1 #include "database-builder.h"
4 #include "turbopfor-encode.h"
10 #include <string_view>
13 #include <sys/types.h>
18 #define P4NENC_BOUND(n) ((n + 127) / 128 + (n + 32) * sizeof(uint32_t))
20 #define NUM_TRIGRAMS 16777216
23 using namespace std::chrono;
25 constexpr unsigned num_overflow_slots = 16;
27 string zstd_compress(const string &src, ZSTD_CDict *cdict, string *tempbuf);
29 static inline uint32_t read_unigram(const string_view s, size_t idx)
32 return (unsigned char)s[idx];
38 static inline uint32_t read_trigram(const string_view s, size_t start)
40 return read_unigram(s, start) |
41 (read_unigram(s, start + 1) << 8) |
42 (read_unigram(s, start + 2) << 16);
45 class PostingListBuilder {
47 inline void add_docid(uint32_t docid);
48 inline void add_first_docid(uint32_t docid);
52 size_t num_docids = 0;
55 void write_header(uint32_t docid);
58 vector<uint32_t> pending_deltas;
60 uint32_t last_docid = -1;
63 void PostingListBuilder::add_docid(uint32_t docid)
65 // Deduplicate against the last inserted value, if any.
66 if (docid == last_docid) {
70 assert(num_docids != 0);
72 pending_deltas.push_back(docid - last_docid - 1);
74 if (pending_deltas.size() == 128) {
76 pending_deltas.clear();
81 void PostingListBuilder::add_first_docid(uint32_t docid)
88 void PostingListBuilder::finish()
90 if (pending_deltas.empty()) {
94 assert(!encoded.empty()); // write_header() should already have run.
96 // No interleaving for partial blocks.
97 unsigned char buf[P4NENC_BOUND(128)];
98 unsigned char *end = encode_pfor_single_block<128>(pending_deltas.data(), pending_deltas.size(), /*interleaved=*/false, buf);
99 encoded.append(reinterpret_cast<char *>(buf), reinterpret_cast<char *>(end));
102 void PostingListBuilder::append_block()
104 unsigned char buf[P4NENC_BOUND(128)];
105 assert(pending_deltas.size() == 128);
106 unsigned char *end = encode_pfor_single_block<128>(pending_deltas.data(), 128, /*interleaved=*/true, buf);
107 encoded.append(reinterpret_cast<char *>(buf), reinterpret_cast<char *>(end));
110 void PostingListBuilder::write_header(uint32_t docid)
112 unsigned char buf[P4NENC_BOUND(1)];
113 unsigned char *end = write_baseval(docid, buf);
114 encoded.append(reinterpret_cast<char *>(buf), end - buf);
117 void DictionaryBuilder::add_file(string filename, dir_time)
119 if (keep_current_block) { // Only bother saving the filenames if we're actually keeping the block.
120 if (!current_block.empty()) {
121 current_block.push_back('\0');
123 current_block += filename;
125 if (++num_files_in_block == block_size) {
130 void DictionaryBuilder::flush_block()
132 if (keep_current_block) {
133 if (slot_for_current_block == -1) {
134 lengths.push_back(current_block.size());
135 sampled_blocks.push_back(move(current_block));
137 lengths[slot_for_current_block] = current_block.size();
138 sampled_blocks[slot_for_current_block] = move(current_block);
141 current_block.clear();
142 num_files_in_block = 0;
145 if (block_num < blocks_to_keep) {
146 keep_current_block = true;
147 slot_for_current_block = -1;
149 // Keep every block with equal probability (reservoir sampling).
150 uint64_t idx = uniform_int_distribution<uint64_t>(0, block_num)(reservoir_rand);
151 keep_current_block = (idx < blocks_to_keep);
152 slot_for_current_block = idx;
156 string DictionaryBuilder::train(size_t buf_size)
158 string dictionary_buf;
159 sort(sampled_blocks.begin(), sampled_blocks.end()); // Seemingly important for decompression speed.
