167afa38eSTim Chen /* 267afa38eSTim Chen * Manage cache of swap slots to be used for and returned from 367afa38eSTim Chen * swap. 467afa38eSTim Chen * 567afa38eSTim Chen * Copyright(c) 2016 Intel Corporation. 667afa38eSTim Chen * 767afa38eSTim Chen * Author: Tim Chen <tim.c.chen@linux.intel.com> 867afa38eSTim Chen * 967afa38eSTim Chen * We allocate the swap slots from the global pool and put 1067afa38eSTim Chen * it into local per cpu caches. This has the advantage 1167afa38eSTim Chen * of no needing to acquire the swap_info lock every time 1267afa38eSTim Chen * we need a new slot. 1367afa38eSTim Chen * 1467afa38eSTim Chen * There is also opportunity to simply return the slot 1567afa38eSTim Chen * to local caches without needing to acquire swap_info 1667afa38eSTim Chen * lock. We do not reuse the returned slots directly but 1767afa38eSTim Chen * move them back to the global pool in a batch. This 1867afa38eSTim Chen * allows the slots to coaellesce and reduce fragmentation. 1967afa38eSTim Chen * 2067afa38eSTim Chen * The swap entry allocated is marked with SWAP_HAS_CACHE 2167afa38eSTim Chen * flag in map_count that prevents it from being allocated 2267afa38eSTim Chen * again from the global pool. 2367afa38eSTim Chen * 2467afa38eSTim Chen * The swap slots cache is protected by a mutex instead of 2567afa38eSTim Chen * a spin lock as when we search for slots with scan_swap_map, 2667afa38eSTim Chen * we can possibly sleep. 2767afa38eSTim Chen */ 2867afa38eSTim Chen 2967afa38eSTim Chen #include <linux/swap_slots.h> 3067afa38eSTim Chen #include <linux/cpu.h> 3167afa38eSTim Chen #include <linux/cpumask.h> 3267afa38eSTim Chen #include <linux/vmalloc.h> 3367afa38eSTim Chen #include <linux/mutex.h> 3467afa38eSTim Chen 3567afa38eSTim Chen #ifdef CONFIG_SWAP 3667afa38eSTim Chen 3767afa38eSTim Chen static DEFINE_PER_CPU(struct swap_slots_cache, swp_slots); 3867afa38eSTim Chen static bool swap_slot_cache_active; 39*ba81f838SHuang Ying bool swap_slot_cache_enabled; 4067afa38eSTim Chen static bool swap_slot_cache_initialized; 4167afa38eSTim Chen DEFINE_MUTEX(swap_slots_cache_mutex); 4267afa38eSTim Chen /* Serialize swap slots cache enable/disable operations */ 4367afa38eSTim Chen DEFINE_MUTEX(swap_slots_cache_enable_mutex); 4467afa38eSTim Chen 4567afa38eSTim Chen static void __drain_swap_slots_cache(unsigned int type); 4667afa38eSTim Chen static void deactivate_swap_slots_cache(void); 4767afa38eSTim Chen static void reactivate_swap_slots_cache(void); 4867afa38eSTim Chen 4967afa38eSTim Chen #define use_swap_slot_cache (swap_slot_cache_active && \ 5067afa38eSTim Chen swap_slot_cache_enabled && swap_slot_cache_initialized) 5167afa38eSTim Chen #define SLOTS_CACHE 0x1 5267afa38eSTim Chen #define SLOTS_CACHE_RET 0x2 5367afa38eSTim Chen 5467afa38eSTim Chen static void deactivate_swap_slots_cache(void) 5567afa38eSTim Chen { 5667afa38eSTim Chen mutex_lock(&swap_slots_cache_mutex); 5767afa38eSTim Chen swap_slot_cache_active = false; 5867afa38eSTim Chen __drain_swap_slots_cache(SLOTS_CACHE|SLOTS_CACHE_RET); 5967afa38eSTim Chen mutex_unlock(&swap_slots_cache_mutex); 6067afa38eSTim Chen } 6167afa38eSTim