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 num_docids = 0;
52 void write_header(uint32_t docid);
55 vector<uint32_t> pending_deltas;
57 uint32_t last_docid = -1;
60 void PostingListBuilder::add_docid(uint32_t docid)
62 // Deduplicate against the last inserted value, if any.
63 if (docid == last_docid) {
67 assert(num_docids != 0);
69 pending_deltas.push_back(docid - last_docid - 1);
71 if (pending_deltas.size() == 128) {
73 pending_deltas.clear();
78 void PostingListBuilder::add_first_docid(uint32_t docid)
85 void PostingListBuilder::finish()
87 if (pending_deltas.empty()) {
91 assert(!encoded.empty()); // write_header() should already have run.
93 // No interleaving for partial blocks.
94 unsigned char buf[P4NENC_BOUND(128)];
95 unsigned char *end = encode_pfor_single_block<128>(pending_deltas.data(), pending_deltas.size(), /*interleaved=*/false, buf);
96 encoded.insert(encoded.end(), buf, end);
99 void PostingListBuilder::append_block()
101 unsigned char buf[P4NENC_BOUND(128)];
102 assert(pending_deltas.size() == 128);
103 unsigned char *end = encode_pfor_single_block<128>(pending_deltas.data(), 128, /*interleaved=*/true, buf);
104 encoded.insert(encoded.end(), buf, end);
107 void PostingListBuilder::write_header(uint32_t docid)
109 unsigned char buf[P4NENC_BOUND(1)];
110 unsigned char *end = write_baseval(docid, buf);
111 encoded.insert(encoded.end(), buf, end);
114 void DictionaryBuilder::add_file(string filename, dir_time)
116 if (keep_current_block) { // Only bother saving the filenames if we're actually keeping the block.
117 if (!current_block.empty()) {
118 current_block.push_back('\0');
120 current_block += filename;
122 if (++num_files_in_block == block_size) {
127 void DictionaryBuilder::flush_block()
129 if (keep_current_block) {
130 if (slot_for_current_block == -1) {
131 lengths.push_back(current_block.size());
132 sampled_blocks.push_back(move(current_block));
134 lengths[slot_for_current_block] = current_block.size();
135 sampled_blocks[slot_for_current_block] = move(current_block);
138 current_block.clear();
139 num_files_in_block = 0;
142 if (block_num < blocks_to_keep) {
143 keep_current_block = true;
144 slot_for_current_block = -1;
146 // Keep every block with equal probability (reservoir sampling).
147 uint64_t idx = uniform_int_distribution<uint64_t>(0, block_num)(reservoir_rand);
148 keep_current_block = (idx < blocks_to_keep);
149 slot_for_current_block = idx;
153 string DictionaryBuilder::train(size_t buf_size)
155 string dictionary_buf;
156 sort(sampled_blocks.begin(), sampled_blocks.end()); // Seemingly important for decompression speed.
