1 #include "database-builder.h"
4 #include "turbopfor-encode.h"
13 #include <string_view>
16 #include <sys/types.h>
20 #include <unicode/coll.h>
21 #include <unicode/locid.h>
23 #define P4NENC_BOUND(n) ((n + 127) / 128 + (n + 32) * sizeof(uint32_t))
25 #define NUM_TRIGRAMS 16777216
28 using namespace std::chrono;
30 constexpr unsigned num_overflow_slots = 16;
32 string zstd_compress(const string &src, ZSTD_CDict *cdict, string *tempbuf);
34 class PostingListBuilder {
36 inline void add_docid(uint32_t docid);
37 inline void add_first_docid(uint32_t docid);
40 vector<unsigned char> encoded;
41 size_t get_num_docids() const
43 // Updated only when we flush, so check that we're finished.
44 assert(pending_deltas.empty());
49 void write_header(uint32_t docid);
52 vector<uint32_t> pending_deltas;
54 uint32_t num_docids = 0; // Should be size_t, except the format only supports 2^32 docids per posting list anyway.
55 uint32_t last_docid = -1;
58 void PostingListBuilder::add_docid(uint32_t docid)
60 // Deduplicate against the last inserted value, if any.
61 if (docid == last_docid) {
65 pending_deltas.push_back(docid - last_docid - 1);
67 if (pending_deltas.size() == 128) {
69 pending_deltas.clear();
74 void PostingListBuilder::add_first_docid(uint32_t docid)
81 void PostingListBuilder::finish()
83 if (pending_deltas.empty()) {
87 assert(!encoded.empty()); // write_header() should already have run.
89 // No interleaving for partial blocks.
90 unsigned char buf[P4NENC_BOUND(128)];
91 unsigned char *end = encode_pfor_single_block<128>(pending_deltas.data(), pending_deltas.size(), /*interleaved=*/false, buf);
92 encoded.insert(encoded.end(), buf, end);
94 num_docids += pending_deltas.size();
95 pending_deltas.clear();
98 void PostingListBuilder::append_block()
100 unsigned char buf[P4NENC_BOUND(128)];
101 assert(pending_deltas.size() == 128);
102 unsigned char *end = encode_pfor_single_block<128>(pending_deltas.data(), 128, /*interleaved=*/true, buf);
103 encoded.insert(encoded.end(), buf, end);
106 void PostingListBuilder::write_header(uint32_t docid)
108 unsigned char buf[P4NENC_BOUND(1)];
109 unsigned char *end = write_baseval(docid, buf);
110 encoded.insert(encoded.end(), buf, end);
113 void DictionaryBuilder::add_file(string filename, dir_time)
115 if (keep_current_block) { // Only bother saving the filenames if we're actually keeping the block.
116 if (!current_block.empty()) {
117 current_block.push_back('\0');
119 current_block += filename;
121 if (++num_files_in_block == block_size) {
126 void DictionaryBuilder::flush_block()
128 if (keep_current_block) {
129 if (slot_for_current_block == -1) {
130 lengths.push_back(current_block.size());
131 sampled_blocks.push_back(move(current_block));
133 lengths[slot_for_current_block] = current_block.size();
134 sampled_blocks[slot_for_current_block] = move(current_block);
137 current_block.clear();
138 num_files_in_block = 0;
141 if (block_num < blocks_to_keep) {
142 keep_current_block = true;
143 slot_for_current_block = -1;
145 // Keep every block with equal probability (reservoir sampling).
146 uint64_t idx = uniform_int_distribution<uint64_t>(0, block_num)(reservoir_rand);
147 keep_current_block = (idx < blocks_to_keep);
148 slot_for_current_block = idx;
152 string DictionaryBuilder::train(size_t buf_size)
154 string dictionary_buf;
155 sort(sampled_blocks.begin(), sampled_blocks.end()); // Seemingly important for decompression speed.
