Back to home page

OSCL-LXR
 
 

     

0001 // SPDX-License-Identifier: GPL-2.0-only
0002 /*
0003  * Copyright © 2006-2009, Intel Corporation.
0004  *
0005  * Author: Anil S Keshavamurthy <anil.s.keshavamurthy@intel.com>
0006  */
0007 
0008 #include <linux/iova.h>
0009 #include <linux/module.h>
0010 #include <linux/slab.h>
0011 #include <linux/smp.h>
0012 #include <linux/bitops.h>
0013 #include <linux/cpu.h>
0014 
0015 /* The anchor node sits above the top of the usable address space */
0016 #define IOVA_ANCHOR ~0UL
0017 
0018 #define IOVA_RANGE_CACHE_MAX_SIZE 6 /* log of max cached IOVA range size (in pages) */
0019 
0020 static bool iova_rcache_insert(struct iova_domain *iovad,
0021                    unsigned long pfn,
0022                    unsigned long size);
0023 static unsigned long iova_rcache_get(struct iova_domain *iovad,
0024                      unsigned long size,
0025                      unsigned long limit_pfn);
0026 static void free_cpu_cached_iovas(unsigned int cpu, struct iova_domain *iovad);
0027 static void free_iova_rcaches(struct iova_domain *iovad);
0028 
0029 unsigned long iova_rcache_range(void)
0030 {
0031     return PAGE_SIZE << (IOVA_RANGE_CACHE_MAX_SIZE - 1);
0032 }
0033 
0034 static int iova_cpuhp_dead(unsigned int cpu, struct hlist_node *node)
0035 {
0036     struct iova_domain *iovad;
0037 
0038     iovad = hlist_entry_safe(node, struct iova_domain, cpuhp_dead);
0039 
0040     free_cpu_cached_iovas(cpu, iovad);
0041     return 0;
0042 }
0043 
0044 static void free_global_cached_iovas(struct iova_domain *iovad);
0045 
0046 static struct iova *to_iova(struct rb_node *node)
0047 {
0048     return rb_entry(node, struct iova, node);
0049 }
0050 
0051 void
0052 init_iova_domain(struct iova_domain *iovad, unsigned long granule,
0053     unsigned long start_pfn)
0054 {
0055     /*
0056      * IOVA granularity will normally be equal to the smallest
0057      * supported IOMMU page size; both *must* be capable of
0058      * representing individual CPU pages exactly.
0059      */
0060     BUG_ON((granule > PAGE_SIZE) || !is_power_of_2(granule));
0061 
0062     spin_lock_init(&iovad->iova_rbtree_lock);
0063     iovad->rbroot = RB_ROOT;
0064     iovad->cached_node = &iovad->anchor.node;
0065     iovad->cached32_node = &iovad->anchor.node;
0066     iovad->granule = granule;
0067     iovad->start_pfn = start_pfn;
0068     iovad->dma_32bit_pfn = 1UL << (32 - iova_shift(iovad));
0069     iovad->max32_alloc_size = iovad->dma_32bit_pfn;
0070     iovad->anchor.pfn_lo = iovad->anchor.pfn_hi = IOVA_ANCHOR;
0071     rb_link_node(&iovad->anchor.node, NULL, &iovad->rbroot.rb_node);
0072     rb_insert_color(&iovad->anchor.node, &iovad->rbroot);
0073 }
0074 EXPORT_SYMBOL_GPL(init_iova_domain);
0075 
0076 static struct rb_node *
0077 __get_cached_rbnode(struct iova_domain *iovad, unsigned long limit_pfn)
0078 {
0079     if (limit_pfn <= iovad->dma_32bit_pfn)
0080         return iovad->cached32_node;
0081 
0082     return iovad->cached_node;
0083 }
0084 
0085 static void
0086 __cached_rbnode_insert_update(struct iova_domain *iovad, struct iova *new)
0087 {
0088     if (new->pfn_hi < iovad->dma_32bit_pfn)
0089         iovad->cached32_node = &new->node;
0090     else
0091         iovad->cached_node = &new->node;
0092 }
0093 
0094 static void
0095 __cached_rbnode_delete_update(struct iova_domain *iovad, struct iova *free)
0096 {
0097     struct iova *cached_iova;
0098 
0099     cached_iova = to_iova(iovad->cached32_node);
0100     if (free == cached_iova ||
0101         (free->pfn_hi < iovad->dma_32bit_pfn &&
0102          free->pfn_lo >= cached_iova->pfn_lo))
0103         iovad->cached32_node = rb_next(&free->node);
0104 
0105     if (free->pfn_lo < iovad->dma_32bit_pfn)
0106         iovad->max32_alloc_size = iovad->dma_32bit_pfn;
0107 
0108     cached_iova = to_iova(iovad->cached_node);
0109     if (free->pfn_lo >= cached_iova->pfn_lo)
0110         iovad->cached_node = rb_next(&free->node);
0111 }
0112 
0113 static struct rb_node *iova_find_limit(struct iova_domain *iovad, unsigned long limit_pfn)
0114 {
0115     struct rb_node *node, *next;
0116     /*
0117      * Ideally what we'd like to judge here is whether limit_pfn is close
0118      * enough to the highest-allocated IOVA that starting the allocation
0119      * walk from the anchor node will be quicker than this initial work to
0120      * find an exact starting point (especially if that ends up being the
0121      * anchor node anyway). This is an incredibly crude approximation which
0122      * only really helps the most likely case, but is at least trivially easy.