Chen 6267afa38eSTim Chen static void reactivate_swap_slots_cache(void) 6367afa38eSTim Chen { 6467afa38eSTim Chen mutex_lock(&swap_slots_cache_mutex); 6567afa38eSTim Chen swap_slot_cache_active = true; 6667afa38eSTim Chen mutex_unlock(&swap_slots_cache_mutex); 6767afa38eSTim Chen } 6867afa38eSTim Chen 6967afa38eSTim Chen /* Must not be called with cpu hot plug lock */ 7067afa38eSTim Chen void disable_swap_slots_cache_lock(void) 7167afa38eSTim Chen { 7267afa38eSTim Chen mutex_lock(&swap_slots_cache_enable_mutex); 7367afa38eSTim Chen swap_slot_cache_enabled = false; 7467afa38eSTim Chen if (swap_slot_cache_initialized) { 7567afa38eSTim Chen /* serialize with cpu hotplug operations */ 7667afa38eSTim Chen get_online_cpus(); 7767afa38eSTim Chen __drain_swap_slots_cache(SLOTS_CACHE|SLOTS_CACHE_RET); 7867afa38eSTim Chen put_online_cpus(); 7967afa38eSTim Chen } 8067afa38eSTim Chen } 8167afa38eSTim Chen 8267afa38eSTim Chen static void __reenable_swap_slots_cache(void) 8367afa38eSTim Chen { 8467afa38eSTim Chen swap_slot_cache_enabled = has_usable_swap(); 8567afa38eSTim Chen } 8667afa38eSTim Chen 8767afa38eSTim Chen void reenable_swap_slots_cache_unlock(void) 8867afa38eSTim Chen { 8967afa38eSTim Chen __reenable_swap_slots_cache(); 9067afa38eSTim Chen mutex_unlock(&swap_slots_cache_enable_mutex); 9167afa38eSTim Chen } 9267afa38eSTim Chen 9367afa38eSTim Chen static bool check_cache_active(void) 9467afa38eSTim Chen { 9567afa38eSTim Chen long pages; 9667afa38eSTim Chen 9767afa38eSTim Chen if (!swap_slot_cache_enabled || !swap_slot_cache_initialized) 9867afa38eSTim Chen return false; 9967afa38eSTim Chen 10067afa38eSTim Chen pages = get_nr_swap_pages(); 10167afa38eSTim Chen if (!swap_slot_cache_active) { 10267afa38eSTim Chen if (pages > num_online_cpus() * 10367afa38eSTim Chen THRESHOLD_ACTIVATE_SWAP_SLOTS_CACHE) 10467afa38eSTim Chen reactivate_swap_slots_cache(); 10567afa38eSTim Chen goto out; 10667afa38eSTim Chen } 10767afa38eSTim Chen 10867afa38eSTim Chen /* if global pool of slot caches too low, deactivate cache */ 10967afa38eSTim Chen if (pages < num_online_cpus() * THRESHOLD_DEACTIVATE_SWAP_SLOTS_CACHE) 11067afa38eSTim Chen deactivate_swap_slots_cache(); 11167afa38eSTim Chen out: 11267afa38eSTim Chen return swap_slot_cache_active; 11367afa38eSTim Chen } 11467afa38eSTim Chen 11567afa38eSTim Chen static int alloc_swap_slot_cache(unsigned int cpu) 11667afa38eSTim Chen { 11767afa38eSTim Chen struct swap_slots_cache *cache; 11867afa38eSTim Chen swp_entry_t *slots, *slots_ret; 11967afa38eSTim Chen 12067afa38eSTim Chen /* 12167afa38eSTim Chen * Do allocation outside swap_slots_cache_mutex 12267afa38eSTim Chen * as vzalloc could trigger reclaim and get_swap_page, 12367afa38eSTim Chen * which can lock swap_slots_cache_mutex. 