1 /* 2 * Copyright (c) 2005 Topspin Communications. All rights reserved. 3 * Copyright (c) 2005 Mellanox Technologies. All rights reserved. 4 * Copyright (c) 2013 Cisco Systems. All rights reserved. 5 * 6 * This software is available to you under a choice of one of two 7 * licenses. You may choose to be licensed under the terms of the GNU 8 * General Public License (GPL) Version 2, available from the file 9 * COPYING in the main directory of this source tree, or the 10 * BSD license below: 11 * 12 * Redistribution and use in source and binary forms, with or 13 * without modification, are permitted provided that the following 14 * conditions are met: 15 * 16 * - Redistributions of source code must retain the above 17 * copyright notice, this list of conditions and the following 18 * disclaimer. 19 * 20 * - Redistributions in binary form must reproduce the above 21 * copyright notice, this list of conditions and the following 22 * disclaimer in the documentation and/or other materials 23 * provided with the distribution. 24 * 25 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, 26 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF 27 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND 28 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS 29 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN 30 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN 31 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 32 * SOFTWARE. 33 */ 34 35 #include <linux/mm.h> 36 #include <linux/dma-mapping.h> 37 #include <linux/sched/signal.h> 38 #include <linux/sched/mm.h> 39 #include <linux/hugetlb.h> 40 #include <linux/iommu.h> 41 #include <linux/workqueue.h> 42 #include <linux/list.h> 43 #include <linux/pci.h> 44 45 #include "usnic_log.h" 46 #include "usnic_uiom.h" 47 #include "usnic_uiom_interval_tree.h" 48 49 static struct workqueue_struct *usnic_uiom_wq; 50 51 #define USNIC_UIOM_PAGE_CHUNK \ 52 ((PAGE_SIZE - offsetof(struct usnic_uiom_chunk, page_list)) /\ 53 ((void *) &((struct usnic_uiom_chunk *) 0)->page_list[1] - \ 54 (void *) &((struct usnic_uiom_chunk *) 0)->page_list[0])) 55 56 static void usnic_uiom_reg_account(struct work_struct *work) 57 { 58 struct usnic_uiom_reg *umem = container_of(work, 59 struct usnic_uiom_reg, work); 60 61 down_write(&umem->mm->mmap_sem); 62 umem->mm->locked_vm -= umem->diff; 63 up_write(&umem->mm->mmap_sem); 64 mmput(umem->mm); 65 kfree(umem); 66 } 67 68 static int usnic_uiom_dma_fault(struct iommu_domain *domain, 69 struct device *dev, 70 unsigned long iova, int flags, 71 void *token) 72 { 73 usnic_err("Device %s iommu fault domain 0x%pK va 0x%lx flags 0x%x\n", 74 dev_name(dev), 75 domain, iova, flags); 76 return -ENOSYS; 77 } 78 79 static void usnic_uiom_put_pages(struct list_head *chunk_list, int dirty) 80 { 81 struct usnic_uiom_chunk *chunk, *tmp; 82 struct page *page; 83 struct scatterlist *sg; 84 int i; 85 dma_addr_t pa; 86 87 list_for_each_entry_safe(chunk, tmp, chunk_list, list) { 88 for_each_sg(chunk->page_list, sg, chunk->nents, i) { 89 page = sg_page(sg); 90 pa = sg_phys(sg); 91 if (dirty) 92 set_page_dirty_lock(page); 93 put_page(page); 94 usnic_dbg("pa: %pa\n", &pa); 95 } 96 kfree(chunk); 97 } 98 } 99 100 static int usnic_uiom_get_pages(unsigned long addr, size_t