drivers/infiniband/hw/usnic/usnic_uiom.c

   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 #include <rdma/ib_verbs.h>
  45
  46 #include "usnic_log.h"
  47 #include "usnic_uiom.h"
  48 #include "usnic_uiom_interval_tree.h"
  49
  50 static struct workqueue_struct *usnic_uiom_wq;
  51
  52 #define USNIC_UIOM_PAGE_CHUNK                                           \
  53         ((PAGE_SIZE - offsetof(struct usnic_uiom_chunk, page_list))     /\
  54         ((void *) &((struct usnic_uiom_chunk *) 0)->page_list[1] -      \
  55         (void *) &((struct usnic_uiom_chunk *) 0)->page_list[0]))
  56
  57 static int usnic_uiom_dma_fault(struct iommu_domain *domain,
  58                                 struct device *dev,
  59                                 unsigned long iova, int flags,
  60                                 void *token)
  61 {
  62         usnic_err("Device %s iommu fault domain 0x%pK va 0x%lx flags 0x%x\n",
  63                 dev_name(dev),
  64                 domain, iova, flags);
  65         return -ENOSYS;
  66 }
  67
  68 static void usnic_uiom_put_pages(struct list_head *chunk_list, int dirty)
  69 {
  70         struct usnic_uiom_chunk *chunk, *tmp;
  71         struct page *page;
  72         struct scatterlist *sg;
  73         int i;
  74         dma_addr_t pa;
  75
  76         list_for_each_entry_safe(chunk, tmp, chunk_list, list) {
  77                 for_each_sg(chunk->page_list, sg, chunk->nents, i) {
  78                         page = sg_page(sg);
  79                         pa = sg_phys(sg);
  80                         if (!PageDirty(page) && dirty)
  81                                 set_page_dirty_lock(page);
  82                         put_page(page);
  83                         usnic_dbg("pa: %pa\n", &pa);
  84                 }
  85                 kfree(chunk);
  86         }
  87 }
  88
  89 static int usnic_uiom_get_pages(unsigned long addr, size_t size, int writable,
  90                                 int dmasync, struct usnic_uiom_reg *uiomr)
  91 {
  92         struct list_head *chunk_list = &uiomr->chunk_list;
  93         struct page **page_list;
  94         struct scatterlist *sg;
  95         struct usnic_uiom_chunk *chunk;
  96         unsigned long locked;
  97         unsigned long lock_limit;
  98         unsigned long cur_base;
  99         unsigned long npages;
 100         int ret;
 101         int off;
 102         int i;
 103         int flags;
 104         dma_addr_t pa;
 105         unsigned int gup_flags;
 106         struct mm_struct *mm;
 107
 108         /*
 109          * If the combination of the addr and size requested for this memory
 110          * region causes an integer overflow, return error.
