X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=plocate-build.cpp;h=62be45dab81e70a1fd94ca8fa29bc6163487a766;hb=e1c30afeca4717fd2d123224f82901f305f92451;hp=90141ffd1df2e14a026f66b242e552a788f28d99;hpb=3dca6d5f3d30118f5b12f3613c7873ed6c31fb2b;p=plocate diff --git a/plocate-build.cpp b/plocate-build.cpp index 90141ff..62be45d 100644 --- a/plocate-build.cpp +++ b/plocate-build.cpp @@ -1,36 +1,38 @@ #include "db.h" -#include "vp4.h" +#include "dprintf.h" +#include "turbopfor-encode.h" #include -#include #include #include -#include -#include +#include +#include #include #include +#include +#include #include +#include #include #include +#include #include -#include -#include -#include +#include #include +#include #include #define P4NENC_BOUND(n) ((n + 127) / 128 + (n + 32) * sizeof(uint32_t)) -#define dprintf(...) -//#define dprintf(...) fprintf(stderr, __VA_ARGS__); #define NUM_TRIGRAMS 16777216 using namespace std; using namespace std::chrono; -string zstd_compress(const string &src, string *tempbuf); +string zstd_compress(const string &src, ZSTD_CDict *cdict, string *tempbuf); constexpr unsigned num_overflow_slots = 16; +bool use_debug = false; static inline uint32_t read_unigram(const string_view s, size_t idx) { @@ -122,7 +124,7 @@ void PostingListBuilder::finish() // No interleaving for partial blocks. unsigned char buf[P4NENC_BOUND(128)]; - unsigned char *end = p4enc32(pending_deltas.data(), pending_deltas.size(), buf); + unsigned char *end = encode_pfor_single_block<128>(pending_deltas.data(), pending_deltas.size(), /*interleaved=*/false, buf); encoded.append(reinterpret_cast(buf), reinterpret_cast(end)); } @@ -130,23 +132,124 @@ void PostingListBuilder::append_block() { unsigned char buf[P4NENC_BOUND(128)]; assert(pending_deltas.size() == 128); - unsigned char *end = p4enc128v32(pending_deltas.data(), 128, buf); + unsigned char *end = encode_pfor_single_block<128>(pending_deltas.data(), 128, /*interleaved=*/true, buf); encoded.append(reinterpret_cast(buf), reinterpret_cast(end)); } void PostingListBuilder::write_header(uint32_t docid) { unsigned char buf[P4NENC_BOUND(1)]; - size_t bytes = p4nd1enc128v32(&docid, 1, buf); - encoded.append(reinterpret_cast(buf), bytes); + unsigned char *end = write_baseval(docid, buf); + encoded.append(reinterpret_cast(buf), end - buf); } -class Corpus { +class DatabaseReceiver { public: - Corpus(FILE *outfp, size_t block_size) - : invindex(new PostingListBuilder *[NUM_TRIGRAMS]), outfp(outfp), block_size(block_size) {} - void add_file(string filename); - void flush_block(); + virtual ~DatabaseReceiver() = default; + virtual void add_file(string filename) = 0; + virtual void flush_block() = 0; +}; + +class DictionaryBuilder : public DatabaseReceiver { +public: + DictionaryBuilder(size_t blocks_to_keep, size_t block_size) + : blocks_to_keep(blocks_to_keep), block_size(block_size) {} + void add_file(string filename) override; + void flush_block() override; + string train(size_t buf_size); + +private: + const size_t blocks_to_keep, block_size; + string current_block; + uint64_t block_num = 0; + size_t num_files_in_block = 0; + + std::mt19937 reservoir_rand{ 1234 }; // Fixed seed for reproducibility. + bool keep_current_block = true; + int64_t slot_for_current_block = -1; + + vector sampled_blocks; + vector lengths; +}; + +void DictionaryBuilder::add_file(string filename) +{ + if (keep_current_block) { // Only bother saving the filenames if we're actually keeping the block. + if (!current_block.empty()) { + current_block.push_back('\0'); + } + current_block += filename; + } + if (++num_files_in_block == block_size) { + flush_block(); + } +} + +void DictionaryBuilder::flush_block() +{ + if (keep_current_block) { + if (slot_for_current_block == -1) { + lengths.push_back(current_block.size()); + sampled_blocks.push_back(move(current_block)); + } else { + lengths[slot_for_current_block] = current_block.size(); + sampled_blocks[slot_for_current_block] = move(current_block); + } + } + current_block.clear(); + num_files_in_block = 0; + ++block_num; + + if (block_num < blocks_to_keep) { + keep_current_block = true; + slot_for_current_block = -1; + } else { + // Keep every block with equal probability (reservoir sampling). + uint64_t idx = uniform_int_distribution(0, block_num)(reservoir_rand); + keep_current_block = (idx < blocks_to_keep); + slot_for_current_block = idx; + } +} + +string DictionaryBuilder::train(size_t buf_size) +{ + string dictionary_buf; + sort(sampled_blocks.begin(), sampled_blocks.end()); // Seemingly important for decompression speed. + for (const string &block : sampled_blocks) { + dictionary_buf += block; + } + + string buf; + buf.resize(buf_size); + size_t ret = ZDICT_trainFromBuffer(&buf[0], buf_size, dictionary_buf.data(), lengths.data(), lengths.size()); + if (ret == size_t(-1)) { + return ""; + } + dprintf("Sampled %zu bytes in %zu blocks, built a dictionary of size %zu\n", dictionary_buf.size(), lengths.size(), ret); + buf.resize(ret); + + sampled_blocks.clear(); + lengths.clear(); + + return buf; +} + +class Corpus : public DatabaseReceiver { +public: + Corpus(FILE *outfp, size_t block_size, ZSTD_CDict *cdict) + : invindex(new PostingListBuilder *[NUM_TRIGRAMS]), outfp(outfp), block_size(block_size), cdict(cdict) + { + fill(invindex.get(), invindex.get() + NUM_TRIGRAMS, nullptr); + } + ~Corpus() override + { + for (unsigned i = 0; i < NUM_TRIGRAMS; ++i) { + delete invindex[i]; + } + } + + void add_file(string filename) override; + void flush_block() override; vector filename_blocks; size_t num_files = 0, num_files_in_block = 0, num_blocks = 0; @@ -161,6 +264,7 @@ public: } return *invindex[trgm]; } + size_t num_trigrams() const; private: unique_ptr invindex; @@ -168,6 +272,7 @@ private: string current_block; string tempbuf; const size_t block_size; + ZSTD_CDict *cdict; }; void Corpus::add_file(string filename) @@ -205,7 +310,7 @@ void Corpus::flush_block() // Compress and add the filename block. filename_blocks.push_back(ftell(outfp)); - string compressed = zstd_compress(current_block, &tempbuf); + string compressed = zstd_compress(current_block, cdict, &tempbuf); if (fwrite(compressed.data(), compressed.size(), 1, outfp) != 1) { perror("fwrite()"); exit(1); @@ -216,6 +321,17 @@ void Corpus::flush_block() ++num_blocks; } +size_t Corpus::num_trigrams() const +{ + size_t num = 0; + for (unsigned trgm = 0; trgm < NUM_TRIGRAMS; ++trgm) { + if (invindex[trgm] != nullptr) { + ++num; + } + } + return num; +} + string read_cstr(FILE *fp) { string ret; @@ -232,7 +348,7 @@ string read_cstr(FILE *fp) } } -void handle_directory(FILE *fp, Corpus *corpus) +void handle_directory(FILE *fp, DatabaseReceiver *receiver) { db_directory dummy; if (fread(&dummy, sizeof(dummy), 1, fp) != 1) { @@ -252,21 +368,47 @@ void handle_directory(FILE *fp, Corpus *corpus) int type = getc(fp); if (type == DBE_NORMAL) { string filename = read_cstr(fp); - corpus->add_file(dir_path + "/" + filename); + receiver->add_file(dir_path + "/" + filename); } else if (type == DBE_DIRECTORY) { string dirname = read_cstr(fp); - corpus->add_file(dir_path + "/" + dirname); + receiver->add_file(dir_path + "/" + dirname); } else { return; // Probably end. } } } -void read_mlocate(const char *filename, Corpus *corpus) +void read_plaintext(FILE *fp, DatabaseReceiver *receiver) +{ + if (fseek(fp, 0, SEEK_SET) != 0) { + perror("fseek"); + exit(1); + } + + while (!feof(fp)) { + char buf[1024]; + if (fgets(buf, sizeof(buf), fp) == nullptr) { + break; + } + string s(buf); + assert(!s.empty()); + while (s.back() != '\n' && !feof(fp)) { + // The string was longer than the buffer, so read again. + if (fgets(buf, sizeof(buf), fp) == nullptr) { + break; + } + s += buf; + } + if (!s.empty() && s.back() == '\n') + s.pop_back(); + receiver->add_file(move(s)); + } +} + +void read_mlocate(FILE *fp, DatabaseReceiver *receiver) { - FILE *fp = fopen(filename, "rb"); - if (fp == nullptr) { - perror(filename); + if (fseek(fp, 0, SEEK_SET) != 0) { + perror("fseek"); exit(1); } @@ -279,18 +421,27 @@ void read_mlocate(const char *filename, Corpus *corpus) // TODO: Care about the base path. string path = read_cstr(fp); while (!feof(fp)) { - handle_directory(fp, corpus); + handle_directory(fp, receiver); } - fclose(fp); } -string zstd_compress(const string &src, string *tempbuf) +string zstd_compress(const string &src, ZSTD_CDict *cdict, string *tempbuf) { + static ZSTD_CCtx *ctx = nullptr; + if (ctx == nullptr) { + ctx = ZSTD_createCCtx(); + } + size_t max_size = ZSTD_compressBound(src.size()); if (tempbuf->size() < max_size) { tempbuf->resize(max_size); } - size_t size = ZSTD_compress(&(*tempbuf)[0], max_size, src.data(), src.size(), /*level=*/6); + size_t size; + if (cdict == nullptr) { + size = ZSTD_compressCCtx(ctx, &(*tempbuf)[0], max_size, src.data(), src.size(), /*level=*/6); + } else { + size = ZSTD_compress_usingCDict(ctx, &(*tempbuf)[0], max_size, src.data(), src.size(), cdict); + } return string(tempbuf->data(), size); } @@ -351,77 +502,100 @@ unique_ptr create_hashtable(Corpus &corpus, const vector &a return ht; } -void do_build(const char *infile, const char *outfile, int block_size) -{ - steady_clock::time_point start __attribute__((unused)) = steady_clock::now(); +class DatabaseBuilder { +public: + DatabaseBuilder(const char *outfile, int block_size, string dictionary); + Corpus *start_corpus(); + void finish_corpus(); +private: + FILE *outfp; + Header hdr; + const int block_size; + steady_clock::time_point corpus_start; + Corpus *corpus = nullptr; + ZSTD_CDict *cdict = nullptr; +}; + +DatabaseBuilder::DatabaseBuilder(const char *outfile, int block_size, string dictionary) + : block_size(block_size) +{ umask(0027); - FILE *outfp = fopen(outfile, "wb"); + outfp = fopen(outfile, "wb"); + if (outfp == nullptr) { + perror(outfile); + exit(1); + } // Write the header. - Header hdr; memcpy(hdr.magic, "\0plocate", 8); hdr.version = -1; // Mark as broken. hdr.hashtable_size = 0; // Not known yet. hdr.extra_ht_slots = num_overflow_slots; + hdr.num_docids = 0; hdr.hash_table_offset_bytes = -1; // We don't know these offsets yet. + hdr.max_version = 1; hdr.filename_index_offset_bytes = -1; + hdr.zstd_dictionary_length_bytes = -1; fwrite(&hdr, sizeof(hdr), 1, outfp); - Corpus corpus(outfp, block_size); - - read_mlocate(infile, &corpus); - if (false) { // To read a plain text file. - FILE *fp = fopen(infile, "r"); - while (!feof(fp)) { - char buf[1024]; - if (fgets(buf, 1024, fp) == nullptr || feof(fp)) { - break; - } - string s(buf); - if (s.back() == '\n') - s.pop_back(); - corpus.add_file(move(s)); - } - fclose(fp); + if (dictionary.empty()) { + hdr.zstd_dictionary_offset_bytes = 0; + hdr.zstd_dictionary_length_bytes = 0; + } else { + hdr.zstd_dictionary_offset_bytes = ftell(outfp); + fwrite(dictionary.data(), dictionary.size(), 1, outfp); + hdr.zstd_dictionary_length_bytes = dictionary.size(); + cdict = ZSTD_createCDict(dictionary.data(), dictionary.size(), /*level=*/6); } - corpus.flush_block(); - dprintf("Read %zu files from %s\n", corpus.num_files, infile); - hdr.num_docids = corpus.filename_blocks.size(); +} + +Corpus *DatabaseBuilder::start_corpus() +{ + corpus_start = steady_clock::now(); + corpus = new Corpus(outfp, block_size, cdict); + return corpus; +} + +void DatabaseBuilder::finish_corpus() +{ + corpus->flush_block(); + hdr.num_docids = corpus->filename_blocks.size(); // Stick an empty block at the end as sentinel. - corpus.filename_blocks.push_back(ftell(outfp)); - const size_t bytes_for_filenames = corpus.filename_blocks.back() - corpus.filename_blocks.front(); + corpus->filename_blocks.push_back(ftell(outfp)); + const size_t bytes_for_filenames = corpus->filename_blocks.back() - corpus->filename_blocks.front(); // Write the offsets to the filenames. hdr.filename_index_offset_bytes = ftell(outfp); - const size_t bytes_for_filename_index = corpus.filename_blocks.size() * sizeof(uint64_t); - fwrite(corpus.filename_blocks.data(), corpus.filename_blocks.size(), sizeof(uint64_t), outfp); - corpus.filename_blocks.clear(); - corpus.filename_blocks.shrink_to_fit(); + const size_t bytes_for_filename_index = corpus->filename_blocks.size() * sizeof(uint64_t); + fwrite(corpus->filename_blocks.data(), corpus->filename_blocks.size(), sizeof(uint64_t), outfp); + corpus->filename_blocks.clear(); + corpus->filename_blocks.shrink_to_fit(); // Finish up encoding the posting lists. size_t trigrams = 0, longest_posting_list = 0; size_t bytes_for_posting_lists = 0; for (unsigned trgm = 0; trgm < NUM_TRIGRAMS; ++trgm) { - if (!corpus.seen_trigram(trgm)) + if (!corpus->seen_trigram(trgm)) continue; - PostingListBuilder &pl_builder = corpus.get_pl_builder(trgm); + PostingListBuilder &pl_builder = corpus->get_pl_builder(trgm); pl_builder.finish(); longest_posting_list = max(longest_posting_list, pl_builder.num_docids); trigrams += pl_builder.num_docids; bytes_for_posting_lists += pl_builder.encoded.size(); } + size_t num_trigrams = corpus->num_trigrams(); dprintf("%zu files, %zu different trigrams, %zu entries, avg len %.2f, longest %zu\n", - corpus.num_files, corpus.invindex.size(), trigrams, double(trigrams) / corpus.invindex.size(), longest_posting_list); + corpus->num_files, num_trigrams, trigrams, double(trigrams) / num_trigrams, longest_posting_list); dprintf("%zu bytes used for posting lists (%.2f bits/entry)\n", bytes_for_posting_lists, 8 * bytes_for_posting_lists / double(trigrams)); - dprintf("Building posting lists took %.1f ms.\n\n", 1e3 * duration(steady_clock::now() - start).count()); + dprintf("Building posting lists took %.1f ms.\n\n", 1e3 * duration(steady_clock::now() - corpus_start).count()); // Find the used trigrams. vector all_trigrams; for (unsigned trgm = 0; trgm < NUM_TRIGRAMS; ++trgm) { - if (corpus.seen_trigram(trgm)) { + if (corpus->seen_trigram(trgm)) { all_trigrams.push_back(trgm); } } @@ -430,7 +604,7 @@ void do_build(const char *infile, const char *outfile, int block_size) unique_ptr hashtable; uint32_t ht_size = next_prime(all_trigrams.size()); for (;;) { - hashtable = create_hashtable(corpus, all_trigrams, ht_size, num_overflow_slots); + hashtable = create_hashtable(*corpus, all_trigrams, ht_size, num_overflow_slots); if (hashtable == nullptr) { dprintf("Failed creating hash table of size %u, increasing by 5%% and trying again.\n", ht_size); ht_size = next_prime(ht_size * 1.05); @@ -449,7 +623,7 @@ void do_build(const char *infile, const char *outfile, int block_size) continue; } - const string &encoded = corpus.get_pl_builder(hashtable[i].trgm).encoded; + const string &encoded = corpus->get_pl_builder(hashtable[i].trgm).encoded; offset += encoded.size(); } @@ -463,19 +637,20 @@ void do_build(const char *infile, const char *outfile, int block_size) if (hashtable[i].num_docids == 0) { continue; } - const string &encoded = corpus.get_pl_builder(hashtable[i].trgm).encoded; + const string &encoded = corpus->get_pl_builder(hashtable[i].trgm).encoded; fwrite(encoded.data(), encoded.size(), 1, outfp); } // Rewind, and write the updated header. - hdr.version = 0; + hdr.version = 1; fseek(outfp, 0, SEEK_SET); fwrite(&hdr, sizeof(hdr), 1, outfp); fclose(outfp); - size_t total_bytes __attribute__((unused)) = (bytes_for_hashtable + bytes_for_posting_lists + bytes_for_filename_index + bytes_for_filenames); + size_t total_bytes = (bytes_for_hashtable + bytes_for_posting_lists + bytes_for_filename_index + bytes_for_filenames); dprintf("Block size: %7d files\n", block_size); + dprintf("Dictionary: %'7.1f MB\n", hdr.zstd_dictionary_length_bytes / 1048576.0); dprintf("Hash table: %'7.1f MB\n", bytes_for_hashtable / 1048576.0); dprintf("Posting lists: %'7.1f MB\n", bytes_for_posting_lists / 1048576.0); dprintf("Filename index: %'7.1f MB\n", bytes_for_filename_index / 1048576.0); @@ -484,8 +659,109 @@ void do_build(const char *infile, const char *outfile, int block_size) dprintf("\n"); } +void do_build(const char *infile, const char *outfile, int block_size, bool plaintext) +{ + FILE *infp = fopen(infile, "rb"); + if (infp == nullptr) { + perror(infile); + exit(1); + } + + // Train the dictionary by sampling real blocks. + // The documentation for ZDICT_trainFromBuffer() claims that a reasonable + // dictionary size is ~100 kB, but 1 kB seems to actually compress better for us, + // and decompress just as fast. + DictionaryBuilder builder(/*blocks_to_keep=*/1000, block_size); + if (plaintext) { + read_plaintext(infp, &builder); + } else { + read_mlocate(infp, &builder); + } + string dictionary = builder.train(1024); + + DatabaseBuilder db(outfile, block_size, dictionary); + Corpus *corpus = db.start_corpus(); + if (plaintext) { + read_plaintext(infp, corpus); + } else { + read_mlocate(infp, corpus); + } + fclose(infp); + + dprintf("Read %zu files from %s\n", corpus->num_files, infile); + db.finish_corpus(); +} + +void usage() +{ + printf( + "Usage: plocate-build MLOCATE_DB PLOCATE_DB\n" + "\n" + "Generate plocate index from mlocate.db, typically /var/lib/mlocate/mlocate.db.\n" + "Normally, the destination should be /var/lib/mlocate/plocate.db.\n" + "\n" + " -b, --block-size SIZE number of filenames to store in each block (default 32)\n" + " -p, --plaintext input is a plaintext file, not an mlocate database\n" + " --help print this help\n" + " --version print version information\n"); +} + +void version() +{ + printf("plocate-build %s\n", PLOCATE_VERSION); + printf("Copyright 2020 Steinar H. Gunderson\n"); + printf("License GPLv2+: GNU GPL version 2 or later .\n"); + printf("This is free software: you are free to change and redistribute it.\n"); + printf("There is NO WARRANTY, to the extent permitted by law.\n"); +} + int main(int argc, char **argv) { - do_build(argv[1], argv[2], 32); + static const struct option long_options[] = { + { "block-size", required_argument, 0, 'b' }, + { "plaintext", no_argument, 0, 'p' }, + { "help", no_argument, 0, 'h' }, + { "version", no_argument, 0, 'V' }, + { "debug", no_argument, 0, 'D' }, // Not documented. + { 0, 0, 0, 0 } + }; + + int block_size = 32; + bool plaintext = false; + + setlocale(LC_ALL, ""); + for (;;) { + int option_index = 0; + int c = getopt_long(argc, argv, "b:hpVD", long_options, &option_index); + if (c == -1) { + break; + } + switch (c) { + case 'b': + block_size = atoi(optarg); + break; + case 'p': + plaintext = true; + break; + case 'h': + usage(); + exit(0); + case 'V': + version(); + exit(0); + case 'D': + use_debug = true; + break; + default: + exit(1); + } + } + + if (argc - optind != 2) { + usage(); + exit(1); + } + + do_build(argv[optind], argv[optind + 1], block_size, plaintext); exit(EXIT_SUCCESS); }