size, int writable, 101 int dmasync, struct list_head *chunk_list) 102 { 103 struct page **page_list; 104 struct scatterlist *sg; 105 struct usnic_uiom_chunk *chunk; 106 unsigned long locked; 107 unsigned long lock_limit; 108 unsigned long cur_base; 109 unsigned long npages; 110 int ret; 111 int off; 112 int i; 113 int flags; 114 dma_addr_t pa; 115 unsigned int gup_flags; 116 117 if (!can_do_mlock()) 118 return -EPERM; 119 120 INIT_LIST_HEAD(chunk_list); 121 122 page_list = (struct page **) __get_free_page(GFP_KERNEL); 123 if (!page_list) 124 return -ENOMEM; 125 126 npages = PAGE_ALIGN(size + (addr & ~PAGE_MASK)) >> PAGE_SHIFT; 127 128 down_write(¤t->mm->mmap_sem); 129 130 locked = npages + current->mm->locked_vm; 131 lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT; 132 133 if ((locked > lock_limit) && !capable(CAP_IPC_LOCK)) { 134 ret = -ENOMEM; 135 goto out; 136 } 137 138 flags = IOMMU_READ | IOMMU_CACHE; 139 flags |= (writable) ? IOMMU_WRITE : 0; 140 gup_flags = FOLL_WRITE; 141 gup_flags |= (writable) ? 0 : FOLL_FORCE; 142 cur_base = addr & PAGE_MASK; 143 ret = 0; 144 145 while (npages) { 146 ret = get_user_pages(cur_base, 147 min_t(unsigned long, npages, 148 PAGE_SIZE / sizeof(struct page *)), 149 gup_flags, page_list, NULL); 150 151 if (ret < 0) 152 goto out; 153 154 npages -= ret; 155 off = 0; 156 157 while (ret) { 158 chunk = kmalloc(sizeof(*chunk) + 159 sizeof(struct scatterlist) * 160 min_t(int, ret, USNIC_UIOM_PAGE_CHUNK), 161 GFP_KERNEL); 162 if (!chunk) { 163 ret = -ENOMEM; 164 goto out; 165 } 166 167 chunk->nents = min_t(int, ret, USNIC_UIOM_PAGE_CHUNK); 168 sg_init_table(chunk->page_list, chunk->nents); 169 for_each_sg(chunk->page_list, sg, chunk->nents, i) { 170 sg_set_page(sg, page_list[i + off], 171 PAGE_SIZE, 0); 172 pa = sg_phys(sg); 173 usnic_dbg("va: 0x%lx pa: %pa\n", 174 cur_base + i*PAGE_SIZE, &pa); 175 } 176 cur_base += chunk->nents * PAGE_SIZE; 177 ret -= chunk->nents; 178 off += chunk->nents; 179 list_add_tail(&chunk->list, chunk_list); 180 } 181 182 ret = 0; 183 } 184 185 out: 186 if (ret < 0) 187 usnic_uiom_put_pages(chunk_list, 0); 188 else 189 current->mm->locked_vm = locked; 190 191 up_write(¤t->mm->mmap_sem); 192 free_page((unsigned long) page_list); 193 return ret; 194 } 195 196 static void usnic_uiom_unmap_sorted_intervals(struct list_head *intervals, 197 struct usnic_uiom_pd *pd) 198 { 199 struct usnic_uiom_interval_node *interval, *tmp; 200 long unsigned va, size; 201 202 list_for_each_entry_safe(interval, tmp, intervals, link) { 203 va = interval->start << PAGE_SHIFT; 204 size = ((interval->last - interval->start) + 1) << PAGE_SHIFT; 205 while (size > 0) { 206 /* Workaround for RH 970401 */ 207 usnic_dbg("va 0x%lx size 0x%lx", va, PAGE_SIZE); 208 iommu_unmap(pd->domain, va, PAGE_SIZE); 209 va += PAGE_SIZE; 210 size -= PAGE_SIZE; 211 } 212 } 213 } 214 215 static void __usnic_uiom_reg_release(struct usnic_uiom_pd *pd, 216 struct usnic_uiom_reg *uiomr, 217 int dirty) 218 { 219 int npages; 220 unsigned long vpn_start, vpn_last; 221 struct usnic_uiom_interval_node *interval, *tmp; 222 int writable = 0; 223 LIST_HEAD(rm_intervals); 224 225 npages = PAGE_ALIGN(uiomr->length + uiomr->offset) >> PAGE_SHIFT; 226 vpn_start = (uiomr->va & PAGE_MASK) >> PAGE_SHIFT; 227 vpn_last = vpn_start + npages - 1; 228 229 spin_lock(&pd->lock); 230 usnic_uiom_remove_interval(&pd->rb_root, vpn_start, 231 vpn_last, &rm_intervals); 232 usnic_uiom_unmap_sorted_intervals(&rm_intervals, pd); 233 234 list_for_each_entry_safe(interval, tmp, &rm_intervals, link) { 235 if (interval->flags & IOMMU_WRITE) 236 writable = 1; 237 list_del(&interval->link); 238 kfree(interval); 239 } 240 241 usnic_uiom_put_pages(&uiomr->chunk_list, dirty & writable); 242 spin_unlock(&pd->lock); 243 } 244 245 static int usnic_uiom_map_sorted_intervals(struct list_head *intervals, 246 struct usnic_uiom_reg *uiomr) 247 { 248 int i, err; 249 size_t size; 250 struct usnic_uiom_chunk *chunk; 251 struct usnic_uiom_interval_node *interval_node; 252 dma_addr_t pa; 253 dma_addr_t pa_start = 0; 254 dma_addr_t pa_end = 0; 255 long int va_start = -EINVAL; 256 struct usnic_uiom_pd *pd = uiomr->pd; 257 long int va = uiomr->va & PAGE_MASK; 258 int flags = IOMMU_READ | IOMMU_CACHE; 259 260 flags |= (uiomr->writable) ? IOMMU_WRITE : 0; 261 chunk = list_first_entry(&uiomr->chunk_list, struct usnic_uiom_chunk, 262 list); 263 list_for_each_entry(interval_node, intervals, link) { 264 iter_chunk: 265 for (i = 0; i < chunk->nents; i++, va += PAGE_SIZE) { 266 pa = sg_phys(&chunk->page_list[i]); 267 if ((va >> PAGE_SHIFT) < interval_node->start) 268 continue; 269 270 if ((va >> PAGE_SHIFT) == interval_node->start) { 271 /* First page of the interval */ 272 va_start = va; 273 pa_start = pa; 274 pa_end = pa; 275 } 276 277 WARN_ON(va_start == -EINVAL); 278 279 if ((pa_end + PAGE_SIZE != pa) && 280 (pa != pa_start)) { 281 /* PAs are not contiguous */ 282 size = pa_end - pa_start + PAGE_SIZE; 283 usnic_dbg("va 0x%lx pa %pa size 0x%zx flags 0x%x", 284 va_start, &pa_start, size, flags); 285 err = iommu_map(pd->domain, va_start, pa_start, 286 size, flags); 287 if (err) { 288 usnic_err("Failed to map va 0x%lx pa %pa size 0x%zx with err %d\n", 289 va_start, &pa_start, size, err); 290 goto err_out; 291 } 292 va_start = va; 293 pa_start = pa; 294 pa_end = pa; 295 } 296 297 if ((va >> PAGE_SHIFT) == interval_node->last) { 298 /* Last page of the interval */ 299 size = pa - pa_start + PAGE_SIZE; 300 usnic_dbg("va 0x%lx pa %pa size 0x%zx flags 0x%x\n", 301 va_start, &pa_start, size, flags); 302 err = iommu_map(pd->domain, va_start, pa_start, 303 size, flags); 304 if (err) { 305 usnic_err("Failed to map va 0x%lx pa %pa size 0x%zx with err %d\n", 306 va_start, &pa_start, size, err); 307 goto err_out; 308 } 309 break; 310 } 311 312 if (pa != pa_start) 313 pa_end += PAGE_SIZE; 314 } 315 316 if (i == chunk->nents) { 317 /* 318 * Hit last entry of the chunk, 319 * hence advance to next chunk 320 */ 321 chunk = list_first_entry(&chunk->list, 322 struct usnic_uiom_chunk, 323 list); 324 goto iter_chunk; 325 } 326 } 327 328 return 0; 329 330 err_out: 331 usnic_uiom_unmap_sorted_intervals(intervals, pd); 332 return err; 333 } 334 335 struct usnic_uiom_reg *usnic_uiom_reg_get(struct usnic_uiom_pd *pd, 336 unsigned long addr, size_t size, 337 int writable, int dmasync) 338 { 339 struct usnic_uiom_reg *uiomr; 340 unsigned long va_base, vpn_start, vpn_last; 341 unsigned long npages; 342 int offset, err; 343 LIST_HEAD(sorted_diff_intervals); 344 345 /* 346 * Intel IOMMU map throws an error if a translation entry is 347 * changed from read to write. This module may not unmap 348 * and then remap the entry after fixing the permission 349 * b/c this open up a small windows where hw DMA may page fault 350 * Hence, make all entries to be writable. 