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/slab.h>
  22
  23 static struct kmem_cache *iommu_iova_cache;
  24
  25 int iommu_iova_cache_init(void)
  26 {
  27         int ret = 0;
  28
  29         iommu_iova_cache = kmem_cache_create("iommu_iova",
  30                                          sizeof(struct iova),
  31                                          0,
  32                                          SLAB_HWCACHE_ALIGN,
  33                                          NULL);
  34         if (!iommu_iova_cache) {
  35                 pr_err("Couldn't create iova cache\n");
  36                 ret = -ENOMEM;
  37         }
  38
  39         return ret;
  40 }
  41
  42 void iommu_iova_cache_destroy(void)
  43 {
  44         kmem_cache_destroy(iommu_iova_cache);
  45 }
  46
  47 struct iova *alloc_iova_mem(void)
  48 {
  49         return kmem_cache_alloc(iommu_iova_cache, GFP_ATOMIC);
  50 }
  51
  52 void free_iova_mem(struct iova *iova)
  53 {
  54         kmem_cache_free(iommu_iova_cache, iova);
  55 }
  56
  57 void
  58 init_iova_domain(struct iova_domain *iovad, unsigned long granule,
  59         unsigned long start_pfn, unsigned long pfn_32bit)
  60 {
  61         /*
  62          * IOVA granularity will normally be equal to the smallest
  63          * supported IOMMU page size; both *must* be capable of
  64          * representing individual CPU pages exactly.
  65          */
  66         BUG_ON((granule > PAGE_SIZE) || !is_power_of_2(granule));
  67
  68         spin_lock_init(&iovad->iova_rbtree_lock);
  69         iovad->rbroot = RB_ROOT;
  70         iovad->cached32_node = NULL;
  71         iovad->granule = granule;
  72         iovad->start_pfn = start_pfn;
  73         iovad->dma_32bit_pfn = pfn_32bit;
  74 }
  75
  76 static struct rb_node *
  77 __get_cached_rbnode(struct iova_domain *iovad, unsigned long *limit_pfn)
  78 {
  79         if ((*limit_pfn != iovad->dma_32bit_pfn) ||
  80                 (iovad->cached32_node == NULL))
  81                 return rb_last(&iovad->rbroot);
  82         else {
  83                 struct rb_node *prev_node = rb_prev(iovad->cached32_node);
  84                 struct iova *curr_iova =
  85                         container_of(iovad->cached32_node, struct iova, node);
  86                 *limit_pfn = curr_iova->pfn_lo - 1;
  87                 return prev_node;
  88         }
  89 }
  90
  91 static void
  92 __cached_rbnode_insert_update(struct iova_domain *iovad,
  93         unsigned long limit_pfn, struct iova *new)
  94 {
  95         if (limit_pfn != iovad->dma_32bit_pfn)
  96                 return;
  97         iovad->cached32_node = &new->node;
  98 }
  99
 100 static void
 101 __cached_rbnode_delete_update(struct iova_domain *iovad, struct iova *free)
 102 {
 103         struct iova *cached_iova;
 104         struct rb_node *curr;
 105
 106         if (!iovad->cached32_node)
 107                 return;
 108         curr = iovad->cached32_node;
 109         cached_iova = container_of(curr, struct iova, node);
 110
 111         if (free->pfn_lo >= cached_iova->pfn_lo) {
 112                 struct rb_node *node = rb_next(&free->node);
 113                 struct iova *iova = container_of(node, struct iova, node);
 114
 115                 /* only cache if it's below 32bit pfn */
 116                 if (node && iova->pfn_lo < iovad->dma_32bit_pfn)
 117                         iovad->cached32_node = node;
 118                 else
 119                         iovad->cached32_node = NULL;
 120         }
 121 }
 122
 123 /* Computes the padding size required, to make the
 124  * the start address naturally aligned on its size
 125  */
 126 static int
 127 iova_get_pad_size(int size, unsigned int limit_pfn)
 128 {
 129         unsigned int pad_size = 0;
 130         unsigned int order = ilog2(size);
 131
 132         if (order)
 133                 pad_size = (limit_pfn + 1) % (1 << order);
 134
 135         return pad_size;
 136 }
 137
 138 static int __alloc_and_insert_iova_range(struct iova_domain *iovad,
 139                 unsigned long size, unsigned long limit_pfn,
 140                         struct iova *new, bool size_aligned)
 141 {
 142         struct rb_node *prev, *curr = NULL;
 143         unsigned long flags;
 144         unsigned long saved_pfn;
 145         unsigned int pad_size = 0;
 146
 147         /* Walk the tree backwards */
 148         spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
 149         saved_pfn = limit_pfn;
 150         curr = __get_cached_rbnode(iovad, &limit_pfn);
 151         prev = curr;
 152         while (curr) {
 153                 struct iova *curr_iova = container_of(curr, struct iova, node);
 154
 155                 if (limit_pfn < curr_iova->pfn_lo)