 111          */
 112         if (((addr + size) < addr) || PAGE_ALIGN(addr + size) < (addr + size))
 113                 return -EINVAL;
 114
 115         if (!size)
 116                 return -EINVAL;
 117
 118         if (!can_do_mlock())
 119                 return -EPERM;
 120
 121         INIT_LIST_HEAD(chunk_list);
 122
 123         page_list = (struct page **) __get_free_page(GFP_KERNEL);
 124         if (!page_list)
 125                 return -ENOMEM;
 126
 127         npages = PAGE_ALIGN(size + (addr & ~PAGE_MASK)) >> PAGE_SHIFT;
 128
 129         uiomr->owning_mm = mm = current->mm;
 130         down_write(&mm->mmap_sem);
 131
 132         locked = npages + current->mm->pinned_vm;
 133         lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
 134
 135         if ((locked > lock_limit) && !capable(CAP_IPC_LOCK)) {
 136                 ret = -ENOMEM;
 137                 goto out;
 138         }
 139
 140         flags = IOMMU_READ | IOMMU_CACHE;
 141         flags |= (writable) ? IOMMU_WRITE : 0;
 142         gup_flags = FOLL_WRITE;
 143         gup_flags |= (writable) ? 0 : FOLL_FORCE;
 144         cur_base = addr & PAGE_MASK;
 145         ret = 0;
 146
 147         while (npages) {
 148                 ret = get_user_pages_longterm(cur_base,
 149                                         min_t(unsigned long, npages,
 150                                         PAGE_SIZE / sizeof(struct page *)),
 151                                         gup_flags, page_list, NULL);
 152
 153                 if (ret < 0)
 154                         goto out;
 155
 156                 npages -= ret;
 157                 off = 0;
 158
 159                 while (ret) {
 160                         chunk = kmalloc(sizeof(*chunk) +
 161                                         sizeof(struct scatterlist) *
 162                                         min_t(int, ret, USNIC_UIOM_PAGE_CHUNK),
 163                                         GFP_KERNEL);
 164                         if (!chunk) {
 165                                 ret = -ENOMEM;
 166                                 goto out;
 167                         }
 168
 169                         chunk->nents = min_t(int, ret, USNIC_UIOM_PAGE_CHUNK);
 170                         sg_init_table(chunk->page_list, chunk->nents);
 171                         for_each_sg(chunk->page_list, sg, chunk->nents, i) {
 172                                 sg_set_page(sg, page_list[i + off],
 173                                                 PAGE_SIZE, 0);
 174                                 pa = sg_phys(sg);
 175                                 usnic_dbg("va: 0x%lx pa: %pa\n",
 176                                                 cur_base + i*PAGE_SIZE, &pa);
 177                         }
 178                         cur_base += chunk->nents * PAGE_SIZE;
 179                         ret -= chunk->nents;
 180                         off += chunk->nents;
 181                         list_add_tail(&chunk->list, chunk_list);
 182                 }
 183
 184                 ret = 0;
 185         }
 186
 187 out:
 188         if (ret < 0)
 189                 usnic_uiom_put_pages(chunk_list, 0);
 190         else {
 191                 mm->pinned_vm = locked;
 192                 mmgrab(uiomr->owning_mm);
 193         }
 194
 195         up_write(&mm->mmap_sem);
 196         free_page((unsigned long) page_list);
 197         return ret;
 198 }
 199
 200 static void usnic_uiom_unmap_sorted_intervals(struct list_head *intervals,
 201                                                 struct usnic_uiom_pd *pd)
 202 {
 203         struct usnic_uiom_interval_node *interval, *tmp;
 204         long unsigned va, size;
 205
 206         list_for_each_entry_safe(interval, tmp, intervals, link) {
 207                 va = interval->start << PAGE_SHIFT;
 208                 size = ((interval->last - interval->start) + 1) << PAGE_SHIFT;
 209                 while (size > 0) {
 210                         /* Workaround for RH 970401 */
