IB which maps to this MR 140 * 141 * Returns the memory region on success, otherwise returns an errno. 142 */ 143struct ib_mr *qib_reg_phys_mr(struct ib_pd *pd, 144 struct ib_phys_buf *buffer_list, 145 int num_phys_buf, int acc, u64 *iova_start) 146{ 147 struct qib_mr *mr; 148 int n, m, i; 149 struct ib_mr *ret; 150 151 mr = alloc_mr(num_phys_buf, &to_idev(pd->device)->lk_table); 152 if (mr == NULL) { 153 ret = ERR_PTR(-ENOMEM); 154 goto bail; 155 } 156 157 mr->mr.pd = pd; 158 mr->mr.user_base = *iova_start; 159 mr->mr.iova = *iova_start; 160 mr->mr.length = 0; 161 mr->mr.offset = 0; 162 mr->mr.access_flags = acc; 163 mr->umem = NULL; 164 165 m = 0; 166 n = 0; 167 for (i = 0; i < num_phys_buf; i++) { 168 mr->mr.map[m]->segs[n].vaddr = (void *) buffer_list[i].addr; 169 mr->mr.map[m]->segs[n].length = buffer_list[i].size; 170 mr->mr.length += buffer_list[i].size; 171 n++; 172 if (n == QIB_SEGSZ) { 173 m++; 174 n = 0; 175 } 176 } 177 178 ret = &mr->ibmr; 179 180bail: 181 return ret; 182} 183 184/** 185 * qib_reg_user_mr - register a userspace memory region 186 * @pd: protection domain for this memory region 187 * @start: starting userspace address 188 * @length: length of region to register 189 * @virt_addr: virtual address to use (from HCA's point of view) 190 * @mr_access_flags: access flags for this memory region 191 * @udata: unused by the QLogic_IB driver 192 * 193 * Returns the memory region on success, otherwise returns an errno. 194 */ 195struct ib_mr *qib_reg_user_mr(struct ib_pd *pd, u64 start, u64 length, 196 u64 virt_addr, int mr_access_flags, 197 struct ib_udata *udata) 198{ 199 struct qib_mr *mr; 200 struct ib_umem *umem; 201 struct ib_umem_chunk *chunk; 202 int n, m, i; 203 struct ib_mr *ret; 204 205 if (length == 0) { 206 ret = ERR_PTR(-EINVAL); 207 goto bail; 208 } 209 210 umem = ib_umem_get(pd->uobject->context, start, length, 211 mr_access_flags, 0); 212 if (IS_ERR(umem)) 213 return (void *) umem; 214 215 n = 0; 216 list_for_each_entry(chunk, &umem->chunk_list, list) 217 n += chunk->nents; 218 219 mr = alloc_mr(n, &to_idev(pd->device)->lk_table); 220 if (!mr) { 221 ret = ERR_PTR(-ENOMEM); 222 ib_umem_release(umem); 223 goto bail; 224 } 225 226 mr->mr.pd = pd; 227 mr->mr.user_base = start; 228 mr->mr.iova = virt_addr; 229 mr->mr.length = length; 230 mr->mr.offset = umem->offset; 231 mr->mr.access_flags = mr_access_flags; 232 mr->umem = umem; 233 234 m = 0; 235 n = 0; 236 list_for_each_entry(chunk, &umem->chunk_list, list) { 237 for (i = 0; i < chunk->nents; i++) { 238 void *vaddr; 239 240 vaddr = page_address(sg_page(&chunk->page_list[i])); 241 if (!vaddr) { 242 ret = ERR_PTR(-EINVAL); 243 goto bail; 244 } 245 mr->mr.map[m]->segs[n].vaddr = vaddr; 246 mr->mr.map[m]->segs[n].length = umem->page_size; 247 n++; 248 if (n == QIB_SEGSZ) { 249 m++; 250 n = 0; 251 } 252 } 253 } 254 ret = &mr->ibmr; 255 256bail: 257 return ret; 258} 259 260/** 261 * qib_dereg_mr - unregister and free a memory region 262 * @ibmr: the memory region to free 263 * 264 * Returns 0 on success. 