 156                         goto move_left;
 157                 else if (limit_pfn < curr_iova->pfn_hi)
 158                         goto adjust_limit_pfn;
 159                 else {
 160                         if (size_aligned)
 161                                 pad_size = iova_get_pad_size(size, limit_pfn);
 162                         if ((curr_iova->pfn_hi + size + pad_size) <= limit_pfn)
 163                                 break;  /* found a free slot */
 164                 }
 165 adjust_limit_pfn:
 166                 limit_pfn = curr_iova->pfn_lo - 1;
 167 move_left:
 168                 prev = curr;
 169                 curr = rb_prev(curr);
 170         }
 171
 172         if (!curr) {
 173                 if (size_aligned)
 174                         pad_size = iova_get_pad_size(size, limit_pfn);
 175                 if ((iovad->start_pfn + size + pad_size) > limit_pfn) {
 176                         spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
 177                         return -ENOMEM;
 178                 }
 179         }
 180
 181         /* pfn_lo will point to size aligned address if size_aligned is set */
 182         new->pfn_lo = limit_pfn - (size + pad_size) + 1;
 183         new->pfn_hi = new->pfn_lo + size - 1;
 184
 185         /* Insert the new_iova into domain rbtree by holding writer lock */
 186         /* Add new node and rebalance tree. */
 187         {
 188                 struct rb_node **entry, *parent = NULL;
 189
 190                 /* If we have 'prev', it's a valid place to start the
 191                    insertion. Otherwise, start from the root. */
 192                 if (prev)
 193                         entry = &prev;
 194                 else
 195                         entry = &iovad->rbroot.rb_node;
 196
 197                 /* Figure out where to put new node */
 198                 while (*entry) {
 199                         struct iova *this = container_of(*entry,
 200                                                         struct iova, node);
 201                         parent = *entry;
 202
 203                         if (new->pfn_lo < this->pfn_lo)
 204                                 entry = &((*entry)->rb_left);
 205                         else if (new->pfn_lo > this->pfn_lo)
 206                                 entry = &((*entry)->rb_right);
 207                         else
 208                                 BUG(); /* this should not happen */
 209                 }
 210
 211                 /* Add new node and rebalance tree. */
 212                 rb_link_node(&new->node, parent, entry);
 213                 rb_insert_color(&new->node, &iovad->rbroot);
 214         }
 215         __cached_rbnode_insert_update(iovad, saved_pfn, new);
 216
 217         spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
 218
 219
 220         return 0;
 221 }
 222
 223 static void
 224 iova_insert_rbtree(struct rb_root *root, struct iova *iova)
 225 {
 226         struct rb_node **new = &(root->rb_node), *parent = NULL;
 227         /* Figure out where to put new node */
 228         while (*new) {
 229                 struct iova *this = container_of(*new, struct iova, node);
 230
 231                 parent = *new;
 232
 233                 if (iova->pfn_lo < this->pfn_lo)
 234                         new = &((*new)->rb_left);
 235                 else if (iova->pfn_lo > this->pfn_lo)
 236                         new = &((*new)->rb_right);
 237                 else
 238                         BUG(); /* this should not happen */
 239         }
 240         /* Add new node and rebalance tree. */
 241         rb_link_node(&iova->node, parent, new);
 242         rb_insert_color(&iova->node, root);
 243 }
 244
 245 /**
 246  * alloc_iova - allocates an iova
 247  * @iovad: - iova domain in question
 248  * @size: - size of page frames to allocate
 249  * @limit_pfn: - max limit address
 250  * @size_aligned: - set if size_aligned address range is required
 251  * This function allocates an iova in the range iovad->start_pfn to limit_pfn,
 252  * searching top-down from limit_pfn to iovad->start_pfn. If the size_aligned
 253  * flag is set then the allocated address iova->pfn_lo will be naturally
 254  * aligned on roundup_power_of_two(size).
 255  */
 256 struct iova *
 257 alloc_iova(struct iova_domain *iovad, unsigned long size,
 258         unsigned long limit_pfn,
 259         bool size_aligned)
 260 {
 261         struct iova *new_iova;
 262         int ret;
 263
 264         new_iova = alloc_iova_mem();
 265         if (!new_iova)
 266                 return NULL;
 267
 268         /* If size aligned is set then round the size to
 269          * to next power of two.