 211                         usnic_dbg("va 0x%lx size 0x%lx", va, PAGE_SIZE);
 212                         iommu_unmap(pd->domain, va, PAGE_SIZE);
 213                         va += PAGE_SIZE;
 214                         size -= PAGE_SIZE;
 215                 }
 216         }
 217 }
 218
 219 static void __usnic_uiom_reg_release(struct usnic_uiom_pd *pd,
 220                                         struct usnic_uiom_reg *uiomr,
 221                                         int dirty)
 222 {
 223         int npages;
 224         unsigned long vpn_start, vpn_last;
 225         struct usnic_uiom_interval_node *interval, *tmp;
 226         int writable = 0;
 227         LIST_HEAD(rm_intervals);
 228
 229         npages = PAGE_ALIGN(uiomr->length + uiomr->offset) >> PAGE_SHIFT;
 230         vpn_start = (uiomr->va & PAGE_MASK) >> PAGE_SHIFT;
 231         vpn_last = vpn_start + npages - 1;
 232
 233         spin_lock(&pd->lock);
 234         usnic_uiom_remove_interval(&pd->root, vpn_start,
 235                                         vpn_last, &rm_intervals);
 236         usnic_uiom_unmap_sorted_intervals(&rm_intervals, pd);
 237
 238         list_for_each_entry_safe(interval, tmp, &rm_intervals, link) {
 239                 if (interval->flags & IOMMU_WRITE)
 240                         writable = 1;
 241                 list_del(&interval->link);
 242                 kfree(interval);
 243         }
 244
 245         usnic_uiom_put_pages(&uiomr->chunk_list, dirty & writable);
 246         spin_unlock(&pd->lock);
 247 }
 248
 249 static int usnic_uiom_map_sorted_intervals(struct list_head *intervals,
 250                                                 struct usnic_uiom_reg *uiomr)
 251 {
 252         int i, err;
 253         size_t size;
 254         struct usnic_uiom_chunk *chunk;
 255         struct usnic_uiom_interval_node *interval_node;
 256         dma_addr_t pa;
 257         dma_addr_t pa_start = 0;
 258         dma_addr_t pa_end = 0;
 259         long int va_start = -EINVAL;
 260         struct usnic_uiom_pd *pd = uiomr->pd;
 261         long int va = uiomr->va & PAGE_MASK;
 262         int flags = IOMMU_READ | IOMMU_CACHE;
 263
 264         flags |= (uiomr->writable) ? IOMMU_WRITE : 0;
 265         chunk = list_first_entry(&uiomr->chunk_list, struct usnic_uiom_chunk,
 266                                                                         list);
 267         list_for_each_entry(interval_node, intervals, link) {
 268 iter_chunk:
 269                 for (i = 0; i < chunk->nents; i++, va += PAGE_SIZE) {
 270                         pa = sg_phys(&chunk->page_list[i]);
 271                         if ((va >> PAGE_SHIFT) < interval_node->start)
 272                                 continue;
 273
 274                         if ((va >> PAGE_SHIFT) == interval_node->start) {
 275                                 /* First page of the interval */
 276                                 va_start = va;
 277                                 pa_start = pa;
 278                                 pa_end = pa;
 279                         }
 280
 281                         WARN_ON(va_start == -EINVAL);
 282
 283                         if ((pa_end + PAGE_SIZE != pa) &&
 284                                         (pa != pa_start)) {
 285                                 /* PAs are not contiguous */
 286                                 size = pa_end - pa_start + PAGE_SIZE;
 287                                 usnic_dbg("va 0x%lx pa %pa size 0x%zx flags 0x%x",
 288                                         va_start, &pa_start, size, flags);
 289                                 err = iommu_map(pd->domain, va_start, pa_start,
 290                                                         size, flags);
 291                                 if (err) {
 292                                         usnic_err("Failed to map va 0x%lx pa %pa size 0x%zx with err %d\n",
 293                                                 va_start, &pa_start, size, err);
 294                                         goto err_out;