12467afa38eSTim Chen */ 12567afa38eSTim Chen slots = vzalloc(sizeof(swp_entry_t) * SWAP_SLOTS_CACHE_SIZE); 12667afa38eSTim Chen if (!slots) 12767afa38eSTim Chen return -ENOMEM; 12867afa38eSTim Chen 12967afa38eSTim Chen slots_ret = vzalloc(sizeof(swp_entry_t) * SWAP_SLOTS_CACHE_SIZE); 13067afa38eSTim Chen if (!slots_ret) { 13167afa38eSTim Chen vfree(slots); 13267afa38eSTim Chen return -ENOMEM; 13367afa38eSTim Chen } 13467afa38eSTim Chen 13567afa38eSTim Chen mutex_lock(&swap_slots_cache_mutex); 13667afa38eSTim Chen cache = &per_cpu(swp_slots, cpu); 13767afa38eSTim Chen if (cache->slots || cache->slots_ret) 13867afa38eSTim Chen /* cache already allocated */ 13967afa38eSTim Chen goto out; 14067afa38eSTim Chen if (!cache->lock_initialized) { 14167afa38eSTim Chen mutex_init(&cache->alloc_lock); 14267afa38eSTim Chen spin_lock_init(&cache->free_lock); 14367afa38eSTim Chen cache->lock_initialized = true; 14467afa38eSTim Chen } 14567afa38eSTim Chen cache->nr = 0; 14667afa38eSTim Chen cache->cur = 0; 14767afa38eSTim Chen cache->n_ret = 0; 14867afa38eSTim Chen cache->slots = slots; 14967afa38eSTim Chen slots = NULL; 15067afa38eSTim Chen cache->slots_ret = slots_ret; 15167afa38eSTim Chen slots_ret = NULL; 15267afa38eSTim Chen out: 15367afa38eSTim Chen mutex_unlock(&swap_slots_cache_mutex); 15467afa38eSTim Chen if (slots) 15567afa38eSTim Chen vfree(slots); 15667afa38eSTim Chen if (slots_ret) 15767afa38eSTim Chen vfree(slots_ret); 15867afa38eSTim Chen return 0; 15967afa38eSTim Chen } 16067afa38eSTim Chen 16167afa38eSTim Chen static void drain_slots_cache_cpu(unsigned int cpu, unsigned int type, 16267afa38eSTim Chen bool free_slots) 16367afa38eSTim Chen { 16467afa38eSTim Chen struct swap_slots_cache *cache; 16567afa38eSTim Chen swp_entry_t *slots = NULL; 16667afa38eSTim Chen 16767afa38eSTim Chen cache = &per_cpu(swp_slots, cpu); 16867afa38eSTim Chen if ((type & SLOTS_CACHE) && cache->slots) { 16967afa38eSTim Chen mutex_lock(&cache->alloc_lock); 17067afa38eSTim Chen swapcache_free_entries(cache->slots + cache->cur, cache->nr); 17167afa38eSTim Chen cache->cur = 0; 17267afa38eSTim Chen cache->nr = 0; 17367afa38eSTim Chen if (free_slots && cache->slots) { 17467afa38eSTim Chen vfree(cache->slots); 17567afa38eSTim Chen cache->slots = NULL; 17667afa38eSTim Chen } 17767afa38eSTim Chen mutex_unlock(&cache->alloc_lock); 17867afa38eSTim Chen } 17967afa38eSTim Chen if ((type & SLOTS_CACHE_RET) && cache->slots_ret) { 18067afa38eSTim Chen spin_lock_irq(&cache->free_lock); 18167afa38eSTim Chen swapcache_free_entries(cache->slots_ret, cache->n_ret); 18267afa38eSTim Chen cache->n_ret = 0; 18367afa38eSTim Chen if (free_slots && cache->slots_ret) { 18467afa38eSTim Chen slots = cache->slots_ret; 18567afa38eSTim Chen cache->slots_ret = NULL; 18667afa38eSTim Chen } 18767afa38eSTim Chen spin_unlock_irq(&cache->free_lock); 18867afa38eSTim Chen if (slots) 18967afa38eSTim Chen vfree(slots); 19067afa38eSTim Chen } 19167afa38eSTim Chen } 19267afa38eSTim Chen 19367afa38eSTim Chen static void __drain_swap_slots_cache(unsigned int type) 19467afa38eSTim Chen { 19567afa38eSTim Chen unsigned int cpu; 19667afa38eSTim Chen 19767afa38eSTim Chen /* 19867afa38eSTim Chen * This function is called during 19967afa38eSTim Chen * 1) swapoff, when we have to make sure no 20067afa38eSTim Chen * left over slots are in cache when we remove 20167afa38eSTim Chen * a swap device; 20267afa38eSTim Chen * 2) disabling of swap slot cache, when we run low 20367afa38eSTim Chen * on swap slots when allocating