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);
51 size_t num_docids = 0;
54 void write_header(uint32_t docid);
57 vector<uint32_t> pending_deltas;
59 uint32_t last_docid = -1;
62 void PostingListBuilder::add_docid(uint32_t docid)
64 // Deduplicate against the last inserted value, if any.
65 if (docid == last_docid) {
69 if (num_docids == 0) {
77 pending_deltas.push_back(docid - last_docid - 1);
79 if (pending_deltas.size() == 128) {
81 pending_deltas.clear();
86 void PostingListBuilder::finish()
88 if (pending_deltas.empty()) {
92 assert(!encoded.empty()); // write_header() should already have run.
94 // No interleaving for partial blocks.
95 unsigned char buf[P4NENC_BOUND(128)];
96 unsigned char *end = encode_pfor_single_block<128>(pending_deltas.data(), pending_deltas.size(), /*interleaved=*/false, buf);
97 encoded.append(reinterpret_cast<char *>(buf), reinterpret_cast<char *>(end));
100 void PostingListBuilder::append_block()
102 unsigned char buf[P4NENC_BOUND(128)];
103 assert(pending_deltas.size() == 128);
104 unsigned char *end = encode_pfor_single_block<128>(pending_deltas.data(), 128, /*interleaved=*/true, buf);
105 encoded.append(reinterpret_cast<char *>(buf), reinterpret_cast<char *>(end));
108 void PostingListBuilder::write_header(uint32_t docid)
110 unsigned char buf[P4NENC_BOUND(1)];
111 unsigned char *end = write_baseval(docid, buf);
112 encoded.append(reinterpret_cast<char *>(buf), end - buf);
115 void DictionaryBuilder::add_file(string filename, dir_time)
117 if (keep_current_block) { // Only bother saving the filenames if we're actually keeping the block.
118 if (!current_block.empty()) {
119 current_block.push_back('\0');
121 current_block += filename;
123 if (++num_files_in_block == block_size) {
128 void DictionaryBuilder::flush_block()
130 if (keep_current_block) {
131 if (slot_for_current_block == -1) {
132 lengths.push_back(current_block.size());
133 sampled_blocks.push_back(move(current_block));
135 lengths[slot_for_current_block] = current_block.size();
136 sampled_blocks[slot_for_current_block] = move(current_block);
139 current_block.clear();
140 num_files_in_block = 0;
143 if (block_num < blocks_to_keep) {
144 keep_current_block = true;
145 slot_for_current_block = -1;
147 // Keep every block with equal probability (reservoir sampling).
148 uint64_t idx = uniform_int_distribution<uint64_t>(0, block_num)(reservoir_rand);
149 keep_current_block = (idx < blocks_to_keep);
150 slot_for_current_block = idx;
154 string DictionaryBuilder::train(size_t buf_size)
156 string dictionary_buf;
157 sort(sampled_blocks.begin(), sampled_blocks.end()); // Seemingly important for decompression speed.
158 for (const string &block : sampled_blocks) {
159 dictionary_buf += block;
163 buf.resize(buf_size);
164 size_t ret = ZDICT_trainFromBuffer(&buf[0], buf_size, dictionary_buf.data(), lengths.data(), lengths.size());
165 if (ZDICT_isError(ret)) {
168 dprintf("Sampled %zu bytes in %zu blocks, built a dictionary of size %zu\n", dictionary_buf.size(), lengths.size(), ret);
171 sampled_blocks.clear();
177 class EncodingCorpus : public DatabaseReceiver {
179 EncodingCorpus(FILE *outfp, size_t block_size, ZSTD_CDict *cdict, bool store_dir_times);
182 void add_file(std::string filename, dir_time dt) override;
183 void flush_block() override;
184 void finish() override;
186 std::vector<uint64_t> filename_blocks;
187 size_t num_files = 0, num_files_in_block = 0, num_blocks = 0;
188 bool seen_trigram(uint32_t trgm)
190 return invindex[trgm] != nullptr;
192 size_t num_files_seen() const override { return num_files; }
193 PostingListBuilder &get_pl_builder(uint32_t trgm)
195 if (invindex[trgm] == nullptr) {
196 invindex[trgm] = new PostingListBuilder;
198 return *invindex[trgm];
201 size_t num_trigrams() const;
202 std::string get_compressed_dir_times();
205 void compress_dir_times(size_t allowed_slop);
207 std::unique_ptr<PostingListBuilder *[]> invindex;
209 std::string current_block;
211 const size_t block_size;
212 const bool store_dir_times;
215 ZSTD_CStream *dir_time_ctx = nullptr;
216 std::string dir_times; // Buffer of still-uncompressed data.
