| 1 | #include "internal.h" |
| 2 | |
| 3 | /* Free_queue function naming semantics: |
| 4 | * |
| 5 | * "Draining" is when the free_queue elements are removed one by one, |
| 6 | * and each element first tries to get put on a given per-core |
| 7 | * cache's magazines. The free_queue elements that do not fit in the |
| 8 | * magazine are then optionally freed from the slab cache or re-enqueued. |
| 9 | * |
| 10 | * "Flushing" is when the free_queue elements are all freed from the |
| 11 | * slab cache. The per-core magazines are not touched */ |
| 12 | |
| 13 | void slab_free_queue_init(struct slab_domain *domain, struct slab_free_queue *q, |
| 14 | size_t capacity) { |
| 15 | q->capacity = capacity; |
| 16 | q->slots = |
| 17 | kmalloc(sizeof(struct slab_free_slot) * capacity, ALLOC_FLAGS_ZERO); |
| 18 | if (!q->slots) |
| 19 | panic("Could not allocate slab free queue slots!" ); |
| 20 | |
| 21 | atomic_store(&q->head, 0); |
| 22 | atomic_store(&q->tail, 0); |
| 23 | |
| 24 | for (size_t i = 0; i < capacity; i++) |
| 25 | atomic_store(&q->slots[i].seq, i); |
| 26 | |
| 27 | q->parent = domain; |
| 28 | q->count = 0; |
| 29 | } |
| 30 | |
| 31 | bool slab_free_queue_ringbuffer_enqueue(struct slab_free_queue *q, |
| 32 | vaddr_t addr) { |
| 33 | uint64_t pos; |
| 34 | struct slab_free_slot *slot; |
| 35 | uint64_t seq; |
| 36 | int64_t diff; |
| 37 | |
| 38 | while (true) { |
| 39 | pos = atomic_load_explicit(&q->head, memory_order_relaxed); |
| 40 | slot = &q->slots[pos % q->capacity]; |
| 41 | seq = atomic_load_explicit(&slot->seq, memory_order_acquire); |
| 42 | diff = (int64_t) seq - (int64_t) pos; |
| 43 | |
| 44 | if (diff == 0) { |
| 45 | if (atomic_compare_exchange_weak_explicit(&q->head, &pos, pos + 1, |
| 46 | memory_order_acq_rel, |
| 47 | memory_order_relaxed)) { |
| 48 | |
| 49 | slot->addr = addr; |
| 50 | |
| 51 | atomic_store_explicit(&slot->seq, pos + 1, |
| 52 | memory_order_release); |
| 53 | |
| 54 | SLAB_FREE_QUEUE_INC_COUNT(q); |
| 55 | return true; |
| 56 | } |
| 57 | } else if (diff < 0) { |
| 58 | return false; |
| 59 | } |
| 60 | } |
| 61 | } |
| 62 | |
| 63 | vaddr_t slab_free_queue_ringbuffer_dequeue(struct slab_free_queue *q) { |
| 64 | uint64_t pos; |
| 65 | struct slab_free_slot *slot; |
| 66 | uint64_t seq; |
| 67 | int64_t diff; |
| 68 | |
| 69 | while (true) { |
| 70 | pos = atomic_load_explicit(&q->tail, memory_order_relaxed); |
| 71 | slot = &q->slots[pos % q->capacity]; |
| 72 | seq = atomic_load_explicit(&slot->seq, memory_order_acquire); |
| 73 | diff = (int64_t) seq - (int64_t) (pos + 1); |
| 74 | |
| 75 | if (diff == 0) { |
| 76 | if (atomic_compare_exchange_weak_explicit(&q->tail, &pos, pos + 1, |
| 77 | memory_order_acq_rel, |
| 78 | memory_order_relaxed)) { |
| 79 | |
| 80 | vaddr_t ret = slot->addr; |
| 81 | slot->addr = 0; |
| 82 | |
| 83 | atomic_store_explicit(&slot->seq, pos + q->capacity, |
| 84 | memory_order_release); |
| 85 | |
| 86 | SLAB_FREE_QUEUE_DEC_COUNT(q); |
| 87 | return ret; |
| 88 | } |
| 89 | } else if (diff < 0) { |
| 90 | return 0x0; |
| 91 | } |
| 92 | } |
| 93 | } |
| 94 | |
| 95 | bool slab_free_queue_enqueue(struct slab_free_queue *q, vaddr_t addr) { |
| 96 | return slab_free_queue_ringbuffer_enqueue(q, addr); |
| 97 | } |
| 98 | |
| 99 | vaddr_t slab_free_queue_dequeue(struct slab_free_queue *q) { |
| 100 | return slab_free_queue_ringbuffer_dequeue(q); |
| 101 | } |
| 102 | |
| 103 | static void slab_free_queue_free(struct slab_domain *d, void *ptr, |
| 104 | enum alloc_behavior bh) { |
| 105 | int32_t class = slab_size_to_index(size: ksize(ptr)); |
| 106 | bool fits_in_slab = class >= 0; |
