X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=plocate-build.cpp;h=e1e2f644ef2ca891f93d48fdaab15e4f681bf418;hb=db80346f0412329379cfde7a6aa0de8fd731ce4a;hp=53cefde64c2149064d43a41fba3639143ec32b2b;hpb=994138463a9df08adbd7759049e8da415cba3e59;p=plocate

diff --git a/plocate-build.cpp b/plocate-build.cpp
index 53cefde..e1e2f64 100644
--- a/plocate-build.cpp
+++ b/plocate-build.cpp
@@ -1,19 +1,21 @@
-#include "vp4.h"
+#include "db.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 <stdint.h>
 #include <stdio.h>
+#include <stdlib.h>
 #include <string.h>
 #include <string>
-#include <sys/mman.h>
+#include <string_view>
 #include <sys/stat.h>
-#include <sys/types.h>
-#include <unistd.h>
-#include <unordered_map>
+#include <utility>
 #include <vector>
 #include <zstd.h>
 
@@ -21,11 +23,15 @@
 #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);
 
+constexpr unsigned num_overflow_slots = 16;
+
 static inline uint32_t read_unigram(const string_view s, size_t idx)
 {
 	if (idx < s.size()) {
@@ -66,7 +72,7 @@ struct db_directory {
 
 class PostingListBuilder {
 public:
-	void add_docid(uint32_t docid);
+	inline void add_docid(uint32_t docid);
 	void finish();
 
 	string encoded;
@@ -76,34 +82,31 @@ private:
 	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;
@@ -111,7 +114,7 @@ void PostingListBuilder::add_docid(uint32_t docid)
 
 void PostingListBuilder::finish()
 {
-	if (pending_docids.empty()) {
+	if (pending_deltas.empty()) {
 		return;
 	}
 
@@ -119,37 +122,59 @@ void PostingListBuilder::finish()
 
 	// 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 Corpus {
 public:
-	Corpus(size_t block_size)
-		: block_size(block_size) {}
+	Corpus(FILE *outfp, size_t block_size)
+		: invindex(new PostingListBuilder *[NUM_TRIGRAMS]), outfp(outfp), block_size(block_size)
+	{
+		fill(invindex.get(), invindex.get() + NUM_TRIGRAMS, nullptr);
+	}
+	~Corpus()
+	{
+		for (unsigned i = 0; i < NUM_TRIGRAMS; ++i) {
+			delete invindex[i];
+		}
+	}
+
 	void add_file(string filename);
 	void flush_block();
 
-	vector<string> filename_blocks;
-	unordered_map<uint32_t, PostingListBuilder> invindex;
+	vector<uint64_t> filename_blocks;
 	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];
+	}
 
 private:
+	unique_ptr<PostingListBuilder *[]> invindex;
+	FILE *outfp;
 	string current_block;
 	string tempbuf;
 	const size_t block_size;
@@ -182,77 +207,91 @@ void Corpus::flush_block()
 		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(zstd_compress(current_block, &tempbuf));
+	filename_blocks.push_back(ftell(outfp));
+	string compressed = zstd_compress(current_block, &tempbuf);
+	if (fwrite(compressed.data(), compressed.size(), 1, outfp) != 1) {
+		perror("fwrite()");
+		exit(1);
+	}
 
 	current_block.clear();
 	num_files_in_block = 0;
 	++num_blocks;
 }
 
-const char *handle_directory(const char *ptr, Corpus *corpus)
+string read_cstr(FILE *fp)
+{
+	string ret;
+	for (;;) {
+		int ch = getc(fp);
+		if (ch == -1) {
+			perror("getc");
+			exit(1);
+		}
+		if (ch == 0) {
+			return ret;
+		}
+		ret.push_back(ch);
+	}
+}
+
+void handle_directory(FILE *fp, Corpus *corpus)
 {
-	ptr += sizeof(db_directory);
+	db_directory dummy;
+	if (fread(&dummy, sizeof(dummy), 1, fp) != 1) {
+		if (feof(fp)) {
+			return;
+		} else {
+			perror("fread");
+		}
+	}
 