217 std::string dir_times_compressed;
221 EncodingCorpus::EncodingCorpus(FILE *outfp, size_t block_size, ZSTD_CDict *cdict, bool store_dir_times)
222 : invindex(new PostingListBuilder *[NUM_TRIGRAMS]), outfp(outfp), block_size(block_size), store_dir_times(store_dir_times), cdict(cdict)
224 fill(invindex.get(), invindex.get() + NUM_TRIGRAMS, nullptr);
225 if (store_dir_times) {
226 dir_time_ctx = ZSTD_createCStream();
227 ZSTD_initCStream(dir_time_ctx, /*level=*/6);
231 EncodingCorpus::~EncodingCorpus()
233 for (unsigned i = 0; i < NUM_TRIGRAMS; ++i) {
238 void EncodingCorpus::add_file(string filename, dir_time dt)
241 if (!current_block.empty()) {
242 current_block.push_back('\0');
244 current_block += filename;
245 if (++num_files_in_block == block_size) {
249 if (store_dir_times) {
252 dir_times.push_back('\0');
254 dir_times.push_back('\1');
255 dir_times.append(reinterpret_cast<char *>(&dt.sec), sizeof(dt.sec));
256 dir_times.append(reinterpret_cast<char *>(&dt.nsec), sizeof(dt.nsec));
258 compress_dir_times(/*allowed_slop=*/4096);
262 void EncodingCorpus::compress_dir_times(size_t allowed_slop)
264 while (dir_times.size() >= allowed_slop) {
265 size_t old_size = dir_times_compressed.size();
266 dir_times_compressed.resize(old_size + 4096);
268 ZSTD_outBuffer outbuf;
269 outbuf.dst = dir_times_compressed.data() + old_size;
274 inbuf.src = dir_times.data();
275 inbuf.size = dir_times.size();
278 int ret = ZSTD_compressStream(dir_time_ctx, &outbuf, &inbuf);
280 fprintf(stderr, "ZSTD_compressStream() failed\n");
284 dir_times_compressed.resize(old_size + outbuf.pos);
285 dir_times.erase(dir_times.begin(), dir_times.begin() + inbuf.pos);
287 if (outbuf.pos == 0 && inbuf.pos == 0) {
288 // Nothing happened (not enough data?), try again later.
294 void EncodingCorpus::flush_block()
296 if (current_block.empty()) {
300 uint32_t docid = num_blocks;
303 const char *ptr = current_block.c_str();
304 while (ptr < current_block.c_str() + current_block.size()) {
307 for (size_t j = 0; j < s.size() - 2; ++j) {
308 uint32_t trgm = read_trigram(s, j);
309 get_pl_builder(trgm).add_docid(docid);
315 // Compress and add the filename block.