0123      */
0124     if (limit_pfn > iovad->dma_32bit_pfn)
0125         return &iovad->anchor.node;
0126 
0127     node = iovad->rbroot.rb_node;
0128     while (to_iova(node)->pfn_hi < limit_pfn)
0129         node = node->rb_right;
0130 
0131 search_left:
0132     while (node->rb_left && to_iova(node->rb_left)->pfn_lo >= limit_pfn)
0133         node = node->rb_left;
0134 
0135     if (!node->rb_left)
0136         return node;
0137 
0138     next = node->rb_left;
0139     while (next->rb_right) {
0140         next = next->rb_right;
0141         if (to_iova(next)->pfn_lo >= limit_pfn) {
0142             node = next;
0143             goto search_left;
0144         }
0145     }
0146 
0147     return node;
0148 }
0149 
0150 /* Insert the iova into domain rbtree by holding writer lock */
0151 static void
0152 iova_insert_rbtree(struct rb_root *root, struct iova *iova,
0153            struct rb_node *start)
0154 {
0155     struct rb_node **new, *parent = NULL;
0156 
0157     new = (start) ? &start : &(root->rb_node);
0158     /* Figure out where to put new node */
0159     while (*new) {
0160         struct iova *this = to_iova(*new);
0161 
0162         parent = *new;
0163 
0164         if (iova->pfn_lo < this->pfn_lo)
0165             new = &((*new)->rb_left);
0166         else if (iova->pfn_lo > this->pfn_lo)
0167             new = &((*new)->rb_right);
0168         else {
0169             WARN_ON(1); /* this should not happen */
0170             return;
0171         }
0172     }
0173     /* Add new node and rebalance tree. */
0174     rb_link_node(&iova->node, parent, new);
0175     rb_insert_color(&iova->node, root);
0176 }
0177 
0178 static int __alloc_and_insert_iova_range(struct iova_domain *iovad,
0179         unsigned long size, unsigned long limit_pfn,
0180             struct iova *new, bool size_aligned)
0181 {
0182     struct rb_node *curr, *prev;
0183     struct iova *curr_iova;
0184     unsigned long flags;
0185     unsigned long new_pfn, retry_pfn;
0186     unsigned long align_mask = ~0UL;
0187     unsigned long high_pfn = limit_pfn, low_pfn = iovad->start_pfn;
0188 
0189     if (size_aligned)
0190         align_mask <<= fls_long(size - 1);
0191 
0192     /* Walk the tree backwards */
0193     spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
0194     if (limit_pfn <= iovad->dma_32bit_pfn &&
0195             size >= iovad->max32_alloc_size)
0196         goto iova32_full;
0197 
0198     curr = __get_cached_rbnode(iovad, limit_pfn);
0199     curr_iova = to_iova(curr);
0200     retry_pfn = curr_iova->pfn_hi + 1;
0201 
0202 retry:
0203     do {
0204         high_pfn = min(high_pfn, curr_iova->pfn_lo);
0205         new_pfn = (high_pfn - size) & align_mask;
0206         prev = curr;
0207         curr = rb_prev(curr);
0208         curr_iova = to_iova(curr);
0209     } while (curr && new_pfn <= curr_iova->pfn_hi && new_pfn >= low_pfn);
0210 
0211     if (high_pfn < size || new_pfn < low_pfn) {
0212         if (low_pfn == iovad->start_pfn && retry_pfn < limit_pfn) {
0213             high_pfn = limit_pfn;
0214             low_pfn = retry_pfn;
0215             curr = iova_find_limit(iovad, limit_pfn);
0216             curr_iova = to_iova(curr);
0217             goto retry;
0218         }
0219         iovad->max32_alloc_size = size;
0220         goto iova32_full;
0221     }
0222 
0223     /* pfn_lo will point to size aligned address if size_aligned is set */
0224     new->pfn_lo = new_pfn;
0225     new->pfn_hi = new->pfn_lo + size - 1;
0226 
0227     /* If we have 'prev', it's a valid place to start the insertion. */
0228     iova_insert_rbtree(&iovad->rbroot, new, prev);
0229     __cached_rbnode_insert_update(iovad, new);
0230 
0231     spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
0232     return 0;
0233 
0234 iova32_full:
0235     spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
0236     return -ENOMEM;
0237 }
0238 
0239 static struct kmem_cache *iova_cache;
