drivers/iommu/iova.c

   1 /*
   2  * Copyright © 2006-2009, Intel Corporation.
   3  *
   4  * This program is free software; you can redistribute it and/or modify it
   5  * under the terms and conditions of the GNU General Public License,
   6  * version 2, as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope it will be useful, but WITHOUT
   9  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  10  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  11  * more details.
  12  *
  13  * You should have received a copy of the GNU General Public License along with
  14  * this program; if not, write to the Free Software Foundation, Inc., 59 Temple
  15  * Place - Suite 330, Boston, MA 02111-1307 USA.
  16  *
  17  * Author: Anil S Keshavamurthy <anil.s.keshavamurthy@intel.com>
  18  */
  19
  20 #include <linux/iova.h>
  21 #include <linux/module.h>
  22 #include <linux/slab.h>
  23 #include <linux/smp.h>
  24 #include <linux/bitops.h>
  25
  26 static bool iova_rcache_insert(struct iova_domain *iovad,
  27                                unsigned long pfn,
  28                                unsigned long size);
  29 static unsigned long iova_rcache_get(struct iova_domain *iovad,
  30                                      unsigned long size,
  31                                      unsigned long limit_pfn);
  32 static void init_iova_rcaches(struct iova_domain *iovad);
  33 static void free_iova_rcaches(struct iova_domain *iovad);
  34
  35 void
  36 init_iova_domain(struct iova_domain *iovad, unsigned long granule,
  37         unsigned long start_pfn, unsigned long pfn_32bit)
  38 {
  39         /*
  40          * IOVA granularity will normally be equal to the smallest
  41          * supported IOMMU page size; both *must* be capable of
  42          * representing individual CPU pages exactly.
  43          */
  44         BUG_ON((granule > PAGE_SIZE) || !is_power_of_2(granule));
  45
  46         spin_lock_init(&iovad->iova_rbtree_lock);
  47         iovad->rbroot = RB_ROOT;
  48         iovad->cached32_node = NULL;
  49         iovad->granule = granule;
  50         iovad->start_pfn = start_pfn;
  51         iovad->dma_32bit_pfn = pfn_32bit;
  52         init_iova_rcaches(iovad);
  53 }
  54 EXPORT_SYMBOL_GPL(init_iova_domain);
  55
  56 static struct rb_node *
  57 __get_cached_rbnode(struct iova_domain *iovad, unsigned long *limit_pfn)
  58 {
  59         if ((*limit_pfn > iovad->dma_32bit_pfn) ||
  60                 (iovad->cached32_node == NULL))
  61                 return rb_last(&iovad->rbroot);
  62         else {
  63                 struct rb_node *prev_node = rb_prev(iovad->cached32_node);
  64                 struct iova *curr_iova =
  65                         rb_entry(iovad->cached32_node, struct iova, node);
  66                 *limit_pfn = curr_iova->pfn_lo - 1;
  67                 return prev_node;
  68         }
  69 }
  70
  71 static void
  72 __cached_rbnode_insert_update(struct iova_domain *iovad,
  73         unsigned long limit_pfn, struct iova *new)
  74 {
  75         if (limit_pfn != iovad->dma_32bit_pfn)
  76                 return;
  77         iovad->cached32_node = &new->node;
  78 }
  79
  80 static void
  81 __cached_rbnode_delete_update(struct iova_domain *iovad, struct iova *free)
  82 {
  83         struct iova *cached_iova;
  84         struct rb_node *curr;
  85
  86         if (!iovad->cached32_node)
  87                 return;
  88         curr = iovad->cached32_node;
  89         cached_iova = rb_entry(curr, struct iova, node);
  90
  91         if (free->pfn_lo >= cached_iova->pfn_lo) {
  92                 struct rb_node *node = rb_next(&free->node);
  93                 struct iova *iova = rb_entry(node, struct iova, node);
  94
  95                 /* only cache if it's below 32bit pfn */
  96                 if (node && iova->pfn_lo < iovad->dma_32bit_pfn)
  97                         iovad->cached32_node = node;
  98                 else
  99                         iovad->cached32_node = NULL;
 100         }
 101 }
 102
 103 /* Insert the iova into domain rbtree by holding writer lock */
 104 static void
 105 iova_insert_rbtree(struct rb_root *root, struct iova *iova,
 106                    struct rb_node *start)
