summaryrefslogtreecommitdiff
path: root/src/mongo/db/exec/and_hash.cpp
blob: dc8197afc8e98b1460435e36d92ec94f33fae0dd (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
/**
 *    Copyright (C) 2018-present MongoDB, Inc.
 *
 *    This program is free software: you can redistribute it and/or modify
 *    it under the terms of the Server Side Public License, version 1,
 *    as published by MongoDB, Inc.
 *
 *    This program is distributed in the hope that it will be useful,
 *    but WITHOUT ANY WARRANTY; without even the implied warranty of
 *    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 *    Server Side Public License for more details.
 *
 *    You should have received a copy of the Server Side Public License
 *    along with this program. If not, see
 *    <http://www.mongodb.com/licensing/server-side-public-license>.
 *
 *    As a special exception, the copyright holders give permission to link the
 *    code of portions of this program with the OpenSSL library under certain
 *    conditions as described in each individual source file and distribute
 *    linked combinations including the program with the OpenSSL library. You
 *    must comply with the Server Side Public License in all respects for
 *    all of the code used other than as permitted herein. If you modify file(s)
 *    with this exception, you may extend this exception to your version of the
 *    file(s), but you are not obligated to do so. If you do not wish to do so,
 *    delete this exception statement from your version. If you delete this
 *    exception statement from all source files in the program, then also delete
 *    it in the license file.
 */

#include "mongo/db/exec/and_hash.h"

#include "mongo/db/exec/and_common.h"
#include "mongo/db/exec/scoped_timer.h"
#include "mongo/db/exec/working_set.h"
#include "mongo/db/exec/working_set_common.h"
#include "mongo/stdx/memory.h"
#include "mongo/util/str.h"

namespace {

// Upper limit for buffered data.
// Stage execution will fail once size of all buffered data exceeds this threshold.
const size_t kDefaultMaxMemUsageBytes = 32 * 1024 * 1024;

}  // namespace

namespace mongo {

using std::unique_ptr;
using std::vector;
using stdx::make_unique;

const size_t AndHashStage::kLookAheadWorks = 10;

// static
const char* AndHashStage::kStageType = "AND_HASH";

AndHashStage::AndHashStage(OperationContext* opCtx, WorkingSet* ws)
    : PlanStage(kStageType, opCtx),
      _ws(ws),
      _hashingChildren(true),
      _currentChild(0),
      _memUsage(0),
      _maxMemUsage(kDefaultMaxMemUsageBytes) {}

AndHashStage::AndHashStage(OperationContext* opCtx, WorkingSet* ws, size_t maxMemUsage)
    : PlanStage(kStageType, opCtx),
      _ws(ws),
      _hashingChildren(true),
      _currentChild(0),
      _memUsage(0),
      _maxMemUsage(maxMemUsage) {}

void AndHashStage::addChild(PlanStage* child) {
    _children.emplace_back(child);
}

size_t AndHashStage::getMemUsage() const {
    return _memUsage;
}

bool AndHashStage::isEOF() {
    // This is empty before calling work() and not-empty after.
    if (_lookAheadResults.empty()) {
        return false;
    }

    // Either we're busy hashing children, in which case we're not done yet.
    if (_hashingChildren) {
        return false;
    }

    // Or we're streaming in results from the last child.

    // If there's nothing to probe against, we're EOF.
    if (_dataMap.empty()) {
        return true;
    }

    // Otherwise, we're done when the last child is done.
    invariant(_children.size() >= 2);
    return (WorkingSet::INVALID_ID == _lookAheadResults[_children.size() - 1]) &&
        _children[_children.size() - 1]->isEOF();
}

PlanStage::StageState AndHashStage::doWork(WorkingSetID* out) {
    if (isEOF()) {
        return PlanStage::IS_EOF;
    }

    // Fast-path for one of our children being EOF immediately.  We work each child a few times.
    // If it hits EOF, the AND cannot output anything.  If it produces a result, we stash that
    // result in _lookAheadResults.
    if (_lookAheadResults.empty()) {
        // INVALID_ID means that the child didn't produce a valid result.

