Use asynchronous filesystem access to speed up extraction

src/dwarfsextract.cpp

@@ -20,8 +20,12 @@
  */
 
 #include <algorithm>
+#include <condition_variable>
 #include <cstring>
 #include <exception>
+#include <future>
+#include <memory>
+#include <mutex>
 #include <vector>
 
 #include <sys/statvfs.h>
@@ -42,6 +46,7 @@
 #include "dwarfs/options.h"
 #include "dwarfs/util.h"
 #include "dwarfs/version.h"
+#include "dwarfs/worker_group.h"
 
 namespace po = boost::program_options;
 
@@ -49,6 +54,30 @@ using namespace dwarfs;
 
 namespace {
 
+class semaphore {
+ public:
+  void post(uint64_t n = 1) {
+    {
+      std::lock_guard lock(mx_);
+      count_ += n;
+    }
+    condition_.notify_one();
+  }
+
+  void wait(uint64_t n = 1) {
+    std::unique_lock lock(mx_);
+    while (count_ < n) {
+      condition_.wait(lock);
+    }
+    count_ -= n;
+  }
+
+ private:
+  std::mutex mx_;
+  std::condition_variable condition_;
+  uint64_t count_{0};
+};
+
 int dwarfsextract(int argc, char** argv) {
   std::string filesystem, output, format, cache_size_str, log_level;
   size_t num_workers;
@@ -66,7 +95,7 @@ int dwarfsextract(int argc, char** argv) {
         po::value<std::string>(&format),
         "output format")
     ("num-workers,n",
-        po::value<size_t>(&num_workers)->default_value(1),
+        po::value<size_t>(&num_workers)->default_value(4),
         "number of worker threads")
     ("cache-size,s",
         po::value<std::string>(&cache_size_str)->default_value("256m"),
@@ -150,15 +179,36 @@ int dwarfsextract(int argc, char** argv) {
 
     ::archive_entry* spare = nullptr;
 
+    worker_group archiver("archiver", 1);
+    semaphore sem;
+    sem.post(fsopts.block_cache.max_bytes);
+
     auto do_archive = [&](::archive_entry* ae, entry_view entry) {
-      check_result(::archive_write_header(a, ae));
+      if (auto size = ::archive_entry_size(ae);
-      if (auto size = ::archive_entry_size(ae); size > 0) {
+          S_ISREG(entry.mode()) && size > 0) {
-        int fh = fs.open(entry);
+        auto fd = fs.open(entry);
-        iovec_read_buf irb;
+        sem.wait(size);
-        fs.readv(fh, irb, size, 0);
+        auto ranges = fs.readv(fd, size, 0);
-        for (auto const& iov : irb.buf) {
+        if (!ranges) {
-          check_result(::archive_write_data(a, iov.iov_base, iov.iov_len));
+          LOG_ERROR << "error reading inode [" << fd
+                    << "]: " << ::strerror(-ranges.error());
+          return;
         }
+        archiver.add_job([&sem, &check_result, ranges = std::move(*ranges), a,
+                          ae, size]() mutable {
+          check_result(::archive_write_header(a, ae));
+          for (auto& r : ranges) {
+            auto br = r.get();
+            check_result(::archive_write_data(a, br.data(), br.size()));
+          }
+          sem.post(size);
+          ::archive_entry_free(ae);
+        });
+      } else {
+        archiver.add_job([&check_result, a, ae] {
+          check_result(::archive_write_header(a, ae));
+          ::archive_entry_free(ae);
+        });
       }
     };
 
@@ -197,7 +247,6 @@ int dwarfsextract(int argc, char** argv) {
 
       if (ae) {
         do_archive(ae, entry);
-        ::archive_entry_free(ae);
       }
 
       if (spare) {
@@ -207,10 +256,11 @@ int dwarfsextract(int argc, char** argv) {
         }
         LOG_DEBUG << "archiving spare " << ::archive_entry_pathname(spare);
         do_archive(spare, *ev);
-        ::archive_entry_free(spare);
       }
     });
 
+    archiver.wait();
+
     // As we're visiting *all* hardlinks, we should never see any deferred
     // entries.
     ::archive_entry* ae = nullptr;