redict/src/sort.c

/* SORT command and helper functions.
 *
 * Copyright (c) 2009-2012, Salvatore Sanfilippo <antirez at gmail dot com>
 * All rights reserved.
 *
 * Redistribution and use in source and binary forms, with or without
 * modification, are permitted provided that the following conditions are met:
 *
 *   * Redistributions of source code must retain the above copyright notice,
 *     this list of conditions and the following disclaimer.
 *   * Redistributions in binary form must reproduce the above copyright
 *     notice, this list of conditions and the following disclaimer in the
 *     documentation and/or other materials provided with the distribution.
 *   * Neither the name of Redis nor the names of its contributors may be used
 *     to endorse or promote products derived from this software without
 *     specific prior written permission.
 *
 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
 * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
 * ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
 * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
 * POSSIBILITY OF SUCH DAMAGE.
 */


#include "server.h"
#include "pqsort.h" /* Partial qsort for SORT+LIMIT */
#include <math.h> /* isnan() */

zskiplistNode* zslGetElementByRank(zskiplist *zsl, unsigned long rank);

redisSortOperation *createSortOperation(int type, robj *pattern) {
    redisSortOperation *so = zmalloc(sizeof(*so));
    so->type = type;
    so->pattern = pattern;
    return so;
}

/* Return the value associated to the key with a name obtained using
 * the following rules:
 *
 * 1) The first occurrence of '*' in 'pattern' is substituted with 'subst'.
 *
 * 2) If 'pattern' matches the "->" string, everything on the left of
 *    the arrow is treated as the name of a hash field, and the part on the
 *    left as the key name containing a hash. The value of the specified
 *    field is returned.
 *
 * 3) If 'pattern' equals "#", the function simply returns 'subst' itself so
 *    that the SORT command can be used like: SORT key GET # to retrieve
 *    the Set/List elements directly.
 *
 * The returned object will always have its refcount increased by 1
 * when it is non-NULL. */
robj *lookupKeyByPattern(redisDb *db, robj *pattern, robj *subst, int writeflag) {
    char *p, *f, *k;
    sds spat, ssub;
    robj *keyobj, *fieldobj = NULL, *o;
    int prefixlen, sublen, postfixlen, fieldlen;

    /* If the pattern is "#" return the substitution object itself in order
     * to implement the "SORT ... GET #" feature. */
    spat = pattern->ptr;
    if (spat[0] == '#' && spat[1] == '\0') {
        incrRefCount(subst);
        return subst;
    }

    /* The substitution object may be specially encoded. If so we create
     * a decoded object on the fly. Otherwise getDecodedObject will just
     * increment the ref count, that we'll decrement later. */
    subst = getDecodedObject(subst);
    ssub = subst->ptr;

    /* If we can't find '*' in the pattern we return NULL as to GET a
     * fixed key does not make sense. */
    p = strchr(spat,'*');
    if (!p) {
        decrRefCount(subst);
        return NULL;
    }

    /* Find out if we're dealing with a hash dereference. */
    if ((f = strstr(p+1, "->")) != NULL && *(f+2) != '\0') {
        fieldlen = sdslen(spat)-(f-spat)-2;
        fieldobj = createStringObject(f+2,fieldlen);
    } else {
        fieldlen = 0;
    }

    /* Perform the '*' substitution. */
    prefixlen = p-spat;
    sublen = sdslen(ssub);
    postfixlen = sdslen(spat)-(prefixlen+1)-(fieldlen ? fieldlen+2 : 0);
    keyobj = createStringObject(NULL,prefixlen+sublen+postfixlen);
    k = keyobj->ptr;
    memcpy(k,spat,prefixlen);
    memcpy(k+prefixlen,ssub,sublen);
    memcpy(k+prefixlen+sublen,p+1,postfixlen);
    decrRefCount(subst); /* Incremented by decodeObject() */

    /* Lookup substituted key */
    if (!writeflag)
        o = lookupKeyRead(db,keyobj);
    else
        o = lookupKeyWrite(db,keyobj);
    if (o == NULL) goto noobj;

    if (fieldobj) {
        if (o->type != OBJ_HASH) goto noobj;

        /* Retrieve value from hash by the field name. The returned object
         * is a new object with refcount already incremented. */
        o = hashTypeGetValueObject(o, fieldobj->ptr);
    } else {
        if (o->type != OBJ_STRING) goto noobj;