0240 static unsigned int iova_cache_users;
0241 static DEFINE_MUTEX(iova_cache_mutex);
0242 
0243 static struct iova *alloc_iova_mem(void)
0244 {
0245     return kmem_cache_zalloc(iova_cache, GFP_ATOMIC | __GFP_NOWARN);
0246 }
0247 
0248 static void free_iova_mem(struct iova *iova)
0249 {
0250     if (iova->pfn_lo != IOVA_ANCHOR)
0251         kmem_cache_free(iova_cache, iova);
0252 }
0253 
0254 int iova_cache_get(void)
0255 {
0256     mutex_lock(&iova_cache_mutex);
0257     if (!iova_cache_users) {
0258         int ret;
0259 
0260         ret = cpuhp_setup_state_multi(CPUHP_IOMMU_IOVA_DEAD, "iommu/iova:dead", NULL,
0261                     iova_cpuhp_dead);
0262         if (ret) {
0263             mutex_unlock(&iova_cache_mutex);
0264             pr_err("Couldn't register cpuhp handler\n");
0265             return ret;
0266         }
0267 
0268         iova_cache = kmem_cache_create(
0269             "iommu_iova", sizeof(struct iova), 0,
0270             SLAB_HWCACHE_ALIGN, NULL);
0271         if (!iova_cache) {
0272             cpuhp_remove_multi_state(CPUHP_IOMMU_IOVA_DEAD);
0273             mutex_unlock(&iova_cache_mutex);
0274             pr_err("Couldn't create iova cache\n");
0275             return -ENOMEM;
0276         }
0277     }
0278 
0279     iova_cache_users++;
0280     mutex_unlock(&iova_cache_mutex);
0281 
0282     return 0;
0283 }
0284 EXPORT_SYMBOL_GPL(iova_cache_get);
0285 
0286 void iova_cache_put(void)
0287 {
0288     mutex_lock(&iova_cache_mutex);
0289     if (WARN_ON(!iova_cache_users)) {
0290         mutex_unlock(&iova_cache_mutex);
0291         return;
0292     }
0293     iova_cache_users--;
0294     if (!iova_cache_users) {
0295         cpuhp_remove_multi_state(CPUHP_IOMMU_IOVA_DEAD);
0296         kmem_cache_destroy(iova_cache);
0297     }
0298     mutex_unlock(&iova_cache_mutex);
0299 }
0300 EXPORT_SYMBOL_GPL(iova_cache_put);
0301 
0302 /**
0303  * alloc_iova - allocates an iova
0304  * @iovad: - iova domain in question
0305  * @size: - size of page frames to allocate
0306  * @limit_pfn: - max limit address
0307  * @size_aligned: - set if size_aligned address range is required
0308  * This function allocates an iova in the range iovad->start_pfn to limit_pfn,
0309  * searching top-down from limit_pfn to iovad->start_pfn. If the size_aligned
0310  * flag is set then the allocated address iova->pfn_lo will be naturally
0311  * aligned on roundup_power_of_two(size).
0312  */
0313 struct iova *
0314 alloc_iova(struct iova_domain *iovad, unsigned long size,
0315     unsigned long limit_pfn,
0316     bool size_aligned)
0317 {
0318     struct iova *new_iova;
0319     int ret;
0320 
0321     new_iova = alloc_iova_mem();
0322     if (!new_iova)
0323         return NULL;
0324 
0325     ret = __alloc_and_insert_iova_range(iovad, size, limit_pfn + 1,
0326             new_iova, size_aligned);
0327 
0328     if (ret) {
0329         free_iova_mem(new_iova);
0330         return NULL;
0331     }
0332 
0333     return new_iova;
0334 }
0335 EXPORT_SYMBOL_GPL(alloc_iova);
0336 
0337 static struct iova *
0338 private_find_iova(struct iova_domain *iovad, unsigned long pfn)
0339 {
0340     struct rb_node *node = iovad->rbroot.rb_node;
0341 
0342     assert_spin_locked(&iovad->iova_rbtree_lock);
0343 
0344     while (node) {
0345         struct iova *iova = to_iova(node);
0346 
0347         if (pfn < iova->pfn_lo)
0348             node = node->rb_left;
0349         else if (pfn > iova->pfn_hi)
0350             node = node->rb_right;
0351         else
0352             return iova;    /* pfn falls within iova's range */
0353     }
0354 
0355     return NULL;
0356 }
0357 
0358 static void remove_iova(struct iova_domain *iovad, struct iova *iova)
0359 {
0360     assert_spin_locked(&iovad->iova_rbtree_lock);
0361     __cached_rbnode_delete_update(iovad, iova);
0362     rb_erase(&iova->node, &iovad->rbroot);
0363 }
0364 
0365 /**
0366  * find_iova - finds an iova for a given pfn
0367  * @iovad: - iova domain in question.