351 */ 352 writable = 1; 353 354 va_base = addr & PAGE_MASK; 355 offset = addr & ~PAGE_MASK; 356 npages = PAGE_ALIGN(size + offset) >> PAGE_SHIFT; 357 vpn_start = (addr & PAGE_MASK) >> PAGE_SHIFT; 358 vpn_last = vpn_start + npages - 1; 359 360 uiomr = kmalloc(sizeof(*uiomr), GFP_KERNEL); 361 if (!uiomr) 362 return ERR_PTR(-ENOMEM); 363 364 uiomr->va = va_base; 365 uiomr->offset = offset; 366 uiomr->length = size; 367 uiomr->writable = writable; 368 uiomr->pd = pd; 369 370 err = usnic_uiom_get_pages(addr, size, writable, dmasync, 371 &uiomr->chunk_list); 372 if (err) { 373 usnic_err("Failed get_pages vpn [0x%lx,0x%lx] err %d\n", 374 vpn_start, vpn_last, err); 375 goto out_free_uiomr; 376 } 377 378 spin_lock(&pd->lock); 379 err = usnic_uiom_get_intervals_diff(vpn_start, vpn_last, 380 (writable) ? IOMMU_WRITE : 0, 381 IOMMU_WRITE, 382 &pd->rb_root, 383 &sorted_diff_intervals); 384 if (err) { 385 usnic_err("Failed disjoint interval vpn [0x%lx,0x%lx] err %d\n", 386 vpn_start, vpn_last, err); 387 goto out_put_pages; 388 } 389 390 err = usnic_uiom_map_sorted_intervals(&sorted_diff_intervals, uiomr); 391 if (err) { 392 usnic_err("Failed map interval vpn [0x%lx,0x%lx] err %d\n", 393 vpn_start, vpn_last, err); 394 goto out_put_intervals; 395 396 } 397 398 err = usnic_uiom_insert_interval(&pd->rb_root, vpn_start, vpn_last, 399 (writable) ? IOMMU_WRITE : 0); 400 if (err) { 401 usnic_err("Failed insert interval vpn [0x%lx,0x%lx] err %d\n", 402 vpn_start, vpn_last, err); 403 goto out_unmap_intervals; 404 } 405 406 usnic_uiom_put_interval_set(&sorted_diff_intervals); 407 spin_unlock(&pd->lock); 408 409 return uiomr; 410 411 out_unmap_intervals: 412 usnic_uiom_unmap_sorted_intervals(&sorted_diff_intervals, pd); 413 out_put_intervals: 414 usnic_uiom_put_interval_set(&sorted_diff_intervals); 415 out_put_pages: 416 usnic_uiom_put_pages(&uiomr->chunk_list, 0); 417 spin_unlock(&pd->lock); 418 out_free_uiomr: 419 kfree(uiomr); 420 return ERR_PTR(err); 421 } 422 423 void usnic_uiom_reg_release(struct usnic_uiom_reg *uiomr, int closing) 424 { 425 struct mm_struct *mm; 426 unsigned long diff; 427 428 __usnic_uiom_reg_release(uiomr->pd, uiomr, 1); 429 430 mm = get_task_mm(current); 431 if (!mm) { 432 kfree(uiomr); 433 return; 434 } 435 436 diff = PAGE_ALIGN(uiomr->length + uiomr->offset) >> PAGE_SHIFT; 437 438 /* 439 * We may be called with the mm's mmap_sem already held. This 440 * can happen when a userspace munmap() is the call that drops 441 * the last reference to our file and calls our release 442 * method. If there are memory regions to destroy, we'll end 443 * up here and not be able to take the mmap_sem. In that case 444 * we defer the vm_locked accounting to the system workqueue. 