-#include "vp4.h"
+#include "db.h"
#include "io_uring_engine.h"
+#include "vp4.h"
#include <algorithm>
#include <arpa/inet.h>
+#include <assert.h>
#include <chrono>
#include <endian.h>
#include <fcntl.h>
#include <functional>
+#include <getopt.h>
+#include <limits.h>
#include <memory>
#include <stdio.h>
#include <string.h>
#define dprintf(...)
//#define dprintf(...) fprintf(stderr, __VA_ARGS__);
+#include "turbopfor.h"
+
+const char *dbpath = "/var/lib/mlocate/plocate.db";
+bool print_nul = false;
+
class Serializer {
public:
- void do_or_wait(int seq, function<void()> cb);
+ bool ready_to_print(int seq) { return next_seq == seq; }
+ void print_delayed(int seq, const vector<string> msg);
+ void release_current();
private:
int next_seq = 0;
struct Element {
int seq;
- function<void()> cb;
+ vector<string> msg;
bool operator<(const Element &other) const
{
priority_queue<Element> pending;
};
-void Serializer::do_or_wait(int seq, function<void()> cb)
+void Serializer::print_delayed(int seq, const vector<string> msg)
{
- if (seq != next_seq) {
- pending.emplace(Element{ seq, move(cb) });
- return;
- }
+ pending.push(Element{ seq, move(msg) });
+}
- cb();
+void Serializer::release_current()
+{
++next_seq;
+ // See if any delayed prints can now be dealt with.
while (!pending.empty() && pending.top().seq == next_seq) {
- pending.top().cb();
+ for (const string &msg : pending.top().msg) {
+ if (print_nul) {
+ printf("%s%c", msg.c_str(), 0);
+ } else {
+ printf("%s\n", msg.c_str());
+ }
+ }
pending.pop();
++next_seq;
}
return true;
}
-struct Trigram {
- uint32_t trgm;
- uint32_t num_docids;
- uint64_t offset;
-
- bool operator==(const Trigram &other) const
- {
- return trgm == other.trgm;
- }
- bool operator<(const Trigram &other) const
- {
- return trgm < other.trgm;
- }
-};
-
class Corpus {
public:
Corpus(int fd, IOUringEngine *engine);
void find_trigram(uint32_t trgm, function<void(const Trigram *trgmptr, size_t len)> cb);
void get_compressed_filename_block(uint32_t docid, function<void(string)> cb) const;
size_t get_num_filename_blocks() const;
- off_t offset_for_block(uint32_t docid) const {
- return filename_index_offset + docid * sizeof(uint64_t);
+ off_t offset_for_block(uint32_t docid) const
+ {
+ return hdr.filename_index_offset_bytes + docid * sizeof(uint64_t);
}
public:
const int fd;
IOUringEngine *const engine;
- off_t len;
- uint64_t filename_index_offset;
-
- uint64_t num_trigrams;
- const off_t trigram_offset = sizeof(uint64_t) * 2;
-
- void binary_search_trigram(uint32_t trgm, uint32_t left, uint32_t right, function<void(const Trigram *trgmptr, size_t len)> cb);
+ Header hdr;
};
Corpus::Corpus(int fd, IOUringEngine *engine)
: fd(fd), engine(engine)
{
- len = lseek(fd, 0, SEEK_END);
- if (len == -1) {
- perror("lseek");
- exit(1);
+ // Enable to test cold-cache behavior (except for access()).
+ if (false) {
+ off_t len = lseek(fd, 0, SEEK_END);
+ if (len == -1) {
+ perror("lseek");
+ exit(1);
+ }
+ posix_fadvise(fd, 0, len, POSIX_FADV_DONTNEED);
}
- // Uncomment to test cold-cache behavior (except for access()).