        /* Every object that this function returns needs to have its refcount
         * increased. sortCommand decreases it again. */
        incrRefCount(o);
    }
    decrRefCount(keyobj);
    if (fieldobj) decrRefCount(fieldobj);
    return o;

noobj:
    decrRefCount(keyobj);
    if (fieldlen) decrRefCount(fieldobj);
    return NULL;
}

/* sortCompare() is used by qsort in sortCommand(). Given that qsort_r with
 * the additional parameter is not standard but a BSD-specific we have to
 * pass sorting parameters via the global 'server' structure */
int sortCompare(const void *s1, const void *s2) {
    const redisSortObject *so1 = s1, *so2 = s2;
    int cmp;

    if (!server.sort_alpha) {
        /* Numeric sorting. Here it's trivial as we precomputed scores */
        if (so1->u.score > so2->u.score) {
            cmp = 1;
        } else if (so1->u.score < so2->u.score) {
            cmp = -1;
        } else {
            /* Objects have the same score, but we don't want the comparison
             * to be undefined, so we compare objects lexicographically.
             * This way the result of SORT is deterministic. */
            cmp = compareStringObjects(so1->obj,so2->obj);
        }
    } else {
        /* Alphanumeric sorting */
        if (server.sort_bypattern) {
            if (!so1->u.cmpobj || !so2->u.cmpobj) {
                /* At least one compare object is NULL */
                if (so1->u.cmpobj == so2->u.cmpobj)
                    cmp = 0;
                else if (so1->u.cmpobj == NULL)
                    cmp = -1;
                else
                    cmp = 1;
            } else {
                /* We have both the objects, compare them. */
                if (server.sort_store) {
                    cmp = compareStringObjects(so1->u.cmpobj,so2->u.cmpobj);
                } else {
                    /* Here we can use strcoll() directly as we are sure that
                     * the objects are decoded string objects. */
                    cmp = strcoll(so1->u.cmpobj->ptr,so2->u.cmpobj->ptr);
                }
            }
        } else {
            /* Compare elements directly. */
            if (server.sort_store) {
                cmp = compareStringObjects(so1->obj,so2->obj);
            } else {
                cmp = collateStringObjects(so1->obj,so2->obj);
            }
        }
    }
    return server.sort_desc ? -cmp : cmp;
}

/* The SORT command is the most complex command in Redis. Warning: this code
 * is optimized for speed and a bit less for readability */
void sortCommand(client *c) {
    list *operations;
    unsigned int outputlen = 0;
    int desc = 0, alpha = 0;
    long limit_start = 0, limit_count = -1, start, end;
    int j, dontsort = 0, vectorlen;
    int getop = 0; /* GET operation counter */
    int int_conversion_error = 0;
    int syntax_error = 0;
    robj *sortval, *sortby = NULL, *storekey = NULL;
    redisSortObject *vector; /* Resulting vector to sort */

    /* Create a list of operations to perform for every sorted element.
     * Operations can be GET */
    operations = listCreate();
    listSetFreeMethod(operations,zfree);
    j = 2; /* options start at argv[2] */

    /* The SORT command has an SQL-alike syntax, parse it */
    while(j < c->argc) {
        int leftargs = c->argc-j-1;
        if (!strcasecmp(c->argv[j]->ptr,"asc")) {
            desc = 0;
        } else if (!strcasecmp(c->argv[j]->ptr,"desc")) {
            desc = 1;
        } else if (!strcasecmp(c->argv[j]->ptr,"alpha")) {
            alpha = 1;
        } else if (!strcasecmp(c->argv[j]->ptr,"limit") && leftargs >= 2) {
            if ((getLongFromObjectOrReply(c, c->argv[j+1], &limit_start, NULL)
                 != C_OK) ||
                (getLongFromObjectOrReply(c, c->argv[j+2], &limit_count, NULL)
                 != C_OK))
            {
                syntax_error++;
                break;
            }
            j+=2;
        } else if (!strcasecmp(c->argv[j]->ptr,"store") && leftargs >= 1) {
            storekey = c->argv[j+1];
            j++;
        } else if (!strcasecmp(c->argv[j]->ptr,"by") && leftargs >= 1) {
            sortby = c->argv[j+1];
            /* If the BY pattern does not contain '*', i.e. it is constant,
             * we don't need to sort nor to lookup the weight keys. */
            if (strchr(c->argv[j+1]->ptr,'*') == NULL) {
                dontsort = 1;
            } else {
                /* If BY is specified with a real patter, we can't accept
                 * it in cluster mode. */
                if (server.cluster_enabled) {
                    addReplyError(c,"BY option of SORT denied in Cluster mode.");
                    syntax_error++;
                    break;
                }
            }
            j++;
        } else if (!strcasecmp(c->argv[j]->ptr,"get") && leftargs >= 1) {
            if (server.cluster_enabled) {
                addReplyError(c,"GET option of SORT denied in Cluster mode.");
                syntax_error++;
                break;
            }
            listAddNodeTail(operations,createSortOperation(
                SORT_OP_GET,c->argv[j+1]));
            getop++;
            j++;
        } else {
            addReplyErrorObject(c,shared.syntaxerr);
            syntax_error++;
            break;
        }
        j++;
    }

