#include "db.h"
-#include "vp4.h"
+#include "dprintf.h"
+#include "turbopfor-encode.h"
#include <algorithm>
-#include <arpa/inet.h>
#include <assert.h>
#include <chrono>
-#include <endian.h>
-#include <fcntl.h>
+#include <getopt.h>
+#include <iosfwd>
#include <math.h>
#include <memory>
+#include <random>
+#include <stdint.h>
#include <stdio.h>
+#include <stdlib.h>
#include <string.h>
#include <string>
+#include <string_view>
#include <sys/stat.h>
-#include <sys/types.h>
-#include <unistd.h>
-#include <unordered_map>
+#include <utility>
#include <vector>
+#include <zdict.h>
#include <zstd.h>
#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)
{
class PostingListBuilder {
public:
- void add_docid(uint32_t docid);
+ inline void add_docid(uint32_t docid);
void finish();
string encoded;
void write_header(uint32_t docid);
void append_block();
- vector<uint32_t> pending_docids;
+ vector<uint32_t> pending_deltas;
- uint32_t last_block_end;
+ uint32_t last_block_end, last_docid = -1;
};
void PostingListBuilder::add_docid(uint32_t docid)
{
// Deduplicate against the last inserted value, if any.
- if (pending_docids.empty()) {
- if (encoded.empty()) {
- // Very first docid.
- write_header(docid);
- ++num_docids;
- last_block_end = docid;
- return;
- } else if (docid == last_block_end) {
- return;
- }
- } else {
- if (docid == pending_docids.back()) {
- return;
- }
+ if (docid == last_docid) {
+ return;
+ }
+
+ if (num_docids == 0) {
+ // Very first docid.
+ write_header(docid);
+ ++num_docids;
+ last_block_end = last_docid = docid;
+ return;
}
- pending_docids.push_back(docid);
- if (pending_docids.size() == 128) {
+ pending_deltas.push_back(docid - last_docid - 1);
+ last_docid = docid;
+ if (pending_deltas.size() == 128) {
append_block();
- pending_docids.clear();
+ pending_deltas.clear();
last_block_end = docid;
}
++num_docids;
void PostingListBuilder::finish()
{
- if (pending_docids.empty()) {
+ if (pending_deltas.empty()) {
return;
}
// No interleaving for partial blocks.
unsigned char buf[P4NENC_BOUND(128)];
- unsigned char *end = p4d1enc32(pending_docids.data(), pending_docids.size(), buf, last_block_end);
+ unsigned char *end = encode_pfor_single_block<128>(pending_deltas.data(), pending_deltas.size(), /*interleaved=*/false, buf);
encoded.append(reinterpret_cast<char *>(buf), reinterpret_cast<char *>(end));
}
void PostingListBuilder::append_block()
{
unsigned char buf[P4NENC_BOUND(128)];
- assert(pending_docids.size() == 128);
- unsigned char *end = p4d1enc128v32(pending_docids.data(), 128, buf, last_block_end);
+ assert(pending_deltas.size() == 128);
+ unsigned char *end = encode_pfor_single_block<128>(pending_deltas.data(), 128, /*interleaved=*/true, buf);
encoded.append(reinterpret_cast<char *>(buf), reinterpret_cast<char *>(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<char *>(buf), bytes);
+ unsigned char *end = write_baseval(docid, buf);
+ encoded.append(reinterpret_cast<char *>(buf), end - buf);
+}
+
+class DatabaseReceiver {
+public:
+ 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<string> sampled_blocks;
+ vector<size_t> 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<uint64_t>(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());
+ 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 {
+class Corpus : public DatabaseReceiver {
public:
- Corpus(FILE *outfp, size_t block_size)
- : outfp(outfp), block_size(block_size) {}
- void add_file(string filename);
- void flush_block();
+ 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<uint64_t> filename_blocks;
- unordered_map<uint32_t, PostingListBuilder> invindex;
size_t num_files = 0, num_files_in_block = 0, num_blocks = 0;
+ bool seen_trigram(uint32_t trgm)
+ {
+ return invindex[trgm] != nullptr;
+ }
+ PostingListBuilder &get_pl_builder(uint32_t trgm)
+ {
+ if (invindex[trgm] == nullptr) {
+ invindex[trgm] = new PostingListBuilder;
+ }
+ return *invindex[trgm];
+ }
+ size_t num_trigrams() const;
private:
+ unique_ptr<PostingListBuilder *[]> invindex;
FILE *outfp;
string current_block;
string tempbuf;
const size_t block_size;
+ ZSTD_CDict *cdict;
};
void Corpus::add_file(string filename)
if (s.size() >= 3) {
for (size_t j = 0; j < s.size() - 2; ++j) {
uint32_t trgm = read_trigram(s, j);
- invindex[trgm].add_docid(docid);
+ get_pl_builder(trgm).add_docid(docid);
}
}
ptr += s.size() + 1;
// 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);
++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;
- for ( ;; ) {
+ for (;;) {
int ch = getc(fp);
if (ch == -1) {
perror("getc");
}
}
-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) {
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_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);
}
// 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);
}
return x;
}
-unique_ptr<Trigram[]> create_hashtable(const Corpus &corpus, const vector<uint32_t> &all_trigrams, uint32_t ht_size, uint32_t num_overflow_slots)
+unique_ptr<Trigram[]> create_hashtable(Corpus &corpus, const vector<uint32_t> &all_trigrams, uint32_t ht_size, uint32_t num_overflow_slots)
{
unique_ptr<Trigram[]> ht(new Trigram[ht_size + num_overflow_slots + 1]); // 1 for the sentinel element at the end.