160 for (const string &block : sampled_blocks) {
161 dictionary_buf += block;
165 buf.resize(buf_size);
166 size_t ret = ZDICT_trainFromBuffer(&buf[0], buf_size, dictionary_buf.data(), lengths.data(), lengths.size());
167 if (ZDICT_isError(ret)) {
170 dprintf("Sampled %zu bytes in %zu blocks, built a dictionary of size %zu\n", dictionary_buf.size(), lengths.size(), ret);
173 sampled_blocks.clear();
179 class EncodingCorpus : public DatabaseReceiver {
181 EncodingCorpus(FILE *outfp, size_t block_size, ZSTD_CDict *cdict, bool store_dir_times);
184 void add_file(std::string filename, dir_time dt) override;
185 void flush_block() override;
186 void finish() override;
188 std::vector<uint64_t> filename_blocks;
189 size_t num_files = 0, num_files_in_block = 0, num_blocks = 0;
190 bool seen_trigram(uint32_t trgm)
192 return invindex[trgm] != nullptr;
194 size_t num_files_seen() const override { return num_files; }
195 PostingListBuilder &get_pl_builder(uint32_t trgm)
197 return *invindex[trgm];
200 void add_docid(uint32_t trgm, uint32_t docid)
202 if (invindex[trgm] == nullptr) {
203 invindex[trgm] = new PostingListBuilder;
204 invindex[trgm]->add_first_docid(docid);
206 invindex[trgm]->add_docid(docid);
210 size_t num_trigrams() const;
211 std::string get_compressed_dir_times();
214 void compress_dir_times(size_t allowed_slop);
216 std::unique_ptr<PostingListBuilder *[]> invindex;
218 std::string current_block;
220 const size_t block_size;
221 const bool store_dir_times;
224 ZSTD_CStream *dir_time_ctx = nullptr;
225 std::string dir_times; // Buffer of still-uncompressed data.
226 std::string dir_times_compressed;
230 EncodingCorpus::EncodingCorpus(FILE *outfp, size_t block_size, ZSTD_CDict *cdict, bool store_dir_times)
231 : invindex(new PostingListBuilder *[NUM_TRIGRAMS]), outfp(outfp), block_size(block_size), store_dir_times(store_dir_times), cdict(cdict)
233 fill(invindex.get(), invindex.get() + NUM_TRIGRAMS, nullptr);
234 if (store_dir_times) {
235 dir_time_ctx = ZSTD_createCStream();
236 ZSTD_initCStream(dir_time_ctx, /*level=*/6);
240 EncodingCorpus::~EncodingCorpus()
242 for (unsigned i = 0; i < NUM_TRIGRAMS; ++i) {
247 void EncodingCorpus::add_file(string filename, dir_time dt)
250 if (!current_block.empty()) {
251 current_block.push_back('\0');
253 current_block += filename;
254 if (++num_files_in_block == block_size) {
258 if (store_dir_times) {
261 dir_times.push_back('\0');
263 dir_times.push_back('\1');
264 dir_times.append(reinterpret_cast<char *>(&dt.sec), sizeof(dt.sec));
265 dir_times.append(reinterpret_cast<char *>(&dt.nsec), sizeof(dt.nsec));
267 compress_dir_times(/*allowed_slop=*/4096);
271 void EncodingCorpus::compress_dir_times(size_t allowed_slop)
273 while (dir_times.size() >= allowed_slop) {
274 size_t old_size = dir_times_compressed.size();
275 dir_times_compressed.resize(old_size + 4096);
277 ZSTD_outBuffer outbuf;
278 outbuf.dst = dir_times_compressed.data() + old_size;
283 inbuf.src = dir_times.data();
284 inbuf.size = dir_times.size();
287 int ret = ZSTD_compressStream(dir_time_ctx, &outbuf, &inbuf);
289 fprintf(stderr, "ZSTD_compressStream() failed\n");
293 dir_times_compressed.resize(old_size + outbuf.pos);
294 dir_times.erase(dir_times.begin(), dir_times.begin() + inbuf.pos);
296 if (outbuf.pos == 0 && inbuf.pos == 0) {
297 // Nothing happened (not enough data?), try again later.
303 void EncodingCorpus::flush_block()
305 if (current_block.empty()) {
309 uint32_t docid = num_blocks;
312 const char *ptr = current_block.c_str();
313 while (ptr < current_block.c_str() + current_block.size()) {
316 for (size_t j = 0; j < s.size() - 2; ++j) {
317 uint32_t trgm = read_trigram(s, j);
318 add_docid(trgm, docid);
324 // Compress and add the filename block.