    /* Handle syntax errors set during options parsing. */
    if (syntax_error) {
        listRelease(operations);
        return;
    }

    /* Lookup the key to sort. It must be of the right types */
    if (!storekey)
        sortval = lookupKeyRead(c->db,c->argv[1]);
    else
        sortval = lookupKeyWrite(c->db,c->argv[1]);
    if (sortval && sortval->type != OBJ_SET &&
                   sortval->type != OBJ_LIST &&
                   sortval->type != OBJ_ZSET)
    {
        listRelease(operations);
        addReplyErrorObject(c,shared.wrongtypeerr);
        return;
    }

    /* Now we need to protect sortval incrementing its count, in the future
     * SORT may have options able to overwrite/delete keys during the sorting
     * and the sorted key itself may get destroyed */
    if (sortval)
        incrRefCount(sortval);
    else
        sortval = createQuicklistObject();


    /* When sorting a set with no sort specified, we must sort the output
     * so the result is consistent across scripting and replication.
     *
     * The other types (list, sorted set) will retain their native order
     * even if no sort order is requested, so they remain stable across
     * scripting and replication. */
    if (dontsort &&
        sortval->type == OBJ_SET &&
        (storekey || c->flags & CLIENT_LUA))
    {
        /* Force ALPHA sorting */
        dontsort = 0;
        alpha = 1;
        sortby = NULL;
    }

    /* Destructively convert encoded sorted sets for SORT. */
    if (sortval->type == OBJ_ZSET)
        zsetConvert(sortval, OBJ_ENCODING_SKIPLIST);

    /* Obtain the length of the object to sort. */
    switch(sortval->type) {
    case OBJ_LIST: vectorlen = listTypeLength(sortval); break;
    case OBJ_SET: vectorlen =  setTypeSize(sortval); break;
    case OBJ_ZSET: vectorlen = dictSize(((zset*)sortval->ptr)->dict); break;
    default: vectorlen = 0; serverPanic("Bad SORT type"); /* Avoid GCC warning */
    }

    /* Perform LIMIT start,count sanity checking. */
    start = (limit_start < 0) ? 0 : limit_start;
    end = (limit_count < 0) ? vectorlen-1 : start+limit_count-1;
    if (start >= vectorlen) {
        start = vectorlen-1;
        end = vectorlen-2;
    }
    if (end >= vectorlen) end = vectorlen-1;

    /* Whenever possible, we load elements into the output array in a more
     * direct way. This is possible if:
     *
     * 1) The object to sort is a sorted set or a list (internally sorted).
     * 2) There is nothing to sort as dontsort is true (BY <constant string>).
     *
     * In this special case, if we have a LIMIT option that actually reduces
     * the number of elements to fetch, we also optimize to just load the
     * range we are interested in and allocating a vector that is big enough
     * for the selected range length. */
    if ((sortval->type == OBJ_ZSET || sortval->type == OBJ_LIST) &&
        dontsort &&
        (start != 0 || end != vectorlen-1))
    {
        vectorlen = end-start+1;
    }