 295                                 }
 296                                 va_start = va;
 297                                 pa_start = pa;
 298                                 pa_end = pa;
 299                         }
 300
 301                         if ((va >> PAGE_SHIFT) == interval_node->last) {
 302                                 /* Last page of the interval */
 303                                 size = pa - pa_start + PAGE_SIZE;
 304                                 usnic_dbg("va 0x%lx pa %pa size 0x%zx flags 0x%x\n",
 305                                         va_start, &pa_start, size, flags);
 306                                 err = iommu_map(pd->domain, va_start, pa_start,
 307                                                 size, flags);
 308                                 if (err) {
 309                                         usnic_err("Failed to map va 0x%lx pa %pa size 0x%zx with err %d\n",
 310                                                 va_start, &pa_start, size, err);
 311                                         goto err_out;
 312                                 }
 313                                 break;
 314                         }
 315
 316                         if (pa != pa_start)
 317                                 pa_end += PAGE_SIZE;
 318                 }
 319
 320                 if (i == chunk->nents) {
 321                         /*
 322                          * Hit last entry of the chunk,
 323                          * hence advance to next chunk
 324                          */
 325                         chunk = list_first_entry(&chunk->list,
 326                                                         struct usnic_uiom_chunk,
 327                                                         list);
 328                         goto iter_chunk;
 329                 }
 330         }
 331
 332         return 0;
 333
 334 err_out:
 335         usnic_uiom_unmap_sorted_intervals(intervals, pd);
 336         return err;
 337 }
 338
 339 struct usnic_uiom_reg *usnic_uiom_reg_get(struct usnic_uiom_pd *pd,
 340                                                 unsigned long addr, size_t size,
 341                                                 int writable, int dmasync)
 342 {
 343         struct usnic_uiom_reg *uiomr;
 344         unsigned long va_base, vpn_start, vpn_last;
 345         unsigned long npages;
 346         int offset, err;
 347         LIST_HEAD(sorted_diff_intervals);
 348
 349         /*
 350          * Intel IOMMU map throws an error if a translation entry is
 351          * changed from read to write.  This module may not unmap
 352          * and then remap the entry after fixing the permission
 353          * b/c this open up a small windows where hw DMA may page fault
 354          * Hence, make all entries to be writable.
 355          */
 356         writable = 1;
 357
 358         va_base = addr & PAGE_MASK;
 359         offset = addr & ~PAGE_MASK;
 360         npages = PAGE_ALIGN(size + offset) >> PAGE_SHIFT;
 361         vpn_start = (addr & PAGE_MASK) >> PAGE_SHIFT;
 362         vpn_last = vpn_start + npages - 1;
 363
 364         uiomr = kmalloc(sizeof(*uiomr), GFP_KERNEL);
 365         if (!uiomr)
 366                 return ERR_PTR(-ENOMEM);
 367
 368         uiomr->va = va_base;
 369         uiomr->offset = offset;
 370         uiomr->length = size;
 371         uiomr->writable = writable;
 372         uiomr->pd = pd;
 373
 374         err = usnic_uiom_get_pages(addr, size, writable, dmasync,
 375                                    uiomr);
 376         if (err) {
 377                 usnic_err("Failed get_pages vpn [0x%lx,0x%lx] err %d\n",
 378                                 vpn_start, vpn_last, err);
 379                 goto out_free_uiomr;
 380         }
 381
 382         spin_lock(&pd->lock);
 383         err = usnic_uiom_get_intervals_diff(vpn_start, vpn_last,
 384                                                 (writable) ? IOMMU_WRITE : 0,
