1 #include "database-builder.h"
4 #include "turbopfor-encode.h"
13 #include <string_view>
16 #include <sys/types.h>
21 #define P4NENC_BOUND(n) ((n + 127) / 128 + (n + 32) * sizeof(uint32_t))
23 #define NUM_TRIGRAMS 16777216
26 using namespace std::chrono;
28 constexpr unsigned num_overflow_slots = 16;
30 string zstd_compress(const string &src, ZSTD_CDict *cdict, string *tempbuf);
32 // NOTE: Will read one byte past the end of the trigram, but it's OK,
33 // since we always call it from contexts where there's a terminating zero byte.
34 static inline uint32_t read_trigram(const string_view s, size_t start)
37 memcpy(&trgm, s.data() + start, sizeof(trgm));
39 return trgm & 0xffffff;
42 class PostingListBuilder {
44 inline void add_docid(uint32_t docid);
45 inline void add_first_docid(uint32_t docid);
48 vector<unsigned char> encoded;
49 size_t get_num_docids() const {
50 // Updated only when we flush, so check that we're finished.
51 assert(pending_deltas.empty());
56 void write_header(uint32_t docid);
59 vector<uint32_t> pending_deltas;
61 uint32_t num_docids = 0; // Should be size_t, except the format only supports 2^32 docids per posting list anyway.
62 uint32_t last_docid = -1;
65 void PostingListBuilder::add_docid(uint32_t docid)
67 // Deduplicate against the last inserted value, if any.
68 if (docid == last_docid) {
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.insert(encoded.end(), buf, end);
101 num_docids += pending_deltas.size();
102 pending_deltas.clear();
105 void PostingListBuilder::append_block()
107 unsigned char buf[P4NENC_BOUND(128)];
108 assert(pending_deltas.size() == 128);
109 unsigned char *end = encode_pfor_single_block<128>(pending_deltas.data(), 128, /*interleaved=*/true, buf);
110 encoded.insert(encoded.end(), buf, end);
113 void PostingListBuilder::write_header(uint32_t docid)
115 unsigned char buf[P4NENC_BOUND(1)];
116 unsigned char *end = write_baseval(docid, buf);
117 encoded.insert(encoded.end(), buf, end);
120 void DictionaryBuilder::add_file(string filename, dir_time)
122 if (keep_current_block) { // Only bother saving the filenames if we're actually keeping the block.
123 if (!current_block.empty()) {
124 current_block.push_back('\0');
126 current_block += filename;
128 if (++num_files_in_block == block_size) {
133 void DictionaryBuilder::flush_block()
135 if (keep_current_block) {
136 if (slot_for_current_block == -1) {
137 lengths.push_back(current_block.size());
138 sampled_blocks.push_back(move(current_block));
140 lengths[slot_for_current_block] = current_block.size();
141 sampled_blocks[slot_for_current_block] = move(current_block);
144 current_block.clear();
145 num_files_in_block = 0;
148 if (block_num < blocks_to_keep) {
149 keep_current_block = true;
150 slot_for_current_block = -1;
152 // Keep every block with equal probability (reservoir sampling).
153 uint64_t idx = uniform_int_distribution<uint64_t>(0, block_num)(reservoir_rand);
154 keep_current_block = (idx < blocks_to_keep);
155 slot_for_current_block = idx;
159 string DictionaryBuilder::train(size_t buf_size)
161 string dictionary_buf;
162 sort(sampled_blocks.begin(), sampled_blocks.end()); // Seemingly important for decompression speed.