316 filename_blocks.push_back(ftell(outfp));
317 string compressed = zstd_compress(current_block, cdict, &tempbuf);
318 if (fwrite(compressed.data(), compressed.size(), 1, outfp) != 1) {
323 current_block.clear();
324 num_files_in_block = 0;
328 void EncodingCorpus::finish()
333 size_t EncodingCorpus::num_trigrams() const
336 for (unsigned trgm = 0; trgm < NUM_TRIGRAMS; ++trgm) {
337 if (invindex[trgm] != nullptr) {
344 string EncodingCorpus::get_compressed_dir_times()
346 if (!store_dir_times) {
349 compress_dir_times(/*allowed_slop=*/0);
350 assert(dir_times.empty());
353 size_t old_size = dir_times_compressed.size();
354 dir_times_compressed.resize(old_size + 4096);
356 ZSTD_outBuffer outbuf;
357 outbuf.dst = dir_times_compressed.data() + old_size;
361 int ret = ZSTD_endStream(dir_time_ctx, &outbuf);
363 fprintf(stderr, "ZSTD_compressStream() failed\n");
367 dir_times_compressed.resize(old_size + outbuf.pos);
375 return dir_times_compressed;
378 string zstd_compress(const string &src, ZSTD_CDict *cdict, string *tempbuf)
380 static ZSTD_CCtx *ctx = nullptr;
381 if (ctx == nullptr) {
382 ctx = ZSTD_createCCtx();
385 size_t max_size = ZSTD_compressBound(src.size());
386 if (tempbuf->size() < max_size) {
387 tempbuf->resize(max_size);
390 if (cdict == nullptr) {
391 size = ZSTD_compressCCtx(ctx, &(*tempbuf)[0], max_size, src.data(), src.size(), /*level=*/6);
393 size = ZSTD_compress_usingCDict(ctx, &(*tempbuf)[0], max_size, src.data(), src.size(), cdict);
395 return string(tempbuf->data(), size);
398 bool is_prime(uint32_t x)
400 if ((x % 2) == 0 || (x % 3) == 0) {
403 uint32_t limit = ceil(sqrt(x));
404 for (uint32_t factor = 5; factor <= limit; ++factor) {
405 if ((x % factor) == 0) {
412 uint32_t next_prime(uint32_t x)
417 while (!is_prime(x)) {
423 unique_ptr<Trigram[]> create_hashtable(EncodingCorpus &corpus, const vector<uint32_t> &all_trigrams, uint32_t ht_size, uint32_t num_overflow_slots)
425 unique_ptr<Trigram[]> ht(new Trigram[ht_size + num_overflow_slots + 1]); // 1 for the sentinel element at the end.
426 for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
427 ht[i].trgm = uint32_t(-1);
428 ht[i].num_docids = 0;
431 for (uint32_t trgm : all_trigrams) {
432 // We don't know offset yet, so set it to zero.
433 Trigram to_insert{ trgm, uint32_t(corpus.get_pl_builder(trgm).num_docids), 0 };
435 uint32_t bucket = hash_trigram(trgm, ht_size);
436 unsigned distance = 0;
437 while (ht[bucket].num_docids != 0) {
438 // Robin Hood hashing; reduces the longest distance by a lot.
439 unsigned other_distance = bucket - hash_trigram(ht[bucket].trgm, ht_size);
440 if (distance > other_distance) {
441 swap(to_insert, ht[bucket]);
442 distance = other_distance;
445 ++bucket, ++distance;
446 if (distance > num_overflow_slots) {
450 ht[bucket] = to_insert;
455 DatabaseBuilder::DatabaseBuilder(const char *outfile, gid_t owner, int block_size, string dictionary, bool check_visibility)
456 : outfile(outfile), block_size(block_size)
460 string path = outfile;
461 path.resize(path.find_last_of('/') + 1);
466 int fd = open(path.c_str(), O_WRONLY | O_TMPFILE, 0640);
468 perror(path.c_str());
472 temp_filename = string(outfile) + ".XXXXXX";
473 int fd = mkstemp(&temp_filename[0]);
475 perror(temp_filename.c_str());
478 if (fchmod(fd, 0640) == -1) {
484 if (owner != (gid_t)-1) {
485 if (fchown(fd, (uid_t)-1, owner) == -1) {
491 outfp = fdopen(fd, "wb");
492 if (outfp == nullptr) {
498 memcpy(hdr.magic, "\0plocate", 8);
499 hdr.version = -1; // Mark as broken.
500 hdr.hashtable_size = 0; // Not known yet.
501 hdr.extra_ht_slots = num_overflow_slots;
503 hdr.hash_table_offset_bytes = -1; // We don't know these offsets yet.