0368  * @pfn: - page frame number
0369  * This function finds and returns an iova belonging to the
0370  * given domain which matches the given pfn.
0371  */
0372 struct iova *find_iova(struct iova_domain *iovad, unsigned long pfn)
0373 {
0374     unsigned long flags;
0375     struct iova *iova;
0376 
0377     /* Take the lock so that no other thread is manipulating the rbtree */
0378     spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
0379     iova = private_find_iova(iovad, pfn);
0380     spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
0381     return iova;
0382 }
0383 EXPORT_SYMBOL_GPL(find_iova);
0384 
0385 /**
0386  * __free_iova - frees the given iova
0387  * @iovad: iova domain in question.
0388  * @iova: iova in question.
0389  * Frees the given iova belonging to the giving domain
0390  */
0391 void
0392 __free_iova(struct iova_domain *iovad, struct iova *iova)
0393 {
0394     unsigned long flags;
0395 
0396     spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
0397     remove_iova(iovad, iova);
0398     spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
0399     free_iova_mem(iova);
0400 }
0401 EXPORT_SYMBOL_GPL(__free_iova);
0402 
0403 /**
0404  * free_iova - finds and frees the iova for a given pfn
0405  * @iovad: - iova domain in question.
0406  * @pfn: - pfn that is allocated previously
0407  * This functions finds an iova for a given pfn and then
0408  * frees the iova from that domain.
0409  */
0410 void
0411 free_iova(struct iova_domain *iovad, unsigned long pfn)
0412 {
0413     unsigned long flags;
0414     struct iova *iova;
0415 
0416     spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
0417     iova = private_find_iova(iovad, pfn);
0418     if (!iova) {
0419         spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
0420         return;
0421     }
0422     remove_iova(iovad, iova);
0423     spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
0424     free_iova_mem(iova);
0425 }
0426 EXPORT_SYMBOL_GPL(free_iova);
0427 
0428 /**
0429  * alloc_iova_fast - allocates an iova from rcache
0430  * @iovad: - iova domain in question
0431  * @size: - size of page frames to allocate
0432  * @limit_pfn: - max limit address
0433  * @flush_rcache: - set to flush rcache on regular allocation failure
0434  * This function tries to satisfy an iova allocation from the rcache,
0435  * and falls back to regular allocation on failure. If regular allocation
0436  * fails too and the flush_rcache flag is set then the rcache will be flushed.
0437 */
0438 unsigned long
0439 alloc_iova_fast(struct iova_domain *iovad, unsigned long size,
0440         unsigned long limit_pfn, bool flush_rcache)
0441 {
0442     unsigned long iova_pfn;
0443     struct iova *new_iova;
0444 
0445     /*
0446      * Freeing non-power-of-two-sized allocations back into the IOVA caches
0447      * will come back to bite us badly, so we have to waste a bit of space
0448      * rounding up anything cacheable to make sure that can't happen. The
0449      * order of the unadjusted size will still match upon freeing.
0450      */
0451     if (size < (1 << (IOVA_RANGE_CACHE_MAX_SIZE - 1)))
0452         size = roundup_pow_of_two(size);
0453 
0454     iova_pfn = iova_rcache_get(iovad, size, limit_pfn + 1);
0455     if (iova_pfn)
0456         return iova_pfn;
0457 
0458 retry:
0459     new_iova = alloc_iova(iovad, size, limit_pfn, true);
0460     if (!new_iova) {
0461         unsigned int cpu;
0462 
0463         if (!flush_rcache)
0464             return 0;
0465 
0466         /* Try replenishing IOVAs by flushing rcache. */
0467         flush_rcache = false;
0468         for_each_online_cpu(cpu)
0469             free_cpu_cached_iovas(cpu, iovad);
0470         free_global_cached_iovas(iovad);
0471         goto retry;
0472     }
0473 
0474     return new_iova->pfn_lo;
0475 }
0476 EXPORT_SYMBOL_GPL(alloc_iova_fast);
0477 
0478 /**
0479  * free_iova_fast - free iova pfn range into rcache
0480  * @iovad: - iova domain in question.
0481  * @pfn: - pfn that is allocated previously
0482  * @size: - # of pages in range
0483  * This functions frees an iova range by trying to put it into the rcache,
0484  * falling back to regular iova deallocation via free_iova() if this fails.
