// SPDX-License-Identifier: GPL-3.0-or-later
#include "cache.h"
/* STATES AND TRANSITIONS
*
* entry | entry
* v v
* HOT -> DIRTY --> CLEAN --> EVICT
* v | v
* flush | evict
* v | v
* save | free
* callback | callback
*
*/
// to use ARAL uncomment the following line:
#if !defined(FSANITIZE_ADDRESS)
#define PGC_WITH_ARAL 1
#endif
#define PGC_QUEUE_LOCK_AS_WAITING_QUEUE 1
typedef enum __attribute__ ((__packed__)) {
// mutually exclusive flags
PGC_PAGE_CLEAN = (1 << 0), // none of the following
PGC_PAGE_DIRTY = (1 << 1), // contains unsaved data
PGC_PAGE_HOT = (1 << 2), // currently being collected
// flags related to various actions on each page
PGC_PAGE_IS_BEING_DELETED = (1 << 3),
PGC_PAGE_IS_BEING_MIGRATED_TO_V2 = (1 << 4),
PGC_PAGE_HAS_NO_DATA_IGNORE_ACCESSES = (1 << 5),
PGC_PAGE_HAS_BEEN_ACCESSED = (1 << 6),
} PGC_PAGE_FLAGS;
#define page_flag_check(page, flag) (__atomic_load_n(&((page)->flags), __ATOMIC_ACQUIRE) & (flag))
#define page_flag_set(page, flag) __atomic_or_fetch(&((page)->flags), flag, __ATOMIC_RELEASE)
#define page_flag_clear(page, flag) __atomic_and_fetch(&((page)->flags), ~(flag), __ATOMIC_RELEASE)
#define page_get_status_flags(page) page_flag_check(page, PGC_PAGE_HOT | PGC_PAGE_DIRTY | PGC_PAGE_CLEAN)
#define is_page_hot(page) (page_get_status_flags(page) == PGC_PAGE_HOT)
#define is_page_dirty(page) (page_get_status_flags(page) == PGC_PAGE_DIRTY)
#define is_page_clean(page) (page_get_status_flags(page) == PGC_PAGE_CLEAN)
struct pgc_page {
// indexing data
Word_t section;
Word_t metric_id;
time_t start_time_s;
time_t end_time_s;
uint32_t update_every_s;
uint32_t assumed_size;
REFCOUNT refcount;
uint16_t accesses; // counts the number of accesses on this page
PGC_PAGE_FLAGS flags;
SPINLOCK transition_spinlock; // when the page changes between HOT, DIRTY, CLEAN, we have to get this lock
struct {
struct pgc_page *next;
struct pgc_page *prev;
} link;
void *data;
uint8_t custom_data[];
// IMPORTANT!
// THIS STRUCTURE NEEDS TO BE INITIALIZED BY HAND!
};
struct pgc_queue {
#if defined(PGC_QUEUE_LOCK_AS_WAITING_QUEUE)
WAITQ wq;
#else
SPINLOCK spinlock;
#endif
union {
PGC_PAGE *base;
Pvoid_t sections_judy;
};
PGC_PAGE_FLAGS flags;
size_t version;
size_t last_version_checked;
bool linked_list_in_sections_judy; // when true, we use 'sections_judy', otherwise we use 'base'
struct pgc_queue_statistics *stats;
};
struct pgc {
struct {
char name[PGC_NAME_MAX + 1];
bool stats; // enable extended statistics
bool use_all_ram;
size_t partitions;
int64_t clean_size;
size_t max_dirty_pages_per_call;
size_t max_pages_per_inline_eviction;
size_t max_skip_pages_per_inline_eviction;
size_t max_flushes_inline;
size_t max_workers_evict_inline;
size_t additional_bytes_per_page;
int64_t out_of_memory_protection_bytes;
free_clean_page_callback pgc_free_clean_cb;
save_dirty_page_callback pgc_save_dirty_cb;
save_dirty_init_callback pgc_save_init_cb;
PGC_OPTIONS options;
ssize_t severe_pressure_per1000;
ssize_t aggressive_evict_per1000;
ssize_t healthy_size_per1000;
ssize_t evict_low_threshold_per1000;
dynamic_target_cache_size_callback dynamic_target_size_cb;
nominal_page_size_callback nominal_page_size_cb;
} config;
struct {
ND_THREAD *thread; // the thread
struct completion completion; // signal the thread to wake up
} evictor;
struct pgc_index {
RW_SPINLOCK rw_spinlock;
Pvoid_t sections_judy;
#ifdef PGC_WITH_ARAL
ARAL *aral;
#endif
} *index;
struct {
SPINLOCK spinlock;
ssize_t per1000;
} usage;
struct pgc_queue clean; // LRU is applied here to free memory from the cache
struct pgc_queue dirty; // in the dirty list, pages are ordered the way they were marked dirty
struct pgc_queue hot; // in the hot list, pages are order the way they were marked hot
struct pgc_statistics stats; // statistics
#ifdef NETDATA_PGC_POINTER_CHECK
netdata_mutex_t global_pointer_registry_mutex;
Pvoid_t global_pointer_registry;
#endif
};
// ----------------------------------------------------------------------------
// validate each pointer is indexed once - internal checks only
static inline void pointer_index_init(PGC *cache __maybe_unused) {
#ifdef NETDATA_PGC_POINTER_CHECK
netdata_mutex_init(&cache->global_pointer_registry_mutex);
#else
;
#endif
}
static inline void pointer_destroy_index(PGC *cache __maybe_unused) {
#ifdef NETDATA_PGC_POINTER_CHECK
netdata_mutex_lock(&cache->global_pointer_registry_mutex);
JudyHSFreeArray(&cache->global_pointer_registry, PJE0);
netdata_mutex_unlock(&cache->global_pointer_registry_mutex);
#else
;
#endif
}
static inline void pointer_add(PGC *cache __maybe_unused, PGC_PAGE *page __maybe_unused) {
#ifdef NETDATA_PGC_POINTER_CHECK
netdata_mutex_lock(&cache->global_pointer_registry_mutex);
Pvoid_t *PValue = JudyHSIns(&cache->global_pointer_registry, &page, sizeof(void *), PJE0);
if(*PValue != NULL)
fatal("pointer already exists in registry");
*PValue = page;
netdata_mutex_unlock(&cache->global_pointer_registry_mutex);
#else
;
#endif
}
static inline void pointer_check(PGC *cache __maybe_unused, PGC_PAGE *page __maybe_unused) {
#ifdef NETDATA_PGC_POINTER_CHECK
netdata_mutex_lock(&cache->global_pointer_registry_mutex);
Pvoid_t *PValue = JudyHSGet(cache->global_pointer_registry, &page, sizeof(void *));
if(PValue == NULL)
fatal("pointer is not found in registry");
netdata_mutex_unlock(&cache->global_pointer_registry_mutex);
#else
;
#endif
}
static inline void pointer_del(PGC *cache __maybe_unused, PGC_PAGE *page __maybe_unused) {
#ifdef NETDATA_PGC_POINTER_CHECK
netdata_mutex_lock(&cache->global_pointer_registry_mutex);
int ret = JudyHSDel(&cache->global_pointer_registry, &page, sizeof(void *), PJE0);
if(!ret)
fatal("pointer to be deleted does not exist in registry");
netdata_mutex_unlock(&cache->global_pointer_registry_mutex);
#else
;
#endif
}
// ----------------------------------------------------------------------------
// helpers
static inline size_t page_assumed_size(PGC *cache, size_t size) {
return size + (sizeof(PGC_PAGE) + cache->config.additional_bytes_per_page + sizeof(Word_t) * 3);
}
static inline size_t page_size_from_assumed_size(PGC *cache, size_t assumed_size) {
return assumed_size - (sizeof(PGC_PAGE) + cache->config.additional_bytes_per_page + sizeof(Word_t) * 3);
}
// ----------------------------------------------------------------------------
// locking
static inline size_t pgc_indexing_partition(PGC *cache, Word_t metric_id) {
static __thread Word_t last_metric_id = 0;
static __thread size_t last_partition = 0;
if(metric_id == last_metric_id || cache->config.partitions == 1)
return last_partition;
last_metric_id = metric_id;
last_partition = indexing_partition(metric_id, cache->config.partitions);
return last_partition;
}
#define pgc_index_read_lock(cache, partition) rw_spinlock_read_lock(&(cache)->index[partition].rw_spinlock)
#define pgc_index_read_unlock(cache, partition) rw_spinlock_read_unlock(&(cache)->index[partition].rw_spinlock)
#define pgc_index_write_lock(cache, partition) rw_spinlock_write_lock(&(cache)->index[partition].rw_spinlock)
#define pgc_index_write_unlock(cache, partition) rw_spinlock_write_unlock(&(cache)->index[partition].rw_spinlock)
#define pgc_index_trywrite_lock(cache, partition, force) ({ \
bool _result; \
if (force) { \
rw_spinlock_write_lock(&(cache)->index[partition].rw_spinlock); \
_result = true; \
} else \
_result = rw_spinlock_trywrite_lock(&(cache)->index[partition].rw_spinlock); \
_result; \
})
#define PGC_QUEUE_LOCK_PRIO_COLLECTORS WAITQ_PRIO_URGENT
#define PGC_QUEUE_LOCK_PRIO_EVICTORS WAITQ_PRIO_HIGH
#define PGC_QUEUE_LOCK_PRIO_FLUSHERS WAITQ_PRIO_NORMAL
#define PGC_QUEUE_LOCK_PRIO_LOW WAITQ_PRIO_LOW
#if defined(PGC_QUEUE_LOCK_AS_WAITING_QUEUE)
#define pgc_queue_trylock(cache, ll, prio) waitq_try_acquire(&((ll)->wq), prio)
#define pgc_queue_lock(cache, ll, prio) waitq_acquire(&((ll)->wq), prio)
#define pgc_queue_unlock(cache, ll) waitq_release(&((ll)->wq))
#else
#define pgc_queue_trylock(cache, ll, prio) spinlock_trylock(&((ll)->spinlock))
#define pgc_queue_lock(cache, ll, prio) spinlock_lock(&((ll)->spinlock))
#define pgc_queue_unlock(cache, ll) spinlock_unlock(&((ll)->spinlock))
#endif
#define page_transition_trylock(cache, page) spinlock_trylock(&(page)->transition_spinlock)
#define page_transition_lock(cache, page) spinlock_lock(&(page)->transition_spinlock)
#define page_transition_unlock(cache, page) spinlock_unlock(&(page)->transition_spinlock)
// ----------------------------------------------------------------------------
// size histogram
static void pgc_size_histogram_init(struct pgc_size_histogram *h) {
// the histogram needs to be all-inclusive for the possible sizes
// so, we start from 0, and the last value is SIZE_MAX.