for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
}
for (uint32_t trgm : all_trigrams) {
// We don't know offset yet, so set it to zero.
- Trigram to_insert{ trgm, uint32_t(corpus.invindex.find(trgm)->second.num_docids), 0 };
+ Trigram to_insert{ trgm, uint32_t(corpus.get_pl_builder(trgm).num_docids), 0 };
uint32_t bucket = hash_trigram(trgm, ht_size);
unsigned distance = 0;
{
steady_clock::time_point start __attribute__((unused)) = steady_clock::now();
+ FILE *infp = fopen(infile, "rb");
+ if (infp == nullptr) {
+ perror(infile);
+ exit(1);
+ }
+
umask(0027);
FILE *outfp = fopen(outfile, "wb");
+ if (outfp == nullptr) {
+ perror(outfile);
+ exit(1);
+ }
// Write the header.
Header hdr;
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);
+ // 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);
+ read_mlocate(infp, &builder);
+ string dictionary = builder.train(1024);
+ ZSTD_CDict *cdict = ZSTD_createCDict(dictionary.data(), dictionary.size(), /*level=*/6);
+
+ hdr.zstd_dictionary_offset_bytes = ftell(outfp);
+ fwrite(dictionary.data(), dictionary.size(), 1, outfp);
+ hdr.zstd_dictionary_length_bytes = dictionary.size();
+
+ Corpus corpus(outfp, block_size, cdict);
+ read_mlocate(infp, &corpus);
+ fclose(infp);
- read_mlocate(infile, &corpus);
if (false) { // To read a plain text file.
FILE *fp = fopen(infile, "r");
while (!feof(fp)) {
}
corpus.flush_block();
dprintf("Read %zu files from %s\n", corpus.num_files, infile);
+ hdr.num_docids = corpus.filename_blocks.size();
// Stick an empty block at the end as sentinel.
corpus.filename_blocks.push_back(ftell(outfp));
// Finish up encoding the posting lists.
size_t trigrams = 0, longest_posting_list = 0;
size_t bytes_for_posting_lists = 0;
- for (auto &[trigram, pl_builder] : corpus.invindex) {
+ for (unsigned trgm = 0; trgm < NUM_TRIGRAMS; ++trgm) {
+ if (!corpus.seen_trigram(trgm))
+ continue;
+ 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<float>(steady_clock::now() - start).count());
- // Sort the trigrams, mostly to get a consistent result every time
- // (the hash table will put things in random order anyway).
+ // Find the used trigrams.
vector<uint32_t> all_trigrams;
- for (auto &[trigram, pl_builder] : corpus.invindex) {
- all_trigrams.push_back(trigram);
+ for (unsigned trgm = 0; trgm < NUM_TRIGRAMS; ++trgm) {
+ if (corpus.seen_trigram(trgm)) {
+ all_trigrams.push_back(trgm);
+ }
}
- sort(all_trigrams.begin(), all_trigrams.end());
// Create the hash table.
unique_ptr<Trigram[]> hashtable;
continue;
}
- const string &encoded = corpus.invindex[hashtable[i].trgm].encoded;
+ const string &encoded = corpus.get_pl_builder(hashtable[i].trgm).encoded;
offset += encoded.size();
}
if (hashtable[i].num_docids == 0) {
continue;
}
- const string &encoded = corpus.invindex[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);
dprintf("Block size: %7d files\n", block_size);
+ dprintf("Dictionary: %'7.1f MB\n", dictionary.size() / 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);
dprintf("\n");
}
+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"
+ " --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 <https://gnu.org/licenses/gpl.html>.\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' },
+ { "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;
+
+ setlocale(LC_ALL, "");
+ for (;;) {
+ int option_index = 0;
+ int c = getopt_long(argc, argv, "b:hVD", long_options, &option_index);
+ if (c == -1) {
+ break;
+ }
+ switch (c) {
+ case 'b':
+ block_size = atoi(optarg);
+ 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);
exit(EXIT_SUCCESS);
}