156 for (const string &block : sampled_blocks) {
157 dictionary_buf += block;
161 buf.resize(buf_size);
162 size_t ret = ZDICT_trainFromBuffer(&buf[0], buf_size, dictionary_buf.data(), lengths.data(), lengths.size());
163 if (ZDICT_isError(ret)) {
166 dprintf("Sampled %zu bytes in %zu blocks, built a dictionary of size %zu\n", dictionary_buf.size(), lengths.size(), ret);
169 sampled_blocks.clear();
175 class EncodingCorpus : public DatabaseReceiver {
177 EncodingCorpus(FILE *outfp, size_t block_size, ZSTD_CDict *cdict, bool store_dir_times);
180 void add_file(std::string filename, dir_time dt) override;
181 void flush_block() override;
182 void finish() override;
184 std::vector<uint64_t> filename_blocks;
185 size_t num_files = 0, num_files_in_block = 0, num_blocks = 0;
186 bool seen_trigram(uint32_t trgm)
188 return invindex[trgm] != nullptr;
190 size_t num_files_seen() const override { return num_files; }
191 PostingListBuilder &get_pl_builder(uint32_t trgm)
193 return *invindex[trgm];
196 void add_docid(uint32_t trgm, uint32_t docid)
198 if (invindex[trgm] == nullptr) {
199 invindex[trgm] = new PostingListBuilder;
200 invindex[trgm]->add_first_docid(docid);
202 invindex[trgm]->add_docid(docid);
206 size_t num_trigrams() const;
207 std::string get_compressed_dir_times();
210 void compress_dir_times(size_t allowed_slop);
212 std::unique_ptr<PostingListBuilder *[]> invindex;
214 off_t outfp_pos; // Cheaper than calling ftell(outfp) all the time.
215 std::string current_block;
217 const size_t block_size;
218 const bool store_dir_times;
221 ZSTD_CStream *dir_time_ctx = nullptr;
222 std::string dir_times; // Buffer of still-uncompressed data.
223 std::string dir_times_compressed;
226 EncodingCorpus::EncodingCorpus(FILE *outfp, size_t block_size, ZSTD_CDict *cdict, bool store_dir_times)
227 : invindex(new PostingListBuilder *[NUM_TRIGRAMS]), outfp(outfp), outfp_pos(ftell(outfp)), block_size(block_size), store_dir_times(store_dir_times), cdict(cdict)
229 fill(invindex.get(), invindex.get() + NUM_TRIGRAMS, nullptr);
230 if (store_dir_times) {
231 dir_time_ctx = ZSTD_createCStream();
232 ZSTD_initCStream(dir_time_ctx, /*level=*/6);
236 EncodingCorpus::~EncodingCorpus()
238 for (unsigned i = 0; i < NUM_TRIGRAMS; ++i) {
243 void EncodingCorpus::add_file(string filename, dir_time dt)
246 if (!current_block.empty()) {
247 current_block.push_back('\0');
249 current_block += filename;
250 if (++num_files_in_block == block_size) {
254 if (store_dir_times) {
257 dir_times.push_back('\0');
259 dir_times.push_back('\1');
260 dir_times.append(reinterpret_cast<char *>(&dt.sec), sizeof(dt.sec));
261 dir_times.append(reinterpret_cast<char *>(&dt.nsec), sizeof(dt.nsec));
263 compress_dir_times(/*allowed_slop=*/4096);
267 void EncodingCorpus::compress_dir_times(size_t allowed_slop)
269 while (dir_times.size() >= allowed_slop) {
270 size_t old_size = dir_times_compressed.size();
271 dir_times_compressed.resize(old_size + 4096);
273 ZSTD_outBuffer outbuf;
274 outbuf.dst = dir_times_compressed.data() + old_size;
279 inbuf.src = dir_times.data();
280 inbuf.size = dir_times.size();
283 int ret = ZSTD_compressStream(dir_time_ctx, &outbuf, &inbuf);
285 fprintf(stderr, "ZSTD_compressStream() failed\n");
289 dir_times_compressed.resize(old_size + outbuf.pos);
290 dir_times.erase(dir_times.begin(), dir_times.begin() + inbuf.pos);
292 if (outbuf.pos == 0 && inbuf.pos == 0) {
293 // Nothing happened (not enough data?), try again later.