157 for (const string &block : sampled_blocks) {
158 dictionary_buf += block;
162 buf.resize(buf_size);
163 size_t ret = ZDICT_trainFromBuffer(&buf[0], buf_size, dictionary_buf.data(), lengths.data(), lengths.size());
164 if (ZDICT_isError(ret)) {
167 dprintf("Sampled %zu bytes in %zu blocks, built a dictionary of size %zu\n", dictionary_buf.size(), lengths.size(), ret);
170 sampled_blocks.clear();
176 class EncodingCorpus : public DatabaseReceiver {
178 EncodingCorpus(FILE *outfp, size_t block_size, ZSTD_CDict *cdict, bool store_dir_times);
181 void add_file(std::string filename, dir_time dt) override;
182 void flush_block() override;
183 void finish() override;
185 std::vector<uint64_t> filename_blocks;
186 size_t num_files = 0, num_files_in_block = 0, num_blocks = 0;
187 bool seen_trigram(uint32_t trgm)
189 return invindex[trgm] != nullptr;
191 size_t num_files_seen() const override { return num_files; }
192 PostingListBuilder &get_pl_builder(uint32_t trgm)
194 return *invindex[trgm];
197 void add_docid(uint32_t trgm, uint32_t docid)
199 if (invindex[trgm] == nullptr) {
200 invindex[trgm] = new PostingListBuilder;
201 invindex[trgm]->add_first_docid(docid);
203 invindex[trgm]->add_docid(docid);
207 size_t num_trigrams() const;
208 std::string get_compressed_dir_times();
211 void compress_dir_times(size_t allowed_slop);
213 std::unique_ptr<PostingListBuilder *[]> invindex;
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;
227 EncodingCorpus::EncodingCorpus(FILE *outfp, size_t block_size, ZSTD_CDict *cdict, bool store_dir_times)
228 : invindex(new PostingListBuilder *[NUM_TRIGRAMS]), outfp(outfp), block_size(block_size), store_dir_times(store_dir_times), cdict(cdict)
230 fill(invindex.get(), invindex.get() + NUM_TRIGRAMS, nullptr);
231 if (store_dir_times) {
232 dir_time_ctx = ZSTD_createCStream();
233 ZSTD_initCStream(dir_time_ctx, /*level=*/6);
237 EncodingCorpus::~EncodingCorpus()
239 for (unsigned i = 0; i < NUM_TRIGRAMS; ++i) {
244 void EncodingCorpus::add_file(string filename, dir_time dt)
247 if (!current_block.empty()) {
248 current_block.push_back('\0');
250 current_block += filename;
251 if (++num_files_in_block == block_size) {
255 if (store_dir_times) {
258 dir_times.push_back('\0');
260 dir_times.push_back('\1');
261 dir_times.append(reinterpret_cast<char *>(&dt.sec), sizeof(dt.sec));
262 dir_times.append(reinterpret_cast<char *>(&dt.nsec), sizeof(dt.nsec));
264 compress_dir_times(/*allowed_slop=*/4096);
268 void EncodingCorpus::compress_dir_times(size_t allowed_slop)
270 while (dir_times.size() >= allowed_slop) {
271 size_t old_size = dir_times_compressed.size();
272 dir_times_compressed.resize(old_size + 4096);
274 ZSTD_outBuffer outbuf;
275 outbuf.dst = dir_times_compressed.data() + old_size;
280 inbuf.src = dir_times.data();
281 inbuf.size = dir_times.size();
284 int ret = ZSTD_compressStream(dir_time_ctx, &outbuf, &inbuf);
286 fprintf(stderr, "ZSTD_compressStream() failed\n");
290 dir_times_compressed.resize(old_size + outbuf.pos);
291 dir_times.erase(dir_times.begin(), dir_times.begin() + inbuf.pos);
293 if (outbuf.pos == 0 && inbuf.pos == 0) {
294 // Nothing happened (not enough data?), try again later.
300 void EncodingCorpus::flush_block()
302 if (current_block.empty()) {
306 uint32_t docid = num_blocks;
309 const char *ptr = current_block.c_str();
310 while (ptr < current_block.c_str() + current_block.size()) {
313 for (size_t j = 0; j < s.size() - 2; ++j) {
314 uint32_t trgm = read_trigram(s, j);
315 add_docid(trgm, docid);
321 // Compress and add the filename block.
