githubmirror/redis

src/evict.c
764 строки32 KB
Ozan Tezcan
Add Atomic Slot Migration (ASM) support (#14414)
22 окт 2025, 15:56
22 окт 2025, 15:562bc4e02
/* Maxmemory directive handling (LRU eviction and other policies).
 *
 * ----------------------------------------------------------------------------
 *
 * Copyright (c) 2009-Present, Redis Ltd.
 * All rights reserved.
 *
 * Licensed under your choice of (a) the Redis Source Available License 2.0
 * (RSALv2); or (b) the Server Side Public License v1 (SSPLv1); or (c) the
 * GNU Affero General Public License v3 (AGPLv3).
 */

#include "server.h"
#include "bio.h"
#include "atomicvar.h"
#include "script.h"
#include "cluster.h"
#include "cluster_asm.h"
#include <math.h>

/* ----------------------------------------------------------------------------
 * Data structures
 * --------------------------------------------------------------------------*/

/* To improve the quality of the LRU approximation we take a set of keys
 * that are good candidate for eviction across performEvictions() calls.
 *
 * Entries inside the eviction pool are taken ordered by idle time, putting
 * greater idle times to the right (ascending order).
 *
 * When an LFU policy is used instead, a reverse frequency indication is used
 * instead of the idle time, so that we still evict by larger value (larger
 * inverse frequency means to evict keys with the least frequent accesses).
 *
 * Empty entries have the key pointer set to NULL. */
#define EVPOOL_SIZE 16
#define EVPOOL_CACHED_SDS_SIZE 255
struct evictionPoolEntry {
    unsigned long long idle;    /* Object idle time (inverse frequency for LFU) */
    sds key;                    /* Key name. */
    sds cached;                 /* Cached SDS object for key name. */
    int dbid;                   /* Key DB number. */
    int slot;                   /* Slot. */
};

static struct evictionPoolEntry *EvictionPoolLRU;

/* ----------------------------------------------------------------------------
 * Implementation of eviction, aging and LRU
 * --------------------------------------------------------------------------*/

/* Return the LRU clock, based on the clock resolution. This is a time
 * in a reduced-bits format that can be used to set and check the
 * object->lru field of redisObject structures. */
unsigned int getLRUClock(void) {
    return (mstime()/LRU_CLOCK_RESOLUTION) & LRU_CLOCK_MAX;
}

/* This function is used to obtain the current LRU clock.
 * If the current resolution is lower than the frequency we refresh the
 * LRU clock (as it should be in production servers) we return the
 * precomputed value, otherwise we need to resort to a system call. */
unsigned int LRU_CLOCK(void) {
    unsigned int lruclock;
    if (1000/server.hz <= LRU_CLOCK_RESOLUTION) {
        lruclock = server.lruclock;
    } else {
        lruclock = getLRUClock();
    }
    return lruclock;
}

/* Given an object returns the min number of milliseconds the object was never
 * requested, using an approximated LRU algorithm. */
unsigned long long estimateObjectIdleTime(robj *o) {
    unsigned long long lruclock = LRU_CLOCK();
    if (lruclock >= o->lru) {
        return (lruclock - o->lru) * LRU_CLOCK_RESOLUTION;
    } else {
        return (lruclock + (LRU_CLOCK_MAX - o->lru)) *
                    LRU_CLOCK_RESOLUTION;
    }
}

/* During atomic slot migration, keys that are being imported are in an
 * intermediate state. We cannot evict them and therefore skip them. */
static int randomEvictionShouldSkipDictIndex(int didx) {
    return !clusterCanAccessKeysInSlot(didx);
}