size_t values[PGC_SIZE_HISTOGRAM_ENTRIES] = {
0, 32, 64, 128, 256, 512, 1024, 2048,
4096, 8192, 16384, 32768, 65536, 128 * 1024, SIZE_MAX
};
size_t last_value = 0;
for(size_t i = 0; i < PGC_SIZE_HISTOGRAM_ENTRIES; i++) {
if(i > 0 && values[i] == 0)
fatal("only the first value in the array can be zero");
if(i > 0 && values[i] <= last_value)
fatal("the values need to be sorted");
h->array[i].upto = values[i];
last_value = values[i];
}
}
static inline size_t pgc_size_histogram_slot(struct pgc_size_histogram *h, size_t size) {
if(size <= h->array[0].upto)
return 0;
if(size >= h->array[_countof(h->array) - 1].upto)
return _countof(h->array) - 1;
// binary search for the right size
size_t low = 0, high = _countof(h->array) - 1;
while (low < high) {
size_t mid = low + (high - low) / 2;
if (size < h->array[mid].upto)
high = mid;
else
low = mid + 1;
}
return low - 1;
}
static inline void pgc_size_histogram_add(PGC *cache, struct pgc_size_histogram *h, PGC_PAGE *page) {
size_t size;
if(cache->config.nominal_page_size_cb)
size = cache->config.nominal_page_size_cb(page->data);
else
size = page_size_from_assumed_size(cache, page->assumed_size);
size_t slot = pgc_size_histogram_slot(h, size);
internal_fatal(slot >= _countof(h->array), "hey!");
__atomic_add_fetch(&h->array[slot].count, 1, __ATOMIC_RELAXED);
}
static inline void pgc_size_histogram_del(PGC *cache, struct pgc_size_histogram *h, PGC_PAGE *page) {
size_t size;
if(cache->config.nominal_page_size_cb)
size = cache->config.nominal_page_size_cb(page->data);
else
size = page_size_from_assumed_size(cache, page->assumed_size);
size_t slot = pgc_size_histogram_slot(h, size);
internal_fatal(slot >= _countof(h->array), "hey!");
__atomic_sub_fetch(&h->array[slot].count, 1, __ATOMIC_RELAXED);
}
// ----------------------------------------------------------------------------
// evictions control
ALWAYS_INLINE
static int64_t pgc_threshold(ssize_t threshold, int64_t wanted, int64_t current, int64_t clean) {
if(current < clean)
current = clean;
if(wanted < current - clean)
wanted = current - clean;
int64_t ret = wanted * threshold / 1000LL;
if(ret < current - clean)
ret = current - clean;
return ret;
}
ALWAYS_INLINE
static int64_t pgc_wanted_size(const int64_t hot, const int64_t hot_max, const int64_t dirty_max, const int64_t index) {
// our promise to users
const int64_t max_size1 = MAX(hot_max, hot) * 2;
// protection against slow flushing
const int64_t max_size2 = hot_max + MAX(dirty_max * 2, hot_max * 2 / 3) + index;
// the final wanted cache size
return MIN(max_size1, max_size2);
}
static ssize_t cache_usage_per1000(PGC *cache, int64_t *size_to_evict) {
if(size_to_evict)
spinlock_lock(&cache->usage.spinlock);
else if(!spinlock_trylock(&cache->usage.spinlock))
return __atomic_load_n(&cache->usage.per1000, __ATOMIC_RELAXED);
int64_t wanted_cache_size;
const int64_t dirty = __atomic_load_n(&cache->dirty.stats->size, __ATOMIC_RELAXED);
const int64_t hot = __atomic_load_n(&cache->hot.stats->size, __ATOMIC_RELAXED);
const int64_t clean = __atomic_load_n(&cache->clean.stats->size, __ATOMIC_RELAXED);
const int64_t evicting = __atomic_load_n(&cache->stats.evicting_size, __ATOMIC_RELAXED);
const int64_t flushing = __atomic_load_n(&cache->stats.flushing_size, __ATOMIC_RELAXED);
const int64_t current_cache_size = __atomic_load_n(&cache->stats.size, __ATOMIC_RELAXED);
const int64_t all_pages_size = hot + dirty + clean + evicting + flushing;
const int64_t index = current_cache_size > all_pages_size ? current_cache_size - all_pages_size : 0;
const int64_t referenced_size = __atomic_load_n(&cache->stats.referenced_size, __ATOMIC_RELAXED);
if(cache->config.options & PGC_OPTIONS_AUTOSCALE) {
const int64_t dirty_max = __atomic_load_n(&cache->dirty.stats->max_size, __ATOMIC_RELAXED);
const int64_t hot_max = __atomic_load_n(&cache->hot.stats->max_size, __ATOMIC_RELAXED);
if(cache->config.dynamic_target_size_cb) {
wanted_cache_size = pgc_wanted_size(hot, hot, dirty, index);
const int64_t wanted_cache_size_cb = cache->config.dynamic_target_size_cb();
if(wanted_cache_size_cb > wanted_cache_size)
wanted_cache_size = wanted_cache_size_cb;
}
else
wanted_cache_size = pgc_wanted_size(hot, hot_max, dirty_max, index);
if (wanted_cache_size < hot + dirty + index + cache->config.clean_size)
wanted_cache_size = hot + dirty + index + cache->config.clean_size;
}
else
wanted_cache_size = hot + dirty + index + cache->config.clean_size;
// calculate the absolute minimum we can go
const int64_t min_cache_size1 = (referenced_size > hot ? referenced_size : hot) + dirty + index;
const int64_t min_cache_size2 = (current_cache_size > clean) ? current_cache_size - clean : min_cache_size1;
const int64_t min_cache_size = MAX(min_cache_size1, min_cache_size2);
if(cache->config.out_of_memory_protection_bytes) {
// out of memory protection
OS_SYSTEM_MEMORY sm = os_system_memory(false);
if(OS_SYSTEM_MEMORY_OK(sm)) {
// when the total exists, ram_available_bytes is also right
const int64_t ram_available_bytes = (int64_t)sm.ram_available_bytes;
const int64_t min_available = cache->config.out_of_memory_protection_bytes;
if (ram_available_bytes < min_available) {
// we must shrink
int64_t must_lose = min_available - ram_available_bytes;
if(current_cache_size > must_lose)
wanted_cache_size = current_cache_size - must_lose;
else
wanted_cache_size = min_cache_size;
}
else if(cache->config.use_all_ram) {
// we can grow
wanted_cache_size = current_cache_size + (ram_available_bytes - min_available);
}
}
}
// never go below our minimum
if(unlikely(wanted_cache_size < min_cache_size))
wanted_cache_size = min_cache_size;
// protection for the case the cache is totally empty
if(unlikely(wanted_cache_size < 65536))
wanted_cache_size = 65536;
const ssize_t per1000 = (ssize_t)(current_cache_size * 1000LL / wanted_cache_size);
__atomic_store_n(&cache->usage.per1000, per1000, __ATOMIC_RELAXED);
__atomic_store_n(&cache->stats.wanted_cache_size, wanted_cache_size, __ATOMIC_RELAXED);
__atomic_store_n(&cache->stats.current_cache_size, current_cache_size, __ATOMIC_RELAXED);
int64_t healthy_target = pgc_threshold(cache->config.healthy_size_per1000, wanted_cache_size, current_cache_size, clean);
if(current_cache_size > healthy_target) {
int64_t low_watermark_target = pgc_threshold(cache->config.evict_low_threshold_per1000, wanted_cache_size, current_cache_size, clean);
int64_t size_to_evict_now = current_cache_size - low_watermark_target;
if(size_to_evict_now > clean)
size_to_evict_now = clean;
if(size_to_evict)
*size_to_evict = size_to_evict_now;
bool signal = false;
if(per1000 >= cache->config.severe_pressure_per1000) {
__atomic_add_fetch(&cache->stats.events_cache_under_severe_pressure, 1, __ATOMIC_RELAXED);
signal = true;
}
else if(per1000 >= cache->config.aggressive_evict_per1000) {
__atomic_add_fetch(&cache->stats.events_cache_needs_space_aggressively, 1, __ATOMIC_RELAXED);
signal = true;
}
if(signal) {
completion_mark_complete_a_job(&cache->evictor.completion);
p2_add_fetch(&cache->stats.p2_waste_evict_thread_signals, 1);
}
}
spinlock_unlock(&cache->usage.spinlock);
return per1000;
}
static inline bool cache_pressure(PGC *cache, ssize_t limit) {
return (cache_usage_per1000(cache, NULL) >= limit);
}
#define cache_under_severe_pressure(cache) cache_pressure(cache, (cache)->config.severe_pressure_per1000)
#define cache_needs_space_aggressively(cache) cache_pressure(cache, (cache)->config.aggressive_evict_per1000)
#define cache_above_healthy_limit(cache) cache_pressure(cache, (cache)->config.healthy_size_per1000)
typedef bool (*evict_filter)(PGC_PAGE *page, void *data);
static bool evict_pages_with_filter(PGC *cache, size_t max_skip, size_t max_evict, bool wait, bool all_of_them, evict_filter filter, void *data);
#define evict_pages(cache, max_skip, max_evict, wait, all_of_them) evict_pages_with_filter(cache, max_skip, max_evict, wait, all_of_them, NULL, NULL)
static inline bool flushing_critical(PGC *cache);
static bool flush_pages(PGC *cache, size_t max_flushes, Word_t section, bool wait, bool all_of_them);
static ALWAYS_INLINE void evict_pages_inline(PGC *cache, bool on_release) {
const ssize_t per1000 = cache_usage_per1000(cache, NULL);
if(!(cache->config.options & PGC_OPTIONS_EVICT_PAGES_NO_INLINE)) {
if (per1000 > cache->config.aggressive_evict_per1000 && !on_release) {
// the threads that add pages, turn into evictors when the cache needs evictions aggressively
p2_add_fetch(&cache->stats.p2_waste_evictions_inline_on_add, 1);
evict_pages(cache,
cache->config.max_skip_pages_per_inline_eviction,
cache->config.max_pages_per_inline_eviction,
false, false);
}
else if (per1000 > cache->config.severe_pressure_per1000 && on_release) {
// the threads that are releasing pages, turn into evictors when the cache is critical
p2_add_fetch(&cache->stats.p2_waste_evictions_inline_on_release, 1);
evict_pages(cache,
cache->config.max_skip_pages_per_inline_eviction,
cache->config.max_pages_per_inline_eviction,
false, false);
}
}
}
static ALWAYS_INLINE void evict_on_clean_page_added(PGC *cache) {
evict_pages_inline(cache, false);
}
static ALWAYS_INLINE void evict_on_page_release_when_permitted(PGC *cache) {
evict_pages_inline(cache, true);
}
static ALWAYS_INLINE void flush_inline(PGC *cache, bool on_release) {
if(!(cache->config.options & PGC_OPTIONS_FLUSH_PAGES_NO_INLINE) && flushing_critical(cache)) {
if (on_release)
p2_add_fetch(&cache->stats.p2_waste_flush_on_release, 1);
else
p2_add_fetch(&cache->stats.p2_waste_flush_on_add, 1);
flush_pages(cache, cache->config.max_flushes_inline, PGC_SECTION_ALL, false, false);
}
}
static ALWAYS_INLINE void flush_on_page_add(PGC *cache) {
flush_inline(cache, false);
}
static ALWAYS_INLINE void flush_on_page_hot_release(PGC *cache) {
flush_inline(cache, true);
}
// ----------------------------------------------------------------------------
// flushing control
static inline bool flushing_critical(PGC *cache) {
if(unlikely(__atomic_load_n(&cache->dirty.stats->size, __ATOMIC_RELAXED) > __atomic_load_n(&cache->hot.stats->max_size, __ATOMIC_RELAXED))) {
__atomic_add_fetch(&cache->stats.events_flush_critical, 1, __ATOMIC_RELAXED);
return true;
}
return false;
}
// ----------------------------------------------------------------------------
// Linked list management
static inline void atomic_set_max_size_t(size_t *max, size_t desired) {
size_t expected;
expected = __atomic_load_n(max, __ATOMIC_RELAXED);
do {
if(expected >= desired)
return;
} while(!__atomic_compare_exchange_n(max, &expected, desired,
false, __ATOMIC_RELAXED, __ATOMIC_RELAXED));
}
static inline void atomic_set_max_int64_t(int64_t *max, int64_t desired) {
int64_t expected;
expected = __atomic_load_n(max, __ATOMIC_RELAXED);
do {
if(expected >= desired)
return;
} while(!__atomic_compare_exchange_n(max, &expected, desired,
false, __ATOMIC_RELAXED, __ATOMIC_RELAXED));
}
struct section_pages {
SPINLOCK migration_to_v2_spinlock;
size_t entries;
size_t size;
PGC_PAGE *base;
};
static struct aral_statistics pgc_aral_statistics = { 0 };
static ARAL *pgc_sections_aral = NULL;
static void pgc_section_pages_static_aral_init(void) {
static SPINLOCK spinlock = SPINLOCK_INITIALIZER;
spinlock_lock(&spinlock);
if(!pgc_sections_aral) {
pgc_sections_aral = aral_create(
"pgc-sections", sizeof(struct section_pages), 0, 0, &pgc_aral_statistics,
NULL, NULL, false, false, false);
pulse_aral_register_statistics(&pgc_aral_statistics, "pgc");
}
spinlock_unlock(&spinlock);
}
static ALWAYS_INLINE void pgc_stats_queue_judy_change(PGC *cache, struct pgc_queue *ll, int64_t delta) {
__atomic_add_fetch(&ll->stats->size, delta, __ATOMIC_RELAXED);
__atomic_add_fetch(&cache->stats.size, delta, __ATOMIC_RELAXED);
}
static ALWAYS_INLINE void pgc_stats_index_judy_change(PGC *cache, int64_t delta) {
__atomic_add_fetch(&cache->stats.size, delta, __ATOMIC_RELAXED);
}
static ALWAYS_INLINE void pgc_queue_add(PGC *cache __maybe_unused, struct pgc_queue *q, PGC_PAGE *page, bool having_lock, WAITQ_PRIORITY prio __maybe_unused) {
if(!having_lock)
pgc_queue_lock(cache, q, prio);
internal_fatal(page_get_status_flags(page) != 0,
"DBENGINE CACHE: invalid page flags, the page has %d, but it is should be %d",
page_get_status_flags(page),
0);
if(q->linked_list_in_sections_judy) {
// HOT and DIRTY pages end up here.
JudyAllocThreadPulseReset();
int64_t mem_delta = 0;
Pvoid_t *section_pages_pptr = JudyLIns(&q->sections_judy, page->section, PJE0);
struct section_pages *sp = *section_pages_pptr;
if(!sp) {
// sp = callocz(1, sizeof(struct section_pages));
sp = aral_mallocz(pgc_sections_aral);
memset(sp, 0, sizeof(struct section_pages));
*section_pages_pptr = sp;
mem_delta += sizeof(struct section_pages);
}
mem_delta += JudyAllocThreadPulseGetAndReset();
pgc_stats_queue_judy_change(cache, q, mem_delta);
sp->entries++;
sp->size += page->assumed_size;
DOUBLE_LINKED_LIST_APPEND_ITEM_UNSAFE(sp->base, page, link.prev, link.next);
if((sp->entries % cache->config.max_dirty_pages_per_call) == 0)
q->version++;
}
else {
// CLEAN pages end up here.
// - New pages created as CLEAN, always have 1 access.
// - DIRTY pages made CLEAN, depending on their accesses may be appended (accesses > 0) or prepended (accesses = 0).