- // posix_fadvise(fd, 0, len, POSIX_FADV_DONTNEED);
-
- uint64_t vals[2];
- complete_pread(fd, vals, sizeof(vals), /*offset=*/0);
-
- num_trigrams = vals[0];
- filename_index_offset = vals[1];
+ complete_pread(fd, &hdr, sizeof(hdr), /*offset=*/0);
+ if (memcmp(hdr.magic, "\0plocate", 8) != 0) {
+ fprintf(stderr, "plocate.db is corrupt or an old version; please rebuild it.\n");
+ exit(1);
+ }
+ if (hdr.version != 0) {
+ fprintf(stderr, "plocate.db has version %u, expected 0; please rebuild it.\n", hdr.version);
+ exit(1);
+ }
}
Corpus::~Corpus()
void Corpus::find_trigram(uint32_t trgm, function<void(const Trigram *trgmptr, size_t len)> cb)
{
- binary_search_trigram(trgm, 0, num_trigrams - 1, move(cb));
-}
-
-void Corpus::binary_search_trigram(uint32_t trgm, uint32_t left, uint32_t right, function<void(const Trigram *trgmptr, size_t len)> cb)
-{
- if (left > right) {
- cb(nullptr, 0);
- return;
- }
- uint32_t mid = (left + right) / 2;
- engine->submit_read(fd, sizeof(Trigram) * 2, trigram_offset + sizeof(Trigram) * mid, [this, trgm, left, mid, right, cb{ move(cb) }](string s) {
+ uint32_t bucket = hash_trigram(trgm, hdr.hashtable_size);
+ engine->submit_read(fd, sizeof(Trigram) * (hdr.extra_ht_slots + 2), hdr.hash_table_offset_bytes + sizeof(Trigram) * bucket, [this, trgm, bucket, cb{ move(cb) }](string s) {
const Trigram *trgmptr = reinterpret_cast<const Trigram *>(s.data());
- const Trigram *next_trgmptr = trgmptr + 1;
- if (trgmptr->trgm < trgm) {
- binary_search_trigram(trgm, mid + 1, right, move(cb));
- } else if (trgmptr->trgm > trgm) {
- binary_search_trigram(trgm, left, mid - 1, move(cb));
- } else {
- cb(trgmptr, next_trgmptr->offset - trgmptr->offset);
+ for (unsigned i = 0; i < hdr.extra_ht_slots + 1; ++i) {
+ if (trgmptr[i].trgm == trgm) {
+ cb(trgmptr + i, trgmptr[i + 1].offset - trgmptr[i].offset);
+ return;
+ }
}
+
+ // Not found.
+ cb(nullptr, 0);
});
}
size_t Corpus::get_num_filename_blocks() const
{
- // The beginning of the filename blocks is the end of the filename index blocks.
- uint64_t end;
- complete_pread(fd, &end, sizeof(end), filename_index_offset);
-
- // Subtract the sentinel block.
- return (end - filename_index_offset) / sizeof(uint64_t) - 1;
+ return hdr.num_docids;
}
-size_t scan_file_block(const string &needle, string_view compressed,
- unordered_map<string, bool> *access_rx_cache)
+size_t scan_file_block(const vector<string> &needles, string_view compressed,
+ unordered_map<string, bool> *access_rx_cache, int seq,
+ Serializer *serializer)
{
size_t matched = 0;
block.resize(uncompressed_len + 1);
size_t err = ZSTD_decompress(&block[0], block.size(), compressed.data(),
- compressed.size());
+ compressed.size());
if (ZSTD_isError(err)) {
fprintf(stderr, "ZSTD_decompress(): %s\n", ZSTD_getErrorName(err));
exit(1);
}
block[block.size() - 1] = '\0';
+ bool immediate_print = (serializer == nullptr || serializer->ready_to_print(seq));
+ vector<string> delayed;
+
for (const char *filename = block.data();
filename != block.data() + block.size();
filename += strlen(filename) + 1) {
- if (strstr(filename, needle.c_str()) == nullptr) {
- continue;
+ bool found = true;
+ for (const string &needle : needles) {
+ if (strstr(filename, needle.c_str()) == nullptr) {
+ found = false;
+ break;
+ }
}
- if (has_access(filename, access_rx_cache)) {
+ if (found && has_access(filename, access_rx_cache)) {
++matched;
- printf("%s\n", filename);
+ if (immediate_print) {
+ if (print_nul) {
+ printf("%s%c", filename, 0);
+ } else {
+ printf("%s\n", filename);
+ }
+ } else {
+ delayed.push_back(filename);
+ }
+ }
+ }
+ if (serializer != nullptr) {
+ if (immediate_print) {