163 for (const string &block : sampled_blocks) {
164 dictionary_buf += block;
168 buf.resize(buf_size);
169 size_t ret = ZDICT_trainFromBuffer(&buf[0], buf_size, dictionary_buf.data(), lengths.data(), lengths.size());
170 if (ZDICT_isError(ret)) {
173 dprintf("Sampled %zu bytes in %zu blocks, built a dictionary of size %zu\n", dictionary_buf.size(), lengths.size(), ret);
176 sampled_blocks.clear();
182 class EncodingCorpus : public DatabaseReceiver {
184 EncodingCorpus(FILE *outfp, size_t block_size, ZSTD_CDict *cdict, bool store_dir_times);
187 void add_file(std::string filename, dir_time dt) override;
188 void flush_block() override;
189 void finish() override;
191 std::vector<uint64_t> filename_blocks;
192 size_t num_files = 0, num_files_in_block = 0, num_blocks = 0;
193 bool seen_trigram(uint32_t trgm)
195 return invindex[trgm] != nullptr;
197 size_t num_files_seen() const override { return num_files; }
198 PostingListBuilder &get_pl_builder(uint32_t trgm)
200 return *invindex[trgm];
203 void add_docid(uint32_t trgm, uint32_t docid)
205 if (invindex[trgm] == nullptr) {
206 invindex[trgm] = new PostingListBuilder;
207 invindex[trgm]->add_first_docid(docid);
209 invindex[trgm]->add_docid(docid);
213 size_t num_trigrams() const;
214 std::string get_compressed_dir_times();
217 void compress_dir_times(size_t allowed_slop);
219 std::unique_ptr<PostingListBuilder *[]> invindex;
221 std::string current_block;
223 const size_t block_size;
224 const bool store_dir_times;
227 ZSTD_CStream *dir_time_ctx = nullptr;
228 std::string dir_times; // Buffer of still-uncompressed data.
229 std::string dir_times_compressed;
233 EncodingCorpus::EncodingCorpus(FILE *outfp, size_t block_size, ZSTD_CDict *cdict, bool store_dir_times)
234 : invindex(new PostingListBuilder *[NUM_TRIGRAMS]), outfp(outfp), block_size(block_size), store_dir_times(store_dir_times), cdict(cdict)
236 fill(invindex.get(), invindex.get() + NUM_TRIGRAMS, nullptr);
237 if (store_dir_times) {
238 dir_time_ctx = ZSTD_createCStream();
239 ZSTD_initCStream(dir_time_ctx, /*level=*/6);
243 EncodingCorpus::~EncodingCorpus()
245 for (unsigned i = 0; i < NUM_TRIGRAMS; ++i) {
250 void EncodingCorpus::add_file(string filename, dir_time dt)
253 if (!current_block.empty()) {
254 current_block.push_back('\0');
256 current_block += filename;
257 if (++num_files_in_block == block_size) {
261 if (store_dir_times) {
264 dir_times.push_back('\0');
266 dir_times.push_back('\1');
267 dir_times.append(reinterpret_cast<char *>(&dt.sec), sizeof(dt.sec));
268 dir_times.append(reinterpret_cast<char *>(&dt.nsec), sizeof(dt.nsec));
270 compress_dir_times(/*allowed_slop=*/4096);
274 void EncodingCorpus::compress_dir_times(size_t allowed_slop)
276 while (dir_times.size() >= allowed_slop) {
277 size_t old_size = dir_times_compressed.size();
278 dir_times_compressed.resize(old_size + 4096);
280 ZSTD_outBuffer outbuf;
281 outbuf.dst = dir_times_compressed.data() + old_size;
286 inbuf.src = dir_times.data();
287 inbuf.size = dir_times.size();
290 int ret = ZSTD_compressStream(dir_time_ctx, &outbuf, &inbuf);
292 fprintf(stderr, "ZSTD_compressStream() failed\n");
296 dir_times_compressed.resize(old_size + outbuf.pos);
297 dir_times.erase(dir_times.begin(), dir_times.begin() + inbuf.pos);
299 if (outbuf.pos == 0 && inbuf.pos == 0) {
300 // Nothing happened (not enough data?), try again later.
306 void EncodingCorpus::flush_block()
308 if (current_block.empty()) {
312 uint32_t docid = num_blocks;
315 const char *ptr = current_block.c_str();
316 while (ptr < current_block.c_str() + current_block.size()) {
319 for (size_t j = 0; j < s.size() - 2; ++j) {
320 uint32_t trgm = read_trigram(s, j);
321 add_docid(trgm, docid);
327 // Compress and add the filename block.
