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 class PostingListBuilder {
34 inline void add_docid(uint32_t docid);
35 inline void add_first_docid(uint32_t docid);
38 vector<unsigned char> encoded;
39 size_t get_num_docids() const {
40 // Updated only when we flush, so check that we're finished.
41 assert(pending_deltas.empty());
46 void write_header(uint32_t docid);
49 vector<uint32_t> pending_deltas;
51 uint32_t num_docids = 0; // Should be size_t, except the format only supports 2^32 docids per posting list anyway.
52 uint32_t last_docid = -1;
55 void PostingListBuilder::add_docid(uint32_t docid)
57 // Deduplicate against the last inserted value, if any.
58 if (docid == last_docid) {
62 pending_deltas.push_back(docid - last_docid - 1);
64 if (pending_deltas.size() == 128) {
66 pending_deltas.clear();
71 void PostingListBuilder::add_first_docid(uint32_t docid)
78 void PostingListBuilder::finish()
80 if (pending_deltas.empty()) {
84 assert(!encoded.empty()); // write_header() should already have run.
86 // No interleaving for partial blocks.
87 unsigned char buf[P4NENC_BOUND(128)];
88 unsigned char *end = encode_pfor_single_block<128>(pending_deltas.data(), pending_deltas.size(), /*interleaved=*/false, buf);
89 encoded.insert(encoded.end(), buf, end);
91 num_docids += pending_deltas.size();
92 pending_deltas.clear();
95 void PostingListBuilder::append_block()
97 unsigned char buf[P4NENC_BOUND(128)];
98 assert(pending_deltas.size() == 128);
99 unsigned char *end = encode_pfor_single_block<128>(pending_deltas.data(), 128, /*interleaved=*/true, buf);
100 encoded.insert(encoded.end(), buf, end);
103 void PostingListBuilder::write_header(uint32_t docid)
105 unsigned char buf[P4NENC_BOUND(1)];
106 unsigned char *end = write_baseval(docid, buf);
107 encoded.insert(encoded.end(), buf, end);
110 void DictionaryBuilder::add_file(string filename, dir_time)
112 if (keep_current_block) { // Only bother saving the filenames if we're actually keeping the block.
113 if (!current_block.empty()) {
114 current_block.push_back('\0');
116 current_block += filename;
118 if (++num_files_in_block == block_size) {
123 void DictionaryBuilder::flush_block()
125 if (keep_current_block) {
126 if (slot_for_current_block == -1) {
127 lengths.push_back(current_block.size());
128 sampled_blocks.push_back(move(current_block));
130 lengths[slot_for_current_block] = current_block.size();
131 sampled_blocks[slot_for_current_block] = move(current_block);
134 current_block.clear();
135 num_files_in_block = 0;
138 if (block_num < blocks_to_keep) {
139 keep_current_block = true;
140 slot_for_current_block = -1;
142 // Keep every block with equal probability (reservoir sampling).
143 uint64_t idx = uniform_int_distribution<uint64_t>(0, block_num)(reservoir_rand);
144 keep_current_block = (idx < blocks_to_keep);
145 slot_for_current_block = idx;
149 string DictionaryBuilder::train(size_t buf_size)
151 string dictionary_buf;
152 sort(sampled_blocks.begin(), sampled_blocks.end()); // Seemingly important for decompression speed.
153 for (const string &block : sampled_blocks) {
154 dictionary_buf += block;
158 buf.resize(buf_size);
159 size_t ret = ZDICT_trainFromBuffer(&buf[0], buf_size, dictionary_buf.data(), lengths.data(), lengths.size());
160 if (ZDICT_isError(ret)) {
163 dprintf("Sampled %zu bytes in %zu blocks, built a dictionary of size %zu\n", dictionary_buf.size(), lengths.size(), ret);
166 sampled_blocks.clear();
172 class EncodingCorpus : public DatabaseReceiver {
174 EncodingCorpus(FILE *outfp, size_t block_size, ZSTD_CDict *cdict, bool store_dir_times);
177 void add_file(std::string filename, dir_time dt) override;
178 void flush_block() override;
179 void finish() override;
181 std::vector<uint64_t> filename_blocks;
182 size_t num_files = 0, num_files_in_block = 0, num_blocks = 0;
183 bool seen_trigram(uint32_t trgm)
185 return invindex[trgm] != nullptr;
187 size_t num_files_seen() const override { return num_files; }
188 PostingListBuilder &get_pl_builder(uint32_t trgm)
190 return *invindex[trgm];
193 void add_docid(uint32_t trgm, uint32_t docid)
195 if (invindex[trgm] == nullptr) {
196 invindex[trgm] = new PostingListBuilder;
197 invindex[trgm]->add_first_docid(docid);
199 invindex[trgm]->add_docid(docid);
203 size_t num_trigrams() const;
204 std::string get_compressed_dir_times();
207 void compress_dir_times(size_t allowed_slop);
209 std::unique_ptr<PostingListBuilder *[]> invindex;
211 std::string current_block;
213 const size_t block_size;
214 const bool store_dir_times;
217 ZSTD_CStream *dir_time_ctx = nullptr;
218 std::string dir_times; // Buffer of still-uncompressed data.