/* LRU approximation algorithm
 *
 * Redis uses an approximation of the LRU algorithm that runs in constant
 * memory. Every time there is a key to expire, we sample N keys (with
 * N very small, usually in around 5) to populate a pool of best keys to
 * evict of M keys (the pool size is defined by EVPOOL_SIZE).
 *
 * The N keys sampled are added in the pool of good keys to expire (the one
 * with an old access time) if they are better than one of the current keys
 * in the pool.
 *
 * After the pool is populated, the best key we have in the pool is expired.
 * However note that we don't remove keys from the pool when they are deleted
 * so the pool may contain keys that no longer exist.
 *
 * When we try to evict a key, and all the entries in the pool don't exist
 * we populate it again. This time we'll be sure that the pool has at least
 * one key that can be evicted, if there is at least one key that can be
 * evicted in the whole database. */

/* Create a new eviction pool. */
void evictionPoolAlloc(void) {
    struct evictionPoolEntry *ep;
    int j;

    ep = zmalloc(sizeof(*ep)*EVPOOL_SIZE);
    for (j = 0; j < EVPOOL_SIZE; j++) {
        ep[j].idle = 0;
        ep[j].key = NULL;
        ep[j].cached = sdsnewlen(NULL,EVPOOL_CACHED_SDS_SIZE);
        ep[j].dbid = 0;
    }
    EvictionPoolLRU = ep;
}

/* This is a helper function for performEvictions(), it is used in order
 * to populate the evictionPool with a few entries every time we want to
 * expire a key. Keys with idle time bigger than one of the current
 * keys are added. Keys are always added if there are free entries.
 *
 * We insert keys on place in ascending order, so keys with the smaller
 * idle time are on the left, and keys with the higher idle time on the
 * right. */
int evictionPoolPopulate(redisDb *db, kvstore *samplekvs, struct evictionPoolEntry *pool) {
    int j, k, count;
    dictEntry *samples[server.maxmemory_samples];

    /* Don't retry, since we will call evictionPoolPopulate multiple times if needed. */
    int slot = kvstoreGetFairRandomDictIndex(samplekvs, randomEvictionShouldSkipDictIndex, 1, 0);
    if (slot == -1) return 0;
    count = kvstoreDictGetSomeKeys(samplekvs,slot,samples,server.maxmemory_samples);
    for (j = 0; j < count; j++) {
        unsigned long long idle;
        
        dictEntry *de = samples[j];
        kvobj *kv = dictGetKV(de);
        sds key = kvobjGetKey(kv);
        
        /* Calculate the idle time according to the policy. This is called
         * idle just because the code initially handled LRU, but is in fact
         * just a score where a higher score means better candidate. */
        if (server.maxmemory_policy & MAXMEMORY_FLAG_LRU) {
            idle = estimateObjectIdleTime(kv);
        } else if (server.maxmemory_policy & MAXMEMORY_FLAG_LFU) {
            /* When we use an LRU policy, we sort the keys by idle time
             * so that we expire keys starting from greater idle time.
             * However when the policy is an LFU one, we have a frequency
             * estimation, and we want to evict keys with lower frequency
             * first. So inside the pool we put objects using the inverted
             * frequency subtracting the actual frequency to the maximum
             * frequency of 255. */
            idle = 255-LFUDecrAndReturn(kv);
        } else if (server.maxmemory_policy == MAXMEMORY_VOLATILE_TTL) {
            /* In this case the sooner the expire the better. */
            idle = ULLONG_MAX - kvobjGetExpire(kv);
        } else {
            serverPanic("Unknown eviction policy in evictionPoolPopulate()");
        }

        /* Insert the element inside the pool.
         * First, find the first empty bucket or the first populated
         * bucket that has an idle time smaller than our idle time. */
        k = 0;
        while (k < EVPOOL_SIZE &&
               pool[k].key &&
               pool[k].idle < idle) k++;
        if (k == 0 && pool[EVPOOL_SIZE-1].key != NULL) {
            /* Can't insert if the element is < the worst element we have
             * and there are no empty buckets. */
            continue;
        } else if (k < EVPOOL_SIZE && pool[k].key == NULL) {
            /* Inserting into empty position. No setup needed before insert. */
        } else {
            /* Inserting in the middle. Now k points to the first element
             * greater than the element to insert.  */
            if (pool[EVPOOL_SIZE-1].key == NULL) {
                /* Free space on the right? Insert at k shifting
                 * all the elements from k to end to the right. */