 107 {
 108         struct rb_node **new, *parent = NULL;
 109
 110         new = (start) ? &start : &(root->rb_node);
 111         /* Figure out where to put new node */
 112         while (*new) {
 113                 struct iova *this = rb_entry(*new, struct iova, node);
 114
 115                 parent = *new;
 116
 117                 if (iova->pfn_lo < this->pfn_lo)
 118                         new = &((*new)->rb_left);
 119                 else if (iova->pfn_lo > this->pfn_lo)
 120                         new = &((*new)->rb_right);
 121                 else {
 122                         WARN_ON(1); /* this should not happen */
 123                         return;
 124                 }
 125         }
 126         /* Add new node and rebalance tree. */
 127         rb_link_node(&iova->node, parent, new);
 128         rb_insert_color(&iova->node, root);
 129 }
 130
 131 /*
 132  * Computes the padding size required, to make the start address
 133  * naturally aligned on the power-of-two order of its size
 134  */
 135 static unsigned int
 136 iova_get_pad_size(unsigned int size, unsigned int limit_pfn)
 137 {
 138         return (limit_pfn + 1 - size) & (__roundup_pow_of_two(size) - 1);
 139 }
 140
 141 static int __alloc_and_insert_iova_range(struct iova_domain *iovad,
 142                 unsigned long size, unsigned long limit_pfn,
 143                         struct iova *new, bool size_aligned)
 144 {
 145         struct rb_node *prev, *curr = NULL;
 146         unsigned long flags;
 147         unsigned long saved_pfn;
 148         unsigned int pad_size = 0;
 149
 150         /* Walk the tree backwards */
 151         spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
 152         saved_pfn = limit_pfn;
 153         curr = __get_cached_rbnode(iovad, &limit_pfn);
 154         prev = curr;
 155         while (curr) {
 156                 struct iova *curr_iova = rb_entry(curr, struct iova, node);
 157
 158                 if (limit_pfn < curr_iova->pfn_lo)
 159                         goto move_left;
 160                 else if (limit_pfn < curr_iova->pfn_hi)
 161                         goto adjust_limit_pfn;
 162                 else {
 163                         if (size_aligned)
 164                                 pad_size = iova_get_pad_size(size, limit_pfn);
 165                         if ((curr_iova->pfn_hi + size + pad_size) <= limit_pfn)
 166                                 break;  /* found a free slot */
 167                 }
 168 adjust_limit_pfn:
 169                 limit_pfn = curr_iova->pfn_lo ? (curr_iova->pfn_lo - 1) : 0;
 170 move_left:
 171                 prev = curr;
 172                 curr = rb_prev(curr);
 173         }
 174
 175         if (!curr) {
 176                 if (size_aligned)
 177                         pad_size = iova_get_pad_size(size, limit_pfn);
 178                 if ((iovad->start_pfn + size + pad_size) > limit_pfn) {
 179                         spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
 180                         return -ENOMEM;
 181                 }
 182         }
 183
 184         /* pfn_lo will point to size aligned address if size_aligned is set */
 185         new->pfn_lo = limit_pfn - (size + pad_size) + 1;
 186         new->pfn_hi = new->pfn_lo + size - 1;
 187
 188         /* If we have 'prev', it's a valid place to start the insertion. */
 189         iova_insert_rbtree(&iovad->rbroot, new, prev);
 190         __cached_rbnode_insert_update(iovad, saved_pfn, new);
 191
 192         spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
 193
 194
 195         return 0;
 196 }
 197
 198 static struct kmem_cache *iova_cache;
 199 static unsigned int iova_cache_users;
 200 static DEFINE_MUTEX(iova_cache_mutex);
 201
 202 struct iova *alloc_iova_mem(void)
 203 {
 204         return kmem_cache_alloc(iova_cache, GFP_ATOMIC);
 205 }
 206 EXPORT_SYMBOL(alloc_iova_mem);
 207
 208 void free_iova_mem(struct iova *iova)
 209 {
 210         kmem_cache_free(iova_cache, iova);
 211 }
 212 EXPORT_SYMBOL(free_iova_mem);
 213
 214 int iova_cache_get(void)
 215 {
 216         mutex_lock(&iova_cache_mutex);
 217         if (!iova_cache_users) {