219 std::string dir_times_compressed;
223 EncodingCorpus::EncodingCorpus(FILE *outfp, size_t block_size, ZSTD_CDict *cdict, bool store_dir_times)
224 : invindex(new PostingListBuilder *[NUM_TRIGRAMS]), outfp(outfp), block_size(block_size), store_dir_times(store_dir_times), cdict(cdict)
226 fill(invindex.get(), invindex.get() + NUM_TRIGRAMS, nullptr);
227 if (store_dir_times) {
228 dir_time_ctx = ZSTD_createCStream();
229 ZSTD_initCStream(dir_time_ctx, /*level=*/6);
233 EncodingCorpus::~EncodingCorpus()
235 for (unsigned i = 0; i < NUM_TRIGRAMS; ++i) {
240 void EncodingCorpus::add_file(string filename, dir_time dt)
243 if (!current_block.empty()) {
244 current_block.push_back('\0');
246 current_block += filename;
247 if (++num_files_in_block == block_size) {
251 if (store_dir_times) {
254 dir_times.push_back('\0');
256 dir_times.push_back('\1');
257 dir_times.append(reinterpret_cast<char *>(&dt.sec), sizeof(dt.sec));
258 dir_times.append(reinterpret_cast<char *>(&dt.nsec), sizeof(dt.nsec));
260 compress_dir_times(/*allowed_slop=*/4096);
264 void EncodingCorpus::compress_dir_times(size_t allowed_slop)
266 while (dir_times.size() >= allowed_slop) {
267 size_t old_size = dir_times_compressed.size();
268 dir_times_compressed.resize(old_size + 4096);
270 ZSTD_outBuffer outbuf;
271 outbuf.dst = dir_times_compressed.data() + old_size;
276 inbuf.src = dir_times.data();
277 inbuf.size = dir_times.size();
280 int ret = ZSTD_compressStream(dir_time_ctx, &outbuf, &inbuf);
282 fprintf(stderr, "ZSTD_compressStream() failed\n");
286 dir_times_compressed.resize(old_size + outbuf.pos);
287 dir_times.erase(dir_times.begin(), dir_times.begin() + inbuf.pos);
289 if (outbuf.pos == 0 && inbuf.pos == 0) {
290 // Nothing happened (not enough data?), try again later.
296 void EncodingCorpus::flush_block()
298 if (current_block.empty()) {
302 uint32_t docid = num_blocks;
305 const char *ptr = current_block.c_str();
306 const char *end = ptr + current_block.size();
307 while (ptr < end - 3) { // Must be at least one filename left, that's at least three bytes.
308 if (ptr[0] == '\0') {
309 // This filename is zero bytes, so skip it (and the zero terminator).
312 } else if (ptr[1] == '\0') {
313 // This filename is one byte, so skip it (and the zero terminator).
316 } else if (ptr[2] == '\0') {
317 // This filename is two bytes, so skip it (and the zero terminator).
322 // NOTE: Will read one byte past the end of the trigram, but it's OK,
323 // since we always call it from contexts where there's a terminating zero byte.
325 memcpy(&trgm, ptr, sizeof(trgm));
327 trgm = le32toh(trgm);
328 add_docid(trgm & 0xffffff, docid);
329 if (trgm <= 0xffffff) {
330 // Terminating zero byte, so we're done with this filename.
331 // Skip the remaining two bytes, and the zero terminator.
338 // Compress and add the filename block.