                /* Save SDS before overwriting. */
                sds cached = pool[EVPOOL_SIZE-1].cached;
                memmove(pool+k+1,pool+k,
                    sizeof(pool[0])*(EVPOOL_SIZE-k-1));
                pool[k].cached = cached;
            } else {
                /* No free space on right? Insert at k-1 */
                k--;
                /* Shift all elements on the left of k (included) to the
                 * left, so we discard the element with smaller idle time. */
                sds cached = pool[0].cached; /* Save SDS before overwriting. */
                if (pool[0].key != pool[0].cached) sdsfree(pool[0].key);
                memmove(pool,pool+1,sizeof(pool[0])*k);
                pool[k].cached = cached;
            }
        }

        /* Try to reuse the cached SDS string allocated in the pool entry,
         * because allocating and deallocating this object is costly
         * (according to the profiler, not my fantasy. Remember:
         * premature optimization bla bla bla. */
        int klen = sdslen(key);
        if (klen > EVPOOL_CACHED_SDS_SIZE) {
            pool[k].key = sdsdup(key);
        } else {
            memcpy(pool[k].cached,key,klen+1);
            sdssetlen(pool[k].cached,klen);
            pool[k].key = pool[k].cached;
        }
        pool[k].idle = idle;
        pool[k].dbid = db->id;
        pool[k].slot = slot;
    }

    return count;
}

/* ----------------------------------------------------------------------------
 * LFU (Least Frequently Used) implementation.

 * We have 24 total bits of space in each object in order to implement
 * an LFU (Least Frequently Used) eviction policy, since we re-use the
 * LRU field for this purpose.
 *
 * We split the 24 bits into two fields:
 *
 *            16 bits       8 bits
 *     +------------------+--------+
 *     + Last access time | LOG_C  |
 *     +------------------+--------+
 *
 * LOG_C is a logarithmic counter that provides an indication of the access
 * frequency. However this field must also be decremented otherwise what used
 * to be a frequently accessed key in the past, will remain ranked like that
 * forever, while we want the algorithm to adapt to access pattern changes.
 *
 * So the remaining 16 bits are used in order to store the "access time",
 * a reduced-precision Unix time (we take 16 bits of the time converted
 * in minutes since we don't care about wrapping around) where the LOG_C
 * counter decays every minute by default (depends on lfu-decay-time).
 *
 * New keys don't start at zero, in order to have the ability to collect
 * some accesses before being trashed away, so they start at LFU_INIT_VAL.
 * The logarithmic increment performed on LOG_C takes care of LFU_INIT_VAL
 * when incrementing the key, so that keys starting at LFU_INIT_VAL
 * (or having a smaller value) have a very high chance of being incremented
 * on access. (The chance depends on counter and lfu-log-factor.)
 *
 * During decrement, the value of the logarithmic counter is decremented by
 * one when lfu-decay-time minutes elapsed.
 * --------------------------------------------------------------------------*/

/* Return the current time in minutes, just taking the least significant
 * 16 bits. The returned time is suitable to be stored as LDT (last access
 * time) for the LFU implementation. */
unsigned long LFUGetTimeInMinutes(void) {
    return (server.unixtime/60) & 65535;
}

/* Given an object ldt (last access time), compute the minimum number of minutes
 * that elapsed since the last access. Handle overflow (ldt greater than
 * the current 16 bits minutes time) considering the time as wrapping
 * exactly once. */
unsigned long LFUTimeElapsed(unsigned long ldt) {
    unsigned long now = LFUGetTimeInMinutes();
    if (now >= ldt) return now-ldt;
    return 65535-ldt+now;
}

