drivers/iommu/iova.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright © 2006-2009, Intel Corporation.
   4  *
   5  * Author: Anil S Keshavamurthy <anil.s.keshavamurthy@intel.com>
   6  */
   7
   8 #include <linux/iova.h>
   9 #include <linux/module.h>
  10 #include <linux/slab.h>
  11 #include <linux/smp.h>
  12 #include <linux/bitops.h>
  13 #include <linux/cpu.h>
  14
  15 /* The anchor node sits above the top of the usable address space */
  16 #define IOVA_ANCHOR     ~0UL
  17
  18 static bool iova_rcache_insert(struct iova_domain *iovad,
  19                                unsigned long pfn,
  20                                unsigned long size);
  21 static unsigned long iova_rcache_get(struct iova_domain *iovad,
  22                                      unsigned long size,
  23                                      unsigned long limit_pfn);
  24 static void init_iova_rcaches(struct iova_domain *iovad);
  25 static void free_iova_rcaches(struct iova_domain *iovad);
  26 static void fq_destroy_all_entries(struct iova_domain *iovad);
  27 static void fq_flush_timeout(struct timer_list *t);
  28 static void free_global_cached_iovas(struct iova_domain *iovad);
  29
  30 void
  31 init_iova_domain(struct iova_domain *iovad, unsigned long granule,
  32         unsigned long start_pfn)
  33 {
  34         /*
  35          * IOVA granularity will normally be equal to the smallest
  36          * supported IOMMU page size; both *must* be capable of
  37          * representing individual CPU pages exactly.
  38          */
  39         BUG_ON((granule > PAGE_SIZE) || !is_power_of_2(granule));
  40
  41         spin_lock_init(&iovad->iova_rbtree_lock);
  42         iovad->rbroot = RB_ROOT;
  43         iovad->cached_node = &iovad->anchor.node;
  44         iovad->cached32_node = &iovad->anchor.node;
  45         iovad->granule = granule;
  46         iovad->start_pfn = start_pfn;
  47         iovad->dma_32bit_pfn = 1UL << (32 - iova_shift(iovad));
  48         iovad->max32_alloc_size = iovad->dma_32bit_pfn;
  49         iovad->flush_cb = NULL;
  50         iovad->fq = NULL;
  51         iovad->anchor.pfn_lo = iovad->anchor.pfn_hi = IOVA_ANCHOR;
  52         rb_link_node(&iovad->anchor.node, NULL, &iovad->rbroot.rb_node);
  53         rb_insert_color(&iovad->anchor.node, &iovad->rbroot);
  54         init_iova_rcaches(iovad);
  55 }
  56 EXPORT_SYMBOL_GPL(init_iova_domain);
  57
  58 bool has_iova_flush_queue(struct iova_domain *iovad)
  59 {
  60         return !!iovad->fq;
  61 }
  62
  63 static void free_iova_flush_queue(struct iova_domain *iovad)
  64 {
  65         if (!has_iova_flush_queue(iovad))
  66                 return;
  67
  68         if (timer_pending(&iovad->fq_timer))
  69                 del_timer(&iovad->fq_timer);
  70
  71         fq_destroy_all_entries(iovad);
  72
  73         free_percpu(iovad->fq);
  74
  75         iovad->fq         = NULL;
  76         iovad->flush_cb   = NULL;
  77         iovad->entry_dtor = NULL;
  78 }
  79
  80 int init_iova_flush_queue(struct iova_domain *iovad,
  81                           iova_flush_cb flush_cb, iova_entry_dtor entry_dtor)
  82 {
  83         struct iova_fq __percpu *queue;
  84         int cpu;
  85
  86         atomic64_set(&iovad->fq_flush_start_cnt,  0);
  87         atomic64_set(&iovad->fq_flush_finish_cnt, 0);
  88
  89         queue = alloc_percpu(struct iova_fq);
  90         if (!queue)
  91                 return -ENOMEM;
  92
  93         iovad->flush_cb   = flush_cb;
  94         iovad->entry_dtor = entry_dtor;
  95
  96         for_each_possible_cpu(cpu) {
  97                 struct iova_fq *fq;
  98
  99                 fq = per_cpu_ptr(queue, cpu);
 100                 fq->head = 0;
 101                 fq->tail = 0;
 102
 103                 spin_lock_init(&fq->lock);
 104         }
 105
 106         smp_wmb();
 107
 108         iovad->fq = queue;
 109
 110         timer_setup(&iovad->fq_timer, fq_flush_timeout, 0);
 111         atomic_set(&iovad->fq_timer_on, 0);
 112
 113         return 0;
 114 }
 115 EXPORT_SYMBOL_GPL(init_iova_flush_queue);
 116
 117 static struct rb_node *
 118 __get_cached_rbnode(struct iova_domain *iovad, unsigned long limit_pfn)
 119 {
 120         if (limit_pfn <= iovad->dma_32bit_pfn)
 121                 return iovad->cached32_node;
 122
 123         return iovad->cached_node;
 124 }
 125
 126 static void
 127 __cached_rbnode_insert_update(struct iova_domain *iovad, struct iova *new)
 128 {
 129         if (new->pfn_hi < iovad->dma_32bit_pfn)
 130                 iovad->cached32_node = &new->node;
 131         else
 132                 iovad->cached_node = &new->node;
 133 }
 134
 135 static void
 136 __cached_rbnode_delete_update(struct iova_domain *iovad, struct iova *free)