 218                 iova_cache = kmem_cache_create(
 219                         "iommu_iova", sizeof(struct iova), 0,
 220                         SLAB_HWCACHE_ALIGN, NULL);
 221                 if (!iova_cache) {
 222                         mutex_unlock(&iova_cache_mutex);
 223                         printk(KERN_ERR "Couldn't create iova cache\n");
 224                         return -ENOMEM;
 225                 }
 226         }
 227
 228         iova_cache_users++;
 229         mutex_unlock(&iova_cache_mutex);
 230
 231         return 0;
 232 }
 233 EXPORT_SYMBOL_GPL(iova_cache_get);
 234
 235 void iova_cache_put(void)
 236 {
 237         mutex_lock(&iova_cache_mutex);
 238         if (WARN_ON(!iova_cache_users)) {
 239                 mutex_unlock(&iova_cache_mutex);
 240                 return;
 241         }
 242         iova_cache_users--;
 243         if (!iova_cache_users)
 244                 kmem_cache_destroy(iova_cache);
 245         mutex_unlock(&iova_cache_mutex);
 246 }
 247 EXPORT_SYMBOL_GPL(iova_cache_put);
 248
 249 /**
 250  * alloc_iova - allocates an iova
 251  * @iovad: - iova domain in question
 252  * @size: - size of page frames to allocate
 253  * @limit_pfn: - max limit address
 254  * @size_aligned: - set if size_aligned address range is required
 255  * This function allocates an iova in the range iovad->start_pfn to limit_pfn,
 256  * searching top-down from limit_pfn to iovad->start_pfn. If the size_aligned
 257  * flag is set then the allocated address iova->pfn_lo will be naturally
 258  * aligned on roundup_power_of_two(size).
 259  */
 260 struct iova *
 261 alloc_iova(struct iova_domain *iovad, unsigned long size,
 262         unsigned long limit_pfn,
 263         bool size_aligned)
 264 {
 265         struct iova *new_iova;
 266         int ret;
 267
 268         new_iova = alloc_iova_mem();
 269         if (!new_iova)
 270                 return NULL;
 271
 272         ret = __alloc_and_insert_iova_range(iovad, size, limit_pfn,
 273                         new_iova, size_aligned);
 274
 275         if (ret) {
 276                 free_iova_mem(new_iova);
 277                 return NULL;
 278         }
 279
 280         return new_iova;
 281 }
 282 EXPORT_SYMBOL_GPL(alloc_iova);
 283
 284 static struct iova *
 285 private_find_iova(struct iova_domain *iovad, unsigned long pfn)
 286 {
 287         struct rb_node *node = iovad->rbroot.rb_node;
 288
 289         assert_spin_locked(&iovad->iova_rbtree_lock);
 290
 291         while (node) {
 292                 struct iova *iova = rb_entry(node, struct iova, node);
 293
 294                 /* If pfn falls within iova's range, return iova */
 295                 if ((pfn >= iova->pfn_lo) && (pfn <= iova->pfn_hi)) {
 296                         return iova;
 297                 }
 298
 299                 if (pfn < iova->pfn_lo)
 300                         node = node->rb_left;
 301                 else if (pfn > iova->pfn_lo)
 302                         node = node->rb_right;
 303         }
 304
 305         return NULL;
 306 }
 307
 308 static void private_free_iova(struct iova_domain *iovad, struct iova *iova)
 309 {
 310         assert_spin_locked(&iovad->iova_rbtree_lock);
 311         __cached_rbnode_delete_update(iovad, iova);
 312         rb_erase(&iova->node, &iovad->rbroot);
 313         free_iova_mem(iova);
 314 }
 315
 316 /**
 317  * find_iova - finds an iova for a given pfn
 318  * @iovad: - iova domain in question.
 319  * @pfn: - page frame number
 320  * This function finds and returns an iova belonging to the
 321  * given doamin which matches the given pfn.
 322  */
 323 struct iova *find_iova(struct iova_domain *iovad, unsigned long pfn)
 324 {
 325         unsigned long flags;
 326         struct iova *iova;
 327
 328         /* Take the lock so that no other thread is manipulating the rbtree */
 329         spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
 330         iova = private_find_iova(iovad, pfn);
 331         spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
 332         return iova;
 333 }
 334 EXPORT_SYMBOL_GPL(find_iova);
 335
 336 /**
 337  * __free_iova - frees the given iova
 338  * @iovad: iova domain in question.
 339  * @iova: iova in question.