328 filename_blocks.push_back(ftell(outfp));
329 string compressed = zstd_compress(current_block, cdict, &tempbuf);
330 if (fwrite(compressed.data(), compressed.size(), 1, outfp) != 1) {
335 current_block.clear();
336 num_files_in_block = 0;
340 void EncodingCorpus::finish()
345 size_t EncodingCorpus::num_trigrams() const
348 for (unsigned trgm = 0; trgm < NUM_TRIGRAMS; ++trgm) {
349 if (invindex[trgm] != nullptr) {
356 string EncodingCorpus::get_compressed_dir_times()
358 if (!store_dir_times) {
361 compress_dir_times(/*allowed_slop=*/0);
362 assert(dir_times.empty());
365 size_t old_size = dir_times_compressed.size();
366 dir_times_compressed.resize(old_size + 4096);
368 ZSTD_outBuffer outbuf;
369 outbuf.dst = dir_times_compressed.data() + old_size;
373 int ret = ZSTD_endStream(dir_time_ctx, &outbuf);
375 fprintf(stderr, "ZSTD_compressStream() failed\n");
379 dir_times_compressed.resize(old_size + outbuf.pos);
387 return dir_times_compressed;
390 string zstd_compress(const string &src, ZSTD_CDict *cdict, string *tempbuf)
392 static ZSTD_CCtx *ctx = nullptr;
393 if (ctx == nullptr) {
394 ctx = ZSTD_createCCtx();
397 size_t max_size = ZSTD_compressBound(src.size());
398 if (tempbuf->size() < max_size) {
399 tempbuf->resize(max_size);
402 if (cdict == nullptr) {
403 size = ZSTD_compressCCtx(ctx, &(*tempbuf)[0], max_size, src.data(), src.size(), /*level=*/6);
405 size = ZSTD_compress_usingCDict(ctx, &(*tempbuf)[0], max_size, src.data(), src.size(), cdict);
407 return string(tempbuf->data(), size);
410 bool is_prime(uint32_t x)
412 if ((x % 2) == 0 || (x % 3) == 0) {
415 uint32_t limit = ceil(sqrt(x));
416 for (uint32_t factor = 5; factor <= limit; ++factor) {
417 if ((x % factor) == 0) {
424 uint32_t next_prime(uint32_t x)
429 while (!is_prime(x)) {
435 unique_ptr<Trigram[]> create_hashtable(EncodingCorpus &corpus, const vector<uint32_t> &all_trigrams, uint32_t ht_size, uint32_t num_overflow_slots)
437 unique_ptr<Trigram[]> ht(new Trigram[ht_size + num_overflow_slots + 1]); // 1 for the sentinel element at the end.
438 for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
439 ht[i].trgm = uint32_t(-1);
440 ht[i].num_docids = 0;
443 for (uint32_t trgm : all_trigrams) {
444 // We don't know offset yet, so set it to zero.
445 Trigram to_insert{ trgm, uint32_t(corpus.get_pl_builder(trgm).get_num_docids()), 0 };
447 uint32_t bucket = hash_trigram(trgm, ht_size);
448 unsigned distance = 0;
449 while (ht[bucket].num_docids != 0) {
450 // Robin Hood hashing; reduces the longest distance by a lot.
451 unsigned other_distance = bucket - hash_trigram(ht[bucket].trgm, ht_size);
452 if (distance > other_distance) {
453 swap(to_insert, ht[bucket]);
454 distance = other_distance;
457 ++bucket, ++distance;
458 if (distance > num_overflow_slots) {
462 ht[bucket] = to_insert;
467 DatabaseBuilder::DatabaseBuilder(const char *outfile, gid_t owner, int block_size, string dictionary, bool check_visibility)
468 : outfile(outfile), block_size(block_size)
472 string path = outfile;
473 path.resize(path.find_last_of('/') + 1);
478 int fd = open(path.c_str(), O_WRONLY | O_TMPFILE, 0640);
480 perror(path.c_str());
484 temp_filename = string(outfile) + ".XXXXXX";
485 int fd = mkstemp(&temp_filename[0]);
487 perror(temp_filename.c_str());
490 if (fchmod(fd, 0640) == -1) {
496 if (owner != (gid_t)-1) {
497 if (fchown(fd, (uid_t)-1, owner) == -1) {
503 outfp = fdopen(fd, "wb");
504 if (outfp == nullptr) {
510 memcpy(hdr.magic, "\0plocate", 8);
511 hdr.version = -1; // Mark as broken.
512 hdr.hashtable_size = 0; // Not known yet.