| 107 | |
| 108 | if (fits_in_slab) |
| 109 | return slab_free(domain: d, obj: ptr); |
| 110 | |
| 111 | struct slab_page_hdr * = slab_page_hdr_for_addr(ptr); |
| 112 | return slab_free_page_hdr(hdr: header, bh); |
| 113 | } |
| 114 | |
| 115 | size_t slab_free_queue_drain(struct slab_percpu_cache *cache, |
| 116 | struct slab_free_queue *queue, size_t target, |
| 117 | enum alloc_behavior bh) { |
| 118 | kassert(cache == slab_percpu_cache_local()); |
| 119 | size_t drained_to_magazine = 0; /* Return value */ |
| 120 | size_t addrs_dequeued = 0; /* Used to check against `target` */ |
| 121 | |
| 122 | while (true) { |
| 123 | if (addrs_dequeued >= target) |
| 124 | break; |
| 125 | |
| 126 | /* Drain an element from our free_queue */ |
| 127 | vaddr_t addr = slab_free_queue_dequeue(q: queue); |
| 128 | if (!addr) |
| 129 | break; |
| 130 | |
| 131 | addrs_dequeued++; |
| 132 | |
| 133 | /* What class? */ |
| 134 | int32_t class = slab_size_to_index(size: slab_allocation_size(addr)); |
| 135 | if (class < 0) |
| 136 | goto flush; |
| 137 | |
| 138 | /* Magazines only cache nonpageable addresses */ |
| 139 | struct slab *slab = slab_for_ptr(ptr: (void *) addr); |
| 140 | if (slab_is_pageable(s: slab)) |
| 141 | goto flush; |
| 142 | |
| 143 | /* Push it onto the magazine */ |
| 144 | enum slab_magazine_type mtype = |
| 145 | slab_is_zeroed(s: slab) ? SLAB_MAGAZINE_ZERO : SLAB_MAGAZINE_NORMAL; |
| 146 | |
| 147 | if (mtype == SLAB_MAGAZINE_ZERO) |
| 148 | memset((void *) addr, 0, slab->parent_cache->obj_size); |
| 149 | |
| 150 | struct slab_magazine *mag = &cache->mags[mtype][class]; |
| 151 | if (!slab_magazine_push(mag, obj: addr)) |
| 152 | goto flush; |
| 153 | |
| 154 | /* Success - pushed onto magazine */ |
| 155 | drained_to_magazine++; |
| 156 | continue; |
| 157 | |
| 158 | flush: |
| 159 | slab_free_queue_free(d: cache->domain, ptr: (void *) addr, bh); |
| 160 | } |
| 161 | |
| 162 | return drained_to_magazine; |
| 163 | } |
| 164 | |
| 165 | size_t slab_free_queue_flush(struct slab_domain *domain, |
| 166 | struct slab_free_queue *queue) { |
| 167 | size_t total_freed = 0; |
| 168 | |
| 169 | /* Drain the ringbuffer one element at a time */ |
| 170 | while (true) { |
| 171 | vaddr_t addr = slab_free_queue_ringbuffer_dequeue(q: queue); |
| 172 | if (addr == 0x0) |
| 173 | break; |
| 174 | |
| 175 | slab_free(domain, obj: (void *) addr); |
| 176 | } |
| 177 | return total_freed; |
| 178 | } |
| 179 | |
| 180 | size_t slab_free_queue_get_target_drain(struct slab_domain *domain, |
| 181 | size_t pct) { |
| 182 | size_t slab_domain_cpus = domain->domain->num_cores; |
| 183 | size_t total_fq_elems = SLAB_FREE_QUEUE_GET_COUNT(&domain->free_queue); |
| 184 | size_t portion = slab_domain_cpus / SLAB_PERCPU_REFILL_PER_CORE_WEIGHT; |
| 185 | if (portion == 0) |
| 186 | portion = 1; |
| 187 | |
| 188 | return (total_fq_elems / portion) * pct / 100; |
| 189 | } |
| 190 | |
| 191 | size_t slab_free_queue_drain_limited(struct slab_percpu_cache *pc, |
| 192 | struct slab_domain *dom, size_t pct, |
| 193 | enum alloc_behavior bh) { |
| 194 | size_t target = slab_free_queue_get_target_drain(domain: dom, pct); |
| 195 | |
| 196 | /* This will also fill up the magazines for other orders. We set the target |
| 197 | * to prevent overly aggressive stealing from the free_queue into our |
| 198 | * percpu cache to allow other CPUs in our domain to get their fair share of |
| 199 | * what remains in the free_queue in the event that they must also refill */ |
| 200 | return slab_free_queue_drain(cache: pc, queue: &dom->free_queue, target, bh); |
| 201 | } |
| 202 | |