339 filename_blocks.push_back(ftell(outfp));
340 string compressed = zstd_compress(current_block, cdict, &tempbuf);
341 if (fwrite(compressed.data(), compressed.size(), 1, outfp) != 1) {
346 current_block.clear();
347 num_files_in_block = 0;
351 void EncodingCorpus::finish()
356 size_t EncodingCorpus::num_trigrams() const
359 for (unsigned trgm = 0; trgm < NUM_TRIGRAMS; ++trgm) {
360 if (invindex[trgm] != nullptr) {
367 string EncodingCorpus::get_compressed_dir_times()
369 if (!store_dir_times) {
372 compress_dir_times(/*allowed_slop=*/0);
373 assert(dir_times.empty());
376 size_t old_size = dir_times_compressed.size();
377 dir_times_compressed.resize(old_size + 4096);
379 ZSTD_outBuffer outbuf;
380 outbuf.dst = dir_times_compressed.data() + old_size;
384 int ret = ZSTD_endStream(dir_time_ctx, &outbuf);
386 fprintf(stderr, "ZSTD_compressStream() failed\n");
390 dir_times_compressed.resize(old_size + outbuf.pos);
398 return dir_times_compressed;
401 string zstd_compress(const string &src, ZSTD_CDict *cdict, string *tempbuf)
403 static ZSTD_CCtx *ctx = nullptr;
404 if (ctx == nullptr) {
405 ctx = ZSTD_createCCtx();
408 size_t max_size = ZSTD_compressBound(src.size());
409 if (tempbuf->size() < max_size) {
410 tempbuf->resize(max_size);
413 if (cdict == nullptr) {
414 size = ZSTD_compressCCtx(ctx, &(*tempbuf)[0], max_size, src.data(), src.size(), /*level=*/6);
416 size = ZSTD_compress_usingCDict(ctx, &(*tempbuf)[0], max_size, src.data(), src.size(), cdict);
418 return string(tempbuf->data(), size);
421 bool is_prime(uint32_t x)
423 if ((x % 2) == 0 || (x % 3) == 0) {
426 uint32_t limit = ceil(sqrt(x));
427 for (uint32_t factor = 5; factor <= limit; ++factor) {
428 if ((x % factor) == 0) {
435 uint32_t next_prime(uint32_t x)
440 while (!is_prime(x)) {
446 unique_ptr<Trigram[]> create_hashtable(EncodingCorpus &corpus, const vector<uint32_t> &all_trigrams, uint32_t ht_size, uint32_t num_overflow_slots)
448 unique_ptr<Trigram[]> ht(new Trigram[ht_size + num_overflow_slots + 1]); // 1 for the sentinel element at the end.
449 for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
450 ht[i].trgm = uint32_t(-1);
451 ht[i].num_docids = 0;
454 for (uint32_t trgm : all_trigrams) {
455 // We don't know offset yet, so set it to zero.
456 Trigram to_insert{ trgm, uint32_t(corpus.get_pl_builder(trgm).get_num_docids()), 0 };
458 uint32_t bucket = hash_trigram(trgm, ht_size);
459 unsigned distance = 0;
460 while (ht[bucket].num_docids != 0) {
461 // Robin Hood hashing; reduces the longest distance by a lot.
462 unsigned other_distance = bucket - hash_trigram(ht[bucket].trgm, ht_size);
463 if (distance > other_distance) {
464 swap(to_insert, ht[bucket]);
465 distance = other_distance;
468 ++bucket, ++distance;
469 if (distance > num_overflow_slots) {
473 ht[bucket] = to_insert;
478 DatabaseBuilder::DatabaseBuilder(const char *outfile, gid_t owner, int block_size, string dictionary, bool check_visibility)
479 : outfile(outfile), block_size(block_size)
483 string path = outfile;
484 path.resize(path.find_last_of('/') + 1);
489 int fd = open(path.c_str(), O_WRONLY | O_TMPFILE, 0640);
491 perror(path.c_str());
495 temp_filename = string(outfile) + ".XXXXXX";
496 int fd = mkstemp(&temp_filename[0]);
498 perror(temp_filename.c_str());
501 if (fchmod(fd, 0640) == -1) {
507 if (owner != (gid_t)-1) {
508 if (fchown(fd, (uid_t)-1, owner) == -1) {
514 outfp = fdopen(fd, "wb");
515 if (outfp == nullptr) {
521 memcpy(hdr.magic, "\0plocate", 8);
522 hdr.version = -1; // Mark as broken.