if(page->accesses || page_flag_check(page, PGC_PAGE_HAS_BEEN_ACCESSED | PGC_PAGE_HAS_NO_DATA_IGNORE_ACCESSES) == PGC_PAGE_HAS_BEEN_ACCESSED) {
DOUBLE_LINKED_LIST_APPEND_ITEM_UNSAFE(q->base, page, link.prev, link.next);
page_flag_clear(page, PGC_PAGE_HAS_BEEN_ACCESSED);
}
else
DOUBLE_LINKED_LIST_PREPEND_ITEM_UNSAFE(q->base, page, link.prev, link.next);
q->version++;
}
page_flag_set(page, q->flags);
if(!having_lock)
pgc_queue_unlock(cache, q);
size_t entries = __atomic_add_fetch(&q->stats->entries, 1, __ATOMIC_RELAXED);
int64_t size = __atomic_add_fetch(&q->stats->size, page->assumed_size, __ATOMIC_RELAXED);
__atomic_add_fetch(&q->stats->added_entries, 1, __ATOMIC_RELAXED);
__atomic_add_fetch(&q->stats->added_size, page->assumed_size, __ATOMIC_RELAXED);
atomic_set_max_size_t(&q->stats->max_entries, entries);
atomic_set_max_int64_t(&q->stats->max_size, size);
if(cache->config.stats)
pgc_size_histogram_add(cache, &q->stats->size_histogram, page);
}
static ALWAYS_INLINE void pgc_queue_del(PGC *cache __maybe_unused, struct pgc_queue *q, PGC_PAGE *page, bool having_lock,
WAITQ_PRIORITY prio __maybe_unused) {
if(cache->config.stats)
pgc_size_histogram_del(cache, &q->stats->size_histogram, page);
__atomic_sub_fetch(&q->stats->entries, 1, __ATOMIC_RELAXED);
__atomic_sub_fetch(&q->stats->size, page->assumed_size, __ATOMIC_RELAXED);
__atomic_add_fetch(&q->stats->removed_entries, 1, __ATOMIC_RELAXED);
__atomic_add_fetch(&q->stats->removed_size, page->assumed_size, __ATOMIC_RELAXED);
if(!having_lock)
pgc_queue_lock(cache, q, prio);
internal_fatal(page_get_status_flags(page) != q->flags,
"DBENGINE CACHE: invalid page flags, the page has %d, but it is should be %d",
page_get_status_flags(page),
q->flags);
page_flag_clear(page, q->flags);
if(q->linked_list_in_sections_judy) {
Pvoid_t *section_pages_pptr = JudyLGet(q->sections_judy, page->section, PJE0);
internal_fatal(!section_pages_pptr, "DBENGINE CACHE: page should be in Judy LL, but it is not");
struct section_pages *sp = *section_pages_pptr;
sp->entries--;
sp->size -= page->assumed_size;
DOUBLE_LINKED_LIST_REMOVE_ITEM_UNSAFE(sp->base, page, link.prev, link.next);
if(!sp->base) {
JudyAllocThreadPulseReset();
int64_t mem_delta = 0;
int rc = JudyLDel(&q->sections_judy, page->section, PJE0);
if(!rc)
fatal("DBENGINE CACHE: cannot delete section from Judy LL");
// freez(sp);
aral_freez(pgc_sections_aral, sp);
mem_delta -= sizeof(struct section_pages);
mem_delta += JudyAllocThreadPulseGetAndReset();
pgc_stats_queue_judy_change(cache, q, mem_delta);
}
}
else {
DOUBLE_LINKED_LIST_REMOVE_ITEM_UNSAFE(q->base, page, link.prev, link.next);
q->version++;
}
if(!having_lock)
pgc_queue_unlock(cache, q);
}
static ALWAYS_INLINE void page_has_been_accessed(PGC *cache, PGC_PAGE *page) {
PGC_PAGE_FLAGS flags = page_flag_check(page, PGC_PAGE_CLEAN | PGC_PAGE_HAS_NO_DATA_IGNORE_ACCESSES);
if (!(flags & PGC_PAGE_HAS_NO_DATA_IGNORE_ACCESSES)) {
__atomic_add_fetch(&page->accesses, 1, __ATOMIC_RELAXED);
if (flags & PGC_PAGE_CLEAN) {
if(pgc_queue_trylock(cache, &cache->clean, PGC_QUEUE_LOCK_PRIO_EVICTORS)) {
DOUBLE_LINKED_LIST_REMOVE_ITEM_UNSAFE(cache->clean.base, page, link.prev, link.next);
DOUBLE_LINKED_LIST_APPEND_ITEM_UNSAFE(cache->clean.base, page, link.prev, link.next);
pgc_queue_unlock(cache, &cache->clean);
page_flag_clear(page, PGC_PAGE_HAS_BEEN_ACCESSED);
}
else
page_flag_set(page, PGC_PAGE_HAS_BEEN_ACCESSED);
}
}
}
// ----------------------------------------------------------------------------
// state transitions
static ALWAYS_INLINE void page_set_clean(PGC *cache, PGC_PAGE *page, bool having_transition_lock, bool having_clean_lock, WAITQ_PRIORITY prio) {
if(!having_transition_lock)
page_transition_lock(cache, page);
PGC_PAGE_FLAGS flags = page_get_status_flags(page);
if(flags & PGC_PAGE_CLEAN) {
if(!having_transition_lock)
page_transition_unlock(cache, page);
return;
}
if(flags & PGC_PAGE_HOT)
pgc_queue_del(cache, &cache->hot, page, false, prio);
if(flags & PGC_PAGE_DIRTY)
pgc_queue_del(cache, &cache->dirty, page, false, prio);
// first add to linked list, the set the flag (required for move_page_last())
pgc_queue_add(cache, &cache->clean, page, having_clean_lock, prio);
if(!having_transition_lock)
page_transition_unlock(cache, page);
}
static ALWAYS_INLINE void page_set_dirty(PGC *cache, PGC_PAGE *page, bool having_hot_lock, WAITQ_PRIORITY prio) {
if(!having_hot_lock)
// to avoid deadlocks, we have to get the hot lock before the page transition
// since this is what all_hot_to_dirty() does
pgc_queue_lock(cache, &cache->hot, prio);
page_transition_lock(cache, page);
PGC_PAGE_FLAGS flags = page_get_status_flags(page);
if(flags & PGC_PAGE_DIRTY) {
page_transition_unlock(cache, page);
if(!having_hot_lock)
// we don't need the hot lock anymore
pgc_queue_unlock(cache, &cache->hot);
return;
}
__atomic_add_fetch(&cache->stats.hot2dirty_entries, 1, __ATOMIC_RELAXED);
__atomic_add_fetch(&cache->stats.hot2dirty_size, page->assumed_size, __ATOMIC_RELAXED);
if(likely(flags & PGC_PAGE_HOT))
pgc_queue_del(cache, &cache->hot, page, true, prio);
if(!having_hot_lock)
// we don't need the hot lock anymore
pgc_queue_unlock(cache, &cache->hot);
if(unlikely(flags & PGC_PAGE_CLEAN))
pgc_queue_del(cache, &cache->clean, page, false, prio);
// first add to linked list, the set the flag (required for move_page_last())
pgc_queue_add(cache, &cache->dirty, page, false, prio);
__atomic_sub_fetch(&cache->stats.hot2dirty_entries, 1, __ATOMIC_RELAXED);
__atomic_sub_fetch(&cache->stats.hot2dirty_size, page->assumed_size, __ATOMIC_RELAXED);
page_transition_unlock(cache, page);
}
static ALWAYS_INLINE void page_set_hot(PGC *cache, PGC_PAGE *page, WAITQ_PRIORITY prio) {
page_transition_lock(cache, page);
PGC_PAGE_FLAGS flags = page_get_status_flags(page);
if(flags & PGC_PAGE_HOT) {
page_transition_unlock(cache, page);
return;
}
if(flags & PGC_PAGE_DIRTY)
pgc_queue_del(cache, &cache->dirty, page, false, prio);
if(flags & PGC_PAGE_CLEAN)
pgc_queue_del(cache, &cache->clean, page, false, prio);
// first add to linked list, the set the flag (required for move_page_last())
pgc_queue_add(cache, &cache->hot, page, false, prio);
page_transition_unlock(cache, page);
}
// ----------------------------------------------------------------------------
// Referencing
static ALWAYS_INLINE size_t PGC_REFERENCED_PAGES(PGC *cache) {
return __atomic_load_n(&cache->stats.referenced_entries, __ATOMIC_RELAXED);
}
static ALWAYS_INLINE void PGC_REFERENCED_PAGES_PLUS1(PGC *cache, PGC_PAGE *page) {
__atomic_add_fetch(&cache->stats.referenced_entries, 1, __ATOMIC_RELAXED);
__atomic_add_fetch(&cache->stats.referenced_size, page->assumed_size, __ATOMIC_RELAXED);
}
static ALWAYS_INLINE void PGC_REFERENCED_PAGES_MINUS1(PGC *cache, int64_t assumed_size) {
__atomic_sub_fetch(&cache->stats.referenced_entries, 1, __ATOMIC_RELAXED);
__atomic_sub_fetch(&cache->stats.referenced_size, assumed_size, __ATOMIC_RELAXED);
}
// If the page is not already acquired,
// YOU HAVE TO HAVE THE QUEUE (hot, dirty, clean - the page is in), LOCKED!
// If you don't have it locked, NOTHING PREVENTS THIS PAGE FROM VANISHING WHILE THIS IS CALLED!
static ALWAYS_INLINE bool page_acquire(PGC *cache, PGC_PAGE *page) {
__atomic_add_fetch(&cache->stats.acquires, 1, __ATOMIC_RELAXED);
REFCOUNT rc = refcount_acquire_advanced(&page->refcount);
if(REFCOUNT_ACQUIRED(rc)) {
if(rc == 1)
PGC_REFERENCED_PAGES_PLUS1(cache, page);
return true;
}
return false;
}
static ALWAYS_INLINE void page_release(PGC *cache, PGC_PAGE *page, bool evict_if_necessary) {
__atomic_add_fetch(&cache->stats.releases, 1, __ATOMIC_RELAXED);
int64_t assumed_size = page->assumed_size; // take the size before we release it
if(refcount_release(&page->refcount) == 0) {
PGC_REFERENCED_PAGES_MINUS1(cache, assumed_size);
if(evict_if_necessary)
evict_on_page_release_when_permitted(cache);
}
}
static ALWAYS_INLINE bool non_acquired_page_get_for_deletion___while_having_clean_locked(PGC *cache __maybe_unused, PGC_PAGE *page) {
__atomic_add_fetch(&cache->stats.acquires_for_deletion, 1, __ATOMIC_RELAXED);
internal_fatal(!is_page_clean(page),
"DBENGINE CACHE: only clean pages can be deleted");
if(refcount_acquire_for_deletion(&page->refcount)) {
// we can delete this page
internal_fatal(page_flag_check(page, PGC_PAGE_IS_BEING_DELETED),
"DBENGINE CACHE: page is already being deleted");
page_flag_set(page, PGC_PAGE_IS_BEING_DELETED);
return true;
}
return false;
}
static ALWAYS_INLINE bool acquired_page_get_for_deletion_or_release_it(PGC *cache __maybe_unused, PGC_PAGE *page) {
__atomic_add_fetch(&cache->stats.acquires_for_deletion, 1, __ATOMIC_RELAXED);
int64_t assumed_size = page->assumed_size; // take the size before we release it
if(refcount_release_and_acquire_for_deletion(&page->refcount)) {
PGC_REFERENCED_PAGES_MINUS1(cache, assumed_size);
// we can delete this page
internal_fatal(page_flag_check(page, PGC_PAGE_IS_BEING_DELETED),
"DBENGINE CACHE: page is already being deleted");
page_flag_set(page, PGC_PAGE_IS_BEING_DELETED);
return true;
}
return false;
}
// ----------------------------------------------------------------------------
// Indexing
static inline void free_this_page(PGC *cache, PGC_PAGE *page, size_t partition __maybe_unused) {
size_t size = page_size_from_assumed_size(cache, page->assumed_size);
// call the callback to free the user supplied memory
cache->config.pgc_free_clean_cb(cache, (PGC_ENTRY){
.section = page->section,
.metric_id = page->metric_id,
.start_time_s = page->start_time_s,
.end_time_s = __atomic_load_n(&page->end_time_s, __ATOMIC_RELAXED),
.update_every_s = page->update_every_s,
.size = size,
.hot = (is_page_hot(page)) ? true : false,
.data = page->data,
.custom_data = (cache->config.additional_bytes_per_page) ? page->custom_data : NULL,
});
timing_dbengine_evict_step(TIMING_STEP_DBENGINE_EVICT_FREE_CB);
// update statistics
__atomic_add_fetch(&cache->stats.removed_entries, 1, __ATOMIC_RELAXED);
__atomic_add_fetch(&cache->stats.removed_size, page->assumed_size, __ATOMIC_RELAXED);
__atomic_sub_fetch(&cache->stats.entries, 1, __ATOMIC_RELAXED);
__atomic_sub_fetch(&cache->stats.size, page->assumed_size, __ATOMIC_RELAXED);
timing_dbengine_evict_step(TIMING_STEP_DBENGINE_EVICT_FREE_ATOMICS2);
// free our memory
#ifdef PGC_WITH_ARAL
aral_freez(cache->index[partition].aral, page);
#else
freez(page);
#endif
timing_dbengine_evict_step(TIMING_STEP_DBENGINE_EVICT_FREE_ARAL);
}
static void remove_this_page_from_index_unsafe(PGC *cache, PGC_PAGE *page, size_t partition) {
// remove it from the Judy arrays
pointer_check(cache, page);
internal_fatal(page_flag_check(page, PGC_PAGE_HOT | PGC_PAGE_DIRTY | PGC_PAGE_CLEAN),
"DBENGINE CACHE: page to be removed from the cache is still in the linked-list");
internal_fatal(!page_flag_check(page, PGC_PAGE_IS_BEING_DELETED),
"DBENGINE CACHE: page to be removed from the index, is not marked for deletion");
internal_fatal(partition != pgc_indexing_partition(cache, page->metric_id),
"DBENGINE CACHE: attempted to remove this page from the wrong partition of the cache");
Pvoid_t *metrics_judy_pptr = JudyLGet(cache->index[partition].sections_judy, page->section, PJE0);
if(unlikely(!metrics_judy_pptr))
fatal("DBENGINE CACHE: section '%lu' should exist, but it does not.", page->section);
Pvoid_t *pages_judy_pptr = JudyLGet(*metrics_judy_pptr, page->metric_id, PJE0);
if(unlikely(!pages_judy_pptr))
fatal("DBENGINE CACHE: metric '%lu' in section '%lu' should exist, but it does not.",
page->metric_id, page->section);
Pvoid_t *page_ptr = JudyLGet(*pages_judy_pptr, page->start_time_s, PJE0);
if(unlikely(!page_ptr))
fatal("DBENGINE CACHE: page with start time '%ld' of metric '%lu' in section '%lu' should exist, but it does not.",
page->start_time_s, page->metric_id, page->section);
PGC_PAGE *found_page = *page_ptr;
if(unlikely(found_page != page))
fatal("DBENGINE CACHE: page with start time '%ld' of metric '%lu' in section '%lu' should exist, but the index returned a different address.",
page->start_time_s, page->metric_id, page->section);
JudyAllocThreadPulseReset();
if(unlikely(!JudyLDel(pages_judy_pptr, page->start_time_s, PJE0)))
fatal("DBENGINE CACHE: page with start time '%ld' of metric '%lu' in section '%lu' exists, but cannot be deleted.",
page->start_time_s, page->metric_id, page->section);
if(!*pages_judy_pptr && !JudyLDel(metrics_judy_pptr, page->metric_id, PJE0))
fatal("DBENGINE CACHE: metric '%lu' in section '%lu' exists and is empty, but cannot be deleted.",
page->metric_id, page->section);
if(!*metrics_judy_pptr && !JudyLDel(&cache->index[partition].sections_judy, page->section, PJE0))
fatal("DBENGINE CACHE: section '%lu' exists and is empty, but cannot be deleted.", page->section);
pgc_stats_index_judy_change(cache, JudyAllocThreadPulseGetAndReset());
pointer_del(cache, page);
}
static inline void remove_and_free_page_not_in_any_queue_and_acquired_for_deletion(PGC *cache, PGC_PAGE *page) {
size_t partition = pgc_indexing_partition(cache, page->metric_id);
pgc_index_write_lock(cache, partition);