0485  */
0486 void
0487 free_iova_fast(struct iova_domain *iovad, unsigned long pfn, unsigned long size)
0488 {
0489     if (iova_rcache_insert(iovad, pfn, size))
0490         return;
0491 
0492     free_iova(iovad, pfn);
0493 }
0494 EXPORT_SYMBOL_GPL(free_iova_fast);
0495 
0496 static void iova_domain_free_rcaches(struct iova_domain *iovad)
0497 {
0498     cpuhp_state_remove_instance_nocalls(CPUHP_IOMMU_IOVA_DEAD,
0499                         &iovad->cpuhp_dead);
0500     free_iova_rcaches(iovad);
0501 }
0502 
0503 /**
0504  * put_iova_domain - destroys the iova domain
0505  * @iovad: - iova domain in question.
0506  * All the iova's in that domain are destroyed.
0507  */
0508 void put_iova_domain(struct iova_domain *iovad)
0509 {
0510     struct iova *iova, *tmp;
0511 
0512     if (iovad->rcaches)
0513         iova_domain_free_rcaches(iovad);
0514 
0515     rbtree_postorder_for_each_entry_safe(iova, tmp, &iovad->rbroot, node)
0516         free_iova_mem(iova);
0517 }
0518 EXPORT_SYMBOL_GPL(put_iova_domain);
0519 
0520 static int
0521 __is_range_overlap(struct rb_node *node,
0522     unsigned long pfn_lo, unsigned long pfn_hi)
0523 {
0524     struct iova *iova = to_iova(node);
0525 
0526     if ((pfn_lo <= iova->pfn_hi) && (pfn_hi >= iova->pfn_lo))
0527         return 1;
0528     return 0;
0529 }
0530 
0531 static inline struct iova *
0532 alloc_and_init_iova(unsigned long pfn_lo, unsigned long pfn_hi)
0533 {
0534     struct iova *iova;
0535 
0536     iova = alloc_iova_mem();
0537     if (iova) {
0538         iova->pfn_lo = pfn_lo;
0539         iova->pfn_hi = pfn_hi;
0540     }
0541 
0542     return iova;
0543 }
0544 
0545 static struct iova *
0546 __insert_new_range(struct iova_domain *iovad,
0547     unsigned long pfn_lo, unsigned long pfn_hi)
0548 {
0549     struct iova *iova;
0550 
0551     iova = alloc_and_init_iova(pfn_lo, pfn_hi);
0552     if (iova)
0553         iova_insert_rbtree(&iovad->rbroot, iova, NULL);
0554 
0555     return iova;
0556 }
0557 
0558 static void
0559 __adjust_overlap_range(struct iova *iova,
0560     unsigned long *pfn_lo, unsigned long *pfn_hi)
0561 {
0562     if (*pfn_lo < iova->pfn_lo)
0563         iova->pfn_lo = *pfn_lo;
0564     if (*pfn_hi > iova->pfn_hi)
0565         *pfn_lo = iova->pfn_hi + 1;
0566 }
0567 
0568 /**
0569  * reserve_iova - reserves an iova in the given range
0570  * @iovad: - iova domain pointer
0571  * @pfn_lo: - lower page frame address
0572  * @pfn_hi:- higher pfn adderss
0573  * This function allocates reserves the address range from pfn_lo to pfn_hi so
0574  * that this address is not dished out as part of alloc_iova.
0575  */
0576 struct iova *
0577 reserve_iova(struct iova_domain *iovad,
0578     unsigned long pfn_lo, unsigned long pfn_hi)
0579 {
0580     struct rb_node *node;
0581     unsigned long flags;
0582     struct iova *iova;
0583     unsigned int overlap = 0;
0584 
0585     /* Don't allow nonsensical pfns */
0586     if (WARN_ON((pfn_hi | pfn_lo) > (ULLONG_MAX >> iova_shift(iovad))))
0587         return NULL;
0588 
0589     spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
0590     for (node = rb_first(&iovad->rbroot); node; node = rb_next(node)) {
0591         if (__is_range_overlap(node, pfn_lo, pfn_hi)) {
0592             iova = to_iova(node);
0593             __adjust_overlap_range(iova, &pfn_lo, &pfn_hi);
0594             if ((pfn_lo >= iova->pfn_lo) &&
0595                 (pfn_hi <= iova->pfn_hi))
0596                 goto finish;
0597             overlap = 1;
0598 
0599         } else if (overlap)
0600                 break;
0601     }
0602 
0603     /* We are here either because this is the first reserver node
0604      * or need to insert remaining non overlap addr range
0605      */
0606     iova = __insert_new_range(iovad, pfn_lo, pfn_hi);
0607 finish:
0608 
0609     spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
0610     return iova;
0611 }
0612 EXPORT_SYMBOL_GPL(reserve_iova);
0613 
0614 /*
0615  * Magazine caches for IOVA ranges.  For an introduction to magazines,
0616  * see the USENIX 2001 paper "Magazines and Vmem: Extending the Slab
0617  * Allocator to Many CPUs and Arbitrary Resources" by Bonwick and Adams.