 137 {
 138         struct iova *cached_iova;
 139
 140         cached_iova = rb_entry(iovad->cached32_node, struct iova, node);
 141         if (free == cached_iova ||
 142             (free->pfn_hi < iovad->dma_32bit_pfn &&
 143              free->pfn_lo >= cached_iova->pfn_lo)) {
 144                 iovad->cached32_node = rb_next(&free->node);
 145                 iovad->max32_alloc_size = iovad->dma_32bit_pfn;
 146         }
 147
 148         cached_iova = rb_entry(iovad->cached_node, struct iova, node);
 149         if (free->pfn_lo >= cached_iova->pfn_lo)
 150                 iovad->cached_node = rb_next(&free->node);
 151 }
 152
 153 /* Insert the iova into domain rbtree by holding writer lock */
 154 static void
 155 iova_insert_rbtree(struct rb_root *root, struct iova *iova,
 156                    struct rb_node *start)
 157 {
 158         struct rb_node **new, *parent = NULL;
 159
 160         new = (start) ? &start : &(root->rb_node);
 161         /* Figure out where to put new node */
 162         while (*new) {
 163                 struct iova *this = rb_entry(*new, struct iova, node);
 164
 165                 parent = *new;
 166
 167                 if (iova->pfn_lo < this->pfn_lo)
 168                         new = &((*new)->rb_left);
 169                 else if (iova->pfn_lo > this->pfn_lo)
 170                         new = &((*new)->rb_right);
 171                 else {
 172                         WARN_ON(1); /* this should not happen */
 173                         return;
 174                 }
 175         }
 176         /* Add new node and rebalance tree. */
 177         rb_link_node(&iova->node, parent, new);
 178         rb_insert_color(&iova->node, root);
 179 }
 180
 181 static int __alloc_and_insert_iova_range(struct iova_domain *iovad,
 182                 unsigned long size, unsigned long limit_pfn,
 183                         struct iova *new, bool size_aligned)
 184 {
 185         struct rb_node *curr, *prev;
 186         struct iova *curr_iova;
 187         unsigned long flags;
 188         unsigned long new_pfn, retry_pfn;
 189         unsigned long align_mask = ~0UL;
 190         unsigned long high_pfn = limit_pfn, low_pfn = iovad->start_pfn;
 191
 192         if (size_aligned)
 193                 align_mask <<= fls_long(size - 1);
 194
 195         /* Walk the tree backwards */
 196         spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
 197         if (limit_pfn <= iovad->dma_32bit_pfn &&
 198                         size >= iovad->max32_alloc_size)
 199                 goto iova32_full;
 200
 201         curr = __get_cached_rbnode(iovad, limit_pfn);
 202         curr_iova = rb_entry(curr, struct iova, node);
 203         retry_pfn = curr_iova->pfn_hi + 1;
 204
 205 retry:
 206         do {
 207                 high_pfn = min(high_pfn, curr_iova->pfn_lo);
 208                 new_pfn = (high_pfn - size) & align_mask;
 209                 prev = curr;
 210                 curr = rb_prev(curr);
 211                 curr_iova = rb_entry(curr, struct iova, node);
 212         } while (curr && new_pfn <= curr_iova->pfn_hi && new_pfn >= low_pfn);
 213
 214         if (high_pfn < size || new_pfn < low_pfn) {
 215                 if (low_pfn == iovad->start_pfn && retry_pfn < limit_pfn) {
 216                         high_pfn = limit_pfn;
 217                         low_pfn = retry_pfn;
 218                         curr = &iovad->anchor.node;
 219                         curr_iova = rb_entry(curr, struct iova, node);
 220                         goto retry;
 221                 }
 222                 iovad->max32_alloc_size = size;
 223                 goto iova32_full;
 224         }
 225
 226         /* pfn_lo will point to size aligned address if size_aligned is set */
 227         new->pfn_lo = new_pfn;
 228         new->pfn_hi = new->pfn_lo + size - 1;
 229
 230         /* If we have 'prev', it's a valid place to start the insertion. */
 231         iova_insert_rbtree(&iovad->rbroot, new, prev);
 232         __cached_rbnode_insert_update(iovad, new);
 233
 234         spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
 235         return 0;
 236
 237 iova32_full:
 238         spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
 239         return -ENOMEM;
 240 }
 241
 242 static struct kmem_cache *iova_cache;
 243 static unsigned int iova_cache_users;
 244 static DEFINE_MUTEX(iova_cache_mutex);
 245
 246 static struct iova *alloc_iova_mem(void)
 247 {
 248         return kmem_cache_zalloc(iova_cache, GFP_ATOMIC | __GFP_NOWARN);
 249 }
 250
 251 static void free_iova_mem(struct iova *iova)
 252 {
 253         if (iova->pfn_lo != IOVA_ANCHOR)
 254                 kmem_cache_free(iova_cache, iova);
 255 }
 256
 257 int iova_cache_get(void)