memory and need 20467afa38eSTim Chen * to return swap slots to global pool. 20567afa38eSTim Chen * 20667afa38eSTim Chen * We cannot acquire cpu hot plug lock here as 20767afa38eSTim Chen * this function can be invoked in the cpu 20867afa38eSTim Chen * hot plug path: 20967afa38eSTim Chen * cpu_up -> lock cpu_hotplug -> cpu hotplug state callback 21067afa38eSTim Chen * -> memory allocation -> direct reclaim -> get_swap_page 21167afa38eSTim Chen * -> drain_swap_slots_cache 21267afa38eSTim Chen * 21367afa38eSTim Chen * Hence the loop over current online cpu below could miss cpu that 21467afa38eSTim Chen * is being brought online but not yet marked as online. 21567afa38eSTim Chen * That is okay as we do not schedule and run anything on a 21667afa38eSTim Chen * cpu before it has been marked online. Hence, we will not 21767afa38eSTim Chen * fill any swap slots in slots cache of such cpu. 21867afa38eSTim Chen * There are no slots on such cpu that need to be drained. 21967afa38eSTim Chen */ 22067afa38eSTim Chen for_each_online_cpu(cpu) 22167afa38eSTim Chen drain_slots_cache_cpu(cpu, type, false); 22267afa38eSTim Chen } 22367afa38eSTim Chen 22467afa38eSTim Chen static int free_slot_cache(unsigned int cpu) 22567afa38eSTim Chen { 22667afa38eSTim Chen mutex_lock(&swap_slots_cache_mutex); 22767afa38eSTim Chen drain_slots_cache_cpu(cpu, SLOTS_CACHE | SLOTS_CACHE_RET, true); 22867afa38eSTim Chen mutex_unlock(&swap_slots_cache_mutex); 22967afa38eSTim Chen return 0; 23067afa38eSTim Chen } 23167afa38eSTim Chen 23267afa38eSTim Chen int enable_swap_slots_cache(void) 23367afa38eSTim Chen { 23467afa38eSTim Chen int ret = 0; 23567afa38eSTim Chen 23667afa38eSTim Chen mutex_lock(&swap_slots_cache_enable_mutex); 23767afa38eSTim Chen if (swap_slot_cache_initialized) { 23867afa38eSTim Chen __reenable_swap_slots_cache(); 23967afa38eSTim Chen goto out_unlock; 24067afa38eSTim Chen } 24167afa38eSTim Chen 24267afa38eSTim Chen ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "swap_slots_cache", 24367afa38eSTim Chen alloc_swap_slot_cache, free_slot_cache); 24467afa38eSTim Chen if (ret < 0) 24567afa38eSTim Chen goto out_unlock; 24667afa38eSTim Chen swap_slot_cache_initialized = true; 24767afa38eSTim Chen __reenable_swap_slots_cache(); 24867afa38eSTim Chen out_unlock: 24967afa38eSTim Chen mutex_unlock(&swap_slots_cache_enable_mutex); 25067afa38eSTim Chen return 0; 25167afa38eSTim Chen } 25267afa38eSTim Chen 25367afa38eSTim Chen /* called with swap slot cache's alloc lock held */ 25467afa38eSTim Chen static int refill_swap_slots_cache(struct swap_slots_cache *cache) 25567afa38eSTim Chen { 25667afa38eSTim Chen if (!use_swap_slot_cache || cache->nr) 25767afa38eSTim Chen return 0; 25867afa38eSTim Chen 25967afa38eSTim Chen cache->cur = 0; 26067afa38eSTim Chen if (swap_slot_cache_active) 26167afa38eSTim Chen cache->nr = get_swap_pages(SWAP_SLOTS_CACHE_SIZE, cache->slots); 26267afa38eSTim Chen 26367afa38eSTim Chen return cache->nr; 26467afa38eSTim Chen } 26567afa38eSTim Chen 26667afa38eSTim Chen int free_swap_slot(swp_entry_t entry) 26767afa38eSTim Chen { 26867afa38eSTim Chen struct swap_slots_cache *cache; 