remove_this_page_from_index_unsafe(cache, page, partition);
pgc_index_write_unlock(cache, partition);
free_this_page(cache, page, partition);
}
static inline bool make_acquired_page_clean_and_evict_or_page_release(PGC *cache, PGC_PAGE *page) {
pointer_check(cache, page);
WAITQ_PRIORITY prio = is_page_clean(page) ? PGC_QUEUE_LOCK_PRIO_EVICTORS : PGC_QUEUE_LOCK_PRIO_COLLECTORS;
page_transition_lock(cache, page);
pgc_queue_lock(cache, &cache->clean, prio);
// make it clean - it does not have any accesses, so it will be prepended
page_set_clean(cache, page, true, true, prio);
if(!acquired_page_get_for_deletion_or_release_it(cache, page)) {
pgc_queue_unlock(cache, &cache->clean);
page_transition_unlock(cache, page);
return false;
}
// remove it from the linked list
pgc_queue_del(cache, &cache->clean, page, true, prio);
pgc_queue_unlock(cache, &cache->clean);
page_transition_unlock(cache, page);
remove_and_free_page_not_in_any_queue_and_acquired_for_deletion(cache, page);
return true;
}
// returns true, when there is potentially more work to do
static bool evict_pages_with_filter(PGC *cache, size_t max_skip, size_t max_evict, bool wait, bool all_of_them, evict_filter filter, void *data) {
ssize_t per1000 = cache_usage_per1000(cache, NULL);
if(!all_of_them && per1000 < cache->config.healthy_size_per1000)
// don't bother - not enough to do anything
return false;
bool under_sever_pressure = per1000 >= cache->config.severe_pressure_per1000;
size_t workers_running = __atomic_add_fetch(&cache->stats.p0_workers_evict, 1, __ATOMIC_RELAXED);
if(!wait && !all_of_them && workers_running > cache->config.max_workers_evict_inline && !under_sever_pressure) {
__atomic_sub_fetch(&cache->stats.p0_workers_evict, 1, __ATOMIC_RELAXED);
return false;
}
internal_fatal(cache->clean.linked_list_in_sections_judy,
"wrong clean pages configuration - clean pages need to have a linked list, not a judy array");
if(unlikely(!max_skip))
max_skip = SIZE_MAX;
else if(unlikely(max_skip < 2))
max_skip = 2;
if(unlikely(!max_evict))
max_evict = SIZE_MAX;
else if(unlikely(max_evict < 2))
max_evict = 2;
size_t this_loop_evicted = 0;
size_t total_pages_evicted = 0;
size_t total_pages_relocated = 0;
bool stopped_before_finishing = false;
size_t spins = 0;
size_t max_pages_to_evict = 0;
do {
int64_t max_size_to_evict = 0;
if (unlikely(all_of_them)) {
// evict them all
max_size_to_evict = SIZE_MAX;
max_pages_to_evict = SIZE_MAX;
under_sever_pressure = true;
}
else if(unlikely(wait)) {
// evict as many as necessary for the cache to go at the predefined threshold
per1000 = cache_usage_per1000(cache, &max_size_to_evict);
if(per1000 >= cache->config.severe_pressure_per1000) {
under_sever_pressure = true;
max_pages_to_evict = max_pages_to_evict ? max_pages_to_evict * 2 : 16;
if(max_pages_to_evict > 64)
max_pages_to_evict = 64;
}
else if(per1000 >= cache->config.aggressive_evict_per1000) {
under_sever_pressure = false;
max_pages_to_evict = max_pages_to_evict ? max_pages_to_evict * 2 : 4;
if(max_pages_to_evict > 16)
max_pages_to_evict = 16;
}
else {
under_sever_pressure = false;
max_pages_to_evict = 1;
}
}
else {
// this is an adder, so evict just 1 page
max_size_to_evict = (cache_above_healthy_limit(cache)) ? 1 : 0;
max_pages_to_evict = 1;
}
if (!max_size_to_evict || !max_pages_to_evict)
break;
// check if we have to stop
if(total_pages_evicted >= max_evict && !all_of_them) {
stopped_before_finishing = true;
break;
}
if(++spins > 1 && !this_loop_evicted)
p2_add_fetch(&cache->stats.p2_waste_evict_useless_spins, 1);
this_loop_evicted = 0;
timing_dbengine_evict_init();
if(!all_of_them && !wait) {
if(!pgc_queue_trylock(cache, &cache->clean, PGC_QUEUE_LOCK_PRIO_EVICTORS)) {
stopped_before_finishing = true;
goto premature_exit;
}
// at this point we have the clean lock
}
else
pgc_queue_lock(cache, &cache->clean, PGC_QUEUE_LOCK_PRIO_EVICTORS);
timing_dbengine_evict_step(TIMING_STEP_DBENGINE_EVICT_LOCK);
// find a page to evict
PGC_PAGE *pages_to_evict = NULL;
int64_t pages_to_evict_size = 0;
size_t pages_to_evict_count = 0;
for(PGC_PAGE *page = cache->clean.base, *next = NULL, *first_page_we_relocated = NULL; page ; page = next) {
next = page->link.next;
if(unlikely(page == first_page_we_relocated))
// we did a complete loop on all pages
break;
if(unlikely(page_flag_check(page, PGC_PAGE_HAS_BEEN_ACCESSED | PGC_PAGE_HAS_NO_DATA_IGNORE_ACCESSES) == PGC_PAGE_HAS_BEEN_ACCESSED)) {
DOUBLE_LINKED_LIST_REMOVE_ITEM_UNSAFE(cache->clean.base, page, link.prev, link.next);
DOUBLE_LINKED_LIST_APPEND_ITEM_UNSAFE(cache->clean.base, page, link.prev, link.next);
page_flag_clear(page, PGC_PAGE_HAS_BEEN_ACCESSED);
continue;
}
if(unlikely(filter && !filter(page, data)))
continue;
if(non_acquired_page_get_for_deletion___while_having_clean_locked(cache, page)) {
// we can delete this page
// remove it from the clean list
pgc_queue_del(cache, &cache->clean, page, true, PGC_QUEUE_LOCK_PRIO_EVICTORS);
__atomic_add_fetch(&cache->stats.evicting_entries, 1, __ATOMIC_RELAXED);
__atomic_add_fetch(&cache->stats.evicting_size, page->assumed_size, __ATOMIC_RELAXED);
DOUBLE_LINKED_LIST_APPEND_ITEM_UNSAFE(pages_to_evict, page, link.prev, link.next);
pages_to_evict_size += page->assumed_size;
pages_to_evict_count++;
timing_dbengine_evict_step(TIMING_STEP_DBENGINE_EVICT_SELECT_PAGE);
if((pages_to_evict_count < max_pages_to_evict && pages_to_evict_size < max_size_to_evict) || all_of_them)
// get more pages
;
else
// one page at a time
break;
}
else {
// we can't delete this page
if(!first_page_we_relocated)
first_page_we_relocated = page;
DOUBLE_LINKED_LIST_REMOVE_ITEM_UNSAFE(cache->clean.base, page, link.prev, link.next);
DOUBLE_LINKED_LIST_APPEND_ITEM_UNSAFE(cache->clean.base, page, link.prev, link.next);
total_pages_relocated++;
timing_dbengine_evict_step(TIMING_STEP_DBENGINE_EVICT_RELOCATE_PAGE);
// check if we have to stop
if(unlikely(total_pages_relocated >= max_skip && !all_of_them)) {
stopped_before_finishing = true;
break;
}
}
}
pgc_queue_unlock(cache, &cache->clean);
timing_dbengine_evict_step(TIMING_STEP_DBENGINE_EVICT_SELECT);
if(likely(pages_to_evict)) {
// remove them from the index
if(unlikely(pages_to_evict->link.next)) {
// we have many pages, let's minimize the index locks we are going to get
PGC_PAGE *pages_per_partition[cache->config.partitions];
memset(pages_per_partition, 0, sizeof(PGC_PAGE *) * cache->config.partitions);
bool partitions_done[cache->config.partitions];
memset(partitions_done, 0, sizeof(bool) * cache->config.partitions);
// sort them by partition
for (PGC_PAGE *page = pages_to_evict, *next = NULL; page; page = next) {
next = page->link.next;
size_t partition = pgc_indexing_partition(cache, page->metric_id);
DOUBLE_LINKED_LIST_REMOVE_ITEM_UNSAFE(pages_to_evict, page, link.prev, link.next);
DOUBLE_LINKED_LIST_APPEND_ITEM_UNSAFE(pages_per_partition[partition], page, link.prev, link.next);
}
timing_dbengine_evict_step(TIMING_STEP_DBENGINE_EVICT_SORT);
// remove them from the index
size_t remaining_partitions = cache->config.partitions;
size_t last_remaining_partitions = remaining_partitions + 1;
while(remaining_partitions) {
bool force = remaining_partitions == last_remaining_partitions;
last_remaining_partitions = remaining_partitions;
remaining_partitions = 0;
for (size_t partition = 0; partition < cache->config.partitions; partition++) {
if (!pages_per_partition[partition] || partitions_done[partition])
continue;
if(pgc_index_trywrite_lock(cache, partition, force)) {
partitions_done[partition] = true;
for (PGC_PAGE *page = pages_per_partition[partition]; page; page = page->link.next)
remove_this_page_from_index_unsafe(cache, page, partition);
pgc_index_write_unlock(cache, partition);
timing_dbengine_evict_step(TIMING_STEP_DBENGINE_EVICT_DEINDEX_PAGE);
}
else
remaining_partitions++;
}
}
timing_dbengine_evict_step(TIMING_STEP_DBENGINE_EVICT_DEINDEX);
// free them
for (size_t partition = 0; partition < cache->config.partitions; partition++) {
if (!pages_per_partition[partition]) continue;
for (PGC_PAGE *page = pages_per_partition[partition], *next = NULL; page; page = next) {
next = page->link.next;
timing_dbengine_evict_step(TIMING_STEP_DBENGINE_EVICT_FREE_LOOP);
int64_t page_size = page->assumed_size;
free_this_page(cache, page, partition);
timing_dbengine_evict_step(TIMING_STEP_DBENGINE_EVICT_FREE_PAGE);
__atomic_sub_fetch(&cache->stats.evicting_entries, 1, __ATOMIC_RELAXED);
__atomic_sub_fetch(&cache->stats.evicting_size, page_size, __ATOMIC_RELAXED);
total_pages_evicted++;
this_loop_evicted++;
timing_dbengine_evict_step(TIMING_STEP_DBENGINE_EVICT_FREE_ATOMICS);
}
}
timing_dbengine_evict_report();
}
else {
// just one page to be evicted
PGC_PAGE *page = pages_to_evict;
int64_t page_size = page->assumed_size;
size_t partition = pgc_indexing_partition(cache, page->metric_id);
pgc_index_write_lock(cache, partition);
remove_this_page_from_index_unsafe(cache, page, partition);
pgc_index_write_unlock(cache, partition);
free_this_page(cache, page, partition);
__atomic_sub_fetch(&cache->stats.evicting_entries, 1, __ATOMIC_RELAXED);
__atomic_sub_fetch(&cache->stats.evicting_size, page_size, __ATOMIC_RELAXED);
total_pages_evicted++;
this_loop_evicted++;
}
}
else
break;
} while(all_of_them || (total_pages_evicted < max_evict && total_pages_relocated < max_skip));
if(all_of_them && !filter) {
pgc_queue_lock(cache, &cache->clean, PGC_QUEUE_LOCK_PRIO_EVICTORS);
size_t entries = __atomic_load_n(&cache->clean.stats->entries, __ATOMIC_RELAXED);
if(entries) {
nd_log_limit_static_global_var(erl, 1, 0);
nd_log_limit(&erl, NDLS_DAEMON, NDLP_NOTICE,
"DBENGINE CACHE: cannot free all clean pages, %zu are still in the clean queue",
entries);
}
pgc_queue_unlock(cache, &cache->clean);
}
premature_exit:
if(unlikely(total_pages_relocated))
p2_add_fetch(&cache->stats.p2_waste_evict_relocated, total_pages_relocated);
__atomic_sub_fetch(&cache->stats.p0_workers_evict, 1, __ATOMIC_RELAXED);
return stopped_before_finishing;
}
static PGC_PAGE *pgc_page_add(PGC *cache, PGC_ENTRY *entry, bool *added) {
internal_fatal(entry->start_time_s < 0 || entry->end_time_s < 0,
"DBENGINE CACHE: timestamps are negative");
p2_add_fetch(&cache->stats.p2_workers_add, 1);
size_t partition = pgc_indexing_partition(cache, entry->metric_id);
#ifdef PGC_WITH_ARAL
PGC_PAGE *allocation = aral_mallocz(cache->index[partition].aral);
#else
PGC_PAGE *allocation = mallocz(sizeof(PGC_PAGE) + cache->config.additional_bytes_per_page);
#endif
allocation->refcount = 1;
allocation->accesses = (entry->hot) ? 0 : 1;
allocation->flags = 0;
allocation->section = entry->section;
allocation->metric_id = entry->metric_id;
allocation->start_time_s = entry->start_time_s;
allocation->end_time_s = entry->end_time_s,
allocation->update_every_s = entry->update_every_s,
allocation->data = entry->data;
allocation->assumed_size = page_assumed_size(cache, entry->size);
spinlock_init(&allocation->transition_spinlock);
allocation->link.prev = NULL;
allocation->link.next = NULL;
if(cache->config.additional_bytes_per_page) {
if(entry->custom_data)
memcpy(allocation->custom_data, entry->custom_data, cache->config.additional_bytes_per_page);
else
memset(allocation->custom_data, 0, cache->config.additional_bytes_per_page);
}
PGC_PAGE *page;
size_t spins = 0;
if(unlikely(entry->start_time_s < 0))
entry->start_time_s = 0;
if(unlikely(entry->end_time_s < 0))
entry->end_time_s = 0;
do {
spins++;
pgc_index_write_lock(cache, partition);
JudyAllocThreadPulseReset();
Pvoid_t *metrics_judy_pptr = JudyLIns(&cache->index[partition].sections_judy, entry->section, PJE0);
if(unlikely(!metrics_judy_pptr || metrics_judy_pptr == PJERR))
fatal("DBENGINE CACHE: corrupted sections judy array");
Pvoid_t *pages_judy_pptr = JudyLIns(metrics_judy_pptr, entry->metric_id, PJE0);
if(unlikely(!pages_judy_pptr || pages_judy_pptr == PJERR))
fatal("DBENGINE CACHE: corrupted pages judy array");
Pvoid_t *page_ptr = JudyLIns(pages_judy_pptr, entry->start_time_s, PJE0);
if(unlikely(!page_ptr || page_ptr == PJERR))
fatal("DBENGINE CACHE: corrupted page in judy array");
pgc_stats_index_judy_change(cache, JudyAllocThreadPulseGetAndReset());
page = *page_ptr;
if (likely(!page)) {
// consume it
page = allocation;
allocation = NULL;
// put it in the index
*page_ptr = page;
pointer_add(cache, page);
pgc_index_write_unlock(cache, partition);
if (entry->hot)
page_set_hot(cache, page, PGC_QUEUE_LOCK_PRIO_COLLECTORS);
else
page_set_clean(cache, page, false, false, PGC_QUEUE_LOCK_PRIO_EVICTORS);
PGC_REFERENCED_PAGES_PLUS1(cache, page);
// update statistics
__atomic_add_fetch(&cache->stats.added_entries, 1, __ATOMIC_RELAXED);
__atomic_add_fetch(&cache->stats.added_size, page->assumed_size, __ATOMIC_RELAXED);
__atomic_add_fetch(&cache->stats.entries, 1, __ATOMIC_RELAXED);
__atomic_add_fetch(&cache->stats.size, page->assumed_size, __ATOMIC_RELAXED);
if(added)
*added = true;
}
else {
if (!page_acquire(cache, page))
page = NULL;
else if(added)
*added = false;
pgc_index_write_unlock(cache, partition);
if(unlikely(!page)) {
// now that we don't have the lock,