+ serializer->release_current();
+ } else {
+ serializer->print_delayed(seq, move(delayed));
}
}
return matched;
}
-size_t scan_docids(const string &needle, const vector<uint32_t> &docids, const Corpus &corpus, IOUringEngine *engine)
+size_t scan_docids(const vector<string> &needles, const vector<uint32_t> &docids, const Corpus &corpus, IOUringEngine *engine)
{
Serializer docids_in_order;
unordered_map<string, bool> access_rx_cache;
size_t matched = 0;
for (size_t i = 0; i < docids.size(); ++i) {
uint32_t docid = docids[i];
- corpus.get_compressed_filename_block(docid, [i, &matched, &needle, &access_rx_cache, &docids_in_order](string compressed) {
- docids_in_order.do_or_wait(i, [&matched, &needle, compressed{ move(compressed) }, &access_rx_cache] {
- matched += scan_file_block(needle, compressed, &access_rx_cache);
- });
+ corpus.get_compressed_filename_block(docid, [i, &matched, &needles, &access_rx_cache, &docids_in_order](string compressed) {
+ matched += scan_file_block(needles, compressed, &access_rx_cache, i, &docids_in_order);
});
}
engine->finish();
// We do this sequentially, as it's faster than scattering
// a lot of I/O through io_uring and hoping the kernel will
// coalesce it plus readahead for us.
-void scan_all_docids(const string &needle, int fd, const Corpus &corpus, IOUringEngine *engine)
+void scan_all_docids(const vector<string> &needles, int fd, const Corpus &corpus, IOUringEngine *engine)
{
unordered_map<string, bool> access_rx_cache;
uint32_t num_blocks = corpus.get_num_filename_blocks();
for (uint32_t docid = io_docid; docid < last_docid; ++docid) {
size_t relative_offset = offsets[docid] - offsets[io_docid];
size_t len = offsets[docid + 1] - offsets[docid];
- scan_file_block(needle, {&compressed[relative_offset], len}, &access_rx_cache);
+ scan_file_block(needles, { &compressed[relative_offset], len }, &access_rx_cache, 0, nullptr);
}
}
}
-void do_search_file(const string &needle, const char *filename)
+void do_search_file(const vector<string> &needles, const char *filename)
{
int fd = open(filename, O_RDONLY);
if (fd == -1) {
IOUringEngine engine;
Corpus corpus(fd, &engine);
- dprintf("Corpus init took %.1f ms.\n", 1e3 * duration<float>(steady_clock::now() - start).count());
+ dprintf("Corpus init done after %.1f ms.\n", 1e3 * duration<float>(steady_clock::now() - start).count());
+
+ vector<pair<Trigram, size_t>> trigrams;
+ uint64_t shortest_so_far = numeric_limits<uint32_t>::max();
+ for (const string &needle : needles) {
+ if (needle.size() < 3)
+ continue;
+ for (size_t i = 0; i < needle.size() - 2; ++i) {
+ uint32_t trgm = read_trigram(needle, i);
+ corpus.find_trigram(trgm, [trgm, &trigrams, &shortest_so_far](const Trigram *trgmptr, size_t len) {
+ if (trgmptr == nullptr) {
+ dprintf("trigram '%c%c%c' isn't found, we abort the search\n",
+ trgm & 0xff, (trgm >> 8) & 0xff, (trgm >> 16) & 0xff);
+ exit(0);
+ }
+ if (trgmptr->num_docids > shortest_so_far * 100) {
+ dprintf("not loading trigram '%c%c%c' with %u docids, it would be ignored later anyway\n",
+ trgm & 0xff, (trgm >> 8) & 0xff, (trgm >> 16) & 0xff,
+ trgmptr->num_docids);
+ } else {
+ trigrams.emplace_back(*trgmptr, len);
+ shortest_so_far = std::min<uint64_t>(shortest_so_far, trgmptr->num_docids);
+ }
+ });
+ }
+ }
+ engine.finish();
+ dprintf("Hashtable lookups done after %.1f ms.\n", 1e3 * duration<float>(steady_clock::now() - start).count());
- if (needle.size() < 3) {
+ if (trigrams.empty()) {
// Too short for trigram matching. Apply brute force.
// (We could have searched through all trigrams that matched
// the pattern and done a union of them, but that's a lot of
// work for fairly unclear gain.)