 258 {
 259         mutex_lock(&iova_cache_mutex);
 260         if (!iova_cache_users) {
 261                 iova_cache = kmem_cache_create(
 262                         "iommu_iova", sizeof(struct iova), 0,
 263                         SLAB_HWCACHE_ALIGN, NULL);
 264                 if (!iova_cache) {
 265                         mutex_unlock(&iova_cache_mutex);
 266                         pr_err("Couldn't create iova cache\n");
 267                         return -ENOMEM;
 268                 }
 269         }
 270
 271         iova_cache_users++;
 272         mutex_unlock(&iova_cache_mutex);
 273
 274         return 0;
 275 }
 276 EXPORT_SYMBOL_GPL(iova_cache_get);
 277
 278 void iova_cache_put(void)
 279 {
 280         mutex_lock(&iova_cache_mutex);
 281         if (WARN_ON(!iova_cache_users)) {
 282                 mutex_unlock(&iova_cache_mutex);
 283                 return;
 284         }
 285         iova_cache_users--;
 286         if (!iova_cache_users)
 287                 kmem_cache_destroy(iova_cache);
 288         mutex_unlock(&iova_cache_mutex);
 289 }
 290 EXPORT_SYMBOL_GPL(iova_cache_put);
 291
 292 /**
 293  * alloc_iova - allocates an iova
 294  * @iovad: - iova domain in question
 295  * @size: - size of page frames to allocate
 296  * @limit_pfn: - max limit address
 297  * @size_aligned: - set if size_aligned address range is required
 298  * This function allocates an iova in the range iovad->start_pfn to limit_pfn,
 299  * searching top-down from limit_pfn to iovad->start_pfn. If the size_aligned
 300  * flag is set then the allocated address iova->pfn_lo will be naturally
 301  * aligned on roundup_power_of_two(size).
 302  */
 303 struct iova *
 304 alloc_iova(struct iova_domain *iovad, unsigned long size,
 305         unsigned long limit_pfn,
 306         bool size_aligned)
 307 {
 308         struct iova *new_iova;
 309         int ret;
 310
 311         new_iova = alloc_iova_mem();
 312         if (!new_iova)
 313                 return NULL;
 314
 315         ret = __alloc_and_insert_iova_range(iovad, size, limit_pfn + 1,
 316                         new_iova, size_aligned);
 317
 318         if (ret) {
 319                 free_iova_mem(new_iova);
 320                 return NULL;
 321         }
 322
 323         return new_iova;
 324 }
 325 EXPORT_SYMBOL_GPL(alloc_iova);
 326
 327 static struct iova *
 328 private_find_iova(struct iova_domain *iovad, unsigned long pfn)
 329 {
 330         struct rb_node *node = iovad->rbroot.rb_node;
 331
 332         assert_spin_locked(&iovad->iova_rbtree_lock);
 333
 334         while (node) {
 335                 struct iova *iova = rb_entry(node, struct iova, node);
 336
 337                 if (pfn < iova->pfn_lo)
 338                         node = node->rb_left;
 339                 else if (pfn > iova->pfn_hi)
 340                         node = node->rb_right;
 341                 else
 342                         return iova;    /* pfn falls within iova's range */
 343         }
 344
 345         return NULL;
 346 }
 347
 348 static void private_free_iova(struct iova_domain *iovad, struct iova *iova)
 349 {
 350         assert_spin_locked(&iovad->iova_rbtree_lock);
 351         __cached_rbnode_delete_update(iovad, iova);
 352         rb_erase(&iova->node, &iovad->rbroot);
 353         free_iova_mem(iova);
 354 }
 355
 356 /**
 357  * find_iova - finds an iova for a given pfn
 358  * @iovad: - iova domain in question.
 359  * @pfn: - page frame number
 360  * This function finds and returns an iova belonging to the
 361  * given domain which matches the given pfn.
 362  */
 363 struct iova *find_iova(struct iova_domain *iovad, unsigned long pfn)
 364 {
 365         unsigned long flags;
 366         struct iova *iova;
 367
 368         /* Take the lock so that no other thread is manipulating the rbtree */
 369         spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
 370         iova = private_find_iova(iovad, pfn);
 371         spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
 372         return iova;
 373 }
 374 EXPORT_SYMBOL_GPL(find_iova);
 375
 376 /**
 377  * __free_iova - frees the given iova
 378  * @iovad: iova domain in question.
 379  * @iova: iova in question.
 380  * Frees the given iova belonging to the giving domain
 381  */
 382 void
 383 __free_iova(struct iova_domain *iovad, struct iova *iova)
 384 {
 385         unsigned long flags;
 386
 387         spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
 388         private_free_iova(iovad, iova);
 389         spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
 390 }
 391 EXPORT_SYMBOL_GPL(__free_iova);
 392
 393 /**
 394  * free_iova - finds and frees the iova for a given pfn
 395  * @iovad: - iova domain in question.