0618  * For simplicity, we use a static magazine size and don't implement the
0619  * dynamic size tuning described in the paper.
0620  */
0621 
0622 /*
0623  * As kmalloc's buffer size is fixed to power of 2, 127 is chosen to
0624  * assure size of 'iova_magazine' to be 1024 bytes, so that no memory
0625  * will be wasted.
0626  */
0627 #define IOVA_MAG_SIZE 127
0628 #define MAX_GLOBAL_MAGS 32  /* magazines per bin */
0629 
0630 struct iova_magazine {
0631     unsigned long size;
0632     unsigned long pfns[IOVA_MAG_SIZE];
0633 };
0634 
0635 struct iova_cpu_rcache {
0636     spinlock_t lock;
0637     struct iova_magazine *loaded;
0638     struct iova_magazine *prev;
0639 };
0640 
0641 struct iova_rcache {
0642     spinlock_t lock;
0643     unsigned long depot_size;
0644     struct iova_magazine *depot[MAX_GLOBAL_MAGS];
0645     struct iova_cpu_rcache __percpu *cpu_rcaches;
0646 };
0647 
0648 static struct iova_magazine *iova_magazine_alloc(gfp_t flags)
0649 {
0650     return kzalloc(sizeof(struct iova_magazine), flags);
0651 }
0652 
0653 static void iova_magazine_free(struct iova_magazine *mag)
0654 {
0655     kfree(mag);
0656 }
0657 
0658 static void
0659 iova_magazine_free_pfns(struct iova_magazine *mag, struct iova_domain *iovad)
0660 {
0661     unsigned long flags;
0662     int i;
0663 
0664     if (!mag)
0665         return;
0666 
0667     spin_lock_irqsave(&iovad->iova_rbtree_lock, flags);
0668 
0669     for (i = 0 ; i < mag->size; ++i) {
0670         struct iova *iova = private_find_iova(iovad, mag->pfns[i]);
0671 
0672         if (WARN_ON(!iova))
0673             continue;
0674 
0675         remove_iova(iovad, iova);
0676         free_iova_mem(iova);
0677     }
0678 
0679     spin_unlock_irqrestore(&iovad->iova_rbtree_lock, flags);
0680 
0681     mag->size = 0;
0682 }
0683 
0684 static bool iova_magazine_full(struct iova_magazine *mag)
0685 {
0686     return (mag && mag->size == IOVA_MAG_SIZE);
0687 }
0688 
0689 static bool iova_magazine_empty(struct iova_magazine *mag)
0690 {
0691     return (!mag || mag->size == 0);
0692 }
0693 
0694 static unsigned long iova_magazine_pop(struct iova_magazine *mag,
0695                        unsigned long limit_pfn)
0696 {
0697     int i;
0698     unsigned long pfn;
0699 
0700     BUG_ON(iova_magazine_empty(mag));
0701 
0702     /* Only fall back to the rbtree if we have no suitable pfns at all */
0703     for (i = mag->size - 1; mag->pfns[i] > limit_pfn; i--)
0704         if (i == 0)
0705             return 0;
0706 
0707     /* Swap it to pop it */
0708     pfn = mag->pfns[i];
0709     mag->pfns[i] = mag->pfns[--mag->size];
0710 
0711     return pfn;
0712 }
0713 
0714 static void iova_magazine_push(struct iova_magazine *mag, unsigned long pfn)
0715 {
0716     BUG_ON(iova_magazine_full(mag));
0717 
0718     mag->pfns[mag->size++] = pfn;
0719 }
0720 
0721 int iova_domain_init_rcaches(struct iova_domain *iovad)
0722 {
0723     unsigned int cpu;
0724     int i, ret;
0725 
0726     iovad->rcaches = kcalloc(IOVA_RANGE_CACHE_MAX_SIZE,
0727                  sizeof(struct iova_rcache),
0728                  GFP_KERNEL);
0729     if (!iovad->rcaches)
0730         return -ENOMEM;
0731 
0732     for (i = 0; i < IOVA_RANGE_CACHE_MAX_SIZE; ++i) {
0733         struct iova_cpu_rcache *cpu_rcache;
0734         struct iova_rcache *rcache;
0735 
0736         rcache = &iovad->rcaches[i];
0737         spin_lock_init(&rcache->lock);
0738         rcache->depot_size = 0;
0739         rcache->cpu_rcaches = __alloc_percpu(sizeof(*cpu_rcache),