299 void EncodingCorpus::flush_block()
301 if (current_block.empty()) {
305 uint32_t docid = num_blocks;
308 vector<uint8_t> sort_key;
311 const icu::Locale* locales = icu::Collator::getAvailableLocales(num_locales);
312 for (int i = 0; i < num_locales; ++i) {
313 const icu::Locale &loc = locales[i];
314 if (strcmp(loc.getName(), "en_US_POSIX") == 0) {
315 continue; // Too weird.
317 UErrorCode status = U_ZERO_ERROR;
318 icu::Collator *coll = icu::Collator::createInstance(loc, status);
319 if (U_FAILURE(status)) {
320 fprintf(stderr, "ERROR: Failed to create collator\n");
323 coll->setStrength(icu::Collator::PRIMARY);
324 const char *ptr = current_block.c_str();
325 const char *end = ptr + current_block.size();
327 size_t len = strlen(ptr);
330 sortkey_len = coll->getSortKey(icu::UnicodeString::fromUTF8(icu::StringPiece(ptr, len)), sort_key.data(), sort_key.size());
331 if (sortkey_len < sort_key.size()) { // Note <, not <=; we need to keep a slop byte.
334 sort_key.resize(sortkey_len * 3 / 2);
337 const uint8_t *keyptr = &sort_key[0];
338 const uint8_t *keyend = keyptr + sortkey_len;
339 while (keyptr < keyend - 3) {
340 // NOTE: Will read one byte past the end of the trigram, but it's OK,
341 // since we always call it from contexts where there's a terminating zero byte.
343 memcpy(&trgm, keyptr, sizeof(trgm));
345 trgm = le32toh(trgm);
346 add_docid(trgm & 0xffffff, docid);
354 const char *ptr = current_block.c_str();
355 const char *end = ptr + current_block.size();
356 while (ptr < end - 3) { // Must be at least one filename left, that's at least three bytes.
357 if (ptr[0] == '\0') {
358 // This filename is zero bytes, so skip it (and the zero terminator).
361 } else if (ptr[1] == '\0') {
362 // This filename is one byte, so skip it (and the zero terminator).
365 } else if (ptr[2] == '\0') {
366 // This filename is two bytes, so skip it (and the zero terminator).
371 // NOTE: Will read one byte past the end of the trigram, but it's OK,
372 // since we always call it from contexts where there's a terminating zero byte.
374 memcpy(&trgm, ptr, sizeof(trgm));
376 trgm = le32toh(trgm);
377 add_docid(trgm & 0xffffff, docid);
378 if (trgm <= 0xffffff) {
379 // Terminating zero byte, so we're done with this filename.
380 // Skip the remaining two bytes, and the zero terminator.
388 // Compress and add the filename block.
