#include "db.h"
+#include "dprintf.h"
#include "io_uring_engine.h"
#include "parse_trigrams.h"
+#include "serializer.h"
#include "turbopfor.h"
#include "unique_sort.h"
#include <algorithm>
#include <assert.h>
+#include <atomic>
#include <chrono>
+#include <condition_variable>
+#include <deque>
#include <fcntl.h>
#include <fnmatch.h>
#include <functional>
#include <getopt.h>
-#include <iosfwd>
+#include <inttypes.h>
#include <iterator>
#include <limits>
+#include <locale.h>
+#include <map>
#include <memory>
-#include <queue>
+#include <mutex>
#include <regex.h>
#include <stdint.h>
#include <stdio.h>
#include <string.h>
#include <string>
#include <string_view>
+#include <thread>
+#include <tuple>
#include <unistd.h>
#include <unordered_map>
#include <unordered_set>
using namespace std;
using namespace std::chrono;
-#define dprintf(...) \
- do { \
- if (use_debug) { \
- fprintf(stderr, __VA_ARGS__); \
- } \
- } while (false)
-
#define DEFAULT_DBPATH "/var/lib/mlocate/plocate.db"
const char *dbpath = DEFAULT_DBPATH;
bool patterns_are_regex = false;
bool use_extended_regex = false;
int64_t limit_matches = numeric_limits<int64_t>::max();
+int64_t limit_left = numeric_limits<int64_t>::max();
-class Serializer {
-public:
- bool ready_to_print(int seq) { return next_seq == seq; }
- void print_delayed(int seq, const vector<string> msg);
- void release_current();
+steady_clock::time_point start;
+ZSTD_DDict *ddict = nullptr;
-private:
- int next_seq = 0;
- struct Element {
- int seq;
- vector<string> msg;
-
- bool operator<(const Element &other) const
- {
- return seq > other.seq;
- }
- };
- priority_queue<Element> pending;
-};
-
-void Serializer::print_delayed(int seq, const vector<string> msg)
-{
- pending.push(Element{ seq, move(msg) });
-}
-
-void Serializer::release_current()
-{
- ++next_seq;
-
- // See if any delayed prints can now be dealt with.
- while (!pending.empty() && pending.top().seq == next_seq) {
- if (limit_matches-- <= 0)
- return;
- 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;
- }
-}
+regex_t compile_regex(const string &needle);
struct Needle {
enum { STRSTR,
}
}
-bool has_access(const char *filename,
- unordered_map<string, bool> *access_rx_cache)
+class AccessRXCache {
+public:
+ AccessRXCache(IOUringEngine *engine)
+ : engine(engine) {}
+ void check_access(const char *filename, bool allow_async, function<void(bool)> cb);
+
+private:
+ unordered_map<string, bool> cache;
+ struct PendingStat {
+ string filename;
+ function<void(bool)> cb;
+ };
+ map<string, vector<PendingStat>> pending_stats;
+ IOUringEngine *engine;
+ mutex mu;
+};
+
+void AccessRXCache::check_access(const char *filename, bool allow_async, function<void(bool)> cb)
{
- const char *end = strchr(filename + 1, '/');
- while (end != nullptr) {
- string parent_path(filename, end);
- auto it = access_rx_cache->find(parent_path);
- bool ok;
- if (it == access_rx_cache->end()) {
- ok = access(parent_path.c_str(), R_OK | X_OK) == 0;
- access_rx_cache->emplace(move(parent_path), ok);
- } else {
- ok = it->second;
+ lock_guard<mutex> lock(mu);
+ if (engine == nullptr || !engine->get_supports_stat()) {
+ allow_async = false;
+ }
+
+ for (const char *end = strchr(filename + 1, '/'); end != nullptr; end = strchr(end + 1, '/')) {
+ string parent_path(filename, end - filename); // string_view from C++20.
+ auto cache_it = cache.find(parent_path);
+ if (cache_it != cache.end()) {
+ // Found in the cache.
+ if (!cache_it->second) {
+ cb(false);
+ return;
+ }
+ continue;
}
- if (!ok) {
- return false;
+
+ if (!allow_async) {
+ bool ok = access(parent_path.c_str(), R_OK | X_OK) == 0;
+ cache.emplace(parent_path, ok);
+ if (!ok) {
+ cb(false);
+ return;
+ }
+ continue;
}
- end = strchr(end + 1, '/');
+
+ // We want to call access(), but it could block on I/O. io_uring doesn't support
+ // access(), but we can do a dummy asynchonous statx() to populate the kernel's cache,
+ // which nearly always makes the next access() instantaneous.
+
+ // See if there's already a pending stat that matches this,
+ // or is a subdirectory.