 270          */
 271         if (size_aligned)
 272                 size = __roundup_pow_of_two(size);
 273
 274         ret = __alloc_and_insert_iova_range(iovad, size, limit_pfn,
 275                         new_iova, size_aligned);
 276
 277         if (ret) {
 278                 free_iova_mem(new_iova);
 279                 return NULL;
 280         }
 281
 282         return new_iova;
 283 }
 284
 285 /**
 286  * find_iova - find's an iova for a given pfn
 287  * @iovad: - iova domain in question.
 288  * @pfn: - page frame number
 289  * This function finds and returns an iova belonging to the
 290  * given doamin which matches the given pfn.
 291  */
 292 struct iova *find_iova(struct iova_domain *iovad, unsigned long pfn)
 293 {
 294         unsigned long flags;
 295         struct rb_node *node;
 296
 297         /* Take the lock so that no other thread is manipulating the rbtree */
 298         spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
 299         node = iovad->rbroot.rb_node;
 300         while (node) {
 301                 struct iova *iova = container_of(node, struct iova, node);
 302
 303                 /* If pfn falls within iova's range, return iova */
 304                 if ((pfn >= iova->pfn_lo) && (pfn <= iova->pfn_hi)) {
 305                         spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
 306                         /* We are not holding the lock while this iova
 307                          * is referenced by the caller as the same thread
 308                          * which called this function also calls __free_iova()
 309                          * and it is by design that only one thread can possibly
 310                          * reference a particular iova and hence no conflict.
 311                          */
 312                         return iova;
 313                 }
 314
 315                 if (pfn < iova->pfn_lo)
 316                         node = node->rb_left;
 317                 else if (pfn > iova->pfn_lo)
 318                         node = node->rb_right;
 319         }
 320
 321         spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
 322         return NULL;
 323 }
 324
 325 /**
 326  * __free_iova - frees the given iova
 327  * @iovad: iova domain in question.
 328  * @iova: iova in question.
 329  * Frees the given iova belonging to the giving domain
 330  */
 331 void
 332 __free_iova(struct iova_domain *iovad, struct iova *iova)
 333 {
 334         unsigned long flags;
 335
 336         spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
 337         __cached_rbnode_delete_update(iovad, iova);
 338         rb_erase(&iova->node, &iovad->rbroot);
 339         spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
 340         free_iova_mem(iova);
 341 }
 342
 343 /**
 344  * free_iova - finds and frees the iova for a given pfn
 345  * @iovad: - iova domain in question.
 346  * @pfn: - pfn that is allocated previously
 347  * This functions finds an iova for a given pfn and then
 348  * frees the iova from that domain.
 349  */
 350 void
 351 free_iova(struct iova_domain *iovad, unsigned long pfn)
 352 {
 353         struct iova *iova = find_iova(iovad, pfn);
 354
 355         if (iova)
 356                 __free_iova(iovad, iova);
 357
 358 }
 359
 360 /**
 361  * put_iova_domain - destroys the iova doamin
 362  * @iovad: - iova domain in question.
 363  * All the iova's in that domain are destroyed.
 364  */
 365 void put_iova_domain(struct iova_domain *iovad)
 366 {
 367         struct rb_node *node;
 368         unsigned long flags;
 369
 370         spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
 371         node = rb_first(&iovad->rbroot);
 372         while (node) {
 373                 struct iova *iova = container_of(node, struct iova, node);
 374
 375                 rb_erase(node, &iovad->rbroot);
 376                 free_iova_mem(iova);
 377                 node = rb_first(&iovad->rbroot);
 378         }
 379         spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
 380 }
 381
 382 static int
 383 __is_range_overlap(struct rb_node *node,
 384         unsigned long pfn_lo, unsigned long pfn_hi)
 385 {
 386         struct iova *iova = container_of(node, struct iova, node);
 387
 388         if ((pfn_lo <= iova->pfn_hi) && (pfn_hi >= iova->pfn_lo))
 389                 return 1;
 390         return 0;
 391 }
 392
 393 static inline struct iova *
 394 alloc_and_init_iova(unsigned long pfn_lo, unsigned long pfn_hi)
 395 {
 396         struct iova *iova;
 397
 398         iova = alloc_iova_mem();
 399         if (iova) {
 400                 iova->pfn_lo = pfn_lo;
 401                 iova->pfn_hi = pfn_hi;
 402         }
 403
 404         return iova;
 405 }
 406
 407 static struct iova *
 408 __insert_new_range(struct iova_domain *iovad,
 409         unsigned long pfn_lo, unsigned long pfn_hi)
 410 {
 411         struct iova *iova;
 412
 413         iova = alloc_and_init_iova(pfn_lo, pfn_hi);
 414         if (iova)
 415                 iova_insert_rbtree(&iovad->rbroot, iova);
 416
 417         return iova;
 418 }
 419
 420 static void
 421 __adjust_overlap_range(struct iova *iova,
 422         unsigned long *pfn_lo, unsigned long *pfn_hi)
 423 {
 424         if (*pfn_lo < iova->pfn_lo)
 425                 iova->pfn_lo = *pfn_lo;
 426         if (*pfn_hi > iova->pfn_hi)
 427                 *pfn_lo = iova->pfn_hi + 1;
 428 }
 429
 430 /**
 431  * reserve_iova - reserves an iova in the given range
 432  * @iovad: - iova domain pointer
 433  * @pfn_lo: - lower page frame address
 434  * @pfn_hi:- higher pfn adderss
 435  * This function allocates reserves the address range from pfn_lo to pfn_hi so
 436  * that this address is not dished out as part of alloc_iova.