 340  * Frees the given iova belonging to the giving domain
 341  */
 342 void
 343 __free_iova(struct iova_domain *iovad, struct iova *iova)
 344 {
 345         unsigned long flags;
 346
 347         spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
 348         private_free_iova(iovad, iova);
 349         spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
 350 }
 351 EXPORT_SYMBOL_GPL(__free_iova);
 352
 353 /**
 354  * free_iova - finds and frees the iova for a given pfn
 355  * @iovad: - iova domain in question.
 356  * @pfn: - pfn that is allocated previously
 357  * This functions finds an iova for a given pfn and then
 358  * frees the iova from that domain.
 359  */
 360 void
 361 free_iova(struct iova_domain *iovad, unsigned long pfn)
 362 {
 363         struct iova *iova = find_iova(iovad, pfn);
 364
 365         if (iova)
 366                 __free_iova(iovad, iova);
 367
 368 }
 369 EXPORT_SYMBOL_GPL(free_iova);
 370
 371 /**
 372  * alloc_iova_fast - allocates an iova from rcache
 373  * @iovad: - iova domain in question
 374  * @size: - size of page frames to allocate
 375  * @limit_pfn: - max limit address
 376  * This function tries to satisfy an iova allocation from the rcache,
 377  * and falls back to regular allocation on failure.
 378 */
 379 unsigned long
 380 alloc_iova_fast(struct iova_domain *iovad, unsigned long size,
 381                 unsigned long limit_pfn)
 382 {
 383         bool flushed_rcache = false;
 384         unsigned long iova_pfn;
 385         struct iova *new_iova;
 386
 387         iova_pfn = iova_rcache_get(iovad, size, limit_pfn);
 388         if (iova_pfn)
 389                 return iova_pfn;
 390
 391 retry:
 392         new_iova = alloc_iova(iovad, size, limit_pfn, true);
 393         if (!new_iova) {
 394                 unsigned int cpu;
 395
 396                 if (flushed_rcache)
 397                         return 0;
 398
 399                 /* Try replenishing IOVAs by flushing rcache. */
 400                 flushed_rcache = true;
 401                 preempt_disable();
 402                 for_each_online_cpu(cpu)
 403                         free_cpu_cached_iovas(cpu, iovad);
 404                 preempt_enable();
 405                 goto retry;
 406         }
 407
 408         return new_iova->pfn_lo;
 409 }
 410 EXPORT_SYMBOL_GPL(alloc_iova_fast);
 411
 412 /**
 413  * free_iova_fast - free iova pfn range into rcache
 414  * @iovad: - iova domain in question.
 415  * @pfn: - pfn that is allocated previously
 416  * @size: - # of pages in range
 417  * This functions frees an iova range by trying to put it into the rcache,
 418  * falling back to regular iova deallocation via free_iova() if this fails.
 419  */
 420 void
 421 free_iova_fast(struct iova_domain *iovad, unsigned long pfn, unsigned long size)
 422 {
 423         if (iova_rcache_insert(iovad, pfn, size))
 424                 return;
 425
 426         free_iova(iovad, pfn);
 427 }
 428 EXPORT_SYMBOL_GPL(free_iova_fast);
 429
 430 /**
 431  * put_iova_domain - destroys the iova doamin
 432  * @iovad: - iova domain in question.
 433  * All the iova's in that domain are destroyed.