325 filename_blocks.push_back(ftell(outfp));
326 string compressed = zstd_compress(current_block, cdict, &tempbuf);
327 if (fwrite(compressed.data(), compressed.size(), 1, outfp) != 1) {
332 current_block.clear();
333 num_files_in_block = 0;
337 void EncodingCorpus::finish()
342 size_t EncodingCorpus::num_trigrams() const
345 for (unsigned trgm = 0; trgm < NUM_TRIGRAMS; ++trgm) {
346 if (invindex[trgm] != nullptr) {
353 string EncodingCorpus::get_compressed_dir_times()
355 if (!store_dir_times) {
358 compress_dir_times(/*allowed_slop=*/0);
359 assert(dir_times.empty());
362 size_t old_size = dir_times_compressed.size();
363 dir_times_compressed.resize(old_size + 4096);
365 ZSTD_outBuffer outbuf;
366 outbuf.dst = dir_times_compressed.data() + old_size;
370 int ret = ZSTD_endStream(dir_time_ctx, &outbuf);
372 fprintf(stderr, "ZSTD_compressStream() failed\n");
376 dir_times_compressed.resize(old_size + outbuf.pos);
384 return dir_times_compressed;
387 string zstd_compress(const string &src, ZSTD_CDict *cdict, string *tempbuf)
389 static ZSTD_CCtx *ctx = nullptr;
390 if (ctx == nullptr) {
391 ctx = ZSTD_createCCtx();
394 size_t max_size = ZSTD_compressBound(src.size());
395 if (tempbuf->size() < max_size) {
396 tempbuf->resize(max_size);
399 if (cdict == nullptr) {
400 size = ZSTD_compressCCtx(ctx, &(*tempbuf)[0], max_size, src.data(), src.size(), /*level=*/6);
402 size = ZSTD_compress_usingCDict(ctx, &(*tempbuf)[0], max_size, src.data(), src.size(), cdict);
404 return string(tempbuf->data(), size);
407 bool is_prime(uint32_t x)
409 if ((x % 2) == 0 || (x % 3) == 0) {
412 uint32_t limit = ceil(sqrt(x));
413 for (uint32_t factor = 5; factor <= limit; ++factor) {
414 if ((x % factor) == 0) {
421 uint32_t next_prime(uint32_t x)
426 while (!is_prime(x)) {
432 unique_ptr<Trigram[]> create_hashtable(EncodingCorpus &corpus, const vector<uint32_t> &all_trigrams, uint32_t ht_size, uint32_t num_overflow_slots)
434 unique_ptr<Trigram[]> ht(new Trigram[ht_size + num_overflow_slots + 1]); // 1 for the sentinel element at the end.
435 for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
436 ht[i].trgm = uint32_t(-1);
437 ht[i].num_docids = 0;
440 for (uint32_t trgm : all_trigrams) {
441 // We don't know offset yet, so set it to zero.
442 Trigram to_insert{ trgm, uint32_t(corpus.get_pl_builder(trgm).num_docids), 0 };
444 uint32_t bucket = hash_trigram(trgm, ht_size);
445 unsigned distance = 0;
446 while (ht[bucket].num_docids != 0) {
447 // Robin Hood hashing; reduces the longest distance by a lot.
448 unsigned other_distance = bucket - hash_trigram(ht[bucket].trgm, ht_size);
449 if (distance > other_distance) {
450 swap(to_insert, ht[bucket]);
451 distance = other_distance;
454 ++bucket, ++distance;
455 if (distance > num_overflow_slots) {
459 ht[bucket] = to_insert;
464 DatabaseBuilder::DatabaseBuilder(const char *outfile, gid_t owner, int block_size, string dictionary, bool check_visibility)
465 : outfile(outfile), block_size(block_size)
469 string path = outfile;
470 path.resize(path.find_last_of('/') + 1);
475 int fd = open(path.c_str(), O_WRONLY | O_TMPFILE, 0640);
477 perror(path.c_str());
481 temp_filename = string(outfile) + ".XXXXXX";
482 int fd = mkstemp(&temp_filename[0]);
484 perror(temp_filename.c_str());
487 if (fchmod(fd, 0640) == -1) {
493 if (owner != (gid_t)-1) {
494 if (fchown(fd, (uid_t)-1, owner) == -1) {
500 outfp = fdopen(fd, "wb");
501 if (outfp == nullptr) {
507 memcpy(hdr.magic, "\0plocate", 8);
508 hdr.version = -1; // Mark as broken.