// give it some time for the old page to go away
tinysleep();
}
}
} while(!page);
if(allocation) {
#ifdef PGC_WITH_ARAL
aral_freez(cache->index[partition].aral, allocation);
#else
freez(allocation);
#endif
}
if(spins > 1)
p2_add_fetch(&cache->stats.p2_waste_insert_spins, spins - 1);
p2_sub_fetch(&cache->stats.p2_workers_add, 1);
if(!entry->hot)
evict_on_clean_page_added(cache);
flush_on_page_add(cache);
return page;
}
static ALWAYS_INLINE PGC_PAGE *page_find_and_acquire_exact_unsafe(PGC *cache, Pvoid_t *pages_judy_pptr, time_t start_time_s) {
Pvoid_t *page_ptr = JudyLGet(*pages_judy_pptr, start_time_s, PJE0);
if(!page_ptr)
return NULL;
if (unlikely(page_ptr == PJERR))
fatal("DBENGINE CACHE: corrupted page in pages judy array");
PGC_PAGE *page = *page_ptr;
if(page && page_acquire(cache, page))
// we have our page acquired
return page;
return NULL;
}
static ALWAYS_INLINE PGC_PAGE *page_find_and_acquire_first_unsafe(PGC *cache, Pvoid_t *pages_judy_pptr, time_t start_time_s) {
Word_t time = start_time_s;
for(Pvoid_t *page_ptr = JudyLFirst(*pages_judy_pptr, &time, PJE0);
page_ptr ;
page_ptr = JudyLNext(*pages_judy_pptr, &time, PJE0)) {
if (unlikely(page_ptr == PJERR))
fatal("DBENGINE CACHE: corrupted page in pages judy array");
PGC_PAGE *page = *page_ptr;
if(page && page_acquire(cache, page))
// we have our page acquired
return page;
}
return NULL;
}
static ALWAYS_INLINE PGC_PAGE *page_find_and_acquire_next_unsafe(PGC *cache, Pvoid_t *pages_judy_pptr, time_t start_time_s) {
Word_t time = start_time_s;
for(Pvoid_t *page_ptr = JudyLNext(*pages_judy_pptr, &time, PJE0);
page_ptr ;
page_ptr = JudyLNext(*pages_judy_pptr, &time, PJE0)) {
if (unlikely(page_ptr == PJERR))
fatal("DBENGINE CACHE: corrupted page in pages judy array");
PGC_PAGE *page = *page_ptr;
if(page && page_acquire(cache, page))
// we have our page acquired
return page;
}
return NULL;
}
static ALWAYS_INLINE PGC_PAGE *page_find_and_acquire_last_unsafe(PGC *cache, Pvoid_t *pages_judy_pptr, time_t start_time_s) {
Word_t time = start_time_s;
for(Pvoid_t *page_ptr = JudyLLast(*pages_judy_pptr, &time, PJE0);
page_ptr ;
page_ptr = JudyLPrev(*pages_judy_pptr, &time, PJE0)) {
if (unlikely(page_ptr == PJERR))
fatal("DBENGINE CACHE: corrupted page in pages judy array");
PGC_PAGE *page = *page_ptr;
if(page && page_acquire(cache, page))
// we have our page acquired
return page;
}
return NULL;
}
static ALWAYS_INLINE PGC_PAGE *page_find_and_acquire_prev_unsafe(PGC *cache, Pvoid_t *pages_judy_pptr, time_t start_time_s) {
Word_t time = start_time_s;
for(Pvoid_t *page_ptr = JudyLPrev(*pages_judy_pptr, &time, PJE0);
page_ptr ;
page_ptr = JudyLPrev(*pages_judy_pptr, &time, PJE0)) {
if (unlikely(page_ptr == PJERR))
fatal("DBENGINE CACHE: corrupted page in pages judy array");
PGC_PAGE *page = *page_ptr;
if(page && page_acquire(cache, page))
// we have our page acquired
return page;
}
return NULL;
}
static ALWAYS_INLINE PGC_PAGE *page_find_and_acquire_once(PGC *cache, Word_t section, Word_t metric_id, time_t start_time_s, PGC_SEARCH method) {
PGC_PAGE *page = NULL;
size_t partition = pgc_indexing_partition(cache, metric_id);
pgc_index_read_lock(cache, partition);
Pvoid_t *metrics_judy_pptr = JudyLGet(cache->index[partition].sections_judy, section, PJE0);
if(unlikely(metrics_judy_pptr == PJERR))
fatal("DBENGINE CACHE: corrupted sections judy array");
if(unlikely(!metrics_judy_pptr)) {
// section does not exist
goto cleanup;
}
Pvoid_t *pages_judy_pptr = JudyLGet(*metrics_judy_pptr, metric_id, PJE0);
if(unlikely(pages_judy_pptr == PJERR))
fatal("DBENGINE CACHE: corrupted pages judy array");
if(unlikely(!pages_judy_pptr)) {
// metric does not exist
goto cleanup;
}
switch(method) {
default:
case PGC_SEARCH_CLOSEST: {
page = page_find_and_acquire_exact_unsafe(cache, pages_judy_pptr, start_time_s);
if(!page) {
page = page_find_and_acquire_prev_unsafe(cache, pages_judy_pptr, start_time_s);
if(page && start_time_s > page->end_time_s) {
// found a page starting before our timestamp
// but our timestamp is not included in it
page_release(cache, page, false);
page = NULL;
}
if(!page)
page = page_find_and_acquire_next_unsafe(cache, pages_judy_pptr, start_time_s);
}
}
break;
case PGC_SEARCH_EXACT:
page = page_find_and_acquire_exact_unsafe(cache, pages_judy_pptr, start_time_s);
break;
case PGC_SEARCH_FIRST:
page = page_find_and_acquire_first_unsafe(cache, pages_judy_pptr, start_time_s);
break;
case PGC_SEARCH_NEXT:
page = page_find_and_acquire_next_unsafe(cache, pages_judy_pptr, start_time_s);
break;
case PGC_SEARCH_LAST:
page = page_find_and_acquire_last_unsafe(cache, pages_judy_pptr, start_time_s);
break;
case PGC_SEARCH_PREV:
page = page_find_and_acquire_prev_unsafe(cache, pages_judy_pptr, start_time_s);
break;
}
#ifdef NETDATA_PGC_POINTER_CHECK
if(page)
pointer_check(cache, page);
#endif
cleanup:
pgc_index_read_unlock(cache, partition);
return page;
}
static void all_hot_pages_to_dirty(PGC *cache, Word_t section) {
pgc_queue_lock(cache, &cache->hot, PGC_QUEUE_LOCK_PRIO_COLLECTORS);
bool first = true;
Word_t last_section = (section == PGC_SECTION_ALL) ? 0 : section;
Pvoid_t *section_pages_pptr;
while ((section_pages_pptr = JudyLFirstThenNext(cache->hot.sections_judy, &last_section, &first))) {
if(section != PGC_SECTION_ALL && last_section != section)
break;
struct section_pages *sp = *section_pages_pptr;
PGC_PAGE *page = sp->base;
while(page) {
PGC_PAGE *next = page->link.next;
if(page_acquire(cache, page)) {
page_set_dirty(cache, page, true, PGC_QUEUE_LOCK_PRIO_COLLECTORS);
page_release(cache, page, false);
// page ptr may be invalid now
}
page = next;
}
}
pgc_queue_unlock(cache, &cache->hot);
}
// returns true when there is more work to do
static bool flush_pages(PGC *cache, size_t max_flushes, Word_t section, bool wait, bool all_of_them) {
internal_fatal(!cache->dirty.linked_list_in_sections_judy,
"wrong dirty pages configuration - dirty pages need to have a judy array, not a linked list");
if(!all_of_them && !wait) {
// we have been called from a data collection thread
// let's not waste its time...
if(!pgc_queue_trylock(cache, &cache->dirty, PGC_QUEUE_LOCK_PRIO_FLUSHERS)) {
// we would block, so give up...
return false;
}
// we got the lock at this point
}
else
pgc_queue_lock(cache, &cache->dirty, PGC_QUEUE_LOCK_PRIO_FLUSHERS);
size_t optimal_flush_size = cache->config.max_dirty_pages_per_call;
size_t dirty_version_at_entry = cache->dirty.version;
size_t entries = __atomic_load_n(&cache->dirty.stats->entries, __ATOMIC_RELAXED);
if(!all_of_them && (entries < optimal_flush_size || cache->dirty.last_version_checked == dirty_version_at_entry)) {
pgc_queue_unlock(cache, &cache->dirty);
return false;
}
p2_add_fetch(&cache->stats.p2_workers_flush, 1);
bool have_dirty_lock = true;
if(all_of_them || !max_flushes)
max_flushes = SIZE_MAX;
Word_t last_section = (section == PGC_SECTION_ALL) ? 0 : section;
size_t flushes_so_far = 0;
Pvoid_t *section_pages_pptr;
bool stopped_before_finishing = false;
bool first = true;
while (have_dirty_lock && (section_pages_pptr = JudyLFirstThenNext(cache->dirty.sections_judy, &last_section, &first))) {
if(section != PGC_SECTION_ALL && last_section != section)
break;
struct section_pages *sp = *section_pages_pptr;
if(!all_of_them && sp->entries < optimal_flush_size)
continue;
if(!all_of_them && flushes_so_far > max_flushes) {
stopped_before_finishing = true;
break;
}
PGC_ENTRY array[optimal_flush_size];
PGC_PAGE *pages[optimal_flush_size];
size_t pages_added = 0,
pages_removed_dirty = 0,
pages_cancelled = 0,
pages_made_clean = 0;
int64_t pages_added_size = 0,
pages_removed_dirty_size = 0,
pages_cancelled_size = 0,
pages_made_clean_size = 0;
PGC_PAGE *page = sp->base;
while (page && pages_added < optimal_flush_size) {
PGC_PAGE *next = page->link.next;
internal_fatal(page_get_status_flags(page) != PGC_PAGE_DIRTY,
"DBENGINE CACHE: page should be in the dirty list before saved");
if (page_acquire(cache, page)) {
internal_fatal(page_get_status_flags(page) != PGC_PAGE_DIRTY,
"DBENGINE CACHE: page should be in the dirty list before saved");
internal_fatal(page->section != last_section,
"DBENGINE CACHE: dirty page is not in the right section (tier)");
if(!page_transition_trylock(cache, page)) {
page_release(cache, page, false);
// page ptr may be invalid now
}
else {
pages[pages_added] = page;
array[pages_added] = (PGC_ENTRY) {
.section = page->section,
.metric_id = page->metric_id,
.start_time_s = page->start_time_s,
.end_time_s = __atomic_load_n(&page->end_time_s, __ATOMIC_RELAXED),
.update_every_s = page->update_every_s,
.size = page_size_from_assumed_size(cache, page->assumed_size),
.data = page->data,
.custom_data = (cache->config.additional_bytes_per_page) ? page->custom_data : NULL,
.hot = false,
};
pages_added_size += page->assumed_size;
pages_added++;
}
}
page = next;
}
// do we have enough to save?
if(all_of_them || pages_added == optimal_flush_size) {
// we should do it
for (size_t i = 0; i < pages_added; i++) {
PGC_PAGE *tpg = pages[i];
internal_fatal(page_get_status_flags(tpg) != PGC_PAGE_DIRTY,
"DBENGINE CACHE: page should be in the dirty list before saved");
__atomic_add_fetch(&cache->stats.flushing_entries, 1, __ATOMIC_RELAXED);
__atomic_add_fetch(&cache->stats.flushing_size, tpg->assumed_size, __ATOMIC_RELAXED);
// remove it from the dirty list
pgc_queue_del(cache, &cache->dirty, tpg, true, PGC_QUEUE_LOCK_PRIO_FLUSHERS);
pages_removed_dirty_size += tpg->assumed_size;
pages_removed_dirty++;
}
// next time, repeat the same section (tier)
first = true;
}
else {
// we can't do it
for (size_t i = 0; i < pages_added; i++) {
PGC_PAGE *tpg = pages[i];
internal_fatal(page_get_status_flags(tpg) != PGC_PAGE_DIRTY,
"DBENGINE CACHE: page should be in the dirty list before saved");
pages_cancelled_size += tpg->assumed_size;
pages_cancelled++;
page_transition_unlock(cache, tpg);
page_release(cache, tpg, false);
// page ptr may be invalid now
}
p2_add_fetch(&cache->stats.p2_waste_flushes_cancelled, pages_cancelled);
p2_add_fetch(&cache->stats.flushes_cancelled_size, pages_cancelled_size);
internal_fatal(pages_added != pages_cancelled || pages_added_size != pages_cancelled_size,
"DBENGINE CACHE: flushing cancel pages mismatch");
// next time, continue to the next section (tier)
first = false;
continue;
}
if(cache->config.pgc_save_init_cb)
cache->config.pgc_save_init_cb(cache, last_section);
pgc_queue_unlock(cache, &cache->dirty);
have_dirty_lock = false;
// call the callback to save them
// it may take some time, so let's release the lock
if(cache->config.pgc_save_dirty_cb)
cache->config.pgc_save_dirty_cb(cache, array, pages, pages_added);
flushes_so_far++;
__atomic_add_fetch(&cache->stats.flushes_completed, pages_added, __ATOMIC_RELAXED);
__atomic_add_fetch(&cache->stats.flushes_completed_size, pages_added_size, __ATOMIC_RELAXED);
size_t pages_to_evict = 0; (void)pages_to_evict;
for (size_t i = 0; i < pages_added; i++) {
PGC_PAGE *tpg = pages[i];
internal_fatal(page_get_status_flags(tpg) != 0,
"DBENGINE CACHE: page should not be in any list while it is being saved");
__atomic_sub_fetch(&cache->stats.flushing_entries, 1, __ATOMIC_RELAXED);
__atomic_sub_fetch(&cache->stats.flushing_size, tpg->assumed_size, __ATOMIC_RELAXED);
pages_made_clean_size += tpg->assumed_size;
pages_made_clean++;
if(!tpg->accesses)
pages_to_evict++;
page_set_clean(cache, tpg, true, false, PGC_QUEUE_LOCK_PRIO_FLUSHERS);
page_transition_unlock(cache, tpg);
page_release(cache, tpg, false);
// tpg ptr may be invalid now
}
internal_fatal(pages_added != pages_made_clean || pages_added != pages_removed_dirty ||
pages_added_size != pages_made_clean_size || pages_added_size != pages_removed_dirty_size
, "DBENGINE CACHE: flushing pages mismatch");
if(!all_of_them && !wait) {
if(pgc_queue_trylock(cache, &cache->dirty, PGC_QUEUE_LOCK_PRIO_FLUSHERS))
have_dirty_lock = true;
else {
stopped_before_finishing = true;
have_dirty_lock = false;
}
}
else {
pgc_queue_lock(cache, &cache->dirty, PGC_QUEUE_LOCK_PRIO_FLUSHERS);
have_dirty_lock = true;
}
}
if(have_dirty_lock) {
if(!stopped_before_finishing && dirty_version_at_entry > cache->dirty.last_version_checked)
cache->dirty.last_version_checked = dirty_version_at_entry;
pgc_queue_unlock(cache, &cache->dirty);
}
p2_sub_fetch(&cache->stats.p2_workers_flush, 1);
return stopped_before_finishing;
}
void free_all_unreferenced_clean_pages(PGC *cache) {
evict_pages(cache, 0, 0, true, true);
}
static void *pgc_evict_thread(void *ptr) {
static usec_t last_malloc_release_ut = 0;
PGC *cache = ptr;
worker_register("PGCEVICT");
worker_register_job_name(0, "signaled");
worker_register_job_name(1, "scheduled");
worker_register_job_name(2, "cleanup");
unsigned job_id = 0;
while (true) {
worker_is_idle();
unsigned new_job_id = completion_wait_for_a_job_with_timeout(
&cache->evictor.completion, job_id, 1000);
worker_is_busy(new_job_id > job_id ? 1 : 0);
job_id = new_job_id;
if (nd_thread_signaled_to_cancel())
break;
int64_t size_to_evict = 0;
bool system_cleanup = false;