/* Logarithmically increment a counter. The greater is the current counter value
 * the less likely is that it gets really incremented. Saturate it at 255. */
uint8_t LFULogIncr(uint8_t counter) {
    if (counter == 255) return 255;
    double r = (double)rand()/RAND_MAX;
    double baseval = counter - LFU_INIT_VAL;
    if (baseval < 0) baseval = 0;
    double p = 1.0/(baseval*server.lfu_log_factor+1);
    if (r < p) counter++;
    return counter;
}

/* If the object's ldt (last access time) is reached, decrement the LFU counter but
 * do not update LFU fields of the object, we update the access time
 * and counter in an explicit way when the object is really accessed.
 * And we will decrement the counter according to the times of
 * elapsed time than server.lfu_decay_time.
 * Return the object frequency counter.
 *
 * This function is used in order to scan the dataset for the best object
 * to fit: as we check for the candidate, we incrementally decrement the
 * counter of the scanned objects if needed. */
unsigned long LFUDecrAndReturn(robj *o) {
    unsigned long ldt = o->lru >> 8;
    unsigned long counter = o->lru & 255;
    unsigned long num_periods = server.lfu_decay_time ? LFUTimeElapsed(ldt) / server.lfu_decay_time : 0;
    if (num_periods)
        counter = (num_periods > counter) ? 0 : counter - num_periods;
    return counter;
}

/* We don't want to count AOF buffers and slaves output buffers as
 * used memory: the eviction should use mostly data size, because
 * it can cause feedback-loop when we push DELs into them, putting
 * more and more DELs will make them bigger, if we count them, we
 * need to evict more keys, and then generate more DELs, maybe cause
 * massive eviction loop, even all keys are evicted.
 *
 * This function returns the sum of AOF and replication buffer. */
size_t freeMemoryGetNotCountedMemory(void) {
    size_t overhead = 0;

    /* Since all replicas and replication backlog share global replication
     * buffer, we think only the part of exceeding backlog size is the extra
     * separate consumption of replicas.
     *
     * Note that although the backlog is also initially incrementally grown
     * (pushing DELs consumes memory), it'll eventually stop growing and
     * remain constant in size, so even if its creation will cause some
     * eviction, it's capped, and also here to stay (no resonance effect)
     *
     * Note that, because we trim backlog incrementally in the background,
     * backlog size may exceeds our setting if slow replicas that reference
     * vast replication buffer blocks disconnect. To avoid massive eviction
     * loop, we don't count the delayed freed replication backlog into used
     * memory even if there are no replicas, i.e. we still regard this memory
     * as replicas'. */
    if ((long long)server.repl_buffer_mem > server.repl_backlog_size) {
        /* We use list structure to manage replication buffer blocks, so backlog
         * also occupies some extra memory, we can't know exact blocks numbers,
         * we only get approximate size according to per block size. */
        size_t extra_approx_size =
            (server.repl_backlog_size/PROTO_REPLY_CHUNK_BYTES + 1) *
            (sizeof(replBufBlock)+sizeof(listNode));
        size_t counted_mem = server.repl_backlog_size + extra_approx_size;
        if (server.repl_buffer_mem > counted_mem) {
            overhead += (server.repl_buffer_mem - counted_mem);
        }
    }

    /* The migrate client is like a replica, we also push DELs into it when
     * evicting keys belonging to the migrating slot, so we don't count its
     * output buffer to avoid eviction loop. */
    overhead += asmGetMigrateOutputBufferSize();

    if (server.aof_state != AOF_OFF) {
        overhead += sdsAllocSize(server.aof_buf);
    }
    return overhead;
}