0740                              cache_line_size());
0741         if (!rcache->cpu_rcaches) {
0742             ret = -ENOMEM;
0743             goto out_err;
0744         }
0745         for_each_possible_cpu(cpu) {
0746             cpu_rcache = per_cpu_ptr(rcache->cpu_rcaches, cpu);
0747 
0748             spin_lock_init(&cpu_rcache->lock);
0749             cpu_rcache->loaded = iova_magazine_alloc(GFP_KERNEL);
0750             cpu_rcache->prev = iova_magazine_alloc(GFP_KERNEL);
0751             if (!cpu_rcache->loaded || !cpu_rcache->prev) {
0752                 ret = -ENOMEM;
0753                 goto out_err;
0754             }
0755         }
0756     }
0757 
0758     ret = cpuhp_state_add_instance_nocalls(CPUHP_IOMMU_IOVA_DEAD,
0759                            &iovad->cpuhp_dead);
0760     if (ret)
0761         goto out_err;
0762     return 0;
0763 
0764 out_err:
0765     free_iova_rcaches(iovad);
0766     return ret;
0767 }
0768 EXPORT_SYMBOL_GPL(iova_domain_init_rcaches);
0769 
0770 /*
0771  * Try inserting IOVA range starting with 'iova_pfn' into 'rcache', and
0772  * return true on success.  Can fail if rcache is full and we can't free
0773  * space, and free_iova() (our only caller) will then return the IOVA
0774  * range to the rbtree instead.
0775  */
0776 static bool __iova_rcache_insert(struct iova_domain *iovad,
0777                  struct iova_rcache *rcache,
0778                  unsigned long iova_pfn)
0779 {
0780     struct iova_magazine *mag_to_free = NULL;
0781     struct iova_cpu_rcache *cpu_rcache;
0782     bool can_insert = false;
0783     unsigned long flags;
0784 
0785     cpu_rcache = raw_cpu_ptr(rcache->cpu_rcaches);
0786     spin_lock_irqsave(&cpu_rcache->lock, flags);
0787 
0788     if (!iova_magazine_full(cpu_rcache->loaded)) {
0789         can_insert = true;
0790     } else if (!iova_magazine_full(cpu_rcache->prev)) {
0791         swap(cpu_rcache->prev, cpu_rcache->loaded);
0792         can_insert = true;
0793     } else {
0794         struct iova_magazine *new_mag = iova_magazine_alloc(GFP_ATOMIC);
0795 
0796         if (new_mag) {
0797             spin_lock(&rcache->lock);
0798             if (rcache->depot_size < MAX_GLOBAL_MAGS) {
0799                 rcache->depot[rcache->depot_size++] =
0800                         cpu_rcache->loaded;
0801             } else {
0802                 mag_to_free = cpu_rcache->loaded;
0803             }
0804             spin_unlock(&rcache->lock);
0805 
0806             cpu_rcache->loaded = new_mag;
0807             can_insert = true;
0808         }
0809     }
0810 
0811     if (can_insert)
0812         iova_magazine_push(cpu_rcache->loaded, iova_pfn);
0813 
0814     spin_unlock_irqrestore(&cpu_rcache->lock, flags);
0815 
0816     if (mag_to_free) {
0817         iova_magazine_free_pfns(mag_to_free, iovad);
0818         iova_magazine_free(mag_to_free);
0819     }
0820 
0821     return can_insert;
0822 }
0823 
0824 static bool iova_rcache_insert(struct iova_domain *iovad, unsigned long pfn,
0825                    unsigned long size)
0826 {
0827     unsigned int log_size = order_base_2(size);
0828 
0829     if (log_size >= IOVA_RANGE_CACHE_MAX_SIZE)
0830         return false;
0831 
0832     return __iova_rcache_insert(iovad, &iovad->rcaches[log_size], pfn);
0833 }
0834 
0835 /*
0836  * Caller wants to allocate a new IOVA range from 'rcache'.  If we can
0837  * satisfy the request, return a matching non-NULL range and remove
0838  * it from the 'rcache'.