505 hdr.filename_index_offset_bytes = -1;
506 hdr.zstd_dictionary_length_bytes = -1;
507 hdr.check_visibility = check_visibility;
508 fwrite(&hdr, sizeof(hdr), 1, outfp);
510 if (dictionary.empty()) {
511 hdr.zstd_dictionary_offset_bytes = 0;
512 hdr.zstd_dictionary_length_bytes = 0;
514 hdr.zstd_dictionary_offset_bytes = ftell(outfp);
515 fwrite(dictionary.data(), dictionary.size(), 1, outfp);
516 hdr.zstd_dictionary_length_bytes = dictionary.size();
517 cdict = ZSTD_createCDict(dictionary.data(), dictionary.size(), /*level=*/6);
520 hdr.directory_data_length_bytes = 0;
521 hdr.directory_data_offset_bytes = 0;
522 hdr.next_zstd_dictionary_length_bytes = 0;
523 hdr.next_zstd_dictionary_offset_bytes = 0;
524 hdr.conf_block_length_bytes = 0;
525 hdr.conf_block_offset_bytes = 0;
528 DatabaseReceiver *DatabaseBuilder::start_corpus(bool store_dir_times)
530 corpus_start = steady_clock::now();
531 corpus = new EncodingCorpus(outfp, block_size, cdict, store_dir_times);
535 void DatabaseBuilder::set_next_dictionary(std::string next_dictionary)
537 this->next_dictionary = move(next_dictionary);
540 void DatabaseBuilder::set_conf_block(std::string conf_block)
542 this->conf_block = move(conf_block);
545 void DatabaseBuilder::finish_corpus()
548 hdr.num_docids = corpus->filename_blocks.size();
550 // Stick an empty block at the end as sentinel.
551 corpus->filename_blocks.push_back(ftell(outfp));
552 const size_t bytes_for_filenames = corpus->filename_blocks.back() - corpus->filename_blocks.front();
554 // Write the offsets to the filenames.
555 hdr.filename_index_offset_bytes = ftell(outfp);
556 const size_t bytes_for_filename_index = corpus->filename_blocks.size() * sizeof(uint64_t);
557 fwrite(corpus->filename_blocks.data(), corpus->filename_blocks.size(), sizeof(uint64_t), outfp);
558 corpus->filename_blocks.clear();
559 corpus->filename_blocks.shrink_to_fit();
561 // Finish up encoding the posting lists.
562 size_t trigrams = 0, longest_posting_list = 0;
563 size_t bytes_for_posting_lists = 0;
564 for (unsigned trgm = 0; trgm < NUM_TRIGRAMS; ++trgm) {
565 if (!corpus->seen_trigram(trgm))
567 PostingListBuilder &pl_builder = corpus->get_pl_builder(trgm);
569 longest_posting_list = max(longest_posting_list, pl_builder.num_docids);
570 trigrams += pl_builder.num_docids;
571 bytes_for_posting_lists += pl_builder.encoded.size();
573 size_t num_trigrams = corpus->num_trigrams();
574 dprintf("%zu files, %zu different trigrams, %zu entries, avg len %.2f, longest %zu\n",
575 corpus->num_files, num_trigrams, trigrams, double(trigrams) / num_trigrams, longest_posting_list);
576 dprintf("%zu bytes used for posting lists (%.2f bits/entry)\n", bytes_for_posting_lists, 8 * bytes_for_posting_lists / double(trigrams));
578 dprintf("Building posting lists took %.1f ms.\n\n", 1e3 * duration<float>(steady_clock::now() - corpus_start).count());
580 // Find the used trigrams.
581 vector<uint32_t> all_trigrams;
582 for (unsigned trgm = 0; trgm < NUM_TRIGRAMS; ++trgm) {
583 if (corpus->seen_trigram(trgm)) {
584 all_trigrams.push_back(trgm);
588 // Create the hash table.
589 unique_ptr<Trigram[]> hashtable;
590 uint32_t ht_size = next_prime(all_trigrams.size());
592 hashtable = create_hashtable(*corpus, all_trigrams, ht_size, num_overflow_slots);
593 if (hashtable == nullptr) {
594 dprintf("Failed creating hash table of size %u, increasing by 5%% and trying again.\n", ht_size);
595 ht_size = next_prime(ht_size * 1.05);
597 dprintf("Created hash table of size %u.\n\n", ht_size);
602 // Find the offsets for each posting list.