 434  */
 435 void put_iova_domain(struct iova_domain *iovad)
 436 {
 437         struct rb_node *node;
 438         unsigned long flags;
 439
 440         free_iova_rcaches(iovad);
 441         spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
 442         node = rb_first(&iovad->rbroot);
 443         while (node) {
 444                 struct iova *iova = rb_entry(node, struct iova, node);
 445
 446                 rb_erase(node, &iovad->rbroot);
 447                 free_iova_mem(iova);
 448                 node = rb_first(&iovad->rbroot);
 449         }
 450         spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
 451 }
 452 EXPORT_SYMBOL_GPL(put_iova_domain);
 453
 454 static int
 455 __is_range_overlap(struct rb_node *node,
 456         unsigned long pfn_lo, unsigned long pfn_hi)
 457 {
 458         struct iova *iova = rb_entry(node, struct iova, node);
 459
 460         if ((pfn_lo <= iova->pfn_hi) && (pfn_hi >= iova->pfn_lo))
 461                 return 1;
 462         return 0;
 463 }
 464
 465 static inline struct iova *
 466 alloc_and_init_iova(unsigned long pfn_lo, unsigned long pfn_hi)
 467 {
 468         struct iova *iova;
 469
 470         iova = alloc_iova_mem();
 471         if (iova) {
 472                 iova->pfn_lo = pfn_lo;
 473                 iova->pfn_hi = pfn_hi;
 474         }
 475
 476         return iova;
 477 }
 478
 479 static struct iova *
 480 __insert_new_range(struct iova_domain *iovad,
 481         unsigned long pfn_lo, unsigned long pfn_hi)
 482 {
 483         struct iova *iova;
 484
 485         iova = alloc_and_init_iova(pfn_lo, pfn_hi);
 486         if (iova)
 487                 iova_insert_rbtree(&iovad->rbroot, iova, NULL);
 488
 489         return iova;
 490 }
 491
 492 static void
 493 __adjust_overlap_range(struct iova *iova,
 494         unsigned long *pfn_lo, unsigned long *pfn_hi)
 495 {
 496         if (*pfn_lo < iova->pfn_lo)
 497                 iova->pfn_lo = *pfn_lo;
 498         if (*pfn_hi > iova->pfn_hi)
 499                 *pfn_lo = iova->pfn_hi + 1;
 500 }
 501
 502 /**
 503  * reserve_iova - reserves an iova in the given range
 504  * @iovad: - iova domain pointer
 505  * @pfn_lo: - lower page frame address
 506  * @pfn_hi:- higher pfn adderss
 507  * This function allocates reserves the address range from pfn_lo to pfn_hi so
 508  * that this address is not dished out as part of alloc_iova.
 509  */
 510 struct iova *
 511 reserve_iova(struct iova_domain *iovad,
 512         unsigned long pfn_lo, unsigned long pfn_hi)
 513 {
 514         struct rb_node *node;
 515         unsigned long flags;
 516         struct iova *iova;
 517         unsigned int overlap = 0;
 518
 519         spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
 520         for (node = rb_first(&iovad->rbroot); node; node = rb_next(node)) {
 521                 if (__is_range_overlap(node, pfn_lo, pfn_hi)) {
 522                         iova = rb_entry(node, struct iova, node);
 523                         __adjust_overlap_range(iova, &pfn_lo, &pfn_hi);
 524                         if ((pfn_lo >= iova->pfn_lo) &&
 525                                 (pfn_hi <= iova->pfn_hi))
 526                                 goto finish;
 527                         overlap = 1;
 528
 529                 } else if (overlap)
 530                                 break;
 531         }
 532
 533         /* We are here either because this is the first reserver node
 534          * or need to insert remaining non overlap addr range
 535          */
 536         iova = __insert_new_range(iovad, pfn_lo, pfn_hi);
 537 finish:
 538
 539         spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
 540         return iova;
 541 }
 542 EXPORT_SYMBOL_GPL(reserve_iova);
 543
 544 /**
 545  * copy_reserved_iova - copies the reserved between domains
 546  * @from: - source doamin from where to copy
 547  * @to: - destination domin where to copy
 548  * This function copies reserved iova's from one doamin to
 549  * other.