+ auto it = pending_stats.lower_bound(parent_path);
+ if (it != pending_stats.end() && it->first.size() >= parent_path.size() &&
+ it->first.compare(0, parent_path.size(), parent_path) == 0) {
+ it->second.emplace_back(PendingStat{ filename, move(cb) });
+ } else {
+ it = pending_stats.emplace(filename, vector<PendingStat>{}).first;
+ engine->submit_stat(filename, [this, it, filename{ strdup(filename) }, cb{ move(cb) }] {
+ // The stat returned, so now do the actual access() calls.
+ // All of them should be in cache, so don't fire off new statx()
+ // calls during that check.
+ check_access(filename, /*allow_async=*/false, move(cb));
+ free(filename);
+
+ // Call all others that waited for the same stat() to finish.
+ // They may fire off new stat() calls if needed.
+ vector<PendingStat> pending = move(it->second);
+ pending_stats.erase(it);
+ for (PendingStat &ps : pending) {
+ check_access(ps.filename.c_str(), /*allow_async=*/true, move(ps.cb));
+ }
+ });
+ }
+ return; // The rest will happen in async context.
}
- return true;
+ // Passed all checks.
+ cb(true);
}
class Corpus {
{
return hdr.filename_index_offset_bytes + docid * sizeof(uint64_t);
}
+ const Header &get_hdr() const { return hdr; }
public:
const int fd;
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);
+ if (hdr.version != 0 && hdr.version != 1) {
+ fprintf(stderr, "plocate.db has version %u, expected 0 or 1; please rebuild it.\n", hdr.version);
exit(1);
}
+ if (hdr.version == 0) {
+ // These will be junk data.
+ hdr.zstd_dictionary_offset_bytes = 0;
+ hdr.zstd_dictionary_length_bytes = 0;
+ }
}
Corpus::~Corpus()
return hdr.num_docids;
}
-uint64_t scan_file_block(const vector<Needle> &needles, string_view compressed,
- unordered_map<string, bool> *access_rx_cache, int seq,
- Serializer *serializer)
+void scan_file_block(const vector<Needle> &needles, string_view compressed,
+ AccessRXCache *access_rx_cache, uint64_t seq, ResultReceiver *serializer,
+ atomic<uint64_t> *matched)
{
- uint64_t matched = 0;
-
unsigned long long uncompressed_len = ZSTD_getFrameContentSize(compressed.data(), compressed.size());
if (uncompressed_len == ZSTD_CONTENTSIZE_UNKNOWN || uncompressed_len == ZSTD_CONTENTSIZE_ERROR) {
fprintf(stderr, "ZSTD_getFrameContentSize() failed\n");
string block;
block.resize(uncompressed_len + 1);
- size_t err = ZSTD_decompress(&block[0], block.size(), compressed.data(),
- compressed.size());
+ static thread_local ZSTD_DCtx *ctx = ZSTD_createDCtx(); // Reused across calls.
+ size_t err;
+
+ if (ddict != nullptr) {
+ err = ZSTD_decompress_usingDDict(ctx, &block[0], block.size(), compressed.data(),
+ compressed.size(), ddict);
+ } else {
+ err = ZSTD_decompressDCtx(ctx, &block[0], block.size(), compressed.data(),
+ 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;
+ auto test_candidate = [&](const char *filename, uint64_t local_seq, uint64_t next_seq) {
+ access_rx_cache->check_access(filename, /*allow_async=*/true, [matched, serializer, local_seq, next_seq, filename{ strdup(filename) }](bool ok) {
+ if (ok) {
+ ++*matched;
+ serializer->print(local_seq, next_seq - local_seq, filename);
+ } else {
+ serializer->print(local_seq, next_seq - local_seq, "");
+ }
+ free(filename);
+ });
+ };
+
+ // We need to know the next sequence number before inserting into Serializer,
+ // so always buffer one candidate.