    /* Load the sorting vector with all the objects to sort */
    vector = zmalloc(sizeof(redisSortObject)*vectorlen);
    j = 0;

    if (sortval->type == OBJ_LIST && dontsort) {
        /* Special handling for a list, if 'dontsort' is true.
         * This makes sure we return elements in the list original
         * ordering, accordingly to DESC / ASC options.
         *
         * Note that in this case we also handle LIMIT here in a direct
         * way, just getting the required range, as an optimization. */
        if (end >= start) {
            listTypeIterator *li;
            listTypeEntry entry;
            li = listTypeInitIterator(sortval,
                    desc ? (long)(listTypeLength(sortval) - start - 1) : start,
                    desc ? LIST_HEAD : LIST_TAIL);

            while(j < vectorlen && listTypeNext(li,&entry)) {
                vector[j].obj = listTypeGet(&entry);
                vector[j].u.score = 0;
                vector[j].u.cmpobj = NULL;
                j++;
            }
            listTypeReleaseIterator(li);
            /* Fix start/end: output code is not aware of this optimization. */
            end -= start;
            start = 0;
        }
    } else if (sortval->type == OBJ_LIST) {
        listTypeIterator *li = listTypeInitIterator(sortval,0,LIST_TAIL);
        listTypeEntry entry;
        while(listTypeNext(li,&entry)) {
            vector[j].obj = listTypeGet(&entry);
            vector[j].u.score = 0;
            vector[j].u.cmpobj = NULL;
            j++;
        }
        listTypeReleaseIterator(li);
    } else if (sortval->type == OBJ_SET) {
        setTypeIterator *si = setTypeInitIterator(sortval);
        sds sdsele;
        while((sdsele = setTypeNextObject(si)) != NULL) {
            vector[j].obj = createObject(OBJ_STRING,sdsele);
            vector[j].u.score = 0;
            vector[j].u.cmpobj = NULL;
            j++;
        }
        setTypeReleaseIterator(si);
    } else if (sortval->type == OBJ_ZSET && dontsort) {
        /* Special handling for a sorted set, if 'dontsort' is true.
         * This makes sure we return elements in the sorted set original
         * ordering, accordingly to DESC / ASC options.
         *
         * Note that in this case we also handle LIMIT here in a direct
         * way, just getting the required range, as an optimization. */

        zset *zs = sortval->ptr;
        zskiplist *zsl = zs->zsl;
        zskiplistNode *ln;
        sds sdsele;
        int rangelen = vectorlen;

        /* Check if starting point is trivial, before doing log(N) lookup. */
        if (desc) {
            long zsetlen = dictSize(((zset*)sortval->ptr)->dict);

            ln = zsl->tail;
            if (start > 0)
                ln = zslGetElementByRank(zsl,zsetlen-start);
        } else {
            ln = zsl->header->level[0].forward;
            if (start > 0)
                ln = zslGetElementByRank(zsl,start+1);
        }

        while(rangelen--) {
            serverAssertWithInfo(c,sortval,ln != NULL);
            sdsele = ln->ele;
            vector[j].obj = createStringObject(sdsele,sdslen(sdsele));
            vector[j].u.score = 0;
            vector[j].u.cmpobj = NULL;
            j++;
            ln = desc ? ln->backward : ln->level[0].forward;
        }
        /* Fix start/end: output code is not aware of this optimization. */
        end -= start;
        start = 0;
    } else if (sortval->type == OBJ_ZSET) {
        dict *set = ((zset*)sortval->ptr)->dict;
        dictIterator *di;
        dictEntry *setele;
        sds sdsele;
        di = dictGetIterator(set);
        while((setele = dictNext(di)) != NULL) {
            sdsele =  dictGetKey(setele);
            vector[j].obj = createStringObject(sdsele,sdslen(sdsele));
            vector[j].u.score = 0;
            vector[j].u.cmpobj = NULL;
            j++;
        }
        dictReleaseIterator(di);
    } else {
        serverPanic("Unknown type");
    }
    serverAssertWithInfo(c,sortval,j == vectorlen);

    /* Now it's time to load the right scores in the sorting vector */
    if (!dontsort) {
        for (j = 0; j < vectorlen; j++) {
            robj *byval;
            if (sortby) {
                /* lookup value to sort by */
                byval = lookupKeyByPattern(c->db,sortby,vector[j].obj,storekey!=NULL);
                if (!byval) continue;
            } else {
                /* use object itself to sort by */
                byval = vector[j].obj;
            }

            if (alpha) {
                if (sortby) vector[j].u.cmpobj = getDecodedObject(byval);
            } else {
                if (sdsEncodedObject(byval)) {
                    char *eptr;

                    vector[j].u.score = strtod(byval->ptr,&eptr);
                    if (eptr[0] != '\0' || errno == ERANGE ||
                        isnan(vector[j].u.score))
                    {
                        int_conversion_error = 1;
                    }
                } else if (byval->encoding == OBJ_ENCODING_INT) {
                    /* Don't need to decode the object if it's
                     * integer-encoded (the only encoding supported) so
                     * far. We can just cast it */
                    vector[j].u.score = (long)byval->ptr;
                } else {
                    serverAssertWithInfo(c,sortval,1 != 1);
                }
            }