 550  */
 551 void
 552 copy_reserved_iova(struct iova_domain *from, struct iova_domain *to)
 553 {
 554         unsigned long flags;
 555         struct rb_node *node;
 556
 557         spin_lock_irqsave(&from->iova_rbtree_lock, flags);
 558         for (node = rb_first(&from->rbroot); node; node = rb_next(node)) {
 559                 struct iova *iova = rb_entry(node, struct iova, node);
 560                 struct iova *new_iova;
 561
 562                 new_iova = reserve_iova(to, iova->pfn_lo, iova->pfn_hi);
 563                 if (!new_iova)
 564                         printk(KERN_ERR "Reserve iova range %lx@%lx failed\n",
 565                                 iova->pfn_lo, iova->pfn_lo);
 566         }
 567         spin_unlock_irqrestore(&from->iova_rbtree_lock, flags);
 568 }
 569 EXPORT_SYMBOL_GPL(copy_reserved_iova);
 570
 571 struct iova *
 572 split_and_remove_iova(struct iova_domain *iovad, struct iova *iova,
 573                       unsigned long pfn_lo, unsigned long pfn_hi)
 574 {
 575         unsigned long flags;
 576         struct iova *prev = NULL, *next = NULL;
 577
 578         spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
 579         if (iova->pfn_lo < pfn_lo) {
 580                 prev = alloc_and_init_iova(iova->pfn_lo, pfn_lo - 1);
 581                 if (prev == NULL)
 582                         goto error;
 583         }
 584         if (iova->pfn_hi > pfn_hi) {
 585                 next = alloc_and_init_iova(pfn_hi + 1, iova->pfn_hi);
 586                 if (next == NULL)
 587                         goto error;
 588         }
 589
 590         __cached_rbnode_delete_update(iovad, iova);
 591         rb_erase(&iova->node, &iovad->rbroot);
 592
 593         if (prev) {
 594                 iova_insert_rbtree(&iovad->rbroot, prev, NULL);
 595                 iova->pfn_lo = pfn_lo;
 596         }
 597         if (next) {
 598                 iova_insert_rbtree(&iovad->rbroot, next, NULL);
 599                 iova->pfn_hi = pfn_hi;
 600         }
 601         spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
 602
 603         return iova;
 604
 605 error:
 606         spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
 607         if (prev)
 608                 free_iova_mem(prev);
 609         return NULL;
 610 }
 611
 612 /*
 613  * Magazine caches for IOVA ranges.  For an introduction to magazines,
 614  * see the USENIX 2001 paper "Magazines and Vmem: Extending the Slab
 615  * Allocator to Many CPUs and Arbitrary Resources" by Bonwick and Adams.
 616  * For simplicity, we use a static magazine size and don't implement the
 617  * dynamic size tuning described in the paper.
 618  */
 619
 620 #define IOVA_MAG_SIZE 128
 621
 622 struct iova_magazine {
 623         unsigned long size;
 624         unsigned long pfns[IOVA_MAG_SIZE];
 625 };
 626
 627 struct iova_cpu_rcache {
 628         spinlock_t lock;
 629         struct iova_magazine *loaded;
 630         struct iova_magazine *prev;
 631 };
 632
 633 static struct iova_magazine *iova_magazine_alloc(gfp_t flags)
 634 {
 635         return kzalloc(sizeof(struct iova_magazine), flags);
 636 }
 637
 638 static void iova_magazine_free(struct iova_magazine *mag)
 639 {
 640         kfree(mag);
 641 }
 642
 643 static void
 644 iova_magazine_free_pfns(struct iova_magazine *mag, struct iova_domain *iovad)
 645 {
 646         unsigned long flags;
 647         int i;
 648
 649         if (!mag)
 650                 return;
 651
 652         spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
 653
 654         for (i = 0 ; i < mag->size; ++i) {
 655                 struct iova *iova = private_find_iova(iovad, mag->pfns[i]);
 656
 657                 BUG_ON(!iova);
 658                 private_free_iova(iovad, iova);
 659         }
 660
 661         spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
 662
 663         mag->size = 0;
 664 }
 665
 666 static bool iova_magazine_full(struct iova_magazine *mag)
 667 {
 668         return (mag && mag->size == IOVA_MAG_SIZE);
 669 }
 670
 671 static bool iova_magazine_empty(struct iova_magazine *mag)
 672 {
 673         return (!mag || mag->size == 0);
 674 }
 675
 676 static unsigned long iova_magazine_pop(struct iova_magazine *mag,
 677                                        unsigned long limit_pfn)
 678 {
 679         BUG_ON(iova_magazine_empty(mag));
 680
 681         if (mag->pfns[mag->size - 1] >= limit_pfn)
 682                 return 0;
 683
 684         return mag->pfns[--mag->size];
 685 }
 686
 687 static void iova_magazine_push(struct iova_magazine *mag, unsigned long pfn)
 688 {
 689         BUG_ON(iova_magazine_full(mag));
 690
 691         mag->pfns[mag->size++] = pfn;
 692 }
 693
 694 static void init_iova_rcaches(struct iova_domain *iovad)
 695 {
 696         struct iova_cpu_rcache *cpu_rcache;
 697         struct iova_rcache *rcache;
 698         unsigned int cpu;
 699         int i;
 700
 701         for (i = 0; i < IOVA_RANGE_CACHE_MAX_SIZE; ++i) {
 702                 rcache = &iovad->rcaches[i];
 703                 spin_lock_init(&rcache->lock);
 704                 rcache->depot_size = 0;
 705                 rcache->cpu_rcaches = __alloc_percpu(sizeof(*cpu_rcache), cache_line_size());
 706                 if (WARN_ON(!rcache->cpu_rcaches))
 707                         continue;
 708                 for_each_possible_cpu(cpu) {
 709                         cpu_rcache = per_cpu_ptr(rcache->cpu_rcaches, cpu);
 710                         spin_lock_init(&cpu_rcache->lock);
 711                         cpu_rcache->loaded = iova_magazine_alloc(GFP_KERNEL);
 712                         cpu_rcache->prev = iova_magazine_alloc(GFP_KERNEL);
 713                 }
 714         }
 715 }
 716
 717 /*
 718  * Try inserting IOVA range starting with 'iova_pfn' into 'rcache', and
 719  * return true on success.  Can fail if rcache is full and we can't free
 720  * space, and free_iova() (our only caller) will then return the IOVA
 721  * range to the rbtree instead.