+ const char *pending_candidate = nullptr;
+ uint64_t local_seq = seq << 32;
for (const char *filename = block.data();
filename != block.data() + block.size();
filename += strlen(filename) + 1) {
break;
}
}
- if (found && has_access(filename, access_rx_cache)) {
- if (limit_matches-- <= 0)
- break;
- ++matched;
- if (only_count)
- continue;
- if (immediate_print) {
- if (print_nul) {
- printf("%s%c", filename, 0);
- } else {
- printf("%s\n", filename);
- }
- } else {
- delayed.push_back(filename);
+ if (found) {
+ if (pending_candidate != nullptr) {
+ test_candidate(pending_candidate, local_seq, local_seq + 1);
+ ++local_seq;
}
+ pending_candidate = filename;
}
}
- if (serializer != nullptr && !only_count) {
- if (immediate_print) {
- serializer->release_current();
- } else {
- serializer->print_delayed(seq, move(delayed));
- }
+ if (pending_candidate == nullptr) {
+ serializer->print(seq << 32, 1ULL << 32, "");
+ } else {
+ test_candidate(pending_candidate, local_seq, (seq + 1) << 32);
}
- return matched;
}
size_t scan_docids(const vector<Needle> &needles, const vector<uint32_t> &docids, const Corpus &corpus, IOUringEngine *engine)
{
Serializer docids_in_order;
- unordered_map<string, bool> access_rx_cache;
- uint64_t matched = 0;
+ AccessRXCache access_rx_cache(engine);
+ atomic<uint64_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, &needles, &access_rx_cache, &docids_in_order](string_view compressed) {
- matched += scan_file_block(needles, compressed, &access_rx_cache, i, &docids_in_order);
+ scan_file_block(needles, compressed, &access_rx_cache, i, &docids_in_order, &matched);
});
}
engine->finish();
return matched;
}
+struct WorkerThread {
+ thread t;
+
+ // We use a result queue instead of synchronizing Serializer,
+ // since a lock on it becomes a huge choke point if there are
+ // lots of threads.
+ mutex result_mu;
+ struct Result {
+ uint64_t seq;
+ uint64_t skip;
+ string msg;
+ };
+ vector<Result> results;
+};
+
+class WorkerThreadReceiver : public ResultReceiver {
+public:
+ WorkerThreadReceiver(WorkerThread *wt)
+ : wt(wt) {}
+
+ void print(uint64_t seq, uint64_t skip, const string msg) override
+ {
+ lock_guard<mutex> lock(wt->result_mu);
+ if (msg.empty() && !wt->results.empty() && wt->results.back().seq + wt->results.back().skip == seq) {
+ wt->results.back().skip += skip;
+ } else {
+ wt->results.emplace_back(WorkerThread::Result{ seq, skip, move(msg) });
+ }
+ }
+
+private:
+ WorkerThread *wt;
+};
+
+void deliver_results(WorkerThread *wt, Serializer *serializer)
+{
+ vector<WorkerThread::Result> results;
+ {
+ lock_guard<mutex> lock(wt->result_mu);
+ results = move(wt->results);
+ }
+ for (const WorkerThread::Result &result : results) {
+ serializer->print(result.seq, result.skip, move(result.msg));
+ }
+}
+
// 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.
-uint64_t scan_all_docids(const vector<Needle> &needles, int fd, const Corpus &corpus, IOUringEngine *engine)
+// coalesce it plus readahead for us. Since we assume that
+// we will primarily be CPU-bound, we'll be firing up one
+// worker thread for each spare core (the last one will
+// only be doing I/O). access() is still synchronous.
+uint64_t scan_all_docids(const vector<Needle> &needles, int fd, const Corpus &corpus)
{
- unordered_map<string, bool> access_rx_cache;
+ {
+ const Header &hdr = corpus.get_hdr();
+ if (hdr.zstd_dictionary_length_bytes > 0) {
+ string dictionary;
+ dictionary.resize(hdr.zstd_dictionary_length_bytes);
+ complete_pread(fd, &dictionary[0], hdr.zstd_dictionary_length_bytes, hdr.zstd_dictionary_offset_bytes);
+ ddict = ZSTD_createDDict(dictionary.data(), dictionary.size());
+ }
+ }
+
+ AccessRXCache access_rx_cache(nullptr);
+ Serializer serializer;
uint32_t num_blocks = corpus.get_num_filename_blocks();
unique_ptr<uint64_t[]> offsets(new uint64_t[num_blocks + 1]);
complete_pread(fd, offsets.get(), (num_blocks + 1) * sizeof(uint64_t), corpus.offset_for_block(0));
+ atomic<uint64_t> matched{ 0 };
+
+ mutex mu;
+ condition_variable queue_added, queue_removed;
+ deque<tuple<int, int, string>> work_queue; // Under mu.
+ bool done = false; // Under mu.
+
+ unsigned num_threads = max<int>(sysconf(_SC_NPROCESSORS_ONLN) - 1, 1);
+ dprintf("Using %u worker threads for linear scan.\n", num_threads);
+ unique_ptr<WorkerThread[]> threads(new WorkerThread[num_threads]);
+ for (unsigned i = 0; i < num_threads; ++i) {
+ threads[i].t = thread([&threads, &mu, &queue_added, &queue_removed, &work_queue, &done, &offsets, &needles, &access_rx_cache, &matched, i] {
+ // regcomp() takes a lock on the regex, so each thread will need its own.