389 filename_blocks.push_back(outfp_pos);
390 string compressed = zstd_compress(current_block, cdict, &tempbuf);
391 if (fwrite(compressed.data(), compressed.size(), 1, outfp) != 1) {
395 outfp_pos += compressed.size();
397 current_block.clear();
398 num_files_in_block = 0;
402 void EncodingCorpus::finish()
407 size_t EncodingCorpus::num_trigrams() const
410 for (unsigned trgm = 0; trgm < NUM_TRIGRAMS; ++trgm) {
411 if (invindex[trgm] != nullptr) {
418 string EncodingCorpus::get_compressed_dir_times()
420 if (!store_dir_times) {
423 compress_dir_times(/*allowed_slop=*/0);
424 assert(dir_times.empty());
427 size_t old_size = dir_times_compressed.size();
428 dir_times_compressed.resize(old_size + 4096);
430 ZSTD_outBuffer outbuf;
431 outbuf.dst = dir_times_compressed.data() + old_size;
435 int ret = ZSTD_endStream(dir_time_ctx, &outbuf);
437 fprintf(stderr, "ZSTD_compressStream() failed\n");
441 dir_times_compressed.resize(old_size + outbuf.pos);
449 return dir_times_compressed;
452 string zstd_compress(const string &src, ZSTD_CDict *cdict, string *tempbuf)
454 static ZSTD_CCtx *ctx = nullptr;
455 if (ctx == nullptr) {
456 ctx = ZSTD_createCCtx();
459 size_t max_size = ZSTD_compressBound(src.size());
460 if (tempbuf->size() < max_size) {
461 tempbuf->resize(max_size);
464 if (cdict == nullptr) {
465 size = ZSTD_compressCCtx(ctx, &(*tempbuf)[0], max_size, src.data(), src.size(), /*level=*/6);
467 size = ZSTD_compress_usingCDict(ctx, &(*tempbuf)[0], max_size, src.data(), src.size(), cdict);
469 return string(tempbuf->data(), size);
472 bool is_prime(uint32_t x)
474 if ((x % 2) == 0 || (x % 3) == 0) {
477 uint32_t limit = ceil(sqrt(x));
478 for (uint32_t factor = 5; factor <= limit; ++factor) {
479 if ((x % factor) == 0) {
486 uint32_t next_prime(uint32_t x)
491 while (!is_prime(x)) {
497 unique_ptr<Trigram[]> create_hashtable(EncodingCorpus &corpus, const vector<uint32_t> &all_trigrams, uint32_t ht_size, uint32_t num_overflow_slots)
499 unique_ptr<Trigram[]> ht(new Trigram[ht_size + num_overflow_slots + 1]); // 1 for the sentinel element at the end.
500 for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
501 ht[i].trgm = uint32_t(-1);
502 ht[i].num_docids = 0;
505 for (uint32_t trgm : all_trigrams) {
506 // We don't know offset yet, so set it to zero.
507 Trigram to_insert{ trgm, uint32_t(corpus.get_pl_builder(trgm).get_num_docids()), 0 };
509 uint32_t bucket = hash_trigram(trgm, ht_size);
510 unsigned distance = 0;
511 while (ht[bucket].num_docids != 0) {
512 // Robin Hood hashing; reduces the longest distance by a lot.
513 unsigned other_distance = bucket - hash_trigram(ht[bucket].trgm, ht_size);
514 if (distance > other_distance) {
515 swap(to_insert, ht[bucket]);
516 distance = other_distance;
519 ++bucket, ++distance;
520 if (distance > num_overflow_slots) {
524 ht[bucket] = to_insert;
529 DatabaseBuilder::DatabaseBuilder(const char *outfile, gid_t owner, int block_size, string dictionary, bool check_visibility)
530 : outfile(outfile), block_size(block_size)
534 string path = outfile;
535 path.resize(path.find_last_of('/') + 1);
541 fd = open(path.c_str(), O_WRONLY | O_TMPFILE, 0640);
542 if (fd == -1 && errno != EOPNOTSUPP && errno != EISDIR) {
543 perror(path.c_str());
548 temp_filename = string(outfile) + ".XXXXXX";
549 fd = mkstemp(&temp_filename[0]);
551 perror(temp_filename.c_str());
554 if (fchmod(fd, 0640) == -1) {
560 if (owner != (gid_t)-1) {
561 if (fchown(fd, (uid_t)-1, owner) == -1) {
567 outfp = fdopen(fd, "wb");
568 if (outfp == nullptr) {
574 memcpy(hdr.magic, "\0plocate", 8);
575 hdr.version = -1; // Mark as broken.
576 hdr.hashtable_size = 0; // Not known yet.
577 hdr.extra_ht_slots = num_overflow_slots;
579 hdr.hash_table_offset_bytes = -1; // We don't know these offsets yet.