509 hdr.hashtable_size = 0; // Not known yet.
510 hdr.extra_ht_slots = num_overflow_slots;
512 hdr.hash_table_offset_bytes = -1; // We don't know these offsets yet.
514 hdr.filename_index_offset_bytes = -1;
515 hdr.zstd_dictionary_length_bytes = -1;
516 hdr.check_visibility = check_visibility;
517 fwrite(&hdr, sizeof(hdr), 1, outfp);
519 if (dictionary.empty()) {
520 hdr.zstd_dictionary_offset_bytes = 0;
521 hdr.zstd_dictionary_length_bytes = 0;
523 hdr.zstd_dictionary_offset_bytes = ftell(outfp);
524 fwrite(dictionary.data(), dictionary.size(), 1, outfp);
525 hdr.zstd_dictionary_length_bytes = dictionary.size();
526 cdict = ZSTD_createCDict(dictionary.data(), dictionary.size(), /*level=*/6);
529 hdr.directory_data_length_bytes = 0;
530 hdr.directory_data_offset_bytes = 0;
531 hdr.next_zstd_dictionary_length_bytes = 0;
532 hdr.next_zstd_dictionary_offset_bytes = 0;
533 hdr.conf_block_length_bytes = 0;
534 hdr.conf_block_offset_bytes = 0;
537 DatabaseReceiver *DatabaseBuilder::start_corpus(bool store_dir_times)
539 corpus_start = steady_clock::now();
540 corpus = new EncodingCorpus(outfp, block_size, cdict, store_dir_times);
544 void DatabaseBuilder::set_next_dictionary(std::string next_dictionary)
546 this->next_dictionary = move(next_dictionary);
549 void DatabaseBuilder::set_conf_block(std::string conf_block)
551 this->conf_block = move(conf_block);
554 void DatabaseBuilder::finish_corpus()
557 hdr.num_docids = corpus->filename_blocks.size();
559 // Stick an empty block at the end as sentinel.
560 corpus->filename_blocks.push_back(ftell(outfp));
561 const size_t bytes_for_filenames = corpus->filename_blocks.back() - corpus->filename_blocks.front();
563 // Write the offsets to the filenames.
564 hdr.filename_index_offset_bytes = ftell(outfp);
565 const size_t bytes_for_filename_index = corpus->filename_blocks.size() * sizeof(uint64_t);
566 fwrite(corpus->filename_blocks.data(), corpus->filename_blocks.size(), sizeof(uint64_t), outfp);
567 corpus->filename_blocks.clear();
568 corpus->filename_blocks.shrink_to_fit();
570 // Finish up encoding the posting lists.
571 size_t trigrams = 0, longest_posting_list = 0;
572 size_t bytes_for_posting_lists = 0;
573 for (unsigned trgm = 0; trgm < NUM_TRIGRAMS; ++trgm) {
574 if (!corpus->seen_trigram(trgm))
576 PostingListBuilder &pl_builder = corpus->get_pl_builder(trgm);
578 longest_posting_list = max(longest_posting_list, pl_builder.num_docids);
579 trigrams += pl_builder.num_docids;
580 bytes_for_posting_lists += pl_builder.encoded.size();
582 size_t num_trigrams = corpus->num_trigrams();
583 dprintf("%zu files, %zu different trigrams, %zu entries, avg len %.2f, longest %zu\n",
584 corpus->num_files, num_trigrams, trigrams, double(trigrams) / num_trigrams, longest_posting_list);
585 dprintf("%zu bytes used for posting lists (%.2f bits/entry)\n", bytes_for_posting_lists, 8 * bytes_for_posting_lists / double(trigrams));
587 dprintf("Building posting lists took %.1f ms.\n\n", 1e3 * duration<float>(steady_clock::now() - corpus_start).count());
589 // Find the used trigrams.
590 vector<uint32_t> all_trigrams;
591 for (unsigned trgm = 0; trgm < NUM_TRIGRAMS; ++trgm) {
592 if (corpus->seen_trigram(trgm)) {
593 all_trigrams.push_back(trgm);
597 // Create the hash table.