 396  * @pfn: - pfn that is allocated previously
 397  * This functions finds an iova for a given pfn and then
 398  * frees the iova from that domain.
 399  */
 400 void
 401 free_iova(struct iova_domain *iovad, unsigned long pfn)
 402 {
 403         unsigned long flags;
 404         struct iova *iova;
 405
 406         spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
 407         iova = private_find_iova(iovad, pfn);
 408         if (iova)
 409                 private_free_iova(iovad, iova);
 410         spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
 411
 412 }
 413 EXPORT_SYMBOL_GPL(free_iova);
 414
 415 /**
 416  * alloc_iova_fast - allocates an iova from rcache
 417  * @iovad: - iova domain in question
 418  * @size: - size of page frames to allocate
 419  * @limit_pfn: - max limit address
 420  * @flush_rcache: - set to flush rcache on regular allocation failure
 421  * This function tries to satisfy an iova allocation from the rcache,
 422  * and falls back to regular allocation on failure. If regular allocation
 423  * fails too and the flush_rcache flag is set then the rcache will be flushed.
 424 */
 425 unsigned long
 426 alloc_iova_fast(struct iova_domain *iovad, unsigned long size,
 427                 unsigned long limit_pfn, bool flush_rcache)
 428 {
 429         unsigned long iova_pfn;
 430         struct iova *new_iova;
 431
 432         iova_pfn = iova_rcache_get(iovad, size, limit_pfn + 1);
 433         if (iova_pfn)
 434                 return iova_pfn;
 435
 436 retry:
 437         new_iova = alloc_iova(iovad, size, limit_pfn, true);
 438         if (!new_iova) {
 439                 unsigned int cpu;
 440
 441                 if (!flush_rcache)
 442                         return 0;
 443
 444                 /* Try replenishing IOVAs by flushing rcache. */
 445                 flush_rcache = false;
 446                 for_each_online_cpu(cpu)
 447                         free_cpu_cached_iovas(cpu, iovad);
 448                 free_global_cached_iovas(iovad);
 449                 goto retry;
 450         }
 451
 452         return new_iova->pfn_lo;
 453 }
 454 EXPORT_SYMBOL_GPL(alloc_iova_fast);
 455
 456 /**
 457  * free_iova_fast - free iova pfn range into rcache
 458  * @iovad: - iova domain in question.
 459  * @pfn: - pfn that is allocated previously
 460  * @size: - # of pages in range
 461  * This functions frees an iova range by trying to put it into the rcache,
 462  * falling back to regular iova deallocation via free_iova() if this fails.
 463  */
 464 void
 465 free_iova_fast(struct iova_domain *iovad, unsigned long pfn, unsigned long size)
 466 {
 467         if (iova_rcache_insert(iovad, pfn, size))
 468                 return;
 469
 470         free_iova(iovad, pfn);
 471 }
 472 EXPORT_SYMBOL_GPL(free_iova_fast);
 473
 474 #define fq_ring_for_each(i, fq) \
 475         for ((i) = (fq)->head; (i) != (fq)->tail; (i) = ((i) + 1) % IOVA_FQ_SIZE)
 476
 477 static inline bool fq_full(struct iova_fq *fq)
 478 {
 479         assert_spin_locked(&fq->lock);
 480         return (((fq->tail + 1) % IOVA_FQ_SIZE) == fq->head);
 481 }
 482
 483 static inline unsigned fq_ring_add(struct iova_fq *fq)
 484 {
 485         unsigned idx = fq->tail;
 486
 487         assert_spin_locked(&fq->lock);
 488
 489         fq->tail = (idx + 1) % IOVA_FQ_SIZE;
 490
 491         return idx;
 492 }
 493
 494 static void fq_ring_free(struct iova_domain *iovad, struct iova_fq *fq)
 495 {
 496         u64 counter = atomic64_read(&iovad->fq_flush_finish_cnt);
 497         unsigned idx;
 498
 499         assert_spin_locked(&fq->lock);
 500
 501         fq_ring_for_each(idx, fq) {
 502
 503                 if (fq->entries[idx].counter >= counter)
 504                         break;
 505
 506                 if (iovad->entry_dtor)
 507                         iovad->entry_dtor(fq->entries[idx].data);
 508
 509                 free_iova_fast(iovad,
 510                                fq->entries[idx].iova_pfn,
 511                                fq->entries[idx].pages);
 512
 513                 fq->head = (fq->head + 1) % IOVA_FQ_SIZE;
 514         }
 515 }
 516
 517 static void iova_domain_flush(struct iova_domain *iovad)
 518 {
 519         atomic64_inc(&iovad->fq_flush_start_cnt);
 520         iovad->flush_cb(iovad);
 521         atomic64_inc(&iovad->fq_flush_finish_cnt);
 522 }
 523
 524 static void fq_destroy_all_entries(struct iova_domain *iovad)
 525 {
 526         int cpu;
 527
 528         /*
 529          * This code runs when the iova_domain is being detroyed, so don't
 530          * bother to free iovas, just call the entry_dtor on all remaining
 531          * entries.