445 */ 446 if (closing) { 447 if (!down_write_trylock(&mm->mmap_sem)) { 448 INIT_WORK(&uiomr->work, usnic_uiom_reg_account); 449 uiomr->mm = mm; 450 uiomr->diff = diff; 451 452 queue_work(usnic_uiom_wq, &uiomr->work); 453 return; 454 } 455 } else 456 down_write(&mm->mmap_sem); 457 458 current->mm->locked_vm -= diff; 459 up_write(&mm->mmap_sem); 460 mmput(mm); 461 kfree(uiomr); 462 } 463 464 struct usnic_uiom_pd *usnic_uiom_alloc_pd(void) 465 { 466 struct usnic_uiom_pd *pd; 467 void *domain; 468 469 pd = kzalloc(sizeof(*pd), GFP_KERNEL); 470 if (!pd) 471 return ERR_PTR(-ENOMEM); 472 473 pd->domain = domain = iommu_domain_alloc(&pci_bus_type); 474 if (!domain) { 475 usnic_err("Failed to allocate IOMMU domain"); 476 kfree(pd); 477 return ERR_PTR(-ENOMEM); 478 } 479 480 iommu_set_fault_handler(pd->domain, usnic_uiom_dma_fault, NULL); 481 482 spin_lock_init(&pd->lock); 483 INIT_LIST_HEAD(&pd->devs); 484 485 return pd; 486 } 487 488 void usnic_uiom_dealloc_pd(struct usnic_uiom_pd *pd) 489 { 490 iommu_domain_free(pd->domain); 491 kfree(pd); 492 } 493 494 int usnic_uiom_attach_dev_to_pd(struct usnic_uiom_pd *pd, struct device *dev) 495 { 496 struct usnic_uiom_dev *uiom_dev; 497 int err; 498 499 uiom_dev = kzalloc(sizeof(*uiom_dev), GFP_ATOMIC); 500 if (!uiom_dev) 501 return -ENOMEM; 502 uiom_dev->dev = dev; 503 504 err = iommu_attach_device(pd->domain, dev); 505 if (err) 506 goto out_free_dev; 507 508 if (!iommu_capable(dev->bus, IOMMU_CAP_CACHE_COHERENCY)) { 509 usnic_err("IOMMU of %s does not support cache coherency\n", 510 dev_name(dev)); 511 err = -EINVAL; 512 goto out_detach_device; 513 } 514 515 spin_lock(&pd->lock); 516 list_add_tail(&uiom_dev->link, &pd->devs); 517 pd->dev_cnt++; 518 spin_unlock(&pd->lock); 519 520 return 0; 521 522 out_detach_device: 523 iommu_detach_device(pd->domain, dev); 524 out_free_dev: 525 kfree(uiom_dev); 526 return err; 527 } 528 529 void usnic_uiom_detach_dev_from_pd(struct usnic_uiom_pd *pd, struct device *dev) 530 { 531 struct usnic_uiom_dev *uiom_dev; 532 int found = 0; 533 534 spin_lock(&pd->lock); 535 list_for_each_entry(uiom_dev, &pd->devs, link) { 536 if (uiom_dev->dev == dev) { 537 found = 1; 538 break; 539 } 540 } 541 542 if (!found) { 543 usnic_err("Unable to free dev %s - not found\n", 544 dev_name(dev)); 545 spin_unlock(&pd->lock); 546 return; 547 } 548 549 list_del(&uiom_dev->link); 550 pd->dev_cnt--; 551 spin_unlock(&pd->lock); 552 553 return iommu_detach_device(pd->domain, dev); 554 } 555 556 struct device **usnic_uiom_get_dev_list(struct usnic_uiom_pd *pd) 557 { 558 struct usnic_uiom_dev *uiom_dev; 559 struct device **devs; 560 int i = 0; 561 562 spin_lock(&pd->lock); 563 devs = kcalloc(pd->dev_cnt + 1, sizeof(*devs), GFP_ATOMIC); 564 if (!devs) { 565 devs = ERR_PTR(-ENOMEM); 566 goto out; 567 } 568 569 list_for_each_entry(uiom_dev, &pd->devs, link) { 570 devs[i++] = uiom_dev->dev; 571 } 572 out: 573 spin_unlock(&pd->lock); 574 return devs; 575 } 576 577 void usnic_uiom_free_dev_list(struct device **devs) 578 { 579 kfree(devs); 580 } 581 582 int usnic_uiom_init(char *drv_name) 583 { 584 if (!iommu_present(&pci_bus_type)) { 585 usnic_err("IOMMU required but not present or enabled. USNIC QPs will not function w/o enabling IOMMU\n"); 586 return -EPERM; 587 } 588 589 usnic_uiom_wq = create_workqueue(drv_name); 590 if (!usnic_uiom_wq) { 591 usnic_err("Unable to alloc wq for drv %s\n", drv_name); 592 return -ENOMEM; 593 } 594 595 return 0; 596 } 597 598 void usnic_uiom_fini(void) 599 { 600 flush_workqueue(usnic_uiom_wq); 601 destroy_workqueue(usnic_uiom_wq); 602 } 603