 385                                                 IOMMU_WRITE,
 386                                                 &pd->root,
 387                                                 &sorted_diff_intervals);
 388         if (err) {
 389                 usnic_err("Failed disjoint interval vpn [0x%lx,0x%lx] err %d\n",
 390                                                 vpn_start, vpn_last, err);
 391                 goto out_put_pages;
 392         }
 393
 394         err = usnic_uiom_map_sorted_intervals(&sorted_diff_intervals, uiomr);
 395         if (err) {
 396                 usnic_err("Failed map interval vpn [0x%lx,0x%lx] err %d\n",
 397                                                 vpn_start, vpn_last, err);
 398                 goto out_put_intervals;
 399
 400         }
 401
 402         err = usnic_uiom_insert_interval(&pd->root, vpn_start, vpn_last,
 403                                         (writable) ? IOMMU_WRITE : 0);
 404         if (err) {
 405                 usnic_err("Failed insert interval vpn [0x%lx,0x%lx] err %d\n",
 406                                                 vpn_start, vpn_last, err);
 407                 goto out_unmap_intervals;
 408         }
 409
 410         usnic_uiom_put_interval_set(&sorted_diff_intervals);
 411         spin_unlock(&pd->lock);
 412
 413         return uiomr;
 414
 415 out_unmap_intervals:
 416         usnic_uiom_unmap_sorted_intervals(&sorted_diff_intervals, pd);
 417 out_put_intervals:
 418         usnic_uiom_put_interval_set(&sorted_diff_intervals);
 419 out_put_pages:
 420         usnic_uiom_put_pages(&uiomr->chunk_list, 0);
 421         spin_unlock(&pd->lock);
 422         mmdrop(uiomr->owning_mm);
 423 out_free_uiomr:
 424         kfree(uiomr);
 425         return ERR_PTR(err);
 426 }
 427
 428 static void __usnic_uiom_release_tail(struct usnic_uiom_reg *uiomr)
 429 {
 430         mmdrop(uiomr->owning_mm);
 431         kfree(uiomr);
 432 }
 433
 434 static inline size_t usnic_uiom_num_pages(struct usnic_uiom_reg *uiomr)
 435 {
 436         return PAGE_ALIGN(uiomr->length + uiomr->offset) >> PAGE_SHIFT;
 437 }
 438
 439 static void usnic_uiom_release_defer(struct work_struct *work)
 440 {
 441         struct usnic_uiom_reg *uiomr =
 442                 container_of(work, struct usnic_uiom_reg, work);
 443
 444         down_write(&uiomr->owning_mm->mmap_sem);
 445         uiomr->owning_mm->pinned_vm -= usnic_uiom_num_pages(uiomr);
 446         up_write(&uiomr->owning_mm->mmap_sem);
 447
 448         __usnic_uiom_release_tail(uiomr);
 449 }
 450
 451 void usnic_uiom_reg_release(struct usnic_uiom_reg *uiomr,
 452                             struct ib_ucontext *context)
 453 {
 454         __usnic_uiom_reg_release(uiomr->pd, uiomr, 1);
 455
 456         /*
 457          * We may be called with the mm's mmap_sem already held.  This
 458          * can happen when a userspace munmap() is the call that drops
 459          * the last reference to our file and calls our release
 460          * method.  If there are memory regions to destroy, we'll end
 461          * up here and not be able to take the mmap_sem.  In that case
 462          * we defer the vm_locked accounting to a workqueue.
 463          */
 464         if (context->closing) {
 465                 if (!down_write_trylock(&uiomr->owning_mm->mmap_sem)) {
 466                         INIT_WORK(&uiomr->work, usnic_uiom_release_defer);
 467                         queue_work(usnic_uiom_wq, &uiomr->work);
 468                         return;
 469                 }
 470         } else {
 471                 down_write(&uiomr->owning_mm->mmap_sem);
 472         }
 473         uiomr->owning_mm->pinned_vm -= usnic_uiom_num_pages(uiomr);
 474         up_write(&uiomr->owning_mm->mmap_sem);
 475
 476         __usnic_uiom_release_tail(uiomr);
 477 }
 478
 479 struct usnic_uiom_pd *usnic_uiom_alloc_pd(void)
 480 {
 481         struct usnic_uiom_pd *pd;
 482         void *domain;
 483
 484         pd = kzalloc(sizeof(*pd), GFP_KERNEL);
 485         if (!pd)