 532          */
 533         if (!iovad->entry_dtor)
 534                 return;
 535
 536         for_each_possible_cpu(cpu) {
 537                 struct iova_fq *fq = per_cpu_ptr(iovad->fq, cpu);
 538                 int idx;
 539
 540                 fq_ring_for_each(idx, fq)
 541                         iovad->entry_dtor(fq->entries[idx].data);
 542         }
 543 }
 544
 545 static void fq_flush_timeout(struct timer_list *t)
 546 {
 547         struct iova_domain *iovad = from_timer(iovad, t, fq_timer);
 548         int cpu;
 549
 550         atomic_set(&iovad->fq_timer_on, 0);
 551         iova_domain_flush(iovad);
 552
 553         for_each_possible_cpu(cpu) {
 554                 unsigned long flags;
 555                 struct iova_fq *fq;
 556
 557                 fq = per_cpu_ptr(iovad->fq, cpu);
 558                 spin_lock_irqsave(&fq->lock, flags);
 559                 fq_ring_free(iovad, fq);
 560                 spin_unlock_irqrestore(&fq->lock, flags);
 561         }
 562 }
 563
 564 void queue_iova(struct iova_domain *iovad,
 565                 unsigned long pfn, unsigned long pages,
 566                 unsigned long data)
 567 {
 568         struct iova_fq *fq = raw_cpu_ptr(iovad->fq);
 569         unsigned long flags;
 570         unsigned idx;
 571
 572         spin_lock_irqsave(&fq->lock, flags);
 573
 574         /*
 575          * First remove all entries from the flush queue that have already been
 576          * flushed out on another CPU. This makes the fq_full() check below less
 577          * likely to be true.
 578          */
 579         fq_ring_free(iovad, fq);
 580
 581         if (fq_full(fq)) {
 582                 iova_domain_flush(iovad);
 583                 fq_ring_free(iovad, fq);
 584         }
 585
 586         idx = fq_ring_add(fq);
 587
 588         fq->entries[idx].iova_pfn = pfn;
 589         fq->entries[idx].pages    = pages;
 590         fq->entries[idx].data     = data;
 591         fq->entries[idx].counter  = atomic64_read(&iovad->fq_flush_start_cnt);
 592
 593         spin_unlock_irqrestore(&fq->lock, flags);
 594
 595         /* Avoid false sharing as much as possible. */
 596         if (!atomic_read(&iovad->fq_timer_on) &&
 597             !atomic_xchg(&iovad->fq_timer_on, 1))
 598                 mod_timer(&iovad->fq_timer,
 599                           jiffies + msecs_to_jiffies(IOVA_FQ_TIMEOUT));
 600 }
 601 EXPORT_SYMBOL_GPL(queue_iova);
 602
 603 /**
 604  * put_iova_domain - destroys the iova domain
 605  * @iovad: - iova domain in question.
 606  * All the iova's in that domain are destroyed.
 607  */
 608 void put_iova_domain(struct iova_domain *iovad)
 609 {
 610         struct iova *iova, *tmp;
 611
 612         free_iova_flush_queue(iovad);
 613         free_iova_rcaches(iovad);
 614         rbtree_postorder_for_each_entry_safe(iova, tmp, &iovad->rbroot, node)
 615                 free_iova_mem(iova);
 616 }
 617 EXPORT_SYMBOL_GPL(put_iova_domain);
 618
 619 static int
 620 __is_range_overlap(struct rb_node *node,
 621         unsigned long pfn_lo, unsigned long pfn_hi)
 622 {
 623         struct iova *iova = rb_entry(node, struct iova, node);
 624
 625         if ((pfn_lo <= iova->pfn_hi) && (pfn_hi >= iova->pfn_lo))
 626                 return 1;
 627         return 0;
 628 }
 629
 630 static inline struct iova *
 631 alloc_and_init_iova(unsigned long pfn_lo, unsigned long pfn_hi)
 632 {
 633         struct iova *iova;
 634
 635         iova = alloc_iova_mem();
 636         if (iova) {
 637                 iova->pfn_lo = pfn_lo;
 638                 iova->pfn_hi = pfn_hi;
 639         }
 640
 641         return iova;
 642 }
 643
 644 static struct iova *
 645 __insert_new_range(struct iova_domain *iovad,
 646         unsigned long pfn_lo, unsigned long pfn_hi)
 647 {
 648         struct iova *iova;
 649
 650         iova = alloc_and_init_iova(pfn_lo, pfn_hi);
 651         if (iova)
 652                 iova_insert_rbtree(&iovad->rbroot, iova, NULL);
 653
 654         return iova;
 655 }
 656
 657 static void
 658 __adjust_overlap_range(struct iova *iova,
 659         unsigned long *pfn_lo, unsigned long *pfn_hi)
 660 {
 661         if (*pfn_lo < iova->pfn_lo)
 662                 iova->pfn_lo = *pfn_lo;
 663         if (*pfn_hi > iova->pfn_hi)
 664                 *pfn_lo = iova->pfn_hi + 1;
 665 }
 666
 667 /**
 668  * reserve_iova - reserves an iova in the given range
 669  * @iovad: - iova domain pointer
 670  * @pfn_lo: - lower page frame address
 671  * @pfn_hi:- higher pfn adderss
 672  * This function allocates reserves the address range from pfn_lo to pfn_hi so
 673  * that this address is not dished out as part of alloc_iova.