513 hdr.extra_ht_slots = num_overflow_slots;
515 hdr.hash_table_offset_bytes = -1; // We don't know these offsets yet.
517 hdr.filename_index_offset_bytes = -1;
518 hdr.zstd_dictionary_length_bytes = -1;
519 hdr.check_visibility = check_visibility;
520 fwrite(&hdr, sizeof(hdr), 1, outfp);
522 if (dictionary.empty()) {
523 hdr.zstd_dictionary_offset_bytes = 0;
524 hdr.zstd_dictionary_length_bytes = 0;
526 hdr.zstd_dictionary_offset_bytes = ftell(outfp);
527 fwrite(dictionary.data(), dictionary.size(), 1, outfp);
528 hdr.zstd_dictionary_length_bytes = dictionary.size();
529 cdict = ZSTD_createCDict(dictionary.data(), dictionary.size(), /*level=*/6);
532 hdr.directory_data_length_bytes = 0;
533 hdr.directory_data_offset_bytes = 0;
534 hdr.next_zstd_dictionary_length_bytes = 0;
535 hdr.next_zstd_dictionary_offset_bytes = 0;
536 hdr.conf_block_length_bytes = 0;
537 hdr.conf_block_offset_bytes = 0;
540 DatabaseReceiver *DatabaseBuilder::start_corpus(bool store_dir_times)
542 corpus_start = steady_clock::now();
543 corpus = new EncodingCorpus(outfp, block_size, cdict, store_dir_times);
547 void DatabaseBuilder::set_next_dictionary(std::string next_dictionary)
549 this->next_dictionary = move(next_dictionary);
552 void DatabaseBuilder::set_conf_block(std::string conf_block)
554 this->conf_block = move(conf_block);
557 void DatabaseBuilder::finish_corpus()
560 hdr.num_docids = corpus->filename_blocks.size();
562 // Stick an empty block at the end as sentinel.
563 corpus->filename_blocks.push_back(ftell(outfp));
564 const size_t bytes_for_filenames = corpus->filename_blocks.back() - corpus->filename_blocks.front();
566 // Write the offsets to the filenames.
567 hdr.filename_index_offset_bytes = ftell(outfp);
568 const size_t bytes_for_filename_index = corpus->filename_blocks.size() * sizeof(uint64_t);
569 fwrite(corpus->filename_blocks.data(), corpus->filename_blocks.size(), sizeof(uint64_t), outfp);
570 corpus->filename_blocks.clear();
571 corpus->filename_blocks.shrink_to_fit();
573 // Finish up encoding the posting lists.
574 size_t trigrams = 0, longest_posting_list = 0;
575 size_t bytes_for_posting_lists = 0;
576 for (unsigned trgm = 0; trgm < NUM_TRIGRAMS; ++trgm) {
577 if (!corpus->seen_trigram(trgm))
579 PostingListBuilder &pl_builder = corpus->get_pl_builder(trgm);
581 longest_posting_list = max(longest_posting_list, pl_builder.get_num_docids());
582 trigrams += pl_builder.get_num_docids();
583 bytes_for_posting_lists += pl_builder.encoded.size();
585 size_t num_trigrams = corpus->num_trigrams();
586 dprintf("%zu files, %zu different trigrams, %zu entries, avg len %.2f, longest %zu\n",
587 corpus->num_files, num_trigrams, trigrams, double(trigrams) / num_trigrams, longest_posting_list);
588 dprintf("%zu bytes used for posting lists (%.2f bits/entry)\n", bytes_for_posting_lists, 8 * bytes_for_posting_lists / double(trigrams));
590 dprintf("Building posting lists took %.1f ms.\n\n", 1e3 * duration<float>(steady_clock::now() - corpus_start).count());
592 // Find the used trigrams.
593 vector<uint32_t> all_trigrams;
594 for (unsigned trgm = 0; trgm < NUM_TRIGRAMS; ++trgm) {
595 if (corpus->seen_trigram(trgm)) {
596 all_trigrams.push_back(trgm);
600 // Create the hash table.