 486                 return ERR_PTR(-ENOMEM);
 487
 488         pd->domain = domain = iommu_domain_alloc(&pci_bus_type);
 489         if (!domain) {
 490                 usnic_err("Failed to allocate IOMMU domain");
 491                 kfree(pd);
 492                 return ERR_PTR(-ENOMEM);
 493         }
 494
 495         iommu_set_fault_handler(pd->domain, usnic_uiom_dma_fault, NULL);
 496
 497         spin_lock_init(&pd->lock);
 498         INIT_LIST_HEAD(&pd->devs);
 499
 500         return pd;
 501 }
 502
 503 void usnic_uiom_dealloc_pd(struct usnic_uiom_pd *pd)
 504 {
 505         iommu_domain_free(pd->domain);
 506         kfree(pd);
 507 }
 508
 509 int usnic_uiom_attach_dev_to_pd(struct usnic_uiom_pd *pd, struct device *dev)
 510 {
 511         struct usnic_uiom_dev *uiom_dev;
 512         int err;
 513
 514         uiom_dev = kzalloc(sizeof(*uiom_dev), GFP_ATOMIC);
 515         if (!uiom_dev)
 516                 return -ENOMEM;
 517         uiom_dev->dev = dev;
 518
 519         err = iommu_attach_device(pd->domain, dev);
 520         if (err)
 521                 goto out_free_dev;
 522
 523         if (!iommu_capable(dev->bus, IOMMU_CAP_CACHE_COHERENCY)) {
 524                 usnic_err("IOMMU of %s does not support cache coherency\n",
 525                                 dev_name(dev));
 526                 err = -EINVAL;
 527                 goto out_detach_device;
 528         }
 529
 530         spin_lock(&pd->lock);
 531         list_add_tail(&uiom_dev->link, &pd->devs);
 532         pd->dev_cnt++;
 533         spin_unlock(&pd->lock);
 534
 535         return 0;
 536
 537 out_detach_device:
 538         iommu_detach_device(pd->domain, dev);
 539 out_free_dev:
 540         kfree(uiom_dev);
 541         return err;
 542 }
 543
 544 void usnic_uiom_detach_dev_from_pd(struct usnic_uiom_pd *pd, struct device *dev)
 545 {
 546         struct usnic_uiom_dev *uiom_dev;
 547         int found = 0;
 548
 549         spin_lock(&pd->lock);
 550         list_for_each_entry(uiom_dev, &pd->devs, link) {
 551                 if (uiom_dev->dev == dev) {
 552                         found = 1;
 553                         break;
 554                 }
 555         }
 556
 557         if (!found) {
 558                 usnic_err("Unable to free dev %s - not found\n",
 559                                 dev_name(dev));
 560                 spin_unlock(&pd->lock);
 561                 return;
 562         }
 563
 564         list_del(&uiom_dev->link);
 565         pd->dev_cnt--;
 566         spin_unlock(&pd->lock);
 567
 568         return iommu_detach_device(pd->domain, dev);
 569 }
 570
 571 struct device **usnic_uiom_get_dev_list(struct usnic_uiom_pd *pd)
 572 {
 573         struct usnic_uiom_dev *uiom_dev;
 574         struct device **devs;
 575         int i = 0;
 576
 577         spin_lock(&pd->lock);
 578         devs = kcalloc(pd->dev_cnt + 1, sizeof(*devs), GFP_ATOMIC);
 579         if (!devs) {
 580                 devs = ERR_PTR(-ENOMEM);
 581                 goto out;
 582         }
 583
 584         list_for_each_entry(uiom_dev, &pd->devs, link) {
 585                 devs[i++] = uiom_dev->dev;
 586         }
 587 out:
 588         spin_unlock(&pd->lock);
 589         return devs;
 590 }
 591
 592 void usnic_uiom_free_dev_list(struct device **devs)
 593 {
 594         kfree(devs);
 595 }
 596
 597 int usnic_uiom_init(char *drv_name)
 598 {
 599         if (!iommu_present(&pci_bus_type)) {
 600                 usnic_err("IOMMU required but not present or enabled.  USNIC QPs will not function w/o enabling IOMMU\n");
 601                 return -EPERM;
 602         }
 603
 604         usnic_uiom_wq = create_workqueue(drv_name);
 605         if (!usnic_uiom_wq) {
 606                 usnic_err("Unable to alloc wq for drv %s\n", drv_name);
 607                 return -ENOMEM;
 608         }
 609
 610         return 0;
 611 }
 612
 613 void usnic_uiom_fini(void)
 614 {
 615         flush_workqueue(usnic_uiom_wq);
 616         destroy_workqueue(usnic_uiom_wq);
 617 }