 674  */
 675 struct iova *
 676 reserve_iova(struct iova_domain *iovad,
 677         unsigned long pfn_lo, unsigned long pfn_hi)
 678 {
 679         struct rb_node *node;
 680         unsigned long flags;
 681         struct iova *iova;
 682         unsigned int overlap = 0;
 683
 684         /* Don't allow nonsensical pfns */
 685         if (WARN_ON((pfn_hi | pfn_lo) > (ULLONG_MAX >> iova_shift(iovad))))
 686                 return NULL;
 687
 688         spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
 689         for (node = rb_first(&iovad->rbroot); node; node = rb_next(node)) {
 690                 if (__is_range_overlap(node, pfn_lo, pfn_hi)) {
 691                         iova = rb_entry(node, struct iova, node);
 692                         __adjust_overlap_range(iova, &pfn_lo, &pfn_hi);
 693                         if ((pfn_lo >= iova->pfn_lo) &&
 694                                 (pfn_hi <= iova->pfn_hi))
 695                                 goto finish;
 696                         overlap = 1;
 697
 698                 } else if (overlap)
 699                                 break;
 700         }
 701
 702         /* We are here either because this is the first reserver node
 703          * or need to insert remaining non overlap addr range
 704          */
 705         iova = __insert_new_range(iovad, pfn_lo, pfn_hi);
 706 finish:
 707
 708         spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
 709         return iova;
 710 }
 711 EXPORT_SYMBOL_GPL(reserve_iova);
 712
 713 /**
 714  * copy_reserved_iova - copies the reserved between domains
 715  * @from: - source domain from where to copy
 716  * @to: - destination domin where to copy
 717  * This function copies reserved iova's from one domain to
 718  * other.
 719  */
 720 void
 721 copy_reserved_iova(struct iova_domain *from, struct iova_domain *to)
 722 {
 723         unsigned long flags;
 724         struct rb_node *node;
 725
 726         spin_lock_irqsave(&from->iova_rbtree_lock, flags);
 727         for (node = rb_first(&from->rbroot); node; node = rb_next(node)) {
 728                 struct iova *iova = rb_entry(node, struct iova, node);
 729                 struct iova *new_iova;
 730
 731                 if (iova->pfn_lo == IOVA_ANCHOR)
 732                         continue;
 733
 734                 new_iova = reserve_iova(to, iova->pfn_lo, iova->pfn_hi);
 735                 if (!new_iova)
 736                         pr_err("Reserve iova range %lx@%lx failed\n",
 737                                iova->pfn_lo, iova->pfn_lo);
 738         }
 739         spin_unlock_irqrestore(&from->iova_rbtree_lock, flags);
 740 }
 741 EXPORT_SYMBOL_GPL(copy_reserved_iova);
 742
 743 /*
 744  * Magazine caches for IOVA ranges.  For an introduction to magazines,
 745  * see the USENIX 2001 paper "Magazines and Vmem: Extending the Slab
 746  * Allocator to Many CPUs and Arbitrary Resources" by Bonwick and Adams.
 747  * For simplicity, we use a static magazine size and don't implement the
 748  * dynamic size tuning described in the paper.
 749  */
 750
 751 #define IOVA_MAG_SIZE 128
 752
 753 struct iova_magazine {
 754         unsigned long size;
 755         unsigned long pfns[IOVA_MAG_SIZE];
 756 };
 757
 758 struct iova_cpu_rcache {
 759         spinlock_t lock;
 760         struct iova_magazine *loaded;
 761         struct iova_magazine *prev;
 762 };
 763
 764 static struct iova_magazine *iova_magazine_alloc(gfp_t flags)
 765 {
 766         return kzalloc(sizeof(struct iova_magazine), flags);
 767 }
 768
 769 static void iova_magazine_free(struct iova_magazine *mag)
 770 {
 771         kfree(mag);
 772 }
 773
 774 static void
 775 iova_magazine_free_pfns(struct iova_magazine *mag, struct iova_domain *iovad)
 776 {
 777         unsigned long flags;
 778         int i;
 779
 780         if (!mag)
 781                 return;
 782
 783         spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
 784
 785         for (i = 0 ; i < mag->size; ++i) {
 786                 struct iova *iova = private_find_iova(iovad, mag->pfns[i]);
 787
 788                 if (WARN_ON(!iova))
 789                         continue;
 790
 791                 private_free_iova(iovad, iova);
 792         }
 793
 794         spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
 795
 796         mag->size = 0;
 797 }
 798
 799 static bool iova_magazine_full(struct iova_magazine *mag)
 800 {
 801         return (mag && mag->size == IOVA_MAG_SIZE);
 802 }
 803
 804 static bool iova_magazine_empty(struct iova_magazine *mag)
 805 {
 806         return (!mag || mag->size == 0);
 807 }
 808