        // We specifically are not using .resize(size, value) here because C++11 builds don't
        // seem to resolve WorkingSet::INVALID_ID during linking.
        _lookAheadResults.resize(_children.size());
        for (size_t i = 0; i < _children.size(); ++i) {
            _lookAheadResults[i] = WorkingSet::INVALID_ID;
        }

        // Work each child some number of times until it's either EOF or produces
        // a result.  If it's EOF this whole stage will be EOF.  If it produces a
        // result we cache it for later.
        for (size_t i = 0; i < _children.size(); ++i) {
            auto& child = _children[i];
            for (size_t j = 0; j < kLookAheadWorks; ++j) {
                // Cache the result in _lookAheadResults[i].
                StageState childStatus = child->work(&_lookAheadResults[i]);

                if (PlanStage::IS_EOF == childStatus) {
                    // A child went right to EOF.  Bail out.
                    _hashingChildren = false;
                    _dataMap.clear();
                    return PlanStage::IS_EOF;
                } else if (PlanStage::ADVANCED == childStatus) {
                    // Ensure that the BSONObj underlying the WorkingSetMember is owned in case we
                    // yield.
                    _ws->get(_lookAheadResults[i])->makeObjOwnedIfNeeded();
                    break;  // Stop looking at this child.
                } else if (PlanStage::FAILURE == childStatus) {
                    // The stage which produces a failure is responsible for allocating a working
                    // set member with error details.
                    invariant(WorkingSet::INVALID_ID != _lookAheadResults[i]);
                    *out = _lookAheadResults[i];
                    _hashingChildren = false;
                    _dataMap.clear();
                    return childStatus;
                }
                // We ignore NEED_TIME. TODO: what do we want to do if we get NEED_YIELD here?
            }
        }

        // We did a bunch of work above, return NEED_TIME to be fair.
        return PlanStage::NEED_TIME;
    }

    // An AND is either reading the first child into the hash table, probing against the hash
    // table with subsequent children, or checking the last child's results to see if they're
    // in the hash table.

    // We read the first child into our hash table.
    if (_hashingChildren) {
        // Check memory usage of previously hashed results.
        if (_memUsage > _maxMemUsage) {
            str::stream ss;
            ss << "hashed AND stage buffered data usage of " << _memUsage
               << " bytes exceeds internal limit of " << kDefaultMaxMemUsageBytes << " bytes";
            Status status(ErrorCodes::Overflow, ss);
            *out = WorkingSetCommon::allocateStatusMember(_ws, status);
            return PlanStage::FAILURE;
        }

        if (0 == _currentChild) {
            return readFirstChild(out);
        } else if (_currentChild < _children.size() - 1) {
            return hashOtherChildren(out);
        } else {
            _hashingChildren = false;
            // We don't hash our last child.  Instead, we probe the table created from the
            // previous children, returning results in the order of the last child.
            // Fall through to below.
        }
    }

    // Returning results.  We read from the last child and return the results that are in our
    // hash map.

    // We should be EOF if we're not hashing results and the dataMap is empty.
    verify(!_dataMap.empty());

    // We probe _dataMap with the last child.
    verify(_currentChild == _children.size() - 1);

    // Get the next result for the (_children.size() - 1)-th child.
    StageState childStatus = workChild(_children.size() - 1, out);
    if (PlanStage::ADVANCED != childStatus) {
        return childStatus;
    }

    // We know that we've ADVANCED.  See if the WSM is in our table.
    WorkingSetMember* member = _ws->get(*out);

    // The child must give us a WorkingSetMember with a record id, since we intersect index keys
    // based on the record id. The planner ensures that the child stage can never produce an WSM
    // with no record id.
    invariant(member->hasRecordId());

    DataMap::iterator it = _dataMap.find(member->recordId);
    if (_dataMap.end() == it) {
        // Child's output wasn't in every previous child.  Throw it out.
        _ws->free(*out);
        return PlanStage::NEED_TIME;
    } else {
        // Child's output was in every previous child.  Merge any key data in
        // the child's output and free the child's just-outputted WSM.
        WorkingSetID hashID = it->second;
        _dataMap.erase(it);

        AndCommon::mergeFrom(_ws, hashID, *member);
        _ws->free(*out);

        *out = hashID;
        return PlanStage::ADVANCED;
    }
}

PlanStage::StageState AndHashStage::workChild(size_t childNo, WorkingSetID* out) {
    if (WorkingSet::INVALID_ID != _lookAheadResults[childNo]) {
        *out = _lookAheadResults[childNo];
        _lookAheadResults[childNo] = WorkingSet::INVALID_ID;
        return PlanStage::ADVANCED;
    } else {
        return _children[childNo]->work(out);
    }
}

PlanStage::StageState AndHashStage::readFirstChild(WorkingSetID* out) {
    verify(_currentChild == 0);

    WorkingSetID id = WorkingSet::INVALID_ID;
    StageState childStatus = workChild(0, &id);

    if (PlanStage::ADVANCED == childStatus) {
        WorkingSetMember* member = _ws->get(id);

        // The child must give us a WorkingSetMember with a record id, since we intersect index keys
        // based on the record id. The planner ensures that the child stage can never produce an WSM
        // with no record id.
        invariant(member->hasRecordId());

        if (!_dataMap.insert(std::make_pair(member->recordId, id)).second) {
            // Didn't insert because we already had this RecordId inside the map. This should only
            // happen if we're seeing a newer copy of the same doc in a more recent snapshot.
            // Throw out the newer copy of the doc.
            _ws->free(id);
            return PlanStage::NEED_TIME;
        }