-	string dir_path = ptr;
-	ptr += dir_path.size() + 1;
+	string dir_path = read_cstr(fp);
 	if (dir_path == "/") {
 		dir_path = "";
 	}
 
 	for (;;) {
-		uint8_t type = *ptr++;
+		int type = getc(fp);
 		if (type == DBE_NORMAL) {
-			string filename = ptr;
+			string filename = read_cstr(fp);
 			corpus->add_file(dir_path + "/" + filename);
-			ptr += filename.size() + 1;
 		} else if (type == DBE_DIRECTORY) {
-			string dirname = ptr;
+			string dirname = read_cstr(fp);
 			corpus->add_file(dir_path + "/" + dirname);
-			ptr += dirname.size() + 1;
 		} else {
-			return ptr;
+			return;  // Probably end.
 		}
 	}
 }
 
 void read_mlocate(const char *filename, Corpus *corpus)
 {
-	int fd = open(filename, O_RDONLY);
-	if (fd == -1) {
+	FILE *fp = fopen(filename, "rb");
+	if (fp == nullptr) {
 		perror(filename);
 		exit(1);
 	}
-	off_t len = lseek(fd, 0, SEEK_END);
-	if (len == -1) {
-		perror("lseek");
-		exit(1);
-	}
-	const char *data = (char *)mmap(nullptr, len, PROT_READ, MAP_SHARED, fd, /*offset=*/0);
-	if (data == MAP_FAILED) {
-		perror("mmap");
+
+	db_header hdr;
+	if (fread(&hdr, sizeof(hdr), 1, fp) != 1) {
+		perror("short read");
 		exit(1);
 	}
 
-	const db_header *hdr = (const db_header *)data;
-
 	// TODO: Care about the base path.
-	string path = data + sizeof(db_header);
-	uint64_t offset = sizeof(db_header) + path.size() + 1 + ntohl(hdr->conf_size);
-
-	const char *ptr = data + offset;
-	while (ptr < data + len) {
-		ptr = handle_directory(ptr, corpus);
+	string path = read_cstr(fp);
+	while (!feof(fp)) {
+		handle_directory(fp, corpus);
 	}
-
-	munmap((void *)data, len);
-	close(fd);
+	fclose(fp);
 }
 
 string zstd_compress(const string &src, string *tempbuf)
@@ -265,11 +304,82 @@ string zstd_compress(const string &src, string *tempbuf)
 	return string(tempbuf->data(), size);
 }
 
+bool is_prime(uint32_t x)
+{
+	if ((x % 2) == 0 || (x % 3) == 0) {
+		return false;
+	}
+	uint32_t limit = ceil(sqrt(x));
+	for (uint32_t factor = 5; factor <= limit; ++factor) {
+		if ((x % factor) == 0) {
+			return false;
+		}
+	}
+	return true;
+}
+
+uint32_t next_prime(uint32_t x)
+{
+	if ((x % 2) == 0) {
+		++x;
+	}
+	while (!is_prime(x)) {
+		x += 2;
+	}
+	return x;
+}
+
+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) {
+		ht[i].trgm = uint32_t(-1);
+		ht[i].num_docids = 0;
+		ht[i].offset = 0;
+	}
+	for (uint32_t trgm : all_trigrams) {
+		// We don't know offset yet, so set it to zero.
+		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;
+		while (ht[bucket].num_docids != 0) {
+			// Robin Hood hashing; reduces the longest distance by a lot.
+			unsigned other_distance = bucket - hash_trigram(ht[bucket].trgm, ht_size);
+			if (distance > other_distance) {
+				swap(to_insert, ht[bucket]);
+				distance = other_distance;
+			}
+
+			++bucket, ++distance;
+			if (distance > num_overflow_slots) {
+				return nullptr;
+			}
+		}
+		ht[bucket] = to_insert;
+	}
+	return ht;
+}
+
 void do_build(const char *infile, const char *outfile, int block_size)
 {
 	steady_clock::time_point start __attribute__((unused)) = steady_clock::now();
 
-	Corpus corpus(block_size);
+	umask(0027);
+	FILE *outfp = fopen(outfile, "wb");
+
+	// 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.filename_index_offset_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.
@@ -288,84 +398,96 @@ void do_build(const char *infile, const char *outfile, int block_size)
 	}
 	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));
+	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();
+
+	// Finish up encoding the posting lists.
 	size_t trigrams = 0, longest_posting_list = 0;
-	size_t bytes_used = 0;
-	for (auto &[trigram, pl_builder] : corpus.invindex) {
+	size_t bytes_for_posting_lists = 0;
+	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_used += pl_builder.encoded.size();
+		bytes_for_posting_lists += pl_builder.encoded.size();
 	}
 	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);
-	dprintf("%zu bytes used for posting lists (%.2f bits/entry)\n", bytes_used, 8 * bytes_used / double(trigrams));
+	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());
 
+	// 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());
 
-	umask(0027);
-	FILE *outfp = fopen(outfile, "wb");
+	// Create the hash table.
+	unique_ptr<Trigram[]> hashtable;
+	uint32_t ht_size = next_prime(all_trigrams.size());
+	for (;;) {
+		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);
+		} else {
+			dprintf("Created hash table of size %u.\n\n", ht_size);
+			break;
+		}
+	}
 