 722  */
 723 static bool __iova_rcache_insert(struct iova_domain *iovad,
 724                                  struct iova_rcache *rcache,
 725                                  unsigned long iova_pfn)
 726 {
 727         struct iova_magazine *mag_to_free = NULL;
 728         struct iova_cpu_rcache *cpu_rcache;
 729         bool can_insert = false;
 730         unsigned long flags;
 731
 732         cpu_rcache = get_cpu_ptr(rcache->cpu_rcaches);
 733         spin_lock_irqsave(&cpu_rcache->lock, flags);
 734
 735         if (!iova_magazine_full(cpu_rcache->loaded)) {
 736                 can_insert = true;
 737         } else if (!iova_magazine_full(cpu_rcache->prev)) {
 738                 swap(cpu_rcache->prev, cpu_rcache->loaded);
 739                 can_insert = true;
 740         } else {
 741                 struct iova_magazine *new_mag = iova_magazine_alloc(GFP_ATOMIC);
 742
 743                 if (new_mag) {
 744                         spin_lock(&rcache->lock);
 745                         if (rcache->depot_size < MAX_GLOBAL_MAGS) {
 746                                 rcache->depot[rcache->depot_size++] =
 747                                                 cpu_rcache->loaded;
 748                         } else {
 749                                 mag_to_free = cpu_rcache->loaded;
 750                         }
 751                         spin_unlock(&rcache->lock);
 752
 753                         cpu_rcache->loaded = new_mag;
 754                         can_insert = true;
 755                 }
 756         }
 757
 758         if (can_insert)
 759                 iova_magazine_push(cpu_rcache->loaded, iova_pfn);
 760
 761         spin_unlock_irqrestore(&cpu_rcache->lock, flags);
 762         put_cpu_ptr(rcache->cpu_rcaches);
 763
 764         if (mag_to_free) {
 765                 iova_magazine_free_pfns(mag_to_free, iovad);
 766                 iova_magazine_free(mag_to_free);
 767         }
 768
 769         return can_insert;
 770 }
 771
 772 static bool iova_rcache_insert(struct iova_domain *iovad, unsigned long pfn,
 773                                unsigned long size)
 774 {
 775         unsigned int log_size = order_base_2(size);
 776
 777         if (log_size >= IOVA_RANGE_CACHE_MAX_SIZE)
 778                 return false;
 779
 780         return __iova_rcache_insert(iovad, &iovad->rcaches[log_size], pfn);
 781 }
 782
 783 /*
 784  * Caller wants to allocate a new IOVA range from 'rcache'.  If we can
 785  * satisfy the request, return a matching non-NULL range and remove
 786  * it from the 'rcache'.
 787  */
 788 static unsigned long __iova_rcache_get(struct iova_rcache *rcache,
 789                                        unsigned long limit_pfn)
 790 {
 791         struct iova_cpu_rcache *cpu_rcache;
 792         unsigned long iova_pfn = 0;
 793         bool has_pfn = false;
 794         unsigned long flags;
 795
 796         cpu_rcache = get_cpu_ptr(rcache->cpu_rcaches);
 797         spin_lock_irqsave(&cpu_rcache->lock, flags);
 798
 799         if (!iova_magazine_empty(cpu_rcache->loaded)) {
 800                 has_pfn = true;
 801         } else if (!iova_magazine_empty(cpu_rcache->prev)) {
 802                 swap(cpu_rcache->prev, cpu_rcache->loaded);
 803                 has_pfn = true;
 804         } else {
 805                 spin_lock(&rcache->lock);
 806                 if (rcache->depot_size > 0) {
 807                         iova_magazine_free(cpu_rcache->loaded);
 808                         cpu_rcache->loaded = rcache->depot[--rcache->depot_size];
 809                         has_pfn = true;
 810                 }
 811                 spin_unlock(&rcache->lock);
 812         }
 813
 814         if (has_pfn)
 815                 iova_pfn = iova_magazine_pop(cpu_rcache->loaded, limit_pfn);
 816
 817         spin_unlock_irqrestore(&cpu_rcache->lock, flags);
 818         put_cpu_ptr(rcache->cpu_rcaches);
 819
 820         return iova_pfn;
 821 }
 822
 823 /*
 824  * Try to satisfy IOVA allocation range from rcache.  Fail if requested
 825  * size is too big or the DMA limit we are given isn't satisfied by the
 826  * top element in the magazine.