581 hdr.filename_index_offset_bytes = -1;
582 hdr.zstd_dictionary_length_bytes = -1;
583 hdr.check_visibility = check_visibility;
584 fwrite(&hdr, sizeof(hdr), 1, outfp);
586 if (dictionary.empty()) {
587 hdr.zstd_dictionary_offset_bytes = 0;
588 hdr.zstd_dictionary_length_bytes = 0;
590 hdr.zstd_dictionary_offset_bytes = ftell(outfp);
591 fwrite(dictionary.data(), dictionary.size(), 1, outfp);
592 hdr.zstd_dictionary_length_bytes = dictionary.size();
593 cdict = ZSTD_createCDict(dictionary.data(), dictionary.size(), /*level=*/6);
596 hdr.directory_data_length_bytes = 0;
597 hdr.directory_data_offset_bytes = 0;
598 hdr.next_zstd_dictionary_length_bytes = 0;
599 hdr.next_zstd_dictionary_offset_bytes = 0;
600 hdr.conf_block_length_bytes = 0;
601 hdr.conf_block_offset_bytes = 0;
604 DatabaseReceiver *DatabaseBuilder::start_corpus(bool store_dir_times)
606 corpus_start = steady_clock::now();
607 corpus = new EncodingCorpus(outfp, block_size, cdict, store_dir_times);
611 void DatabaseBuilder::set_next_dictionary(std::string next_dictionary)
613 this->next_dictionary = move(next_dictionary);
616 void DatabaseBuilder::set_conf_block(std::string conf_block)
618 this->conf_block = move(conf_block);
621 void DatabaseBuilder::finish_corpus()
624 hdr.num_docids = corpus->filename_blocks.size();
626 // Stick an empty block at the end as sentinel.
627 corpus->filename_blocks.push_back(ftell(outfp));
628 const size_t bytes_for_filenames = corpus->filename_blocks.back() - corpus->filename_blocks.front();
630 // Write the offsets to the filenames.
631 hdr.filename_index_offset_bytes = ftell(outfp);
632 const size_t bytes_for_filename_index = corpus->filename_blocks.size() * sizeof(uint64_t);
633 fwrite(corpus->filename_blocks.data(), corpus->filename_blocks.size(), sizeof(uint64_t), outfp);
634 corpus->filename_blocks.clear();
635 corpus->filename_blocks.shrink_to_fit();
637 // Finish up encoding the posting lists.
638 size_t trigrams = 0, longest_posting_list = 0;
639 size_t bytes_for_posting_lists = 0;
640 for (unsigned trgm = 0; trgm < NUM_TRIGRAMS; ++trgm) {
641 if (!corpus->seen_trigram(trgm))
643 PostingListBuilder &pl_builder = corpus->get_pl_builder(trgm);
645 longest_posting_list = max(longest_posting_list, pl_builder.get_num_docids());
646 trigrams += pl_builder.get_num_docids();
647 bytes_for_posting_lists += pl_builder.encoded.size();
649 size_t num_trigrams = corpus->num_trigrams();
650 dprintf("%zu files, %zu different trigrams, %zu entries, avg len %.2f, longest %zu\n",
651 corpus->num_files, num_trigrams, trigrams, double(trigrams) / num_trigrams, longest_posting_list);
652 dprintf("%zu bytes used for posting lists (%.2f bits/entry)\n", bytes_for_posting_lists, 8 * bytes_for_posting_lists / double(trigrams));
654 dprintf("Building posting lists took %.1f ms.\n\n", 1e3 * duration<float>(steady_clock::now() - corpus_start).count());
656 // Find the used trigrams.
657 vector<uint32_t> all_trigrams;
658 for (unsigned trgm = 0; trgm < NUM_TRIGRAMS; ++trgm) {
659 if (corpus->seen_trigram(trgm)) {
660 all_trigrams.push_back(trgm);
664 // Create the hash table.
665 unique_ptr<Trigram[]> hashtable;
666 uint32_t ht_size = next_prime(all_trigrams.size());
668 hashtable = create_hashtable(*corpus, all_trigrams, ht_size, num_overflow_slots);
669 if (hashtable == nullptr) {
670 dprintf("Failed creating hash table of size %u, increasing by 5%% and trying again.\n", ht_size);
671 ht_size = next_prime(ht_size * 1.05);
673 dprintf("Created hash table of size %u.\n\n", ht_size);
678 // Find the offsets for each posting list.