 809 static unsigned long iova_magazine_pop(struct iova_magazine *mag,
 810                                        unsigned long limit_pfn)
 811 {
 812         int i;
 813         unsigned long pfn;
 814
 815         BUG_ON(iova_magazine_empty(mag));
 816
 817         /* Only fall back to the rbtree if we have no suitable pfns at all */
 818         for (i = mag->size - 1; mag->pfns[i] > limit_pfn; i--)
 819                 if (i == 0)
 820                         return 0;
 821
 822         /* Swap it to pop it */
 823         pfn = mag->pfns[i];
 824         mag->pfns[i] = mag->pfns[--mag->size];
 825
 826         return pfn;
 827 }
 828
 829 static void iova_magazine_push(struct iova_magazine *mag, unsigned long pfn)
 830 {
 831         BUG_ON(iova_magazine_full(mag));
 832
 833         mag->pfns[mag->size++] = pfn;
 834 }
 835
 836 static void init_iova_rcaches(struct iova_domain *iovad)
 837 {
 838         struct iova_cpu_rcache *cpu_rcache;
 839         struct iova_rcache *rcache;
 840         unsigned int cpu;
 841         int i;
 842
 843         for (i = 0; i < IOVA_RANGE_CACHE_MAX_SIZE; ++i) {
 844                 rcache = &iovad->rcaches[i];
 845                 spin_lock_init(&rcache->lock);
 846                 rcache->depot_size = 0;
 847                 rcache->cpu_rcaches = __alloc_percpu(sizeof(*cpu_rcache), cache_line_size());
 848                 if (WARN_ON(!rcache->cpu_rcaches))
 849                         continue;
 850                 for_each_possible_cpu(cpu) {
 851                         cpu_rcache = per_cpu_ptr(rcache->cpu_rcaches, cpu);
 852                         spin_lock_init(&cpu_rcache->lock);
 853                         cpu_rcache->loaded = iova_magazine_alloc(GFP_KERNEL);
 854                         cpu_rcache->prev = iova_magazine_alloc(GFP_KERNEL);
 855                 }
 856         }
 857 }
 858
 859 /*
 860  * Try inserting IOVA range starting with 'iova_pfn' into 'rcache', and
 861  * return true on success.  Can fail if rcache is full and we can't free
 862  * space, and free_iova() (our only caller) will then return the IOVA
 863  * range to the rbtree instead.
 864  */
 865 static bool __iova_rcache_insert(struct iova_domain *iovad,
 866                                  struct iova_rcache *rcache,
 867                                  unsigned long iova_pfn)
 868 {
 869         struct iova_magazine *mag_to_free = NULL;
 870         struct iova_cpu_rcache *cpu_rcache;
 871         bool can_insert = false;
 872         unsigned long flags;
 873
 874         cpu_rcache = raw_cpu_ptr(rcache->cpu_rcaches);
 875         spin_lock_irqsave(&cpu_rcache->lock, flags);
 876
 877         if (!iova_magazine_full(cpu_rcache->loaded)) {
 878                 can_insert = true;
 879         } else if (!iova_magazine_full(cpu_rcache->prev)) {
 880                 swap(cpu_rcache->prev, cpu_rcache->loaded);
 881                 can_insert = true;
 882         } else {
 883                 struct iova_magazine *new_mag = iova_magazine_alloc(GFP_ATOMIC);
 884
 885                 if (new_mag) {
 886                         spin_lock(&rcache->lock);
 887                         if (rcache->depot_size < MAX_GLOBAL_MAGS) {
 888                                 rcache->depot[rcache->depot_size++] =
 889                                                 cpu_rcache->loaded;
 890                         } else {
 891                                 mag_to_free = cpu_rcache->loaded;
 892                         }
 893                         spin_unlock(&rcache->lock);
 894
 895                         cpu_rcache->loaded = new_mag;
 896                         can_insert = true;
 897                 }
 898         }
 899
 900         if (can_insert)
 901                 iova_magazine_push(cpu_rcache->loaded, iova_pfn);
 902
 903         spin_unlock_irqrestore(&cpu_rcache->lock, flags);
 904
 905         if (mag_to_free) {
 906                 iova_magazine_free_pfns(mag_to_free, iovad);
 907                 iova_magazine_free(mag_to_free);
 908         }
 909
 910         return can_insert;
 911 }
 912
 913 static bool iova_rcache_insert(struct iova_domain *iovad, unsigned long pfn,
 914                                unsigned long size)
 915 {
 916         unsigned int log_size = order_base_2(size);
 917
 918         if (log_size >= IOVA_RANGE_CACHE_MAX_SIZE)
 919                 return false;
 920
 921         return __iova_rcache_insert(iovad, &iovad->rcaches[log_size], pfn);
 922 }
 923
 924 /*
 925  * Caller wants to allocate a new IOVA range from 'rcache'.  If we can
 926  * satisfy the request, return a matching non-NULL range and remove
 927  * it from the 'rcache'.