        // Ensure that the BSONObj underlying the WorkingSetMember is owned in case we yield.
        member->makeObjOwnedIfNeeded();

        // Update memory stats.
        _memUsage += member->getMemUsage();

        return PlanStage::NEED_TIME;
    } else if (PlanStage::IS_EOF == childStatus) {
        // Done reading child 0.
        _currentChild = 1;

        // If our first child was empty, don't scan any others, no possible results.
        if (_dataMap.empty()) {
            _hashingChildren = false;
            return PlanStage::IS_EOF;
        }

        _specificStats.mapAfterChild.push_back(_dataMap.size());

        return PlanStage::NEED_TIME;
    } else if (PlanStage::FAILURE == childStatus) {
        // The stage which produces a failure is responsible for allocating a working set member
        // with error details.
        invariant(WorkingSet::INVALID_ID != id);
        *out = id;
        return childStatus;
    } else {
        if (PlanStage::NEED_YIELD == childStatus) {
            *out = id;
        }

        return childStatus;
    }
}

PlanStage::StageState AndHashStage::hashOtherChildren(WorkingSetID* out) {
    verify(_currentChild > 0);

    WorkingSetID id = WorkingSet::INVALID_ID;
    StageState childStatus = workChild(_currentChild, &id);

    if (PlanStage::ADVANCED == childStatus) {
        WorkingSetMember* member = _ws->get(id);

        // The child must give us a WorkingSetMember with a record id, since we intersect index keys
        // based on the record id. The planner ensures that the child stage can never produce an
        // WSM with no record id.
        invariant(member->hasRecordId());

        if (_dataMap.end() == _dataMap.find(member->recordId)) {
            // Ignore.  It's not in any previous child.
        } else {
            // We have a hit.  Copy data into the WSM we already have.
            _seenMap.insert(member->recordId);
            WorkingSetID olderMemberID = _dataMap[member->recordId];
            WorkingSetMember* olderMember = _ws->get(olderMemberID);
            size_t memUsageBefore = olderMember->getMemUsage();

            AndCommon::mergeFrom(_ws, olderMemberID, *member);

            // Update memory stats.
            _memUsage += olderMember->getMemUsage() - memUsageBefore;
        }
        _ws->free(id);
        return PlanStage::NEED_TIME;
    } else if (PlanStage::IS_EOF == childStatus) {
        // Finished with a child.
        ++_currentChild;

        // Keep elements of _dataMap that are in _seenMap.
        DataMap::iterator it = _dataMap.begin();
        while (it != _dataMap.end()) {
            if (_seenMap.end() == _seenMap.find(it->first)) {
                DataMap::iterator toErase = it;
                ++it;

                // Update memory stats.
                WorkingSetMember* member = _ws->get(toErase->second);
                _memUsage -= member->getMemUsage();

                _ws->free(toErase->second);
                _dataMap.erase(toErase);
            } else {
                ++it;
            }
        }

        _specificStats.mapAfterChild.push_back(_dataMap.size());

        _seenMap.clear();

        // _dataMap is now the intersection of the first _currentChild nodes.

        // If we have nothing to AND with after finishing any child, stop.
        if (_dataMap.empty()) {
            _hashingChildren = false;
            return PlanStage::IS_EOF;
        }

        // We've finished scanning all children.  Return results with the next call to work().
        if (_currentChild == _children.size()) {
            _hashingChildren = false;
        }

        return PlanStage::NEED_TIME;
    } else if (PlanStage::FAILURE == childStatus) {
        // The stage which produces a failure is responsible for allocating a working set member
        // with error details.
        invariant(WorkingSet::INVALID_ID != id);
        *out = id;
        return childStatus;
    } else {
        if (PlanStage::NEED_YIELD == childStatus) {
            *out = id;
        }

        return childStatus;
    }
}

unique_ptr<PlanStageStats> AndHashStage::getStats() {
    _commonStats.isEOF = isEOF();

    _specificStats.memLimit = _maxMemUsage;
    _specificStats.memUsage = _memUsage;

    unique_ptr<PlanStageStats> ret = make_unique<PlanStageStats>(_commonStats, STAGE_AND_HASH);
    ret->specific = make_unique<AndHashStats>(_specificStats);
    for (size_t i = 0; i < _children.size(); ++i) {
        ret->children.emplace_back(_children[i]->getStats());
    }

    return ret;
}

const SpecificStats* AndHashStage::getSpecificStats() const {
    return &_specificStats;
}

}  // namespace mongo