0839  */
0840 static unsigned long __iova_rcache_get(struct iova_rcache *rcache,
0841                        unsigned long limit_pfn)
0842 {
0843     struct iova_cpu_rcache *cpu_rcache;
0844     unsigned long iova_pfn = 0;
0845     bool has_pfn = false;
0846     unsigned long flags;
0847 
0848     cpu_rcache = raw_cpu_ptr(rcache->cpu_rcaches);
0849     spin_lock_irqsave(&cpu_rcache->lock, flags);
0850 
0851     if (!iova_magazine_empty(cpu_rcache->loaded)) {
0852         has_pfn = true;
0853     } else if (!iova_magazine_empty(cpu_rcache->prev)) {
0854         swap(cpu_rcache->prev, cpu_rcache->loaded);
0855         has_pfn = true;
0856     } else {
0857         spin_lock(&rcache->lock);
0858         if (rcache->depot_size > 0) {
0859             iova_magazine_free(cpu_rcache->loaded);
0860             cpu_rcache->loaded = rcache->depot[--rcache->depot_size];
0861             has_pfn = true;
0862         }
0863         spin_unlock(&rcache->lock);
0864     }
0865 
0866     if (has_pfn)
0867         iova_pfn = iova_magazine_pop(cpu_rcache->loaded, limit_pfn);
0868 
0869     spin_unlock_irqrestore(&cpu_rcache->lock, flags);
0870 
0871     return iova_pfn;
0872 }
0873 
0874 /*
0875  * Try to satisfy IOVA allocation range from rcache.  Fail if requested
0876  * size is too big or the DMA limit we are given isn't satisfied by the
0877  * top element in the magazine.
0878  */
0879 static unsigned long iova_rcache_get(struct iova_domain *iovad,
0880                      unsigned long size,
0881                      unsigned long limit_pfn)
0882 {
0883     unsigned int log_size = order_base_2(size);
0884 
0885     if (log_size >= IOVA_RANGE_CACHE_MAX_SIZE || !iovad->rcaches)
0886         return 0;
0887 
0888     return __iova_rcache_get(&iovad->rcaches[log_size], limit_pfn - size);
0889 }
0890 
0891 /*
0892  * free rcache data structures.
0893  */
0894 static void free_iova_rcaches(struct iova_domain *iovad)
0895 {
0896     struct iova_rcache *rcache;
0897     struct iova_cpu_rcache *cpu_rcache;
0898     unsigned int cpu;
0899     int i, j;
0900 
0901     for (i = 0; i < IOVA_RANGE_CACHE_MAX_SIZE; ++i) {
0902         rcache = &iovad->rcaches[i];
0903         if (!rcache->cpu_rcaches)
0904             break;
0905         for_each_possible_cpu(cpu) {
0906             cpu_rcache = per_cpu_ptr(rcache->cpu_rcaches, cpu);
0907             iova_magazine_free(cpu_rcache->loaded);
0908             iova_magazine_free(cpu_rcache->prev);
0909         }
0910         free_percpu(rcache->cpu_rcaches);
0911         for (j = 0; j < rcache->depot_size; ++j)
0912             iova_magazine_free(rcache->depot[j]);
0913     }
0914 
0915     kfree(iovad->rcaches);
0916     iovad->rcaches = NULL;
0917 }
0918 
0919 /*
0920  * free all the IOVA ranges cached by a cpu (used when cpu is unplugged)
0921  */
0922 static void free_cpu_cached_iovas(unsigned int cpu, struct iova_domain *iovad)
0923 {
0924     struct iova_cpu_rcache *cpu_rcache;
0925     struct iova_rcache *rcache;
0926     unsigned long flags;
0927     int i;
0928 
0929     for (i = 0; i < IOVA_RANGE_CACHE_MAX_SIZE; ++i) {
0930         rcache = &iovad->rcaches[i];
0931         cpu_rcache = per_cpu_ptr(rcache->cpu_rcaches, cpu);
0932         spin_lock_irqsave(&cpu_rcache->lock, flags);
0933         iova_magazine_free_pfns(cpu_rcache->loaded, iovad);
0934         iova_magazine_free_pfns(cpu_rcache->prev, iovad);
0935         spin_unlock_irqrestore(&cpu_rcache->lock, flags);
0936     }
0937 }
0938 
0939 /*
0940  * free all the IOVA ranges of global cache
0941  */
0942 static void free_global_cached_iovas(struct iova_domain *iovad)
0943 {
0944     struct iova_rcache *rcache;
0945     unsigned long flags;
0946     int i, j;
0947 
0948     for (i = 0; i < IOVA_RANGE_CACHE_MAX_SIZE; ++i) {
0949         rcache = &iovad->rcaches[i];
0950         spin_lock_irqsave(&rcache->lock, flags);
0951         for (j = 0; j < rcache->depot_size; ++j) {
0952             iova_magazine_free_pfns(rcache->depot[j], iovad);
0953             iova_magazine_free(rcache->depot[j]);
0954         }
0955         rcache->depot_size = 0;
0956         spin_unlock_irqrestore(&rcache->lock, flags);
0957     }
0958 }
0959 MODULE_AUTHOR("Anil S Keshavamurthy <anil.s.keshavamurthy@intel.com>");
0960 MODULE_LICENSE("GPL");
[ Source navigation ] [ Diff markup ] [ Identifier search ] [ general search ]

This page was automatically generated by the 2.1.0 LXR engine.

The LXR team
Valid CSS 2.1! Valid HTML 4.01!