322 filename_blocks.push_back(ftell(outfp));
323 string compressed = zstd_compress(current_block, cdict, &tempbuf);
324 if (fwrite(compressed.data(), compressed.size(), 1, outfp) != 1) {
329 current_block.clear();
330 num_files_in_block = 0;
334 void EncodingCorpus::finish()
339 size_t EncodingCorpus::num_trigrams() const
342 for (unsigned trgm = 0; trgm < NUM_TRIGRAMS; ++trgm) {
343 if (invindex[trgm] != nullptr) {
350 string EncodingCorpus::get_compressed_dir_times()
352 if (!store_dir_times) {
355 compress_dir_times(/*allowed_slop=*/0);
356 assert(dir_times.empty());
359 size_t old_size = dir_times_compressed.size();
360 dir_times_compressed.resize(old_size + 4096);
362 ZSTD_outBuffer outbuf;
363 outbuf.dst = dir_times_compressed.data() + old_size;
367 int ret = ZSTD_endStream(dir_time_ctx, &outbuf);
369 fprintf(stderr, "ZSTD_compressStream() failed\n");
373 dir_times_compressed.resize(old_size + outbuf.pos);
381 return dir_times_compressed;
384 string zstd_compress(const string &src, ZSTD_CDict *cdict, string *tempbuf)
386 static ZSTD_CCtx *ctx = nullptr;
387 if (ctx == nullptr) {
388 ctx = ZSTD_createCCtx();
391 size_t max_size = ZSTD_compressBound(src.size());
392 if (tempbuf->size() < max_size) {
393 tempbuf->resize(max_size);
396 if (cdict == nullptr) {
397 size = ZSTD_compressCCtx(ctx, &(*tempbuf)[0], max_size, src.data(), src.size(), /*level=*/6);
399 size = ZSTD_compress_usingCDict(ctx, &(*tempbuf)[0], max_size, src.data(), src.size(), cdict);
401 return string(tempbuf->data(), size);
404 bool is_prime(uint32_t x)
406 if ((x % 2) == 0 || (x % 3) == 0) {
409 uint32_t limit = ceil(sqrt(x));
410 for (uint32_t factor = 5; factor <= limit; ++factor) {
411 if ((x % factor) == 0) {
418 uint32_t next_prime(uint32_t x)
423 while (!is_prime(x)) {
429 unique_ptr<Trigram[]> create_hashtable(EncodingCorpus &corpus, const vector<uint32_t> &all_trigrams, uint32_t ht_size, uint32_t num_overflow_slots)
431 unique_ptr<Trigram[]> ht(new Trigram[ht_size + num_overflow_slots + 1]); // 1 for the sentinel element at the end.
432 for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
433 ht[i].trgm = uint32_t(-1);
434 ht[i].num_docids = 0;
437 for (uint32_t trgm : all_trigrams) {
438 // We don't know offset yet, so set it to zero.
439 Trigram to_insert{ trgm, uint32_t(corpus.get_pl_builder(trgm).num_docids), 0 };
441 uint32_t bucket = hash_trigram(trgm, ht_size);
442 unsigned distance = 0;
443 while (ht[bucket].num_docids != 0) {
444 // Robin Hood hashing; reduces the longest distance by a lot.
445 unsigned other_distance = bucket - hash_trigram(ht[bucket].trgm, ht_size);
446 if (distance > other_distance) {
447 swap(to_insert, ht[bucket]);
448 distance = other_distance;
451 ++bucket, ++distance;
452 if (distance > num_overflow_slots) {
456 ht[bucket] = to_insert;
461 DatabaseBuilder::DatabaseBuilder(const char *outfile, gid_t owner, int block_size, string dictionary, bool check_visibility)
462 : outfile(outfile), block_size(block_size)
466 string path = outfile;
467 path.resize(path.find_last_of('/') + 1);
472 int fd = open(path.c_str(), O_WRONLY | O_TMPFILE, 0640);
474 perror(path.c_str());
478 temp_filename = string(outfile) + ".XXXXXX";
479 int fd = mkstemp(&temp_filename[0]);
481 perror(temp_filename.c_str());
484 if (fchmod(fd, 0640) == -1) {
490 if (owner != (gid_t)-1) {
491 if (fchown(fd, (uid_t)-1, owner) == -1) {
497 outfp = fdopen(fd, "wb");
498 if (outfp == nullptr) {
504 memcpy(hdr.magic, "\0plocate", 8);
505 hdr.version = -1; // Mark as broken.
506 hdr.hashtable_size = 0; // Not known yet.
507 hdr.extra_ht_slots = num_overflow_slots;
509 hdr.hash_table_offset_bytes = -1; // We don't know these offsets yet.