603 size_t bytes_for_hashtable = (ht_size + num_overflow_slots + 1) * sizeof(Trigram);
604 uint64_t offset = ftell(outfp) + bytes_for_hashtable;
605 for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
606 hashtable[i].offset = offset; // Needs to be there even for empty slots.
607 if (hashtable[i].num_docids == 0) {
611 const string &encoded = corpus->get_pl_builder(hashtable[i].trgm).encoded;
612 offset += encoded.size();
615 // Write the hash table.
616 hdr.hash_table_offset_bytes = ftell(outfp);
617 hdr.hashtable_size = ht_size;
618 fwrite(hashtable.get(), ht_size + num_overflow_slots + 1, sizeof(Trigram), outfp);
620 // Write the actual posting lists.
621 for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
622 if (hashtable[i].num_docids == 0) {
625 const string &encoded = corpus->get_pl_builder(hashtable[i].trgm).encoded;
626 fwrite(encoded.data(), encoded.size(), 1, outfp);
629 // Finally, write the directory times (for updatedb).
630 string compressed_dir_times = corpus->get_compressed_dir_times();
631 size_t bytes_for_compressed_dir_times = 0;
632 if (!compressed_dir_times.empty()) {
633 hdr.directory_data_offset_bytes = ftell(outfp);
634 hdr.directory_data_length_bytes = compressed_dir_times.size();
635 fwrite(compressed_dir_times.data(), compressed_dir_times.size(), 1, outfp);
636 bytes_for_compressed_dir_times = compressed_dir_times.size();
637 compressed_dir_times.clear();
640 // Write the recommended dictionary for next update.
641 if (!next_dictionary.empty()) {
642 hdr.next_zstd_dictionary_offset_bytes = ftell(outfp);
643 hdr.next_zstd_dictionary_length_bytes = next_dictionary.size();
644 fwrite(next_dictionary.data(), next_dictionary.size(), 1, outfp);
647 // And the configuration block.
648 if (!conf_block.empty()) {
649 hdr.conf_block_offset_bytes = ftell(outfp);
650 hdr.conf_block_length_bytes = conf_block.size();
651 fwrite(conf_block.data(), conf_block.size(), 1, outfp);
654 // Rewind, and write the updated header.
656 fseek(outfp, 0, SEEK_SET);
657 fwrite(&hdr, sizeof(hdr), 1, outfp);
660 // Give the file a proper name, making it visible in the file system.
661 // TODO: It would be nice to be able to do this atomically, like with rename.
662 unlink(outfile.c_str());
664 snprintf(procpath, sizeof(procpath), "/proc/self/fd/%d", fileno(outfp));
665 if (linkat(AT_FDCWD, procpath, AT_FDCWD, outfile.c_str(), AT_SYMLINK_FOLLOW) == -1) {
670 if (rename(temp_filename.c_str(), outfile.c_str()) == -1) {
678 size_t total_bytes = (bytes_for_hashtable + bytes_for_posting_lists + bytes_for_filename_index + bytes_for_filenames + bytes_for_compressed_dir_times);
680 dprintf("Block size: %7d files\n", block_size);
681 dprintf("Dictionary: %'7.1f MB\n", hdr.zstd_dictionary_length_bytes / 1048576.0);
682 dprintf("Hash table: %'7.1f MB\n", bytes_for_hashtable / 1048576.0);
683 dprintf("Posting lists: %'7.1f MB\n", bytes_for_posting_lists / 1048576.0);
684 dprintf("Filename index: %'7.1f MB\n", bytes_for_filename_index / 1048576.0);
685 dprintf("Filenames: %'7.1f MB\n", bytes_for_filenames / 1048576.0);
686 if (bytes_for_compressed_dir_times != 0) {
687 dprintf("Modify times: %'7.1f MB\n", bytes_for_compressed_dir_times / 1048576.0);
689 dprintf("Total: %'7.1f MB\n", total_bytes / 1048576.0);