if(cache_usage_per1000(cache, &size_to_evict) > cache->config.aggressive_evict_per1000)
system_cleanup = true;
evict_pages(cache, 0, 0, true, false);
if(system_cleanup) {
usec_t now_ut = now_monotonic_usec();
if(__atomic_load_n(&last_malloc_release_ut, __ATOMIC_RELAXED) + USEC_PER_SEC <= now_ut) {
__atomic_store_n(&last_malloc_release_ut, now_ut, __ATOMIC_RELAXED);
worker_is_busy(2);
mallocz_release_as_much_memory_to_the_system();
}
}
}
worker_unregister();
return NULL;
}
// ----------------------------------------------------------------------------
// public API
PGC *pgc_create(const char *name,
size_t clean_size_bytes,
free_clean_page_callback pgc_free_cb,
size_t max_dirty_pages_per_flush,
save_dirty_init_callback pgc_save_init_cb,
save_dirty_page_callback pgc_save_dirty_cb,
size_t max_pages_per_inline_eviction,
size_t max_inline_evictors,
size_t max_skip_pages_per_inline_eviction,
size_t max_flushes_inline,
PGC_OPTIONS options,
size_t partitions,
size_t additional_bytes_per_page) {
if(max_pages_per_inline_eviction < 1)
max_pages_per_inline_eviction = 1;
if(max_dirty_pages_per_flush < 1)
max_dirty_pages_per_flush = 1;
if(max_flushes_inline * max_dirty_pages_per_flush < 2)
max_flushes_inline = 2;
PGC *cache = callocz(1, sizeof(PGC));
strncpyz(cache->config.name, name, PGC_NAME_MAX);
cache->config.options = options;
cache->config.additional_bytes_per_page = additional_bytes_per_page;
cache->config.stats = pulse_enabled;
// flushing
cache->config.max_flushes_inline = (max_flushes_inline == 0) ? 2 : max_flushes_inline;
cache->config.max_dirty_pages_per_call = max_dirty_pages_per_flush;
cache->config.pgc_save_init_cb = pgc_save_init_cb;
cache->config.pgc_save_dirty_cb = pgc_save_dirty_cb;
// eviction strategy
cache->config.clean_size = (clean_size_bytes < 1 * 1024 * 1024) ? 1 * 1024 * 1024 : (int64_t)clean_size_bytes;
cache->config.pgc_free_clean_cb = pgc_free_cb;
cache->config.max_workers_evict_inline = max_inline_evictors;
cache->config.max_pages_per_inline_eviction = max_pages_per_inline_eviction;
cache->config.max_skip_pages_per_inline_eviction = (max_skip_pages_per_inline_eviction < 2) ? 2 : max_skip_pages_per_inline_eviction;
cache->config.severe_pressure_per1000 = 1010; // INLINE: use releasers to evict pages (up to max_pages_per_inline_eviction)
cache->config.aggressive_evict_per1000 = 990; // INLINE: use adders to evict pages (up to max_pages_per_inline_eviction)
cache->config.healthy_size_per1000 = 980; // no evictions happen below this threshold
cache->config.evict_low_threshold_per1000 = 970; // when evicting, bring the size down to this threshold
// the eviction thread is signaled ONLY if we run out of memory
// otherwise, it runs by itself every 100ms
// use all ram and protection from out of memory
cache->config.use_all_ram = dbengine_use_all_ram_for_caches;
cache->config.out_of_memory_protection_bytes = (int64_t)dbengine_out_of_memory_protection;
// partitions
if(partitions == 0) partitions = netdata_conf_cpus() * 2;
if(partitions <= 4) partitions = 4;
if(partitions > 256) partitions = 256;
cache->config.partitions = partitions;
cache->index = callocz(cache->config.partitions, sizeof(struct pgc_index));
pgc_section_pages_static_aral_init();
for(size_t part = 0; part < cache->config.partitions ; part++) {
rw_spinlock_init(&cache->index[part].rw_spinlock);
#ifdef PGC_WITH_ARAL
{
char buf[100];
snprintfz(buf, sizeof(buf), "%s", name);
cache->index[part].aral = aral_create(
buf,
sizeof(PGC_PAGE) + cache->config.additional_bytes_per_page,
0,
0,
&pgc_aral_statistics,
NULL, NULL,
false, false, false);
}
#endif
}
#if defined(PGC_QUEUE_LOCK_AS_WAITING_QUEUE)
waitq_init(&cache->hot.wq);
waitq_init(&cache->dirty.wq);
waitq_init(&cache->clean.wq);
#else
spinlock_init(&cache->hot.spinlock);
spinlock_init(&cache->dirty.spinlock);
spinlock_init(&cache->clean.spinlock);
#endif
cache->hot.flags = PGC_PAGE_HOT;
cache->hot.linked_list_in_sections_judy = true;
cache->hot.stats = &cache->stats.queues[PGC_QUEUE_HOT];
cache->dirty.flags = PGC_PAGE_DIRTY;
cache->dirty.linked_list_in_sections_judy = true;
cache->dirty.stats = &cache->stats.queues[PGC_QUEUE_DIRTY];
cache->clean.flags = PGC_PAGE_CLEAN;
cache->clean.linked_list_in_sections_judy = false;
cache->clean.stats = &cache->stats.queues[PGC_QUEUE_CLEAN];
pointer_index_init(cache);
pgc_size_histogram_init(&cache->hot.stats->size_histogram);
pgc_size_histogram_init(&cache->dirty.stats->size_histogram);
pgc_size_histogram_init(&cache->clean.stats->size_histogram);
// last create the eviction thread
{
completion_init(&cache->evictor.completion);
cache->evictor.thread = nd_thread_create(name, NETDATA_THREAD_OPTION_JOINABLE, pgc_evict_thread, cache);
}
return cache;
}
struct aral_statistics *pgc_aral_stats(void) {
return &pgc_aral_statistics;
}
void pgc_flush_dirty_pages(PGC *cache, Word_t section) {
flush_pages(cache, 0, section, true, true);
}
void pgc_flush_all_hot_and_dirty_pages(PGC *cache, Word_t section) {
all_hot_pages_to_dirty(cache, section);
// save all dirty pages to make them clean
flush_pages(cache, 0, section, true, true);
}
void pgc_destroy(PGC *cache, bool flush) {
if(!cache)
return;
if(!flush) {
cache->config.pgc_save_init_cb = NULL;
cache->config.pgc_save_dirty_cb = NULL;
}
// convert all hot pages to dirty
all_hot_pages_to_dirty(cache, PGC_SECTION_ALL);
// save all dirty pages to make them clean
flush_pages(cache, 0, PGC_SECTION_ALL, true, true);
// free all unreferenced clean pages
free_all_unreferenced_clean_pages(cache);
// stop the eviction thread
nd_thread_signal_cancel(cache->evictor.thread);
completion_mark_complete_a_job(&cache->evictor.completion);
nd_thread_join(cache->evictor.thread);
completion_destroy(&cache->evictor.completion);
if(PGC_REFERENCED_PAGES(cache))
netdata_log_error("DBENGINE CACHE: there are %zu referenced cache pages - leaving the cache allocated", PGC_REFERENCED_PAGES(cache));
else {
pointer_destroy_index(cache);
for(size_t part = 0; part < cache->config.partitions ;part++) {
// netdata_rwlock_destroy(&cache->index[part].rw_spinlock);
#ifdef PGC_WITH_ARAL
aral_destroy(cache->index[part].aral);
#endif
}
#if defined(PGC_QUEUE_LOCK_AS_WAITING_QUEUE)
waitq_destroy(&cache->hot.wq);
waitq_destroy(&cache->dirty.wq);
waitq_destroy(&cache->clean.wq);
#endif
freez(cache->index);
freez(cache);
}
}
ALWAYS_INLINE PGC_PAGE *pgc_page_add_and_acquire(PGC *cache, PGC_ENTRY entry, bool *added) {
return pgc_page_add(cache, &entry, added);
}
ALWAYS_INLINE PGC_PAGE *pgc_page_dup(PGC *cache, PGC_PAGE *page) {
if(!page_acquire(cache, page))
fatal("DBENGINE CACHE: tried to dup a page that is not acquired!");
return page;
}
ALWAYS_INLINE void pgc_page_release(PGC *cache, PGC_PAGE *page) {
page_release(cache, page, is_page_clean(page));
}
ALWAYS_INLINE void pgc_page_hot_to_dirty_and_release(PGC *cache, PGC_PAGE *page, bool never_flush) {
p2_add_fetch(&cache->stats.p2_workers_hot2dirty, 1);
//#ifdef NETDATA_INTERNAL_CHECKS
// page_transition_lock(cache, page);
// internal_fatal(!is_page_hot(page), "DBENGINE CACHE: called %s() but page is not hot", __FUNCTION__ );
// page_transition_unlock(cache, page);
//#endif
// make page dirty
page_set_dirty(cache, page, false, PGC_QUEUE_LOCK_PRIO_COLLECTORS);
// release the page
page_release(cache, page, true);
// page ptr may be invalid now
p2_sub_fetch(&cache->stats.p2_workers_hot2dirty, 1);
// flush, if we have to
if(!never_flush)
flush_on_page_hot_release(cache);
}
bool pgc_page_to_clean_evict_or_release(PGC *cache, PGC_PAGE *page) {
bool ret;
p2_add_fetch(&cache->stats.p2_workers_hot2dirty, 1);
// prevent accesses from increasing the accesses counter
page_flag_set(page, PGC_PAGE_HAS_NO_DATA_IGNORE_ACCESSES);
// zero the accesses counter
__atomic_store_n(&page->accesses, 0, __ATOMIC_RELEASE);
// if there are no other references to it, evict it immediately
if(make_acquired_page_clean_and_evict_or_page_release(cache, page)) {
__atomic_add_fetch(&cache->stats.hot_empty_pages_evicted_immediately, 1, __ATOMIC_RELAXED);
ret = true;
}
else {
__atomic_add_fetch(&cache->stats.hot_empty_pages_evicted_later, 1, __ATOMIC_RELAXED);
ret = false;
}
p2_sub_fetch(&cache->stats.p2_workers_hot2dirty, 1);
return ret;
}
Word_t pgc_page_section(PGC_PAGE *page) {
return page->section;
}
Word_t pgc_page_metric(PGC_PAGE *page) {
return page->metric_id;
}
time_t pgc_page_start_time_s(PGC_PAGE *page) {
return page->start_time_s;
}
time_t pgc_page_end_time_s(PGC_PAGE *page) {
return page->end_time_s;
}
uint32_t pgc_page_update_every_s(PGC_PAGE *page) {
return page->update_every_s;
}
uint32_t pgc_page_fix_update_every(PGC_PAGE *page, uint32_t update_every_s) {
if(page->update_every_s == 0)
page->update_every_s = update_every_s;
return page->update_every_s;
}
time_t pgc_page_fix_end_time_s(PGC_PAGE *page, time_t end_time_s) {
page->end_time_s = end_time_s;
return page->end_time_s;
}
void *pgc_page_data(PGC_PAGE *page) {
return page->data;
}
void *pgc_page_custom_data(PGC *cache, PGC_PAGE *page) {
if(cache->config.additional_bytes_per_page)
return page->custom_data;
return NULL;
}
size_t pgc_page_data_size(PGC *cache, PGC_PAGE *page) {
return page_size_from_assumed_size(cache, page->assumed_size);
}
bool pgc_is_page_hot(PGC_PAGE *page) {
return is_page_hot(page);
}
bool pgc_is_page_dirty(PGC_PAGE *page) {
return is_page_dirty(page);
}
bool pgc_is_page_clean(PGC_PAGE *page) {
return is_page_clean(page);
}
void pgc_reset_hot_max(PGC *cache) {
size_t entries = __atomic_load_n(&cache->hot.stats->entries, __ATOMIC_RELAXED);
int64_t size = __atomic_load_n(&cache->hot.stats->size, __ATOMIC_RELAXED);
__atomic_store_n(&cache->hot.stats->max_entries, entries, __ATOMIC_RELAXED);
__atomic_store_n(&cache->hot.stats->max_size, size, __ATOMIC_RELAXED);
int64_t size_to_evict = 0;
cache_usage_per1000(cache, &size_to_evict);
evict_pages(cache, 0, 0, true, false);
}
void pgc_set_dynamic_target_cache_size_callback(PGC *cache, dynamic_target_cache_size_callback callback) {
cache->config.dynamic_target_size_cb = callback;
cache->config.out_of_memory_protection_bytes = 0;
cache->config.use_all_ram = false;
int64_t size_to_evict = 0;
cache_usage_per1000(cache, &size_to_evict);
evict_pages(cache, 0, 0, true, false);
}
void pgc_set_nominal_page_size_callback(PGC *cache, nominal_page_size_callback callback) {
cache->config.nominal_page_size_cb = callback;
}
int64_t pgc_get_current_cache_size(PGC *cache) {
return __atomic_load_n(&cache->stats.current_cache_size, __ATOMIC_RELAXED);
}
int64_t pgc_get_wanted_cache_size(PGC *cache) {
return __atomic_load_n(&cache->stats.wanted_cache_size, __ATOMIC_RELAXED);
}
bool pgc_evict_pages(PGC *cache, size_t max_skip, size_t max_evict) {
bool under_pressure = cache_needs_space_aggressively(cache);
return evict_pages(cache,
under_pressure ? 0 : max_skip,
under_pressure ? 0 : max_evict,
true, false);
}
bool pgc_flush_pages(PGC *cache) {
return flush_pages(cache, 0, PGC_SECTION_ALL, true, false);
}
void pgc_page_hot_set_end_time_s(PGC *cache __maybe_unused, PGC_PAGE *page, time_t end_time_s, size_t additional_bytes) {
internal_fatal(!is_page_hot(page) && !exit_initiated,
"DBENGINE CACHE: end_time_s update on non-hot page");
internal_fatal(end_time_s < __atomic_load_n(&page->end_time_s, __ATOMIC_RELAXED),
"DBENGINE CACHE: end_time_s is not bigger than existing");
__atomic_store_n(&page->end_time_s, end_time_s, __ATOMIC_RELAXED);
if(additional_bytes) {
page_transition_lock(cache, page);
struct pgc_queue_statistics *queue_stats = NULL;
if(page->flags & PGC_PAGE_HOT)
queue_stats = cache->hot.stats;
else if(page->flags & PGC_PAGE_DIRTY)
queue_stats = cache->dirty.stats;
else if(page->flags & PGC_PAGE_CLEAN)
queue_stats = cache->clean.stats;
if(queue_stats && cache->config.stats)
pgc_size_histogram_del(cache, &queue_stats->size_histogram, page);
int64_t old_assumed_size = page->assumed_size;
size_t old_size = page_size_from_assumed_size(cache, old_assumed_size);
size_t size = old_size + additional_bytes;
page->assumed_size = page_assumed_size(cache, size);
int64_t delta = page->assumed_size - old_assumed_size;
__atomic_add_fetch(&cache->stats.size, delta, __ATOMIC_RELAXED);
__atomic_add_fetch(&cache->stats.added_size, delta, __ATOMIC_RELAXED);
__atomic_add_fetch(&cache->stats.referenced_size, delta, __ATOMIC_RELAXED);
if(queue_stats) {
__atomic_add_fetch(&queue_stats->size, delta, __ATOMIC_RELAXED);
__atomic_add_fetch(&queue_stats->added_size, delta, __ATOMIC_RELAXED);
if(cache->config.stats)
pgc_size_histogram_add(cache, &queue_stats->size_histogram, page);
}
page_transition_unlock(cache, page);
}
#ifdef PGC_COUNT_POINTS_COLLECTED
__atomic_add_fetch(&cache->stats.points_collected, 1, __ATOMIC_RELAXED);
#endif
}
PGC_PAGE *pgc_page_get_and_acquire(PGC *cache, Word_t section, Word_t metric_id, time_t start_time_s, PGC_SEARCH method) {
PGC_PAGE *page = NULL;
p2_add_fetch(&cache->stats.p2_workers_search, 1);
size_t *stats_hit_ptr, *stats_miss_ptr;
if(method == PGC_SEARCH_CLOSEST) {
__atomic_add_fetch(&cache->stats.searches_closest, 1, __ATOMIC_RELAXED);
stats_hit_ptr = &cache->stats.searches_closest_hits;