-	// Write the header.
-	uint64_t num_trigrams = corpus.invindex.size();  // 64 to get alignment.
-	fwrite(&num_trigrams, sizeof(num_trigrams), 1, outfp);
+	// Find the offsets for each posting list.
+	size_t bytes_for_hashtable = (ht_size + num_overflow_slots + 1) * sizeof(Trigram);
+	uint64_t offset = ftell(outfp) + bytes_for_hashtable;
+	for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
+		hashtable[i].offset = offset;  // Needs to be there even for empty slots.
+		if (hashtable[i].num_docids == 0) {
+			continue;
+		}
 
-	// Find out where the posting lists will end.
-	uint64_t offset = sizeof(uint64_t) * 2 + 16 * all_trigrams.size();
-	uint64_t filename_index_offset = offset;
-	for (auto &[trigram, pl_builder] : corpus.invindex) {
-		filename_index_offset += pl_builder.encoded.size();
+		const string &encoded = corpus.get_pl_builder(hashtable[i].trgm).encoded;
+		offset += encoded.size();
 	}
-	fwrite(&filename_index_offset, sizeof(filename_index_offset), 1, outfp);
 
-	size_t bytes_for_trigrams = 0, bytes_for_posting_lists = 0, bytes_for_filename_index = 0, bytes_for_filenames = 0;
-	for (uint32_t trgm : all_trigrams) {
-		// 16 bytes: trgm (4), number of docids (4), byte offset (8)
-		fwrite(&trgm, sizeof(trgm), 1, outfp);  // One byte wasted, but OK.
-		const PostingListBuilder &pl_builder = corpus.invindex[trgm];
-		uint32_t num_docids = pl_builder.num_docids;
-		fwrite(&num_docids, sizeof(num_docids), 1, outfp);
-		fwrite(&offset, sizeof(offset), 1, outfp);
-		offset += pl_builder.encoded.size();
-		bytes_for_trigrams += 16;
-	}
+	// Write the hash table.
+	hdr.hash_table_offset_bytes = ftell(outfp);
+	hdr.hashtable_size = ht_size;
+	fwrite(hashtable.get(), ht_size + num_overflow_slots + 1, sizeof(Trigram), outfp);
 
 	// Write the actual posting lists.
-	for (uint32_t trgm : all_trigrams) {
-		const string &encoded = corpus.invindex[trgm].encoded;
+	for (unsigned i = 0; i < ht_size + num_overflow_slots + 1; ++i) {
+		if (hashtable[i].num_docids == 0) {
+			continue;
+		}
+		const string &encoded = corpus.get_pl_builder(hashtable[i].trgm).encoded;
 		fwrite(encoded.data(), encoded.size(), 1, outfp);
-		bytes_for_posting_lists += encoded.size();
-	}
-
-	// Stick an empty block at the end as sentinel.
-	corpus.filename_blocks.push_back("");
-
-	// Write the offsets to the filenames.
-	offset = filename_index_offset + corpus.filename_blocks.size() * sizeof(offset);
-	for (const string &filename : corpus.filename_blocks) {
-		fwrite(&offset, sizeof(offset), 1, outfp);
-		offset += filename.size();
-		bytes_for_filename_index += sizeof(offset);
-		bytes_for_filenames += filename.size();
-	}
-
-	// Write the actual filenames.
-	for (const string &filename : corpus.filename_blocks) {
-		fwrite(filename.data(), filename.size(), 1, outfp);
 	}
 
+	// Rewind, and write the updated header.
+	hdr.version = 0;
+	fseek(outfp, 0, SEEK_SET);
+	fwrite(&hdr, sizeof(hdr), 1, outfp);
 	fclose(outfp);
 
-	size_t total_bytes __attribute__((unused)) = (bytes_for_trigrams + bytes_for_posting_lists + bytes_for_filename_index + bytes_for_filenames);
+	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("Trigrams:       %'7.1f MB\n", bytes_for_trigrams / 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("Filenames:      %'7.1f MB\n", bytes_for_filenames / 1048576.0);
@@ -373,8 +495,66 @@ void do_build(const char *infile, const char *outfile, int block_size)
 	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' },
+		{ 0, 0, 0, 0 }
+	};
+
+	int block_size = 32;
+
+	setlocale(LC_ALL, "");
+	for (;;) {
+		int option_index = 0;
+		int c = getopt_long(argc, argv, "b:hV", 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);
+		default:
+			exit(1);
+		}
+	}
+
+	if (argc - optind != 2) {
+		usage();
+		exit(1);
+	}
+
+	do_build(argv[optind], argv[optind + 1], block_size);
 	exit(EXIT_SUCCESS);
 }