/* Get the memory status from the point of view of the maxmemory directive:
 * if the memory used is under the maxmemory setting then C_OK is returned.
 * Otherwise, if we are over the memory limit, the function returns
 * C_ERR.
 *
 * The function may return additional info via reference, only if the
 * pointers to the respective arguments is not NULL. Certain fields are
 * populated only when C_ERR is returned:
 *
 *  'total'     total amount of bytes used.
 *              (Populated both for C_ERR and C_OK)
 *
 *  'logical'   the amount of memory used minus the slaves/AOF buffers.
 *              (Populated when C_ERR is returned)
 *
 *  'tofree'    the amount of memory that should be released
 *              in order to return back into the memory limits.
 *              (Populated when C_ERR is returned)
 *
 *  'level'     this usually ranges from 0 to 1, and reports the amount of
 *              memory currently used. May be > 1 if we are over the memory
 *              limit.
 *              (Populated both for C_ERR and C_OK)
 */
int getMaxmemoryState(size_t *total, size_t *logical, size_t *tofree, float *level) {
    size_t mem_reported, mem_used, mem_tofree;

    /* Check if we are over the memory usage limit. If we are not, no need
     * to subtract the slaves output buffers. We can just return ASAP. */
    mem_reported = zmalloc_used_memory();
    if (total) *total = mem_reported;

    /* We may return ASAP if there is no need to compute the level. */
    if (!server.maxmemory) {
        if (level) *level = 0;
        return C_OK;
    }
    if (mem_reported <= server.maxmemory && !level) return C_OK;

    /* Remove the size of slaves output buffers and AOF buffer from the
     * count of used memory. */
    mem_used = mem_reported;
    size_t overhead = freeMemoryGetNotCountedMemory();
    mem_used = (mem_used > overhead) ? mem_used-overhead : 0;

    /* Compute the ratio of memory usage. */
    if (level) *level = (float)mem_used / (float)server.maxmemory;

    if (mem_reported <= server.maxmemory) return C_OK;

    /* Check if we are still over the memory limit. */
    if (mem_used <= server.maxmemory) return C_OK;

    /* Compute how much memory we need to free. */
    mem_tofree = mem_used - server.maxmemory;

    if (logical) *logical = mem_used;
    if (tofree) *tofree = mem_tofree;

    return C_ERR;
}

/* Return 1 if used memory is more than maxmemory after allocating more memory,
 * return 0 if not. Redis may reject user's requests or evict some keys if used
 * memory exceeds maxmemory, especially, when we allocate huge memory at once. */
int overMaxmemoryAfterAlloc(size_t moremem) {
    if (!server.maxmemory) return  0; /* No limit. */

    /* Check quickly. */
    size_t mem_used = zmalloc_used_memory();
    if (mem_used + moremem <= server.maxmemory) return 0;

    size_t overhead = freeMemoryGetNotCountedMemory();
    mem_used = (mem_used > overhead) ? mem_used - overhead : 0;
    return mem_used + moremem > server.maxmemory;
}

/* The evictionTimeProc is started when "maxmemory" has been breached and
 * could not immediately be resolved.  This will spin the event loop with short
 * eviction cycles until the "maxmemory" condition has resolved or there are no
 * more evictable items.  */
static int isEvictionProcRunning = 0;
static int evictionTimeProc(
        struct aeEventLoop *eventLoop, long long id, void *clientData) {
    UNUSED(eventLoop);
    UNUSED(id);
    UNUSED(clientData);

    if (performEvictions() == EVICT_RUNNING) return 0;  /* keep evicting */

    /* For EVICT_OK - things are good, no need to keep evicting.
     * For EVICT_FAIL - there is nothing left to evict.  */
    isEvictionProcRunning = 0;
    return AE_NOMORE;
}

void startEvictionTimeProc(void) {
    if (!isEvictionProcRunning) {
        isEvictionProcRunning = 1;
        aeCreateTimeEvent(server.el, 0,
                evictionTimeProc, NULL, NULL);
    }
}

/* Check if it's safe to perform evictions.
 *   Returns 1 if evictions can be performed
 *   Returns 0 if eviction processing should be skipped
 */
static int isSafeToPerformEvictions(void) {
    /* - There must be no script in timeout condition.
     * - Nor we are loading data right now.  */
    if (isInsideYieldingLongCommand() || server.loading) return 0;