523 hdr.hashtable_size = 0; // Not known yet.
524 hdr.extra_ht_slots = num_overflow_slots;
526 hdr.hash_table_offset_bytes = -1; // We don't know these offsets yet.
528 hdr.filename_index_offset_bytes = -1;
529 hdr.zstd_dictionary_length_bytes = -1;
530 hdr.check_visibility = check_visibility;
531 fwrite(&hdr, sizeof(hdr), 1, outfp);
533 if (dictionary.empty()) {
534 hdr.zstd_dictionary_offset_bytes = 0;
535 hdr.zstd_dictionary_length_bytes = 0;
537 hdr.zstd_dictionary_offset_bytes = ftell(outfp);
538 fwrite(dictionary.data(), dictionary.size(), 1, outfp);
539 hdr.zstd_dictionary_length_bytes = dictionary.size();
540 cdict = ZSTD_createCDict(dictionary.data(), dictionary.size(), /*level=*/6);
543 hdr.directory_data_length_bytes = 0;
544 hdr.directory_data_offset_bytes = 0;
545 hdr.next_zstd_dictionary_length_bytes = 0;
546 hdr.next_zstd_dictionary_offset_bytes = 0;
547 hdr.conf_block_length_bytes = 0;
548 hdr.conf_block_offset_bytes = 0;
551 DatabaseReceiver *DatabaseBuilder::start_corpus(bool store_dir_times)
553 corpus_start = steady_clock::now();
554 corpus = new EncodingCorpus(outfp, block_size, cdict, store_dir_times);
558 void DatabaseBuilder::set_next_dictionary(std::string next_dictionary)
560 this->next_dictionary = move(next_dictionary);
563 void DatabaseBuilder::set_conf_block(std::string conf_block)
565 this->conf_block = move(conf_block);
568 void DatabaseBuilder::finish_corpus()
571 hdr.num_docids = corpus->filename_blocks.size();
573 // Stick an empty block at the end as sentinel.
574 corpus->filename_blocks.push_back(ftell(outfp));
575 const size_t bytes_for_filenames = corpus->filename_blocks.back() - corpus->filename_blocks.front();
577 // Write the offsets to the filenames.
578 hdr.filename_index_offset_bytes = ftell(outfp);
579 const size_t bytes_for_filename_index = corpus->filename_blocks.size() * sizeof(uint64_t);
580 fwrite(corpus->filename_blocks.data(), corpus->filename_blocks.size(), sizeof(uint64_t), outfp);
581 corpus->filename_blocks.clear();
582 corpus->filename_blocks.shrink_to_fit();
584 // Finish up encoding the posting lists.
585 size_t trigrams = 0, longest_posting_list = 0;
586 size_t bytes_for_posting_lists = 0;
587 for (unsigned trgm = 0; trgm < NUM_TRIGRAMS; ++trgm) {
588 if (!corpus->seen_trigram(trgm))
590 PostingListBuilder &pl_builder = corpus->get_pl_builder(trgm);
592 longest_posting_list = max(longest_posting_list, pl_builder.get_num_docids());
593 trigrams += pl_builder.get_num_docids();
594 bytes_for_posting_lists += pl_builder.encoded.size();
596 size_t num_trigrams = corpus->num_trigrams();
597 dprintf("%zu files, %zu different trigrams, %zu entries, avg len %.2f, longest %zu\n",
598 corpus->num_files, num_trigrams, trigrams, double(trigrams) / num_trigrams, longest_posting_list);
599 dprintf("%zu bytes used for posting lists (%.2f bits/entry)\n", bytes_for_posting_lists, 8 * bytes_for_posting_lists / double(trigrams));
601 dprintf("Building posting lists took %.1f ms.\n\n", 1e3 * duration<float>(steady_clock::now() - corpus_start).count());
603 // Find the used trigrams.
604 vector<uint32_t> all_trigrams;
605 for (unsigned trgm = 0; trgm < NUM_TRIGRAMS; ++trgm) {
606 if (corpus->seen_trigram(trgm)) {
607 all_trigrams.push_back(trgm);
611 // Create the hash table.