511 hdr.filename_index_offset_bytes = -1;
512 hdr.zstd_dictionary_length_bytes = -1;
513 hdr.check_visibility = check_visibility;
514 fwrite(&hdr, sizeof(hdr), 1, outfp);
516 if (dictionary.empty()) {
517 hdr.zstd_dictionary_offset_bytes = 0;
518 hdr.zstd_dictionary_length_bytes = 0;
520 hdr.zstd_dictionary_offset_bytes = ftell(outfp);
521 fwrite(dictionary.data(), dictionary.size(), 1, outfp);
522 hdr.zstd_dictionary_length_bytes = dictionary.size();
523 cdict = ZSTD_createCDict(dictionary.data(), dictionary.size(), /*level=*/6);
526 hdr.directory_data_length_bytes = 0;
527 hdr.directory_data_offset_bytes = 0;
528 hdr.next_zstd_dictionary_length_bytes = 0;
529 hdr.next_zstd_dictionary_offset_bytes = 0;
530 hdr.conf_block_length_bytes = 0;
531 hdr.conf_block_offset_bytes = 0;
534 DatabaseReceiver *DatabaseBuilder::start_corpus(bool store_dir_times)
536 corpus_start = steady_clock::now();
537 corpus = new EncodingCorpus(outfp, block_size, cdict, store_dir_times);
541 void DatabaseBuilder::set_next_dictionary(std::string next_dictionary)
543 this->next_dictionary = move(next_dictionary);
546 void DatabaseBuilder::set_conf_block(std::string conf_block)
548 this->conf_block = move(conf_block);
551 void DatabaseBuilder::finish_corpus()
554 hdr.num_docids = corpus->filename_blocks.size();
556 // Stick an empty block at the end as sentinel.
557 corpus->filename_blocks.push_back(ftell(outfp));
558 const size_t bytes_for_filenames = corpus->filename_blocks.back() - corpus->filename_blocks.front();
560 // Write the offsets to the filenames.
561 hdr.filename_index_offset_bytes = ftell(outfp);
562 const size_t bytes_for_filename_index = corpus->filename_blocks.size() * sizeof(uint64_t);
563 fwrite(corpus->filename_blocks.data(), corpus->filename_blocks.size(), sizeof(uint64_t), outfp);
564 corpus->filename_blocks.clear();
565 corpus->filename_blocks.shrink_to_fit();
567 // Finish up encoding the posting lists.
568 size_t trigrams = 0, longest_posting_list = 0;
569 size_t bytes_for_posting_lists = 0;
570 for (unsigned trgm = 0; trgm < NUM_TRIGRAMS; ++trgm) {
571 if (!corpus->seen_trigram(trgm))
573 PostingListBuilder &pl_builder = corpus->get_pl_builder(trgm);
575 longest_posting_list = max(longest_posting_list, pl_builder.num_docids);
576 trigrams += pl_builder.num_docids;
577 bytes_for_posting_lists += pl_builder.encoded.size();
579 size_t num_trigrams = corpus->num_trigrams();
580 dprintf("%zu files, %zu different trigrams, %zu entries, avg len %.2f, longest %zu\n",
581 corpus->num_files, num_trigrams, trigrams, double(trigrams) / num_trigrams, longest_posting_list);
582 dprintf("%zu bytes used for posting lists (%.2f bits/entry)\n", bytes_for_posting_lists, 8 * bytes_for_posting_lists / double(trigrams));
584 dprintf("Building posting lists took %.1f ms.\n\n", 1e3 * duration<float>(steady_clock::now() - corpus_start).count());
586 // Find the used trigrams.
587 vector<uint32_t> all_trigrams;
588 for (unsigned trgm = 0; trgm < NUM_TRIGRAMS; ++trgm) {
589 if (corpus->seen_trigram(trgm)) {
590 all_trigrams.push_back(trgm);
594 // Create the hash table.
595 unique_ptr<Trigram[]> hashtable;
596 uint32_t ht_size = next_prime(all_trigrams.size());
598 hashtable = create_hashtable(*corpus, all_trigrams, ht_size, num_overflow_slots);
599 if (hashtable == nullptr) {
600 dprintf("Failed creating hash table of size %u, increasing by 5%% and trying again.\n", ht_size);
601 ht_size = next_prime(ht_size * 1.05);
603 dprintf("Created hash table of size %u.\n\n", ht_size);
608 // Find the offsets for each posting list.