    /* By default replicas should ignore maxmemory
     * and just be masters exact copies. */
    if (server.masterhost && server.repl_slave_ignore_maxmemory) return 0;

    /* Disable eviction during slot migration import to avoid delays and errors
     * caused by failed evictions. A special client is created for data import,
     * identified by the CLIENT_MASTER and CLIENT_ASM_IMPORTING flags. */
    if (server.current_client && server.current_client->flags & CLIENT_MASTER &&
        server.current_client->flags & CLIENT_ASM_IMPORTING)
        return 0;

    /* If 'evict' action is paused, for whatever reason, then return false */
    if (isPausedActionsWithUpdate(PAUSE_ACTION_EVICT)) return 0;

    return 1;
}

/* Algorithm for converting tenacity (0-100) to a time limit.  */
static unsigned long evictionTimeLimitUs(void) {
    serverAssert(server.maxmemory_eviction_tenacity >= 0);
    serverAssert(server.maxmemory_eviction_tenacity <= 100);

    if (server.maxmemory_eviction_tenacity <= 10) {
        /* A linear progression from 0..500us */
        return 50uL * server.maxmemory_eviction_tenacity;
    }

    if (server.maxmemory_eviction_tenacity < 100) {
        /* A 15% geometric progression, resulting in a limit of ~2 min at tenacity==99  */
        return (unsigned long)(500.0 * pow(1.15, server.maxmemory_eviction_tenacity - 10.0));
    }

    return ULONG_MAX;   /* No limit to eviction time */
}

/* Check that memory usage is within the current "maxmemory" limit.  If over
 * "maxmemory", attempt to free memory by evicting data (if it's safe to do so).
 *
 * It's possible for Redis to suddenly be significantly over the "maxmemory"
 * setting.  This can happen if there is a large allocation (like a hash table
 * resize) or even if the "maxmemory" setting is manually adjusted.  Because of
 * this, it's important to evict for a managed period of time - otherwise Redis
 * would become unresponsive while evicting.
 *
 * The goal of this function is to improve the memory situation - not to
 * immediately resolve it.  In the case that some items have been evicted but
 * the "maxmemory" limit has not been achieved, an aeTimeProc will be started
 * which will continue to evict items until memory limits are achieved or
 * nothing more is evictable.
 *
 * This should be called before execution of commands.  If EVICT_FAIL is
 * returned, commands which will result in increased memory usage should be
 * rejected.
 *
 * Returns:
 *   EVICT_OK       - memory is OK or it's not possible to perform evictions now
 *   EVICT_RUNNING  - memory is over the limit, but eviction is still processing
 *   EVICT_FAIL     - memory is over the limit, and there's nothing to evict
 * */
int performEvictions(void) {
    /* Note, we don't goto update_metrics here because this check skips eviction
     * as if it wasn't triggered. it's a fake EVICT_OK. */
    if (!isSafeToPerformEvictions()) return EVICT_OK;

    int keys_freed = 0;
    size_t mem_reported, mem_tofree;
    long long mem_freed; /* May be negative */
    mstime_t latency;
    int slaves = listLength(server.slaves);
    int result = EVICT_FAIL;

    if (getMaxmemoryState(&mem_reported,NULL,&mem_tofree,NULL) == C_OK) {
        result = EVICT_OK;
        goto update_metrics;
    }

    if (server.maxmemory_policy == MAXMEMORY_NO_EVICTION) {
        result = EVICT_FAIL;  /* We need to free memory, but policy forbids. */
        goto update_metrics;
    }

    unsigned long eviction_time_limit_us = evictionTimeLimitUs();

    mem_freed = 0;

    latencyStartMonitor(latency);

    monotime evictionTimer;
    elapsedStart(&evictionTimer);