601 unique_ptr<Trigram[]> hashtable;
602 uint32_t ht_size = next_prime(all_trigrams.size());
604 hashtable = create_hashtable(*corpus, all_trigrams, ht_size, num_overflow_slots);
605 if (hashtable == nullptr) {
606 dprintf("Failed creating hash table of size %u, increasing by 5%% and trying again.\n", ht_size);
607 ht_size = next_prime(ht_size * 1.05);
609 dprintf("Created hash table of size %u.\n\n", ht_size);
614 // Find the offsets for each posting list.
615 size_t bytes_for_hashtable = (ht_size + num_overflow_slots + 1) * sizeof(Trigram);
616 uint64_t offset = ftell(outfp) + bytes_for_hashtable;
617 for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
618 hashtable[i].offset = offset; // Needs to be there even for empty slots.
619 if (hashtable[i].num_docids == 0) {
623 const vector<unsigned char> &encoded = corpus->get_pl_builder(hashtable[i].trgm).encoded;
624 offset += encoded.size();
627 // Write the hash table.
628 hdr.hash_table_offset_bytes = ftell(outfp);
629 hdr.hashtable_size = ht_size;
630 fwrite(hashtable.get(), ht_size + num_overflow_slots + 1, sizeof(Trigram), outfp);
632 // Write the actual posting lists.
633 for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
634 if (hashtable[i].num_docids == 0) {
637 const vector<unsigned char> &encoded = corpus->get_pl_builder(hashtable[i].trgm).encoded;
638 fwrite(encoded.data(), encoded.size(), 1, outfp);
641 // Finally, write the directory times (for updatedb).
642 string compressed_dir_times = corpus->get_compressed_dir_times();
643 size_t bytes_for_compressed_dir_times = 0;
644 if (!compressed_dir_times.empty()) {
645 hdr.directory_data_offset_bytes = ftell(outfp);
646 hdr.directory_data_length_bytes = compressed_dir_times.size();
647 fwrite(compressed_dir_times.data(), compressed_dir_times.size(), 1, outfp);
648 bytes_for_compressed_dir_times = compressed_dir_times.size();
649 compressed_dir_times.clear();
652 // Write the recommended dictionary for next update.
653 if (!next_dictionary.empty()) {
654 hdr.next_zstd_dictionary_offset_bytes = ftell(outfp);
655 hdr.next_zstd_dictionary_length_bytes = next_dictionary.size();
656 fwrite(next_dictionary.data(), next_dictionary.size(), 1, outfp);
659 // And the configuration block.
660 if (!conf_block.empty()) {
661 hdr.conf_block_offset_bytes = ftell(outfp);
662 hdr.conf_block_length_bytes = conf_block.size();
663 fwrite(conf_block.data(), conf_block.size(), 1, outfp);
666 // Rewind, and write the updated header.
668 fseek(outfp, 0, SEEK_SET);
669 fwrite(&hdr, sizeof(hdr), 1, outfp);
672 // Give the file a proper name, making it visible in the file system.
673 // TODO: It would be nice to be able to do this atomically, like with rename.
674 unlink(outfile.c_str());
676 snprintf(procpath, sizeof(procpath), "/proc/self/fd/%d", fileno(outfp));
677 if (linkat(AT_FDCWD, procpath, AT_FDCWD, outfile.c_str(), AT_SYMLINK_FOLLOW) == -1) {
682 if (rename(temp_filename.c_str(), outfile.c_str()) == -1) {
690 size_t total_bytes = (bytes_for_hashtable + bytes_for_posting_lists + bytes_for_filename_index + bytes_for_filenames + bytes_for_compressed_dir_times);
692 dprintf("Block size: %7d files\n", block_size);
693 dprintf("Dictionary: %'7.1f MB\n", hdr.zstd_dictionary_length_bytes / 1048576.0);
694 dprintf("Hash table: %'7.1f MB\n", bytes_for_hashtable / 1048576.0);
695 dprintf("Posting lists: %'7.1f MB\n", bytes_for_posting_lists / 1048576.0);
696 dprintf("Filename index: %'7.1f MB\n", bytes_for_filename_index / 1048576.0);
697 dprintf("Filenames: %'7.1f MB\n", bytes_for_filenames / 1048576.0);
698 if (bytes_for_compressed_dir_times != 0) {
699 dprintf("Modify times: %'7.1f MB\n", bytes_for_compressed_dir_times / 1048576.0);
701 dprintf("Total: %'7.1f MB\n", total_bytes / 1048576.0);