stats_miss_ptr = &cache->stats.searches_closest_misses;
}
else {
__atomic_add_fetch(&cache->stats.searches_exact, 1, __ATOMIC_RELAXED);
stats_hit_ptr = &cache->stats.searches_exact_hits;
stats_miss_ptr = &cache->stats.searches_exact_misses;
}
page = page_find_and_acquire_once(cache, section, metric_id, start_time_s, method);
if(page) {
__atomic_add_fetch(stats_hit_ptr, 1, __ATOMIC_RELAXED);
page_has_been_accessed(cache, page);
}
else
__atomic_add_fetch(stats_miss_ptr, 1, __ATOMIC_RELAXED);
p2_sub_fetch(&cache->stats.p2_workers_search, 1);
return page;
}
struct pgc_statistics pgc_get_statistics(PGC *cache) {
// FIXME - get the statistics atomically
return cache->stats;
}
size_t pgc_hot_and_dirty_entries(PGC *cache) {
size_t entries = 0;
entries += __atomic_load_n(&cache->hot.stats->entries, __ATOMIC_RELAXED);
entries += __atomic_load_n(&cache->dirty.stats->entries, __ATOMIC_RELAXED);
entries += __atomic_load_n(&cache->stats.flushing_entries, __ATOMIC_RELAXED);
entries += __atomic_load_n(&cache->stats.hot2dirty_entries, __ATOMIC_RELAXED);
return entries;
}
void pgc_open_cache_to_journal_v2(PGC *cache, Word_t section, unsigned datafile_fileno, uint8_t type, migrate_to_v2_callback cb, void *data) {
__atomic_add_fetch(&rrdeng_cache_efficiency_stats.journal_v2_indexing_started, 1, __ATOMIC_RELAXED);
p2_add_fetch(&cache->stats.p2_workers_jv2_flush, 1);
pgc_queue_lock(cache, &cache->hot, PGC_QUEUE_LOCK_PRIO_LOW);
Pvoid_t JudyL_metrics = NULL;
Pvoid_t JudyL_extents_pos = NULL;
size_t count_of_unique_extents = 0;
size_t count_of_unique_metrics = 0;
size_t count_of_unique_pages = 0;
size_t master_extent_index_id = 0;
Pvoid_t *section_pages_pptr = JudyLGet(cache->hot.sections_judy, section, PJE0);
if(!section_pages_pptr) {
pgc_queue_unlock(cache, &cache->hot);
return;
}
struct section_pages *sp = *section_pages_pptr;
if(!spinlock_trylock(&sp->migration_to_v2_spinlock)) {
netdata_log_info("DBENGINE: migration to journal v2 for datafile %u is postponed, another jv2 indexer is already running for this section", datafile_fileno);
pgc_queue_unlock(cache, &cache->hot);
return;
}
ARAL *ar_mi = aral_by_size_acquire(sizeof(struct jv2_metrics_info));
ARAL *ar_pi = aral_by_size_acquire(sizeof(struct jv2_page_info));
ARAL *ar_ei = aral_by_size_acquire(sizeof(struct jv2_extents_info));
for(PGC_PAGE *page = sp->base; page ; page = page->link.next) {
struct extent_io_data *xio = (struct extent_io_data *)page->custom_data;
if(xio->fileno != datafile_fileno) continue;
if(page_flag_check(page, PGC_PAGE_IS_BEING_MIGRATED_TO_V2)) {
internal_fatal(true, "Migration to journal v2: page has already been migrated to v2");
continue;
}
if(!page_transition_trylock(cache, page)) {
internal_fatal(true, "Migration to journal v2: cannot get page transition lock");
continue;
}
if(!page_acquire(cache, page)) {
internal_fatal(true, "Migration to journal v2: cannot acquire page for migration to v2");
page_transition_unlock(cache, page);
continue;
}
page_flag_set(page, PGC_PAGE_IS_BEING_MIGRATED_TO_V2);
pgc_queue_unlock(cache, &cache->hot);
// update the extents JudyL
size_t current_extent_index_id;
Pvoid_t *PValue = JudyLIns(&JudyL_extents_pos, xio->pos, PJE0);
if(!PValue || *PValue == PJERR)
fatal("Corrupted JudyL extents pos");
struct jv2_extents_info *ei;
if(!*PValue) {
ei = aral_mallocz(ar_ei); // callocz(1, sizeof(struct jv2_extents_info));
ei->pos = xio->pos;
ei->bytes = xio->bytes;
ei->number_of_pages = 1;
ei->index = master_extent_index_id++;
*PValue = ei;
count_of_unique_extents++;
}
else {
ei = *PValue;
ei->number_of_pages++;
}
current_extent_index_id = ei->index;
// update the metrics JudyL
PValue = JudyLIns(&JudyL_metrics, page->metric_id, PJE0);
if(!PValue || *PValue == PJERR)
fatal("Corrupted JudyL metrics");
struct jv2_metrics_info *mi;
if(!*PValue) {
mi = aral_mallocz(ar_mi); // callocz(1, sizeof(struct jv2_metrics_info));
mi->uuid = mrg_metric_uuid(main_mrg, (METRIC *)page->metric_id);
mi->first_time_s = page->start_time_s;
mi->last_time_s = page->end_time_s;
mi->number_of_pages = 1;
mi->page_list_header = 0;
mi->JudyL_pages_by_start_time = NULL;
*PValue = mi;
count_of_unique_metrics++;
}
else {
mi = *PValue;
mi->number_of_pages++;
if(page->start_time_s < mi->first_time_s)
mi->first_time_s = page->start_time_s;
if(page->end_time_s > mi->last_time_s)
mi->last_time_s = page->end_time_s;
}
PValue = JudyLIns(&mi->JudyL_pages_by_start_time, page->start_time_s, PJE0);
if(!PValue || *PValue == PJERR)
fatal("Corrupted JudyL metric pages");
if(!*PValue) {
struct jv2_page_info *pi = aral_mallocz(ar_pi); // callocz(1, (sizeof(struct jv2_page_info)));
pi->start_time_s = page->start_time_s;
pi->end_time_s = page->end_time_s;
pi->update_every_s = page->update_every_s;
pi->page_length = page_size_from_assumed_size(cache, page->assumed_size);
pi->page = page;
pi->extent_index = current_extent_index_id;
pi->custom_data = (cache->config.additional_bytes_per_page) ? page->custom_data : NULL;
*PValue = pi;
count_of_unique_pages++;
}
else {
// impossible situation
internal_fatal(true, "Page is already in JudyL metric pages");
page_flag_clear(page, PGC_PAGE_IS_BEING_MIGRATED_TO_V2);
page_transition_unlock(cache, page);
page_release(cache, page, false);
}
yield_the_processor(); // do not lock too aggressively
pgc_queue_lock(cache, &cache->hot, PGC_QUEUE_LOCK_PRIO_LOW);
}
spinlock_unlock(&sp->migration_to_v2_spinlock);
pgc_queue_unlock(cache, &cache->hot);
// callback
cb(section, datafile_fileno, type, JudyL_metrics, JudyL_extents_pos, count_of_unique_extents, count_of_unique_metrics, count_of_unique_pages, data);
{
Pvoid_t *PValue1;
bool metric_id_first = true;
Word_t metric_id = 0;
while ((PValue1 = JudyLFirstThenNext(JudyL_metrics, &metric_id, &metric_id_first))) {
struct jv2_metrics_info *mi = *PValue1;
Pvoid_t *PValue2;
bool start_time_first = true;
Word_t start_time = 0;
while ((PValue2 = JudyLFirstThenNext(mi->JudyL_pages_by_start_time, &start_time, &start_time_first))) {
struct jv2_page_info *pi = *PValue2;
// balance-parents: transition from hot to clean directly
yield_the_processor(); // do not lock too aggressively
page_set_clean(cache, pi->page, true, false, PGC_QUEUE_LOCK_PRIO_LOW);
page_transition_unlock(cache, pi->page);
page_release(cache, pi->page, true);
// before balance-parents:
// page_transition_unlock(cache, pi->page);
// pgc_page_hot_to_dirty_and_release(cache, pi->page, true);
// old test - don't enable:
// make_acquired_page_clean_and_evict_or_page_release(cache, pi->page);
aral_freez(ar_pi, pi);
}
JudyLFreeArray(&mi->JudyL_pages_by_start_time, PJE0);
aral_freez(ar_mi, mi);
}
JudyLFreeArray(&JudyL_metrics, PJE0);
}
{
Pvoid_t *PValue;
bool extent_pos_first = true;
Word_t extent_pos = 0;
while ((PValue = JudyLFirstThenNext(JudyL_extents_pos, &extent_pos, &extent_pos_first))) {
struct jv2_extents_info *ei = *PValue;
aral_freez(ar_ei, ei);
}
JudyLFreeArray(&JudyL_extents_pos, PJE0);
}
aral_by_size_release(ar_ei);
aral_by_size_release(ar_pi);
aral_by_size_release(ar_mi);
p2_sub_fetch(&cache->stats.p2_workers_jv2_flush, 1);
// balance-parents: do not flush, there is nothing dirty
// flush_pages(cache, cache->config.max_flushes_inline, PGC_SECTION_ALL, false, false);
}
static bool match_page_data(PGC_PAGE *page, void *data) {
return (page->data == data);
}
void pgc_open_evict_clean_pages_of_datafile(PGC *cache, struct rrdengine_datafile *datafile) {
evict_pages_with_filter(cache, 0, 0, true, true, match_page_data, datafile);
}
size_t pgc_count_clean_pages_having_data_ptr(PGC *cache, Word_t section, void *ptr) {
size_t found = 0;
pgc_queue_lock(cache, &cache->clean, PGC_QUEUE_LOCK_PRIO_LOW);
for(PGC_PAGE *page = cache->clean.base; page ;page = page->link.next)
found += (page->data == ptr && page->section == section) ? 1 : 0;
pgc_queue_unlock(cache, &cache->clean);
return found;
}
size_t pgc_count_hot_pages_having_data_ptr(PGC *cache, Word_t section, void *ptr) {
size_t found = 0;
pgc_queue_lock(cache, &cache->hot, PGC_QUEUE_LOCK_PRIO_LOW);
Pvoid_t *section_pages_pptr = JudyLGet(cache->hot.sections_judy, section, PJE0);
if(section_pages_pptr) {
struct section_pages *sp = *section_pages_pptr;
for(PGC_PAGE *page = sp->base; page ;page = page->link.next)
found += (page->data == ptr) ? 1 : 0;
}
pgc_queue_unlock(cache, &cache->hot);
return found;
}
// ----------------------------------------------------------------------------
// unittest
static void unittest_free_clean_page_callback(PGC *cache __maybe_unused, PGC_ENTRY entry __maybe_unused) {
;
}
static void unittest_save_dirty_page_callback(PGC *cache __maybe_unused, PGC_ENTRY *entries_array __maybe_unused, PGC_PAGE **pages_array __maybe_unused, size_t entries __maybe_unused) {
;
}
#ifdef PGC_STRESS_TEST
struct {
bool stop;
PGC *cache;
PGC_PAGE **metrics;
size_t clean_metrics;
size_t hot_metrics;
time_t first_time_t;
time_t last_time_t;
size_t cache_size;
size_t query_threads;
size_t collect_threads;
size_t partitions;
size_t points_per_page;
time_t time_per_collection_ut;
time_t time_per_query_ut;
time_t time_per_flush_ut;
PGC_OPTIONS options;
char rand_statebufs[1024];
struct random_data *random_data;
} pgc_uts = {
.stop = false,
.metrics = NULL,
.clean_metrics = 100000,
.hot_metrics = 1000000,
.first_time_t = 100000000,
.last_time_t = 0,
.cache_size = 0, // get the default (8MB)
.collect_threads = 16,
.query_threads = 16,
.partitions = 0, // get the default (system cpus)
.options = PGC_OPTIONS_AUTOSCALE,/* PGC_OPTIONS_FLUSH_PAGES_INLINE | PGC_OPTIONS_EVICT_PAGES_INLINE,*/
.points_per_page = 10,
.time_per_collection_ut = 1000000,
.time_per_query_ut = 250,
.time_per_flush_ut = 100,
.rand_statebufs = {},
.random_data = NULL,
};
void *unittest_stress_test_collector(void *ptr) {
size_t id = *((size_t *)ptr);
size_t metric_start = pgc_uts.clean_metrics;
size_t metric_end = pgc_uts.clean_metrics + pgc_uts.hot_metrics;
size_t number_of_metrics = metric_end - metric_start;
size_t per_collector_metrics = number_of_metrics / pgc_uts.collect_threads;
metric_start = metric_start + per_collector_metrics * id + 1;
metric_end = metric_start + per_collector_metrics - 1;
time_t start_time_t = pgc_uts.first_time_t + 1;
heartbeat_t hb;
heartbeat_init(&hb, pgc_uts.time_per_collection_ut);
while(!__atomic_load_n(&pgc_uts.stop, __ATOMIC_RELAXED)) {
// netdata_log_info("COLLECTOR %zu: collecting metrics %zu to %zu, from %ld to %lu", id, metric_start, metric_end, start_time_t, start_time_t + pgc_uts.points_per_page);
for (size_t i = metric_start; i < metric_end; i++) {
bool added;
pgc_uts.metrics[i] = pgc_page_add_and_acquire(pgc_uts.cache, (PGC_ENTRY) {
.section = 1,
.metric_id = i,
.start_time_t = start_time_t,
.end_time_t = start_time_t,
.update_every = 1,
.size = 4096,
.data = NULL,
.hot = true,
}, &added);
if(!pgc_is_page_hot(pgc_uts.metrics[i]) || !added) {
pgc_page_release(pgc_uts.cache, pgc_uts.metrics[i]);
pgc_uts.metrics[i] = NULL;
}
}
time_t end_time_t = start_time_t + (time_t)pgc_uts.points_per_page;
while(++start_time_t <= end_time_t && !__atomic_load_n(&pgc_uts.stop, __ATOMIC_RELAXED)) {
heartbeat_next(&hb);
for (size_t i = metric_start; i < metric_end; i++) {
if(pgc_uts.metrics[i])
pgc_page_hot_set_end_time_t(pgc_uts.cache, pgc_uts.metrics[i], start_time_t);
}
__atomic_store_n(&pgc_uts.last_time_t, start_time_t, __ATOMIC_RELAXED);
}
for (size_t i = metric_start; i < metric_end; i++) {
if (pgc_uts.metrics[i]) {
if(i % 10 == 0)
pgc_page_to_clean_evict_or_release(pgc_uts.cache, pgc_uts.metrics[i]);
else
pgc_page_hot_to_dirty_and_release(pgc_uts.cache, pgc_uts.metrics[i], false);
}
}
}
return ptr;
}
void *unittest_stress_test_queries(void *ptr) {
size_t id = *((size_t *)ptr);
struct random_data *random_data = &pgc_uts.random_data[id];
size_t start = 0;
size_t end = pgc_uts.clean_metrics + pgc_uts.hot_metrics;
while(!__atomic_load_n(&pgc_uts.stop, __ATOMIC_RELAXED)) {
int32_t random_number;
random_r(random_data, &random_number);
size_t metric_id = random_number % (end - start);
time_t start_time_t = pgc_uts.first_time_t;
time_t end_time_t = __atomic_load_n(&pgc_uts.last_time_t, __ATOMIC_RELAXED);
if(end_time_t <= start_time_t)
end_time_t = start_time_t + 1;
size_t pages = (end_time_t - start_time_t) / pgc_uts.points_per_page + 1;
PGC_PAGE *array[pages];
for(size_t i = 0; i < pages ;i++)
array[i] = NULL;
// find the pages the cache has
for(size_t i = 0; i < pages ;i++) {
time_t page_start_time = start_time_t + (time_t)(i * pgc_uts.points_per_page);
array[i] = pgc_page_get_and_acquire(pgc_uts.cache, 1, metric_id,
page_start_time, (i < pages - 1)?PGC_SEARCH_EXACT:PGC_SEARCH_CLOSEST);
}
// load the rest of the pages
for(size_t i = 0; i < pages ;i++) {
if(array[i]) continue;
time_t page_start_time = start_time_t + (time_t)(i * pgc_uts.points_per_page);
array[i] = pgc_page_add_and_acquire(pgc_uts.cache, (PGC_ENTRY) {
.section = 1,
.metric_id = metric_id,
.start_time_t = page_start_time,
.end_time_t = page_start_time + (time_t)pgc_uts.points_per_page,
.update_every = 1,
.size = 4096,
.data = NULL,
.hot = false,
}, NULL);
}
// do the query
// ...