    /* Try to smoke-out bugs (server.also_propagate should be empty here) */
    serverAssert(server.also_propagate.numops == 0);
    /* Evictions are performed on random keys that have nothing to do with the current command slot. */

    while (mem_freed < (long long)mem_tofree) {
        int j, k, i;
        static unsigned int next_db = 0;
        sds bestkey = NULL;
        int bestdbid;
        redisDb *db;
        dictEntry *de;

        if (server.maxmemory_policy & (MAXMEMORY_FLAG_LRU|MAXMEMORY_FLAG_LFU) ||
            server.maxmemory_policy == MAXMEMORY_VOLATILE_TTL)
        {
            struct evictionPoolEntry *pool = EvictionPoolLRU;
            while (bestkey == NULL) {
                unsigned long total_keys = 0;
                unsigned long total_sampled_keys = 0;

                /* We don't want to make local-db choices when expiring keys,
                 * so to start populate the eviction pool sampling keys from
                 * every DB. */
                for (i = 0; i < server.dbnum; i++) {
                    db = server.db+i;
                    kvstore *kvs;
                    if (server.maxmemory_policy & MAXMEMORY_FLAG_ALLKEYS) {
                        kvs = db->keys;
                    } else {
                        kvs = db->expires;
                    }
                    unsigned long sampled_keys = 0;
                    unsigned long current_db_keys = kvstoreSize(kvs);
                    if (current_db_keys == 0) continue;

                    total_keys += current_db_keys;
                    int l = kvstoreNumNonEmptyDicts(kvs);
                    /* Do not exceed the number of non-empty slots when looping. */
                    while (l--) {
                        sampled_keys += evictionPoolPopulate(db, kvs, pool);
                        total_sampled_keys += sampled_keys;
                        /* We have sampled enough keys in the current db, exit the loop. */
                        if (sampled_keys >= (unsigned long) server.maxmemory_samples)
                            break;
                        /* If there are not a lot of keys in the current db, dict/s may be very
                         * sparsely populated, exit the loop without meeting the sampling
                         * requirement. */
                        if (current_db_keys < (unsigned long) server.maxmemory_samples*10)
                            break;
                    }
                }
                if (!total_keys) break; /* No keys to evict. */

                /* If we iterated all the DBs and all non-empty slot dicts, then
                 * did not sample any key, stop sampling. */
                if (!total_sampled_keys) break;

                /* Go backward from best to worst element to evict. */
                for (k = EVPOOL_SIZE-1; k >= 0; k--) {
                    if (pool[k].key == NULL) continue;
                    bestdbid = pool[k].dbid;

                    kvstore *kvs;
                    if (server.maxmemory_policy & MAXMEMORY_FLAG_ALLKEYS) {
                        kvs = server.db[bestdbid].keys;
                    } else {
                        kvs = server.db[bestdbid].expires;
                    }
                    de = kvstoreDictFind(kvs, pool[k].slot, pool[k].key);

                    /* Remove the entry from the pool. */
                    if (pool[k].key != pool[k].cached)
                        sdsfree(pool[k].key);
                    pool[k].key = NULL;
                    pool[k].idle = 0;

                    /* If the key exists, is our pick. Otherwise it is
                     * a ghost and we need to try the next element. */
                    if (de) {
                        bestkey = kvobjGetKey(dictGetKV(de));
                        break;
                    } else {
                        /* Ghost... Iterate again. */
                    }
                }
            }
        }

        /* volatile-random and allkeys-random policy */
        else if (server.maxmemory_policy == MAXMEMORY_ALLKEYS_RANDOM ||
                 server.maxmemory_policy == MAXMEMORY_VOLATILE_RANDOM)
        {
            /* When evicting a random key, we try to evict a key for
             * each DB, so we use the static 'next_db' variable to
             * incrementally visit all DBs. */
            for (i = 0; i < server.dbnum; i++) {
                j = (++next_db) % server.dbnum;
                db = server.db+j;
                kvstore *kvs;
                if (server.maxmemory_policy == MAXMEMORY_ALLKEYS_RANDOM) {
                    kvs = db->keys;
                } else {
                    kvs = db->expires;
                }
                int slot = kvstoreGetFairRandomDictIndex(kvs, randomEvictionShouldSkipDictIndex, 16, 0);
                if (slot == -1) continue;
                de = kvstoreDictGetRandomKey(kvs, slot);
                if (de) {
                    kvobj *kv = dictGetKV(de);
                    bestkey = kvobjGetKey(kv);
                    bestdbid = j;
                    break;
                }
            }
        }