+ const vector<Needle> *use_needles = &needles;
+ vector<Needle> recompiled_needles;
+ if (i != 0 && patterns_are_regex) {
+ recompiled_needles = needles;
+ for (Needle &needle : recompiled_needles) {
+ needle.re = compile_regex(needle.str);
+ }
+ use_needles = &recompiled_needles;
+ }
+
+ WorkerThreadReceiver receiver(&threads[i]);
+ for (;;) {
+ uint32_t io_docid, last_docid;
+ string compressed;
+
+ {
+ unique_lock<mutex> lock(mu);
+ queue_added.wait(lock, [&work_queue, &done] { return !work_queue.empty() || done; });
+ if (done && work_queue.empty()) {
+ return;
+ }
+ tie(io_docid, last_docid, compressed) = move(work_queue.front());
+ work_queue.pop_front();
+ queue_removed.notify_all();
+ }
+
+ 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(*use_needles, { &compressed[relative_offset], len }, &access_rx_cache, docid, &receiver, &matched);
+ }
+ }
+ });
+ }
+
string compressed;
- uint64_t matched = 0;
for (uint32_t io_docid = 0; io_docid < num_blocks; io_docid += 32) {
uint32_t last_docid = std::min(io_docid + 32, num_blocks);
size_t io_len = offsets[last_docid] - offsets[io_docid];
}
complete_pread(fd, &compressed[0], io_len, offsets[io_docid]);
- 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];
- matched += scan_file_block(needles, { &compressed[relative_offset], len }, &access_rx_cache, 0, nullptr);
- if (limit_matches <= 0)
- return matched;
+ {
+ unique_lock<mutex> lock(mu);
+ queue_removed.wait(lock, [&work_queue] { return work_queue.size() < 256; }); // Allow ~2MB of data queued up.
+ work_queue.emplace_back(io_docid, last_docid, move(compressed));
+ queue_added.notify_one(); // Avoid the thundering herd.
}
+
+ // Pick up some results, so that we are sure that we won't just overload.
+ // (Seemingly, going through all of these causes slowness with many threads,
+ // but taking only one is OK.)
+ unsigned i = io_docid / 32;
+ deliver_results(&threads[i % num_threads], &serializer);
+ }
+ {
+ lock_guard<mutex> lock(mu);
+ done = true;
+ queue_added.notify_all();
+ }
+ for (unsigned i = 0; i < num_threads; ++i) {
+ threads[i].t.join();
+ deliver_results(&threads[i], &serializer);
}
return matched;
}
exit(EXIT_FAILURE);
}
- steady_clock::time_point start __attribute__((unused)) = steady_clock::now();
+ start = steady_clock::now();
if (access("/", R_OK | X_OK)) {
// We can't find anything, no need to bother...
return;
// (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.)
- uint64_t matched = scan_all_docids(needles, fd, corpus, &engine);
+ uint64_t matched = scan_all_docids(needles, fd, corpus);
if (only_count) {
- printf("%zu\n", matched);
+ printf("%" PRId64 "\n", matched);
}
return;
}
+ // Sneak in fetching the dictionary, if present. It's not necessarily clear
+ // exactly where it would be cheapest to get it, but it needs to be present
+ // before we can decode any of the posting lists. Most likely, it's
+ // in the same filesystem block as the header anyway, so it should be
+ // present in the cache.
+ {
+ const Header &hdr = corpus.get_hdr();
+ if (hdr.zstd_dictionary_length_bytes > 0) {
+ engine.submit_read(fd, hdr.zstd_dictionary_length_bytes, hdr.zstd_dictionary_offset_bytes, [](string_view s) {
+ ddict = ZSTD_createDDict(s.data(), s.size());
+ dprintf("Dictionary initialized after %.1f ms.\n", 1e3 * duration<float>(steady_clock::now() - start).count());
+ });
+ }
+ }
+
// Look them all up on disk.
for (auto &[trgm, trigram_groups] : trigrams_to_lookup) {
corpus.find_trigram(trgm, [trgm{ trgm }, trigram_groups{ &trigram_groups }](const Trigram *trgmptr, size_t len) {
1e3 * duration<float>(steady_clock::now() - start).count());
uint64_t matched = scan_docids(needles, cur_candidates, corpus, &engine);
- dprintf("Done in %.1f ms, found %zu matches.\n",
+ dprintf("Done in %.1f ms, found %" PRId64 " matches.\n",
1e3 * duration<float>(steady_clock::now() - start).count(), matched);
if (only_count) {
- printf("%zu\n", matched);
+ printf("%" PRId64 "\n", matched);
}
}
break;
case 'l':
case 'n':
- limit_matches = atoll(optarg);
+ limit_matches = limit_left = atoll(optarg);
if (limit_matches <= 0) {
fprintf(stderr, "Error: limit must be a strictly positive number.\n");
exit(1);
projects / plocate / blobdiff