679 size_t bytes_for_hashtable = (ht_size + num_overflow_slots + 1) * sizeof(Trigram);
680 uint64_t offset = ftell(outfp) + bytes_for_hashtable;
681 for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
682 hashtable[i].offset = offset; // Needs to be there even for empty slots.
683 if (hashtable[i].num_docids == 0) {
687 const vector<unsigned char> &encoded = corpus->get_pl_builder(hashtable[i].trgm).encoded;
688 offset += encoded.size();
691 // Write the hash table.
692 hdr.hash_table_offset_bytes = ftell(outfp);
693 hdr.hashtable_size = ht_size;
694 fwrite(hashtable.get(), ht_size + num_overflow_slots + 1, sizeof(Trigram), outfp);
696 // Write the actual posting lists.
697 for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
698 if (hashtable[i].num_docids == 0) {
701 const vector<unsigned char> &encoded = corpus->get_pl_builder(hashtable[i].trgm).encoded;
702 fwrite(encoded.data(), encoded.size(), 1, outfp);
705 // Finally, write the directory times (for updatedb).
706 string compressed_dir_times = corpus->get_compressed_dir_times();
707 size_t bytes_for_compressed_dir_times = 0;
708 if (!compressed_dir_times.empty()) {
709 hdr.directory_data_offset_bytes = ftell(outfp);
710 hdr.directory_data_length_bytes = compressed_dir_times.size();
711 fwrite(compressed_dir_times.data(), compressed_dir_times.size(), 1, outfp);
712 bytes_for_compressed_dir_times = compressed_dir_times.size();
713 compressed_dir_times.clear();
716 // Write the recommended dictionary for next update.
717 if (!next_dictionary.empty()) {
718 hdr.next_zstd_dictionary_offset_bytes = ftell(outfp);
719 hdr.next_zstd_dictionary_length_bytes = next_dictionary.size();
720 fwrite(next_dictionary.data(), next_dictionary.size(), 1, outfp);
723 // And the configuration block.
724 if (!conf_block.empty()) {
725 hdr.conf_block_offset_bytes = ftell(outfp);
726 hdr.conf_block_length_bytes = conf_block.size();
727 fwrite(conf_block.data(), conf_block.size(), 1, outfp);
730 // Rewind, and write the updated header.
732 fseek(outfp, 0, SEEK_SET);
733 fwrite(&hdr, sizeof(hdr), 1, outfp);
735 if (!temp_filename.empty()) {
736 if (rename(temp_filename.c_str(), outfile.c_str()) == -1) {
742 // Give the file a proper name, making it visible in the file system.
743 // TODO: It would be nice to be able to do this atomically, like with rename.
744 unlink(outfile.c_str());
746 snprintf(procpath, sizeof(procpath), "/proc/self/fd/%d", fileno(outfp));
747 if (linkat(AT_FDCWD, procpath, AT_FDCWD, outfile.c_str(), AT_SYMLINK_FOLLOW) == -1) {
756 size_t total_bytes = (bytes_for_hashtable + bytes_for_posting_lists + bytes_for_filename_index + bytes_for_filenames + bytes_for_compressed_dir_times);
758 dprintf("Block size: %7d files\n", block_size);
759 dprintf("Dictionary: %'7.1f MB\n", hdr.zstd_dictionary_length_bytes / 1048576.0);
760 dprintf("Hash table: %'7.1f MB\n", bytes_for_hashtable / 1048576.0);
761 dprintf("Posting lists: %'7.1f MB\n", bytes_for_posting_lists / 1048576.0);
762 dprintf("Filename index: %'7.1f MB\n", bytes_for_filename_index / 1048576.0);
763 dprintf("Filenames: %'7.1f MB\n", bytes_for_filenames / 1048576.0);
764 if (bytes_for_compressed_dir_times != 0) {
765 dprintf("Modify times: %'7.1f MB\n", bytes_for_compressed_dir_times / 1048576.0);
767 dprintf("Total: %'7.1f MB\n", total_bytes / 1048576.0);