598 unique_ptr<Trigram[]> hashtable;
599 uint32_t ht_size = next_prime(all_trigrams.size());
601 hashtable = create_hashtable(*corpus, all_trigrams, ht_size, num_overflow_slots);
602 if (hashtable == nullptr) {
603 dprintf("Failed creating hash table of size %u, increasing by 5%% and trying again.\n", ht_size);
604 ht_size = next_prime(ht_size * 1.05);
606 dprintf("Created hash table of size %u.\n\n", ht_size);
611 // Find the offsets for each posting list.
612 size_t bytes_for_hashtable = (ht_size + num_overflow_slots + 1) * sizeof(Trigram);
613 uint64_t offset = ftell(outfp) + bytes_for_hashtable;
614 for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
615 hashtable[i].offset = offset; // Needs to be there even for empty slots.
616 if (hashtable[i].num_docids == 0) {
620 const string &encoded = corpus->get_pl_builder(hashtable[i].trgm).encoded;
621 offset += encoded.size();
624 // Write the hash table.
625 hdr.hash_table_offset_bytes = ftell(outfp);
626 hdr.hashtable_size = ht_size;
627 fwrite(hashtable.get(), ht_size + num_overflow_slots + 1, sizeof(Trigram), outfp);
629 // Write the actual posting lists.
630 for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
631 if (hashtable[i].num_docids == 0) {
634 const string &encoded = corpus->get_pl_builder(hashtable[i].trgm).encoded;
635 fwrite(encoded.data(), encoded.size(), 1, outfp);
638 // Finally, write the directory times (for updatedb).
639 string compressed_dir_times = corpus->get_compressed_dir_times();
640 size_t bytes_for_compressed_dir_times = 0;
641 if (!compressed_dir_times.empty()) {
642 hdr.directory_data_offset_bytes = ftell(outfp);
643 hdr.directory_data_length_bytes = compressed_dir_times.size();
644 fwrite(compressed_dir_times.data(), compressed_dir_times.size(), 1, outfp);
645 bytes_for_compressed_dir_times = compressed_dir_times.size();
646 compressed_dir_times.clear();
649 // Write the recommended dictionary for next update.
650 if (!next_dictionary.empty()) {
651 hdr.next_zstd_dictionary_offset_bytes = ftell(outfp);
652 hdr.next_zstd_dictionary_length_bytes = next_dictionary.size();
653 fwrite(next_dictionary.data(), next_dictionary.size(), 1, outfp);
656 // And the configuration block.
657 if (!conf_block.empty()) {
658 hdr.conf_block_offset_bytes = ftell(outfp);
659 hdr.conf_block_length_bytes = conf_block.size();
660 fwrite(conf_block.data(), conf_block.size(), 1, outfp);
663 // Rewind, and write the updated header.
665 fseek(outfp, 0, SEEK_SET);
666 fwrite(&hdr, sizeof(hdr), 1, outfp);
669 // Give the file a proper name, making it visible in the file system.
670 // TODO: It would be nice to be able to do this atomically, like with rename.
671 unlink(outfile.c_str());
673 snprintf(procpath, sizeof(procpath), "/proc/self/fd/%d", fileno(outfp));
674 if (linkat(AT_FDCWD, procpath, AT_FDCWD, outfile.c_str(), AT_SYMLINK_FOLLOW) == -1) {
679 if (rename(temp_filename.c_str(), outfile.c_str()) == -1) {
687 size_t total_bytes = (bytes_for_hashtable + bytes_for_posting_lists + bytes_for_filename_index + bytes_for_filenames + bytes_for_compressed_dir_times);
689 dprintf("Block size: %7d files\n", block_size);
690 dprintf("Dictionary: %'7.1f MB\n", hdr.zstd_dictionary_length_bytes / 1048576.0);
691 dprintf("Hash table: %'7.1f MB\n", bytes_for_hashtable / 1048576.0);
692 dprintf("Posting lists: %'7.1f MB\n", bytes_for_posting_lists / 1048576.0);
693 dprintf("Filename index: %'7.1f MB\n", bytes_for_filename_index / 1048576.0);
694 dprintf("Filenames: %'7.1f MB\n", bytes_for_filenames / 1048576.0);
695 if (bytes_for_compressed_dir_times != 0) {
696 dprintf("Modify times: %'7.1f MB\n", bytes_for_compressed_dir_times / 1048576.0);
698 dprintf("Total: %'7.1f MB\n", total_bytes / 1048576.0);