 928  */
 929 static unsigned long __iova_rcache_get(struct iova_rcache *rcache,
 930                                        unsigned long limit_pfn)
 931 {
 932         struct iova_cpu_rcache *cpu_rcache;
 933         unsigned long iova_pfn = 0;
 934         bool has_pfn = false;
 935         unsigned long flags;
 936
 937         cpu_rcache = raw_cpu_ptr(rcache->cpu_rcaches);
 938         spin_lock_irqsave(&cpu_rcache->lock, flags);
 939
 940         if (!iova_magazine_empty(cpu_rcache->loaded)) {
 941                 has_pfn = true;
 942         } else if (!iova_magazine_empty(cpu_rcache->prev)) {
 943                 swap(cpu_rcache->prev, cpu_rcache->loaded);
 944                 has_pfn = true;
 945         } else {
 946                 spin_lock(&rcache->lock);
 947                 if (rcache->depot_size > 0) {
 948                         iova_magazine_free(cpu_rcache->loaded);
 949                         cpu_rcache->loaded = rcache->depot[--rcache->depot_size];
 950                         has_pfn = true;
 951                 }
 952                 spin_unlock(&rcache->lock);
 953         }
 954
 955         if (has_pfn)
 956                 iova_pfn = iova_magazine_pop(cpu_rcache->loaded, limit_pfn);
 957
 958         spin_unlock_irqrestore(&cpu_rcache->lock, flags);
 959
 960         return iova_pfn;
 961 }
 962
 963 /*
 964  * Try to satisfy IOVA allocation range from rcache.  Fail if requested
 965  * size is too big or the DMA limit we are given isn't satisfied by the
 966  * top element in the magazine.
 967  */
 968 static unsigned long iova_rcache_get(struct iova_domain *iovad,
 969                                      unsigned long size,
 970                                      unsigned long limit_pfn)
 971 {
 972         unsigned int log_size = order_base_2(size);
 973
 974         if (log_size >= IOVA_RANGE_CACHE_MAX_SIZE)
 975                 return 0;
 976
 977         return __iova_rcache_get(&iovad->rcaches[log_size], limit_pfn - size);
 978 }
 979
 980 /*
 981  * free rcache data structures.
 982  */
 983 static void free_iova_rcaches(struct iova_domain *iovad)
 984 {
 985         struct iova_rcache *rcache;
 986         struct iova_cpu_rcache *cpu_rcache;
 987         unsigned int cpu;
 988         int i, j;
 989
 990         for (i = 0; i < IOVA_RANGE_CACHE_MAX_SIZE; ++i) {
 991                 rcache = &iovad->rcaches[i];
 992                 for_each_possible_cpu(cpu) {
 993                         cpu_rcache = per_cpu_ptr(rcache->cpu_rcaches, cpu);
 994                         iova_magazine_free(cpu_rcache->loaded);
 995                         iova_magazine_free(cpu_rcache->prev);
 996                 }
 997                 free_percpu(rcache->cpu_rcaches);
 998                 for (j = 0; j < rcache->depot_size; ++j)
 999                         iova_magazine_free(rcache->depot[j]);
1000         }
1001 }
1002
1003 /*
1004  * free all the IOVA ranges cached by a cpu (used when cpu is unplugged)
1005  */
1006 void free_cpu_cached_iovas(unsigned int cpu, struct iova_domain *iovad)
1007 {
1008         struct iova_cpu_rcache *cpu_rcache;
1009         struct iova_rcache *rcache;
1010         unsigned long flags;
1011         int i;
1012
1013         for (i = 0; i < IOVA_RANGE_CACHE_MAX_SIZE; ++i) {
1014                 rcache = &iovad->rcaches[i];
1015                 cpu_rcache = per_cpu_ptr(rcache->cpu_rcaches, cpu);
1016                 spin_lock_irqsave(&cpu_rcache->lock, flags);
1017                 iova_magazine_free_pfns(cpu_rcache->loaded, iovad);
1018                 iova_magazine_free_pfns(cpu_rcache->prev, iovad);
1019                 spin_unlock_irqrestore(&cpu_rcache->lock, flags);
1020         }
1021 }
1022
1023 /*
1024  * free all the IOVA ranges of global cache
1025  */
1026 static void free_global_cached_iovas(struct iova_domain *iovad)
1027 {
1028         struct iova_rcache *rcache;
1029         unsigned long flags;
1030         int i, j;
1031
1032         for (i = 0; i < IOVA_RANGE_CACHE_MAX_SIZE; ++i) {
1033                 rcache = &iovad->rcaches[i];
1034                 spin_lock_irqsave(&rcache->lock, flags);
1035                 for (j = 0; j < rcache->depot_size; ++j) {
1036                         iova_magazine_free_pfns(rcache->depot[j], iovad);
1037                         iova_magazine_free(rcache->depot[j]);
1038                 }
1039                 rcache->depot_size = 0;
1040                 spin_unlock_irqrestore(&rcache->lock, flags);
1041         }
1042 }
1043 MODULE_AUTHOR("Anil S Keshavamurthy <anil.s.keshavamurthy@intel.com>");
1044 MODULE_LICENSE("GPL");