            /* when the object was retrieved using lookupKeyByPattern,
             * its refcount needs to be decreased. */
            if (sortby) {
                decrRefCount(byval);
            }
        }

        server.sort_desc = desc;
        server.sort_alpha = alpha;
        server.sort_bypattern = sortby ? 1 : 0;
        server.sort_store = storekey ? 1 : 0;
        if (sortby && (start != 0 || end != vectorlen-1))
            pqsort(vector,vectorlen,sizeof(redisSortObject),sortCompare, start,end);
        else
            qsort(vector,vectorlen,sizeof(redisSortObject),sortCompare);
    }

    /* Send command output to the output buffer, performing the specified
     * GET/DEL/INCR/DECR operations if any. */
    outputlen = getop ? getop*(end-start+1) : end-start+1;
    if (int_conversion_error) {
        addReplyError(c,"One or more scores can't be converted into double");
    } else if (storekey == NULL) {
        /* STORE option not specified, sent the sorting result to client */
        addReplyArrayLen(c,outputlen);
        for (j = start; j <= end; j++) {
            listNode *ln;
            listIter li;

            if (!getop) addReplyBulk(c,vector[j].obj);
            listRewind(operations,&li);
            while((ln = listNext(&li))) {
                redisSortOperation *sop = ln->value;
                robj *val = lookupKeyByPattern(c->db,sop->pattern,
                    vector[j].obj,storekey!=NULL);

                if (sop->type == SORT_OP_GET) {
                    if (!val) {
                        addReplyNull(c);
                    } else {
                        addReplyBulk(c,val);
                        decrRefCount(val);
                    }
                } else {
                    /* Always fails */
                    serverAssertWithInfo(c,sortval,sop->type == SORT_OP_GET);
                }
            }
        }
    } else {
        robj *sobj = createQuicklistObject();

        /* STORE option specified, set the sorting result as a List object */
        for (j = start; j <= end; j++) {
            listNode *ln;
            listIter li;

            if (!getop) {
                listTypePush(sobj,vector[j].obj,LIST_TAIL);
            } else {
                listRewind(operations,&li);
                while((ln = listNext(&li))) {
                    redisSortOperation *sop = ln->value;
                    robj *val = lookupKeyByPattern(c->db,sop->pattern,
                        vector[j].obj,storekey!=NULL);

                    if (sop->type == SORT_OP_GET) {
                        if (!val) val = createStringObject("",0);

                        /* listTypePush does an incrRefCount, so we should take care
                         * care of the incremented refcount caused by either
                         * lookupKeyByPattern or createStringObject("",0) */
                        listTypePush(sobj,val,LIST_TAIL);
                        decrRefCount(val);
                    } else {
                        /* Always fails */
                        serverAssertWithInfo(c,sortval,sop->type == SORT_OP_GET);
                    }
                }
            }
        }
        if (outputlen) {
            setKey(c,c->db,storekey,sobj);
            notifyKeyspaceEvent(NOTIFY_LIST,"sortstore",storekey,
                                c->db->id);
            server.dirty += outputlen;
        } else if (dbDelete(c->db,storekey)) {
            signalModifiedKey(c,c->db,storekey);
            notifyKeyspaceEvent(NOTIFY_GENERIC,"del",storekey,c->db->id);
            server.dirty++;
        }
        decrRefCount(sobj);
        addReplyLongLong(c,outputlen);
    }

    /* Cleanup */
    for (j = 0; j < vectorlen; j++)
        decrRefCount(vector[j].obj);

    decrRefCount(sortval);
    listRelease(operations);
    for (j = 0; j < vectorlen; j++) {
        if (alpha && vector[j].u.cmpobj)
            decrRefCount(vector[j].u.cmpobj);
    }
    zfree(vector);
}