265 * 266 * Note that this is called to free MRs created by qib_get_dma_mr() 267 * or qib_reg_user_mr(). 268 */ 269int qib_dereg_mr(struct ib_mr *ibmr) 270{ 271 struct qib_mr *mr = to_imr(ibmr); 272 struct qib_ibdev *dev = to_idev(ibmr->device); 273 int ret; 274 int i; 275 276 ret = qib_free_lkey(dev, &mr->mr); 277 if (ret) 278 return ret; 279 280 i = mr->mr.mapsz; 281 while (i) 282 kfree(mr->mr.map[--i]); 283 if (mr->umem) 284 ib_umem_release(mr->umem); 285 kfree(mr); 286 return 0; 287} 288 289/* 290 * Allocate a memory region usable with the 291 * IB_WR_FAST_REG_MR send work request. 292 * 293 * Return the memory region on success, otherwise return an errno. 294 */ 295struct ib_mr *qib_alloc_fast_reg_mr(struct ib_pd *pd, int max_page_list_len) 296{ 297 struct qib_mr *mr; 298 299 mr = alloc_mr(max_page_list_len, &to_idev(pd->device)->lk_table); 300 if (mr == NULL) 301 return ERR_PTR(-ENOMEM); 302 303 mr->mr.pd = pd; 304 mr->mr.user_base = 0; 305 mr->mr.iova = 0; 306 mr->mr.length = 0; 307 mr->mr.offset = 0; 308 mr->mr.access_flags = 0; 309 mr->umem = NULL; 310 311 return &mr->ibmr; 312} 313 314struct ib_fast_reg_page_list * 315qib_alloc_fast_reg_page_list(struct ib_device *ibdev, int page_list_len) 316{ 317 unsigned size = page_list_len * sizeof(u64); 318 struct ib_fast_reg_page_list *pl; 319 320 if (size > PAGE_SIZE) 321 return ERR_PTR(-EINVAL); 322 323 pl = kmalloc(sizeof *pl, GFP_KERNEL); 324 if (!pl) 325 return ERR_PTR(-ENOMEM); 326 327 pl->page_list = kmalloc(size, GFP_KERNEL); 328 if (!pl->page_list) 329 goto err_free; 330 331 return pl; 332 333err_free: 334 kfree(pl); 335 return ERR_PTR(-ENOMEM); 336} 337 338void qib_free_fast_reg_page_list(struct ib_fast_reg_page_list *pl) 339{ 340 kfree(pl->page_list); 341 kfree(pl); 342} 343 344/** 345 * qib_alloc_fmr - allocate a fast memory region 346 * @pd: the protection domain for this memory region 347 * @mr_access_flags: access flags for this memory region 348 * @fmr_attr: fast memory region attributes 349 * 350 * Returns the memory region on success, otherwise returns an errno. 351 */ 352struct ib_fmr *qib_alloc_fmr(struct ib_pd *pd, int mr_access_flags, 353 struct ib_fmr_attr *fmr_attr) 354{ 355 struct qib_fmr *fmr; 356 int m, i = 0; 357 struct ib_fmr *ret; 358 359 /* Allocate struct plus pointers to first level page tables. */ 360 m = (fmr_attr->max_pages + QIB_SEGSZ - 1) / QIB_SEGSZ; 361 fmr = kmalloc(sizeof *fmr + m * sizeof fmr->mr.map[0], GFP_KERNEL); 362 if (!fmr) 363 goto bail; 364 365 /* Allocate first level page tables. */ 366 for (; i < m; i++) { 367 fmr->mr.map[i] = kmalloc(sizeof *fmr->mr.map[0], 368 GFP_KERNEL); 369 if (!fmr->mr.map[i]) 370 goto bail; 371 } 372 fmr->mr.mapsz = m; 373 374 /* 375 * ib_alloc_fmr() will initialize fmr->ibfmr except for lkey & 376 * rkey. 377 */ 378 if (!qib_alloc_lkey(&to_idev(pd->device)->lk_table, &fmr->mr)) 379 goto bail; 380 fmr->ibfmr.rkey = fmr->mr.lkey; 381 fmr->ibfmr.lkey = fmr->mr.lkey; 382 /* 383 * Resources are allocated but no valid mapping (RKEY can't be 384 * used). 