609 size_t bytes_for_hashtable = (ht_size + num_overflow_slots + 1) * sizeof(Trigram);
610 uint64_t offset = ftell(outfp) + bytes_for_hashtable;
611 for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
612 hashtable[i].offset = offset; // Needs to be there even for empty slots.
613 if (hashtable[i].num_docids == 0) {
617 const vector<unsigned char> &encoded = corpus->get_pl_builder(hashtable[i].trgm).encoded;
618 offset += encoded.size();
621 // Write the hash table.
622 hdr.hash_table_offset_bytes = ftell(outfp);
623 hdr.hashtable_size = ht_size;
624 fwrite(hashtable.get(), ht_size + num_overflow_slots + 1, sizeof(Trigram), outfp);
626 // Write the actual posting lists.
627 for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
628 if (hashtable[i].num_docids == 0) {
631 const vector<unsigned char> &encoded = corpus->get_pl_builder(hashtable[i].trgm).encoded;
632 fwrite(encoded.data(), encoded.size(), 1, outfp);
635 // Finally, write the directory times (for updatedb).
636 string compressed_dir_times = corpus->get_compressed_dir_times();
637 size_t bytes_for_compressed_dir_times = 0;
638 if (!compressed_dir_times.empty()) {
639 hdr.directory_data_offset_bytes = ftell(outfp);
640 hdr.directory_data_length_bytes = compressed_dir_times.size();
641 fwrite(compressed_dir_times.data(), compressed_dir_times.size(), 1, outfp);
642 bytes_for_compressed_dir_times = compressed_dir_times.size();
643 compressed_dir_times.clear();
646 // Write the recommended dictionary for next update.
647 if (!next_dictionary.empty()) {
648 hdr.next_zstd_dictionary_offset_bytes = ftell(outfp);
649 hdr.next_zstd_dictionary_length_bytes = next_dictionary.size();
650 fwrite(next_dictionary.data(), next_dictionary.size(), 1, outfp);
653 // And the configuration block.
654 if (!conf_block.empty()) {
655 hdr.conf_block_offset_bytes = ftell(outfp);
656 hdr.conf_block_length_bytes = conf_block.size();
657 fwrite(conf_block.data(), conf_block.size(), 1, outfp);
660 // Rewind, and write the updated header.
662 fseek(outfp, 0, SEEK_SET);
663 fwrite(&hdr, sizeof(hdr), 1, outfp);
666 // Give the file a proper name, making it visible in the file system.
667 // TODO: It would be nice to be able to do this atomically, like with rename.
668 unlink(outfile.c_str());
670 snprintf(procpath, sizeof(procpath), "/proc/self/fd/%d", fileno(outfp));
671 if (linkat(AT_FDCWD, procpath, AT_FDCWD, outfile.c_str(), AT_SYMLINK_FOLLOW) == -1) {
676 if (rename(temp_filename.c_str(), outfile.c_str()) == -1) {
684 size_t total_bytes = (bytes_for_hashtable + bytes_for_posting_lists + bytes_for_filename_index + bytes_for_filenames + bytes_for_compressed_dir_times);
686 dprintf("Block size: %7d files\n", block_size);
687 dprintf("Dictionary: %'7.1f MB\n", hdr.zstd_dictionary_length_bytes / 1048576.0);
688 dprintf("Hash table: %'7.1f MB\n", bytes_for_hashtable / 1048576.0);
689 dprintf("Posting lists: %'7.1f MB\n", bytes_for_posting_lists / 1048576.0);
690 dprintf("Filename index: %'7.1f MB\n", bytes_for_filename_index / 1048576.0);
691 dprintf("Filenames: %'7.1f MB\n", bytes_for_filenames / 1048576.0);
692 if (bytes_for_compressed_dir_times != 0) {
693 dprintf("Modify times: %'7.1f MB\n", bytes_for_compressed_dir_times / 1048576.0);
695 dprintf("Total: %'7.1f MB\n", total_bytes / 1048576.0);