 827  */
 828 static unsigned long iova_rcache_get(struct iova_domain *iovad,
 829                                      unsigned long size,
 830                                      unsigned long limit_pfn)
 831 {
 832         unsigned int log_size = order_base_2(size);
 833
 834         if (log_size >= IOVA_RANGE_CACHE_MAX_SIZE)
 835                 return 0;
 836
 837         return __iova_rcache_get(&iovad->rcaches[log_size], limit_pfn);
 838 }
 839
 840 /*
 841  * Free a cpu's rcache.
 842  */
 843 static void free_cpu_iova_rcache(unsigned int cpu, struct iova_domain *iovad,
 844                                  struct iova_rcache *rcache)
 845 {
 846         struct iova_cpu_rcache *cpu_rcache = per_cpu_ptr(rcache->cpu_rcaches, cpu);
 847         unsigned long flags;
 848
 849         spin_lock_irqsave(&cpu_rcache->lock, flags);
 850
 851         iova_magazine_free_pfns(cpu_rcache->loaded, iovad);
 852         iova_magazine_free(cpu_rcache->loaded);
 853
 854         iova_magazine_free_pfns(cpu_rcache->prev, iovad);
 855         iova_magazine_free(cpu_rcache->prev);
 856
 857         spin_unlock_irqrestore(&cpu_rcache->lock, flags);
 858 }
 859
 860 /*
 861  * free rcache data structures.
 862  */
 863 static void free_iova_rcaches(struct iova_domain *iovad)
 864 {
 865         struct iova_rcache *rcache;
 866         unsigned long flags;
 867         unsigned int cpu;
 868         int i, j;
 869
 870         for (i = 0; i < IOVA_RANGE_CACHE_MAX_SIZE; ++i) {
 871                 rcache = &iovad->rcaches[i];
 872                 for_each_possible_cpu(cpu)
 873                         free_cpu_iova_rcache(cpu, iovad, rcache);
 874                 spin_lock_irqsave(&rcache->lock, flags);
 875                 free_percpu(rcache->cpu_rcaches);
 876                 for (j = 0; j < rcache->depot_size; ++j) {
 877                         iova_magazine_free_pfns(rcache->depot[j], iovad);
 878                         iova_magazine_free(rcache->depot[j]);
 879                 }
 880                 spin_unlock_irqrestore(&rcache->lock, flags);
 881         }
 882 }
 883
 884 /*
 885  * free all the IOVA ranges cached by a cpu (used when cpu is unplugged)
 886  */
 887 void free_cpu_cached_iovas(unsigned int cpu, struct iova_domain *iovad)
 888 {
 889         struct iova_cpu_rcache *cpu_rcache;
 890         struct iova_rcache *rcache;
 891         unsigned long flags;
 892         int i;
 893
 894         for (i = 0; i < IOVA_RANGE_CACHE_MAX_SIZE; ++i) {
 895                 rcache = &iovad->rcaches[i];
 896                 cpu_rcache = per_cpu_ptr(rcache->cpu_rcaches, cpu);
 897                 spin_lock_irqsave(&cpu_rcache->lock, flags);
 898                 iova_magazine_free_pfns(cpu_rcache->loaded, iovad);
 899                 iova_magazine_free_pfns(cpu_rcache->prev, iovad);
 900                 spin_unlock_irqrestore(&cpu_rcache->lock, flags);
 901         }
 902 }
 903
 904 MODULE_AUTHOR("Anil S Keshavamurthy <anil.s.keshavamurthy@intel.com>");
 905 MODULE_LICENSE("GPL");