        /* Finally remove the selected key. */
        if (bestkey) {
            long long key_mem_freed;
            db = server.db+bestdbid;

            enterExecutionUnit(1, 0);
            robj *keyobj = createStringObject(bestkey,sdslen(bestkey));
            deleteEvictedKeyAndPropagate(db, keyobj, &key_mem_freed);
            decrRefCount(keyobj);
            exitExecutionUnit();
            /* Propagate the DEL command */
            postExecutionUnitOperations();

            mem_freed += key_mem_freed;
            keys_freed++;

            if (keys_freed % 16 == 0) {
                /* When the memory to free starts to be big enough, we may
                 * start spending so much time here that is impossible to
                 * deliver data to the replicas fast enough, so we force the
                 * transmission here inside the loop. */
                if (slaves) flushSlavesOutputBuffers();

                /* Normally our stop condition is the ability to release
                 * a fixed, pre-computed amount of memory. However when we
                 * are deleting objects in another thread, it's better to
                 * check, from time to time, if we already reached our target
                 * memory, since the "mem_freed" amount is computed only
                 * across the dbAsyncDelete() call, while the thread can
                 * release the memory all the time. */
                if (server.lazyfree_lazy_eviction) {
                    if (getMaxmemoryState(NULL,NULL,NULL,NULL) == C_OK) {
                        break;
                    }
                }

                /* After some time, exit the loop early - even if memory limit
                 * hasn't been reached.  If we suddenly need to free a lot of
                 * memory, don't want to spend too much time here.  */
                if (elapsedUs(evictionTimer) > eviction_time_limit_us) {
                    // We still need to free memory - start eviction timer proc
                    startEvictionTimeProc();
                    break;
                }
            }
        } else {
            goto cant_free; /* nothing to free... */
        }
    }
    /* at this point, the memory is OK, or we have reached the time limit */
    result = (isEvictionProcRunning) ? EVICT_RUNNING : EVICT_OK;

cant_free:
    if (result == EVICT_FAIL) {
        /* At this point, we have run out of evictable items.  It's possible
         * that some items are being freed in the lazyfree thread.  Perform a
         * short wait here if such jobs exist, but don't wait long.  */
        mstime_t lazyfree_latency;
        latencyStartMonitor(lazyfree_latency);
        while (bioPendingJobsOfType(BIO_LAZY_FREE) &&
              elapsedUs(evictionTimer) < eviction_time_limit_us) {
            if (getMaxmemoryState(NULL,NULL,NULL,NULL) == C_OK) {
                result = EVICT_OK;
                break;
            }
            usleep(eviction_time_limit_us < 1000 ? eviction_time_limit_us : 1000);
        }
        latencyEndMonitor(lazyfree_latency);
        latencyAddSampleIfNeeded("eviction-lazyfree",lazyfree_latency);
    }

    latencyEndMonitor(latency);
    latencyAddSampleIfNeeded("eviction-cycle",latency);

update_metrics:
    if (result == EVICT_RUNNING || result == EVICT_FAIL) {
        if (server.stat_last_eviction_exceeded_time == 0)
            elapsedStart(&server.stat_last_eviction_exceeded_time);
    } else if (result == EVICT_OK) {
        if (server.stat_last_eviction_exceeded_time != 0) {
            server.stat_total_eviction_exceeded_time += elapsedUs(server.stat_last_eviction_exceeded_time);
            server.stat_last_eviction_exceeded_time = 0;
        }
    }
    return result;
}