26967afa38eSTim Chen 27067afa38eSTim Chen BUG_ON(!swap_slot_cache_initialized); 27167afa38eSTim Chen 27267afa38eSTim Chen cache = &get_cpu_var(swp_slots); 27367afa38eSTim Chen if (use_swap_slot_cache && cache->slots_ret) { 27467afa38eSTim Chen spin_lock_irq(&cache->free_lock); 27567afa38eSTim Chen /* Swap slots cache may be deactivated before acquiring lock */ 27667afa38eSTim Chen if (!use_swap_slot_cache) { 27767afa38eSTim Chen spin_unlock_irq(&cache->free_lock); 27867afa38eSTim Chen goto direct_free; 27967afa38eSTim Chen } 28067afa38eSTim Chen if (cache->n_ret >= SWAP_SLOTS_CACHE_SIZE) { 28167afa38eSTim Chen /* 28267afa38eSTim Chen * Return slots to global pool. 28367afa38eSTim Chen * The current swap_map value is SWAP_HAS_CACHE. 28467afa38eSTim Chen * Set it to 0 to indicate it is available for 28567afa38eSTim Chen * allocation in global pool 28667afa38eSTim Chen */ 28767afa38eSTim Chen swapcache_free_entries(cache->slots_ret, cache->n_ret); 28867afa38eSTim Chen cache->n_ret = 0; 28967afa38eSTim Chen } 29067afa38eSTim Chen cache->slots_ret[cache->n_ret++] = entry; 29167afa38eSTim Chen spin_unlock_irq(&cache->free_lock); 29267afa38eSTim Chen } else { 29367afa38eSTim Chen direct_free: 29467afa38eSTim Chen swapcache_free_entries(&entry, 1); 29567afa38eSTim Chen } 29667afa38eSTim Chen put_cpu_var(swp_slots); 29767afa38eSTim Chen 29867afa38eSTim Chen return 0; 29967afa38eSTim Chen } 30067afa38eSTim Chen 30167afa38eSTim Chen swp_entry_t get_swap_page(void) 30267afa38eSTim Chen { 30367afa38eSTim Chen swp_entry_t entry, *pentry; 30467afa38eSTim Chen struct swap_slots_cache *cache; 30567afa38eSTim Chen 30667afa38eSTim Chen /* 30767afa38eSTim Chen * Preemption is allowed here, because we may sleep 30867afa38eSTim Chen * in refill_swap_slots_cache(). But it is safe, because 30967afa38eSTim Chen * accesses to the per-CPU data structure are protected by the 31067afa38eSTim Chen * mutex cache->alloc_lock. 31167afa38eSTim Chen * 31267afa38eSTim Chen * The alloc path here does not touch cache->slots_ret 31367afa38eSTim Chen * so cache->free_lock is not taken. 31467afa38eSTim Chen */ 31567afa38eSTim Chen cache = raw_cpu_ptr(&swp_slots); 31667afa38eSTim Chen 31767afa38eSTim Chen entry.val = 0; 31867afa38eSTim Chen if (check_cache_active()) { 31967afa38eSTim Chen mutex_lock(&cache->alloc_lock); 32067afa38eSTim Chen if (cache->slots) { 32167afa38eSTim Chen repeat: 32267afa38eSTim Chen if (cache->nr) { 32367afa38eSTim Chen pentry = &cache->slots[cache->cur++]; 32467afa38eSTim Chen entry = *pentry; 32567afa38eSTim Chen pentry->val = 0; 32667afa38eSTim Chen cache->nr--; 32767afa38eSTim Chen } else { 32867afa38eSTim Chen if (refill_swap_slots_cache(cache)) 32967afa38eSTim Chen goto repeat; 33067afa38eSTim Chen } 33167afa38eSTim Chen } 33267afa38eSTim Chen mutex_unlock(&cache->alloc_lock); 33367afa38eSTim Chen if (entry.val) 33467afa38eSTim Chen return entry; 33567afa38eSTim Chen } 33667afa38eSTim Chen 33767afa38eSTim Chen get_swap_pages(1, &entry); 33867afa38eSTim Chen 33967afa38eSTim Chen return entry; 34067afa38eSTim Chen } 34167afa38eSTim Chen 34267afa38eSTim Chen #endif /* CONFIG_SWAP */ 343