struct timespec work_duration = {.tv_sec = 0, .tv_nsec = pgc_uts.time_per_query_ut * NSEC_PER_USEC };
nanosleep(&work_duration, NULL);
// release the pages
for(size_t i = 0; i < pages ;i++) {
if(!array[i]) continue;
pgc_page_release(pgc_uts.cache, array[i]);
array[i] = NULL;
}
}
return ptr;
}
void *unittest_stress_test_service(void *ptr) {
heartbeat_t hb;
heartbeat_init(&hb, USEC_PER_SEC);
while(!__atomic_load_n(&pgc_uts.stop, __ATOMIC_RELAXED)) {
heartbeat_next(&hb);
pgc_flush_pages(pgc_uts.cache, 1000);
pgc_evict_pages(pgc_uts.cache, 0, 0);
}
return ptr;
}
static void unittest_stress_test_save_dirty_page_callback(PGC *cache __maybe_unused, PGC_ENTRY *entries_array __maybe_unused, PGC_PAGE **pages_array __maybe_unused, size_t entries __maybe_unused) {
// netdata_log_info("SAVE %zu pages", entries);
if(!pgc_uts.stop) {
usec_t t = pgc_uts.time_per_flush_ut;
if(t > 0) {
struct timespec work_duration = {
.tv_sec = t / USEC_PER_SEC,
.tv_nsec = (long) ((t % USEC_PER_SEC) * NSEC_PER_USEC)
};
nanosleep(&work_duration, NULL);
}
}
}
void unittest_stress_test(void) {
pgc_uts.cache = pgc_create(pgc_uts.cache_size * 1024 * 1024,
unittest_free_clean_page_callback,
64, unittest_stress_test_save_dirty_page_callback,
1000, 10000, 1,
pgc_uts.options, pgc_uts.partitions, 0);
pgc_uts.metrics = callocz(pgc_uts.clean_metrics + pgc_uts.hot_metrics, sizeof(PGC_PAGE *));
pthread_t service_thread;
nd_thread_create(&service_thread, "SERVICE",
NETDATA_THREAD_OPTION_JOINABLE | NETDATA_THREAD_OPTION_DONT_LOG,
unittest_stress_test_service, NULL);
pthread_t collect_threads[pgc_uts.collect_threads];
size_t collect_thread_ids[pgc_uts.collect_threads];
for(size_t i = 0; i < pgc_uts.collect_threads ;i++) {
collect_thread_ids[i] = i;
char buffer[100 + 1];
snprintfz(buffer, sizeof(buffer) - 1, "COLLECT_%zu", i);
nd_thread_create(&collect_threads[i], buffer,
NETDATA_THREAD_OPTION_JOINABLE | NETDATA_THREAD_OPTION_DONT_LOG,
unittest_stress_test_collector, &collect_thread_ids[i]);
}
pthread_t queries_threads[pgc_uts.query_threads];
size_t query_thread_ids[pgc_uts.query_threads];
pgc_uts.random_data = callocz(pgc_uts.query_threads, sizeof(struct random_data));
for(size_t i = 0; i < pgc_uts.query_threads ;i++) {
query_thread_ids[i] = i;
char buffer[100 + 1];
snprintfz(buffer, sizeof(buffer) - 1, "QUERY_%zu", i);
initstate_r(1, pgc_uts.rand_statebufs, 1024, &pgc_uts.random_data[i]);
nd_thread_create(&queries_threads[i], buffer,
NETDATA_THREAD_OPTION_JOINABLE | NETDATA_THREAD_OPTION_DONT_LOG,
unittest_stress_test_queries, &query_thread_ids[i]);
}
heartbeat_t hb;
heartbeat_init(&hb, USEC_PER_SEC);
struct {
size_t entries;
size_t added;
size_t deleted;
size_t referenced;
size_t hot_entries;
size_t hot_added;
size_t hot_deleted;
size_t dirty_entries;
size_t dirty_added;
size_t dirty_deleted;
size_t clean_entries;
size_t clean_added;
size_t clean_deleted;
size_t searches_exact;
size_t searches_exact_hits;
size_t searches_closest;
size_t searches_closest_hits;
size_t collections;
size_t events_cache_under_severe_pressure;
size_t events_cache_needs_space_90;
size_t events_flush_critical;
} stats = {}, old_stats = {};
for(int i = 0; i < 86400 ;i++) {
heartbeat_next(&hb);
old_stats = stats;
stats.entries = __atomic_load_n(&pgc_uts.cache->stats.entries, __ATOMIC_RELAXED);
stats.added = __atomic_load_n(&pgc_uts.cache->stats.added_entries, __ATOMIC_RELAXED);
stats.deleted = __atomic_load_n(&pgc_uts.cache->stats.removed_entries, __ATOMIC_RELAXED);
stats.referenced = __atomic_load_n(&pgc_uts.cache->stats.referenced_entries, __ATOMIC_RELAXED);
stats.hot_entries = __atomic_load_n(&pgc_uts.cache->hot.stats->entries, __ATOMIC_RELAXED);
stats.hot_added = __atomic_load_n(&pgc_uts.cache->hot.stats->added_entries, __ATOMIC_RELAXED);
stats.hot_deleted = __atomic_load_n(&pgc_uts.cache->hot.stats->removed_entries, __ATOMIC_RELAXED);
stats.dirty_entries = __atomic_load_n(&pgc_uts.cache->dirty.stats->entries, __ATOMIC_RELAXED);
stats.dirty_added = __atomic_load_n(&pgc_uts.cache->dirty.stats->added_entries, __ATOMIC_RELAXED);
stats.dirty_deleted = __atomic_load_n(&pgc_uts.cache->dirty.stats->removed_entries, __ATOMIC_RELAXED);
stats.clean_entries = __atomic_load_n(&pgc_uts.cache->clean.stats->entries, __ATOMIC_RELAXED);
stats.clean_added = __atomic_load_n(&pgc_uts.cache->clean.stats->added_entries, __ATOMIC_RELAXED);
stats.clean_deleted = __atomic_load_n(&pgc_uts.cache->clean.stats->removed_entries, __ATOMIC_RELAXED);
stats.searches_exact = __atomic_load_n(&pgc_uts.cache->stats.searches_exact, __ATOMIC_RELAXED);
stats.searches_exact_hits = __atomic_load_n(&pgc_uts.cache->stats.searches_exact_hits, __ATOMIC_RELAXED);
stats.searches_closest = __atomic_load_n(&pgc_uts.cache->stats.searches_closest, __ATOMIC_RELAXED);
stats.searches_closest_hits = __atomic_load_n(&pgc_uts.cache->stats.searches_closest_hits, __ATOMIC_RELAXED);
stats.events_cache_under_severe_pressure = __atomic_load_n(&pgc_uts.cache->stats.events_cache_under_severe_pressure, __ATOMIC_RELAXED);
stats.events_cache_needs_space_90 = __atomic_load_n(&pgc_uts.cache->stats.events_cache_needs_space_aggressively, __ATOMIC_RELAXED);
stats.events_flush_critical = __atomic_load_n(&pgc_uts.cache->stats.events_flush_critical, __ATOMIC_RELAXED);
size_t searches_exact = stats.searches_exact - old_stats.searches_exact;
size_t searches_closest = stats.searches_closest - old_stats.searches_closest;
size_t hit_exact = stats.searches_exact_hits - old_stats.searches_exact_hits;
size_t hit_closest = stats.searches_closest_hits - old_stats.searches_closest_hits;
double hit_exact_pc = (searches_exact > 0) ? (double)hit_exact * 100.0 / (double)searches_exact : 0.0;
double hit_closest_pc = (searches_closest > 0) ? (double)hit_closest * 100.0 / (double)searches_closest : 0.0;
#ifdef PGC_COUNT_POINTS_COLLECTED
stats.collections = __atomic_load_n(&pgc_uts.cache->stats.points_collected, __ATOMIC_RELAXED);
#endif
char *cache_status = "N";
if(stats.events_cache_under_severe_pressure > old_stats.events_cache_under_severe_pressure)
cache_status = "F";
else if(stats.events_cache_needs_space_90 > old_stats.events_cache_needs_space_90)
cache_status = "f";
char *flushing_status = "N";
if(stats.events_flush_critical > old_stats.events_flush_critical)
flushing_status = "F";
netdata_log_info("PGS %5zuk +%4zuk/-%4zuk "
"| RF %5zuk "
"| HOT %5zuk +%4zuk -%4zuk "
"| DRT %s %5zuk +%4zuk -%4zuk "
"| CLN %s %5zuk +%4zuk -%4zuk "
"| SRCH %4zuk %4zuk, HIT %4.1f%% %4.1f%% "
#ifdef PGC_COUNT_POINTS_COLLECTED
"| CLCT %8.4f Mps"
#endif
, stats.entries / 1000
, (stats.added - old_stats.added) / 1000, (stats.deleted - old_stats.deleted) / 1000
, stats.referenced / 1000
, stats.hot_entries / 1000, (stats.hot_added - old_stats.hot_added) / 1000, (stats.hot_deleted - old_stats.hot_deleted) / 1000
, flushing_status
, stats.dirty_entries / 1000
, (stats.dirty_added - old_stats.dirty_added) / 1000, (stats.dirty_deleted - old_stats.dirty_deleted) / 1000
, cache_status
, stats.clean_entries / 1000
, (stats.clean_added - old_stats.clean_added) / 1000, (stats.clean_deleted - old_stats.clean_deleted) / 1000
, searches_exact / 1000, searches_closest / 1000
, hit_exact_pc, hit_closest_pc
#ifdef PGC_COUNT_POINTS_COLLECTED
, (double)(stats.collections - old_stats.collections) / 1000.0 / 1000.0
#endif
);
}
netdata_log_info("Waiting for threads to stop...");
__atomic_store_n(&pgc_uts.stop, true, __ATOMIC_RELAXED);
nd_thread_join(service_thread, NULL);
for(size_t i = 0; i < pgc_uts.collect_threads ;i++)
nd_thread_join(collect_threads[i],NULL);
for(size_t i = 0; i < pgc_uts.query_threads ;i++)
nd_thread_join(queries_threads[i],NULL);
pgc_destroy(pgc_uts.cache);
freez(pgc_uts.metrics);
freez(pgc_uts.random_data);
}
#endif
int pgc_unittest(void) {
PGC *cache = pgc_create("test",
32 * 1024 * 1024, unittest_free_clean_page_callback,
64, NULL, unittest_save_dirty_page_callback,
10, 10, 1000, 10,
PGC_OPTIONS_DEFAULT, 1, 11);
// FIXME - unit tests
// - add clean page
// - add clean page again (should not add it)
// - release page (should decrement counters)
// - add hot page
// - add hot page again (should not add it)
// - turn hot page to dirty, with and without a reference counter to it
// - dirty pages are saved once there are enough of them
// - find page exact
// - find page (should return last)
// - find page (should return next)
// - page cache full (should evict)
// - on destroy, turn hot pages to dirty and save them
PGC_PAGE *page1 = pgc_page_add_and_acquire(cache, (PGC_ENTRY){
.section = 1,
.metric_id = 10,
.start_time_s = 100,
.end_time_s = 1000,
.size = 4096,
.data = NULL,
.hot = false,
.custom_data = (uint8_t *)"0123456789",
}, NULL);
if(strcmp(pgc_page_custom_data(cache, page1), "0123456789") != 0)
fatal("custom data do not work");
memcpy(pgc_page_custom_data(cache, page1), "ABCDEFGHIJ", 11);
if(strcmp(pgc_page_custom_data(cache, page1), "ABCDEFGHIJ") != 0)
fatal("custom data do not work");
pgc_page_release(cache, page1);
PGC_PAGE *page2 = pgc_page_add_and_acquire(cache, (PGC_ENTRY){
.section = 2,
.metric_id = 10,
.start_time_s = 1001,
.end_time_s = 2000,
.size = 4096,
.data = NULL,
.hot = true,
}, NULL);
pgc_page_hot_set_end_time_s(cache, page2, 2001, 0);
pgc_page_hot_to_dirty_and_release(cache, page2, false);
PGC_PAGE *page3 = pgc_page_add_and_acquire(cache, (PGC_ENTRY){
.section = 3,
.metric_id = 10,
.start_time_s = 1001,
.end_time_s = 2000,
.size = 4096,
.data = NULL,
.hot = true,
}, NULL);
pgc_page_hot_set_end_time_s(cache, page3, 2001, 0);
pgc_page_hot_to_dirty_and_release(cache, page3, false);
pgc_destroy(cache, true);
#ifdef PGC_STRESS_TEST
unittest_stress_test();
#endif
return 0;
}