612 unique_ptr<Trigram[]> hashtable;
613 uint32_t ht_size = next_prime(all_trigrams.size());
615 hashtable = create_hashtable(*corpus, all_trigrams, ht_size, num_overflow_slots);
616 if (hashtable == nullptr) {
617 dprintf("Failed creating hash table of size %u, increasing by 5%% and trying again.\n", ht_size);
618 ht_size = next_prime(ht_size * 1.05);
620 dprintf("Created hash table of size %u.\n\n", ht_size);
625 // Find the offsets for each posting list.
626 size_t bytes_for_hashtable = (ht_size + num_overflow_slots + 1) * sizeof(Trigram);
627 uint64_t offset = ftell(outfp) + bytes_for_hashtable;
628 for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
629 hashtable[i].offset = offset; // Needs to be there even for empty slots.
630 if (hashtable[i].num_docids == 0) {
634 const vector<unsigned char> &encoded = corpus->get_pl_builder(hashtable[i].trgm).encoded;
635 offset += encoded.size();
638 // Write the hash table.
639 hdr.hash_table_offset_bytes = ftell(outfp);
640 hdr.hashtable_size = ht_size;
641 fwrite(hashtable.get(), ht_size + num_overflow_slots + 1, sizeof(Trigram), outfp);
643 // Write the actual posting lists.
644 for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
645 if (hashtable[i].num_docids == 0) {
648 const vector<unsigned char> &encoded = corpus->get_pl_builder(hashtable[i].trgm).encoded;
649 fwrite(encoded.data(), encoded.size(), 1, outfp);
652 // Finally, write the directory times (for updatedb).
653 string compressed_dir_times = corpus->get_compressed_dir_times();
654 size_t bytes_for_compressed_dir_times = 0;
655 if (!compressed_dir_times.empty()) {
656 hdr.directory_data_offset_bytes = ftell(outfp);
657 hdr.directory_data_length_bytes = compressed_dir_times.size();
658 fwrite(compressed_dir_times.data(), compressed_dir_times.size(), 1, outfp);
659 bytes_for_compressed_dir_times = compressed_dir_times.size();
660 compressed_dir_times.clear();
663 // Write the recommended dictionary for next update.
664 if (!next_dictionary.empty()) {
665 hdr.next_zstd_dictionary_offset_bytes = ftell(outfp);
666 hdr.next_zstd_dictionary_length_bytes = next_dictionary.size();
667 fwrite(next_dictionary.data(), next_dictionary.size(), 1, outfp);
670 // And the configuration block.
671 if (!conf_block.empty()) {
672 hdr.conf_block_offset_bytes = ftell(outfp);
673 hdr.conf_block_length_bytes = conf_block.size();
674 fwrite(conf_block.data(), conf_block.size(), 1, outfp);
677 // Rewind, and write the updated header.
679 fseek(outfp, 0, SEEK_SET);
680 fwrite(&hdr, sizeof(hdr), 1, outfp);
683 // Give the file a proper name, making it visible in the file system.
684 // TODO: It would be nice to be able to do this atomically, like with rename.
685 unlink(outfile.c_str());
687 snprintf(procpath, sizeof(procpath), "/proc/self/fd/%d", fileno(outfp));
688 if (linkat(AT_FDCWD, procpath, AT_FDCWD, outfile.c_str(), AT_SYMLINK_FOLLOW) == -1) {
693 if (rename(temp_filename.c_str(), outfile.c_str()) == -1) {
701 size_t total_bytes = (bytes_for_hashtable + bytes_for_posting_lists + bytes_for_filename_index + bytes_for_filenames + bytes_for_compressed_dir_times);
703 dprintf("Block size: %7d files\n", block_size);
704 dprintf("Dictionary: %'7.1f MB\n", hdr.zstd_dictionary_length_bytes / 1048576.0);
705 dprintf("Hash table: %'7.1f MB\n", bytes_for_hashtable / 1048576.0);
706 dprintf("Posting lists: %'7.1f MB\n", bytes_for_posting_lists / 1048576.0);
707 dprintf("Filename index: %'7.1f MB\n", bytes_for_filename_index / 1048576.0);
708 dprintf("Filenames: %'7.1f MB\n", bytes_for_filenames / 1048576.0);
709 if (bytes_for_compressed_dir_times != 0) {
710 dprintf("Modify times: %'7.1f MB\n", bytes_for_compressed_dir_times / 1048576.0);
712 dprintf("Total: %'7.1f MB\n", total_bytes / 1048576.0);