- scan_all_docids(needle, fd, corpus, &engine);
+ scan_all_docids(needles, fd, corpus, &engine);
return;
}
-
- vector<pair<Trigram, size_t>> trigrams;
- for (size_t i = 0; i < needle.size() - 2; ++i) {
- uint32_t trgm = read_trigram(needle, i);
- corpus.find_trigram(trgm, [trgm, &trigrams](const Trigram *trgmptr, size_t len) {
- if (trgmptr == nullptr) {
- dprintf("trigram %06x isn't found, we abort the search\n", trgm);
- return;
- }
- trigrams.emplace_back(*trgmptr, len);
- });
- }
- engine.finish();
- dprintf("Binary search took %.1f ms.\n", 1e3 * duration<float>(steady_clock::now() - start).count());
-
sort(trigrams.begin(), trigrams.end());
{
auto last = unique(trigrams.begin(), trigrams.end());
break;
}
engine.submit_read(fd, len, trgmptr.offset, [trgmptr, len, &done, &in1, &in2, &out](string s) {
+ if (done)
+ return;
uint32_t trgm __attribute__((unused)) = trgmptr.trgm;
size_t num = trgmptr.num_docids;
unsigned char *pldata = reinterpret_cast<unsigned char *>(s.data());
if (in1.empty()) {
in1.resize(num + 128);
- p4nd1dec128v32(pldata, num, &in1[0]);
+ decode_pfor_delta1<128>(pldata, num, /*interleaved=*/true, &in1[0]);
in1.resize(num);
dprintf("trigram '%c%c%c' (%zu bytes) decoded to %zu entries\n", trgm & 0xff,
(trgm >> 8) & 0xff, (trgm >> 16) & 0xff, len, num);
if (in2.size() < num + 128) {
in2.resize(num + 128);
}
- p4nd1dec128v32(pldata, num, &in2[0]);
+ decode_pfor_delta1<128>(pldata, num, /*interleaved=*/true, &in2[0]);
out.clear();
set_intersection(in1.begin(), in1.end(), in2.begin(), in2.begin() + num,
});
}
engine.finish();
- dprintf("Intersection took %.1f ms. Doing final verification and printing:\n",
+ if (done) {
+ return;
+ }
+ dprintf("Intersection done after %.1f ms. Doing final verification and printing:\n",
1e3 * duration<float>(steady_clock::now() - start).count());
- size_t matched __attribute__((unused)) = scan_docids(needle, in1, corpus, &engine);
+ size_t matched __attribute__((unused)) = scan_docids(needles, in1, corpus, &engine);
dprintf("Done in %.1f ms, found %zu matches.\n",
1e3 * duration<float>(steady_clock::now() - start).count(), matched);
}
+void usage()
+{
+ // The help text comes from mlocate.
+ printf("Usage: plocate [OPTION]... PATTERN...\n");
+ printf("\n");
+ printf(" -d, --database DBPATH use DBPATH instead of default database (which is\n");
+ printf(" %s)\n", dbpath);
+ printf(" -h, --help print this help\n");
+ printf(" -0, --null separate entries with NUL on output\n");
+}
+
int main(int argc, char **argv)
{
- do_search_file(argv[1], "/var/lib/mlocate/plocate.db");
+ static const struct option long_options[] = {
+ { "help", no_argument, 0, 'h' },
+ { "database", required_argument, 0, 'd' },
+ { "null", no_argument, 0, '0' },
+ { 0, 0, 0, 0 }
+ };
+
+ for (;;) {
+ int option_index = 0;
+ int c = getopt_long(argc, argv, "d:h0", long_options, &option_index);
+ if (c == -1) {
+ break;
+ }
+ switch (c) {
+ case 'd':
+ dbpath = strdup(optarg);
+ break;
+ case 'h':
+ usage();
+ exit(0);
+ case '0':
+ print_nul = true;
+ break;
+ default:
+ exit(1);
+ }
+ }
+
+ vector<string> needles;
+ for (int i = optind; i < argc; ++i) {
+ needles.push_back(argv[i]);
+ }
+ if (needles.empty()) {
+ fprintf(stderr, "plocate: no pattern to search for specified\n");
+ exit(0);
+ }
+ do_search_file(needles, dbpath);
}