 437  */
 438 struct iova *
 439 reserve_iova(struct iova_domain *iovad,
 440         unsigned long pfn_lo, unsigned long pfn_hi)
 441 {
 442         struct rb_node *node;
 443         unsigned long flags;
 444         struct iova *iova;
 445         unsigned int overlap = 0;
 446
 447         spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
 448         for (node = rb_first(&iovad->rbroot); node; node = rb_next(node)) {
 449                 if (__is_range_overlap(node, pfn_lo, pfn_hi)) {
 450                         iova = container_of(node, struct iova, node);
 451                         __adjust_overlap_range(iova, &pfn_lo, &pfn_hi);
 452                         if ((pfn_lo >= iova->pfn_lo) &&
 453                                 (pfn_hi <= iova->pfn_hi))
 454                                 goto finish;
 455                         overlap = 1;
 456
 457                 } else if (overlap)
 458                                 break;
 459         }
 460
 461         /* We are here either because this is the first reserver node
 462          * or need to insert remaining non overlap addr range
 463          */
 464         iova = __insert_new_range(iovad, pfn_lo, pfn_hi);
 465 finish:
 466
 467         spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
 468         return iova;
 469 }
 470
 471 /**
 472  * copy_reserved_iova - copies the reserved between domains
 473  * @from: - source doamin from where to copy
 474  * @to: - destination domin where to copy
 475  * This function copies reserved iova's from one doamin to
 476  * other.
 477  */
 478 void
 479 copy_reserved_iova(struct iova_domain *from, struct iova_domain *to)
 480 {
 481         unsigned long flags;
 482         struct rb_node *node;
 483
 484         spin_lock_irqsave(&from->iova_rbtree_lock, flags);
 485         for (node = rb_first(&from->rbroot); node; node = rb_next(node)) {
 486                 struct iova *iova = container_of(node, struct iova, node);
 487                 struct iova *new_iova;
 488
 489                 new_iova = reserve_iova(to, iova->pfn_lo, iova->pfn_hi);
 490                 if (!new_iova)
 491                         printk(KERN_ERR "Reserve iova range %lx@%lx failed\n",
 492                                 iova->pfn_lo, iova->pfn_lo);
 493         }
 494         spin_unlock_irqrestore(&from->iova_rbtree_lock, flags);
 495 }
 496
 497 struct iova *
 498 split_and_remove_iova(struct iova_domain *iovad, struct iova *iova,
 499                       unsigned long pfn_lo, unsigned long pfn_hi)
 500 {
 501         unsigned long flags;
 502         struct iova *prev = NULL, *next = NULL;
 503
 504         spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
 505         if (iova->pfn_lo < pfn_lo) {
 506                 prev = alloc_and_init_iova(iova->pfn_lo, pfn_lo - 1);
 507                 if (prev == NULL)
 508                         goto error;
 509         }
 510         if (iova->pfn_hi > pfn_hi) {
 511                 next = alloc_and_init_iova(pfn_hi + 1, iova->pfn_hi);
 512                 if (next == NULL)
 513                         goto error;
 514         }
 515
 516         __cached_rbnode_delete_update(iovad, iova);
 517         rb_erase(&iova->node, &iovad->rbroot);
 518
 519         if (prev) {
 520                 iova_insert_rbtree(&iovad->rbroot, prev);
 521                 iova->pfn_lo = pfn_lo;
 522         }
 523         if (next) {
 524                 iova_insert_rbtree(&iovad->rbroot, next);
 525                 iova->pfn_hi = pfn_hi;
 526         }
 527         spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
 528
 529         return iova;
 530
 531 error:
 532         spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
 533         if (prev)
 534                 free_iova_mem(prev);
 535         return NULL;
 536 }