385 */ 386 fmr->mr.pd = pd; 387 fmr->mr.user_base = 0; 388 fmr->mr.iova = 0; 389 fmr->mr.length = 0; 390 fmr->mr.offset = 0; 391 fmr->mr.access_flags = mr_access_flags; 392 fmr->mr.max_segs = fmr_attr->max_pages; 393 fmr->page_shift = fmr_attr->page_shift; 394 395 atomic_set(&fmr->mr.refcount, 0); 396 ret = &fmr->ibfmr; 397 goto done; 398 399bail: 400 while (i) 401 kfree(fmr->mr.map[--i]); 402 kfree(fmr); 403 ret = ERR_PTR(-ENOMEM); 404 405done: 406 return ret; 407} 408 409/** 410 * qib_map_phys_fmr - set up a fast memory region 411 * @ibmfr: the fast memory region to set up 412 * @page_list: the list of pages to associate with the fast memory region 413 * @list_len: the number of pages to associate with the fast memory region 414 * @iova: the virtual address of the start of the fast memory region 415 * 416 * This may be called from interrupt context. 417 */ 418 419int qib_map_phys_fmr(struct ib_fmr *ibfmr, u64 *page_list, 420 int list_len, u64 iova) 421{ 422 struct qib_fmr *fmr = to_ifmr(ibfmr); 423 struct qib_lkey_table *rkt; 424 unsigned long flags; 425 int m, n, i; 426 u32 ps; 427 int ret; 428 429 if (atomic_read(&fmr->mr.refcount)) 430 return -EBUSY; 431 432 if (list_len > fmr->mr.max_segs) { 433 ret = -EINVAL; 434 goto bail; 435 } 436 rkt = &to_idev(ibfmr->device)->lk_table; 437 spin_lock_irqsave(&rkt->lock, flags); 438 fmr->mr.user_base = iova; 439 fmr->mr.iova = iova; 440 ps = 1 << fmr->page_shift; 441 fmr->mr.length = list_len * ps; 442 m = 0; 443 n = 0; 444 for (i = 0; i < list_len; i++) { 445 fmr->mr.map[m]->segs[n].vaddr = (void *) page_list[i]; 446 fmr->mr.map[m]->segs[n].length = ps; 447 if (++n == QIB_SEGSZ) { 448 m++; 449 n = 0; 450 } 451 } 452 spin_unlock_irqrestore(&rkt->lock, flags); 453 ret = 0; 454 455bail: 456 return ret; 457} 458 459/** 460 * qib_unmap_fmr - unmap fast memory regions 461 * @fmr_list: the list of fast memory regions to unmap 462 * 463 * Returns 0 on success. 464 */ 465int qib_unmap_fmr(struct list_head *fmr_list) 466{ 467 struct qib_fmr *fmr; 468 struct qib_lkey_table *rkt; 469 unsigned long flags; 470 471 list_for_each_entry(fmr, fmr_list, ibfmr.list) { 472 rkt = &to_idev(fmr->ibfmr.device)->lk_table; 473 spin_lock_irqsave(&rkt->lock, flags); 474 fmr->mr.user_base = 0; 475 fmr->mr.iova = 0; 476 fmr->mr.length = 0; 477 spin_unlock_irqrestore(&rkt->lock, flags); 478 } 479 return 0; 480} 481 482/** 483 * qib_dealloc_fmr - deallocate a fast memory region 484 * @ibfmr: the fast memory region to deallocate 485 * 486 * Returns 0 on success. 487 */ 488int qib_dealloc_fmr(struct ib_fmr *ibfmr) 489{ 490 struct qib_fmr *fmr = to_ifmr(ibfmr); 491 int ret; 492 int i; 493 494 ret = qib_free_lkey(to_idev(ibfmr->device), &fmr->mr); 495 if (ret) 496 return ret; 497 498 i = fmr->mr.mapsz; 499 while (i) 500 kfree(fmr->mr.map[--i]); 501 kfree(fmr); 502 return 0; 503} 504