X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=plocate.cpp;h=0d59076e8fc99efb4829384da1894e99c4399867;hb=d5ba26d705460a7e37213eeb4954b2efed8bebf0;hp=e31345852341f2a7218a435b873ec88303888a72;hpb=75c42d06f8eb513afc1976e82033ffb893b101b8;p=plocate

diff --git a/plocate.cpp b/plocate.cpp
index e313458..0d59076 100644
--- a/plocate.cpp
+++ b/plocate.cpp
@@ -1,27 +1,28 @@
+#include "access_rx_cache.h"
 #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 <atomic>
 #include <assert.h>
+#include <atomic>
 #include <chrono>
 #include <condition_variable>
+#include <deque>
 #include <fcntl.h>
 #include <fnmatch.h>
 #include <functional>
 #include <getopt.h>
 #include <inttypes.h>
-#include <iosfwd>
 #include <iterator>
 #include <limits>
-#include <map>
+#include <locale.h>
 #include <memory>
 #include <mutex>
-#include <queue>
 #include <regex.h>
 #include <stdint.h>
 #include <stdio.h>
@@ -30,6 +31,7 @@
 #include <string>
 #include <string_view>
 #include <thread>
+#include <tuple>
 #include <unistd.h>
 #include <unordered_map>
 #include <unordered_set>
@@ -57,83 +59,6 @@ ZSTD_DDict *ddict = nullptr;
 
 regex_t compile_regex(const string &needle);
 
-void apply_limit()
-{
-	if (--limit_left > 0) {
-		return;
-	}
-	dprintf("Done in %.1f ms, found %" PRId64 " matches.\n",
-	        1e3 * duration<float>(steady_clock::now() - start).count(), limit_matches);
-	if (only_count) {
-		printf("%" PRId64 "\n", limit_matches);
-	}
-	exit(0);
-}
-
-class ResultReceiver {
-public:
-	virtual ~ResultReceiver() = default;
-	virtual void print(uint64_t seq, uint64_t skip, const string msg) = 0;
-};
-
-class Serializer : public ResultReceiver {
-public:
-	~Serializer() { assert(limit_left <= 0 || pending.empty()); }
-	void print(uint64_t seq, uint64_t skip, const string msg) override;
-
-private:
-	uint64_t next_seq = 0;
-	struct Element {
-		uint64_t seq, skip;
-		string msg;
-
-		bool operator<(const Element &other) const
-		{
-			return seq > other.seq;
-		}
-	};
-	priority_queue<Element> pending;
-};
-
-void Serializer::print(uint64_t seq, uint64_t skip, const string msg)
-{
-	if (only_count) {
-		if (!msg.empty()) {
-			apply_limit();
-		}
-		return;
-	}
-
-	if (next_seq != seq) {
-		pending.push(Element{ seq, skip, move(msg) });
-		return;
-	}
-
-	if (!msg.empty()) {
-		if (print_nul) {
-			printf("%s%c", msg.c_str(), 0);
-		} else {
-			printf("%s\n", msg.c_str());
-		}
-		apply_limit();
-	}
-	next_seq += skip;
-
-	// See if any delayed prints can now be dealt with.
-	while (!pending.empty() && pending.top().seq == next_seq) {
-		if (!pending.top().msg.empty()) {
-			if (print_nul) {
-				printf("%s%c", pending.top().msg.c_str(), 0);
-			} else {
-				printf("%s\n", pending.top().msg.c_str());
-			}
-			apply_limit();
-		}
-		next_seq += pending.top().skip;
-		pending.pop();
-	}
-}
-
 struct Needle {
 	enum { STRSTR,
 	       REGEX,
@@ -155,87 +80,6 @@ bool matches(const Needle &needle, const char *haystack)
 	}
 }
 
-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)
-{
-	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 (!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;
-		}
-
-		// 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.
-	}
-
-	// Passed all checks.
-	cb(true);
-}
-
 class Corpus {
 public:
 	Corpus(int fd, IOUringEngine *engine);
@@ -399,7 +243,7 @@ size_t scan_docids(const vector<Needle> &needles, const vector<uint32_t> &docids
 {
 	Serializer docids_in_order;
 	AccessRXCache access_rx_cache(engine);
-	atomic<uint64_t> matched{0};
+	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) {
@@ -417,17 +261,27 @@ struct WorkerThread {
 	// since a lock on it becomes a huge choke point if there are
 	// lots of threads.
 	mutex result_mu;
-	vector<tuple<uint64_t, uint64_t, string>> results;
+	struct Result {
+		uint64_t seq;
+		uint64_t skip;
+		string msg;
+	};
+	vector<Result> results;
 };
 
 class WorkerThreadReceiver : public ResultReceiver {
 public:
-	WorkerThreadReceiver(WorkerThread *wt) : wt(wt) {}
+	WorkerThreadReceiver(WorkerThread *wt)
+		: wt(wt) {}
 
 	void print(uint64_t seq, uint64_t skip, const string msg) override
 	{
 		lock_guard<mutex> lock(wt->result_mu);
-		wt->results.emplace_back(seq, skip, move(msg));
+		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:
@@ -436,13 +290,13 @@ private:
 
 void deliver_results(WorkerThread *wt, Serializer *serializer)
 {
-	vector<tuple<uint64_t, uint64_t, string>> results;
+	vector<WorkerThread::Result> results;
 	{
 		lock_guard<mutex> lock(wt->result_mu);
 		results = move(wt->results);
 	}
-	for (const auto &result : results) {
-		serializer->print(get<0>(result), get<1>(result), move(get<2>(result)));
+	for (const WorkerThread::Result &result : results) {
+		serializer->print(result.seq, result.skip, move(result.msg));
 	}
 }
 
@@ -469,7 +323,7 @@ uint64_t scan_all_docids(const vector<Needle> &needles, int fd, const Corpus &co
 	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};
+	atomic<uint64_t> matched{ 0 };
 
 	mutex mu;
 	condition_variable queue_added, queue_removed;