Move iovecs around instead of having single data buffers. Hopefully a tad more effici...

diff --git a/stream.cpp b/stream.cpp
index 4be673c864410e9ddfda17675e9d9137b672da71..17b5949172a531cae5a37eb4f8948dbf177bb4a9 100644 (file)
--- a/stream.cpp
+++ b/stream.cpp
@@ -111,7 +111,13 @@ void Stream::set_backlog_size(size_t new_size)
 
        // Now cheat a bit by rewinding, and adding all the old data back.
        bytes_received -= existing_data.size();
 
        // Now cheat a bit by rewinding, and adding all the old data back.
        bytes_received -= existing_data.size();

-       add_data_raw(existing_data.data(), existing_data.size());
+       iovec iov;
+       iov.iov_base = const_cast<char *>(existing_data.data());
+       iov.iov_len = existing_data.size();
+
+       vector<iovec> iovs;
+       iovs.push_back(iov);
+       add_data_raw(iovs);

 }
 
 void Stream::put_client_to_sleep(Client *client)
 }
 
 void Stream::put_client_to_sleep(Client *client)


@@ -119,65 +125,101 @@ void Stream::put_client_to_sleep(Client *client)
        sleeping_clients.push_back(client);
 }
 
        sleeping_clients.push_back(client);
 }
 

-void Stream::add_data_raw(const char *data, ssize_t bytes)
+// Return a new set of iovecs that contains only the first <bytes_wanted> bytes of <data>.
+vector<iovec> collect_iovecs(const vector<iovec> &data, size_t bytes_wanted)

 {
 {

-       size_t pos = bytes_received % backlog_size;
-       bytes_received += bytes;
-
-       if (pos + bytes > backlog_size) {
-               ssize_t to_copy = backlog_size - pos;
-               while (to_copy > 0) {
-                       int ret = pwrite(data_fd, data, to_copy, pos);
-                       if (ret == -1 && errno == EINTR) {
-                               continue;
-                       }
-                       if (ret == -1) {
-                               log_perror("pwrite");
-                               // Dazed and confused, but trying to continue...
-                               break;
-                       }
-                       pos += ret;
-                       data += ret;
-                       to_copy -= ret;
-                       bytes -= ret;
+       vector<iovec> ret;
+       for (size_t i = 0; i < data.size() && bytes_wanted > 0; ++i) {
+               if (data[i].iov_len <= bytes_wanted) {
+                       // Consume the entire iovec.
+                       ret.push_back(data[i]);
+                       bytes_wanted -= data[i].iov_len;
+               } else {
+                       // Take only parts of this iovec.
+                       iovec iov;
+                       iov.iov_base = data[i].iov_base;
+                       iov.iov_len = bytes_wanted;     
+                       ret.push_back(iov);
+                       bytes_wanted = 0;

                }
                }

-               pos = 0;

        }
        }

+       return ret;
+}

 
 

-       while (bytes > 0) {
-               int ret = pwrite(data_fd, data, bytes, pos);
-               if (ret == -1 && errno == EINTR) {
-                       continue;
+// Return a new set of iovecs that contains all of <data> except the first <bytes_wanted> bytes.
+vector<iovec> remove_iovecs(const vector<iovec> &data, size_t bytes_wanted)
+{
+       vector<iovec> ret;
+       size_t i;
+       for (i = 0; i < data.size() && bytes_wanted > 0; ++i) {
+               if (data[i].iov_len <= bytes_wanted) {
+                       // Consume the entire iovec.
+                       bytes_wanted -= data[i].iov_len;
+               } else {
+                       // Take only parts of this iovec.
+                       iovec iov;
+                       iov.iov_base = reinterpret_cast<char *>(data[i].iov_base) + bytes_wanted;
+                       iov.iov_len = data[i].iov_len - bytes_wanted;
+                       ret.push_back(iov);
+                       bytes_wanted = 0;

                }
                }

+       }
+
+       // Add the rest of the iovecs unchanged.
+       ret.insert(ret.end(), data.begin() + i, data.end());
+       return ret;
+}
+
+void Stream::add_data_raw(const vector<iovec> &orig_data)
+{
+       vector<iovec> data = orig_data;
+       while (!data.empty()) {
+               size_t pos = bytes_received % backlog_size;
+
+               // Collect as many iovecs as we can before we hit the point
+               // where the circular buffer wraps around.
+               vector<iovec> to_write = collect_iovecs(data, backlog_size - pos);
+               ssize_t ret;
+               do {
+                       ret = pwritev(data_fd, to_write.data(), to_write.size(), pos);
+               } while (ret == -1 && errno == EINTR);
+

                if (ret == -1) {
                        log_perror("pwrite");
                        // Dazed and confused, but trying to continue...
                if (ret == -1) {
                        log_perror("pwrite");
                        // Dazed and confused, but trying to continue...

-                       break;
+                       return;

                }
                }

-               pos += ret;
-               data += ret;
-               bytes -= ret;
+               bytes_received += ret;
+
+               // Remove the data that was actually written from the set of iovecs.
+               data = remove_iovecs(data, ret);

        }
 }
 
 void Stream::add_data_deferred(const char *data, size_t bytes)
 {
        }
 }
 
 void Stream::add_data_deferred(const char *data, size_t bytes)
 {

-       if (encoding == Stream::STREAM_ENCODING_RAW) {
-               queued_data.append(string(data, data + bytes));
-       } else if (encoding == STREAM_ENCODING_METACUBE) {
+       if (encoding == Stream::STREAM_ENCODING_METACUBE) {
+               // Add a Metacube block header before the data.

                metacube_block_header hdr;
                memcpy(hdr.sync, METACUBE_SYNC, sizeof(hdr.sync));
                hdr.size = htonl(bytes);
                hdr.flags = htonl(0);
 
                metacube_block_header hdr;
                memcpy(hdr.sync, METACUBE_SYNC, sizeof(hdr.sync));
                hdr.size = htonl(bytes);
                hdr.flags = htonl(0);
 

-               char *block = new char[bytes + sizeof(hdr)];
-               memcpy(block, &hdr, sizeof(hdr));
-               memcpy(block + sizeof(hdr), data, bytes);
-               queued_data.append(string(block, block + bytes + sizeof(hdr)));
-               delete[] block;
+               iovec iov;
+               iov.iov_base = new char[sizeof(hdr)];
+               memcpy(iov.iov_base, &hdr, sizeof(hdr));
+               iov.iov_len = sizeof(hdr);
+               queued_data.push_back(iov);

        } else {
        } else {

-               assert(false);
+               assert(encoding == Stream::STREAM_ENCODING_RAW);

        }
        }

+
+       // Add the data itself.
+       iovec iov;
+       iov.iov_base = new char[bytes];
+       memcpy(iov.iov_base, data, bytes);
+       iov.iov_len = bytes;
+       queued_data.push_back(iov);

 }
 
 void Stream::process_queued_data()
 }
 
 void Stream::process_queued_data()


@@ -186,7 +228,11 @@ void Stream::process_queued_data()
                return;
        }
 
                return;
        }
 

-       add_data_raw(queued_data.data(), queued_data.size());
+       add_data_raw(queued_data);
+       for (size_t i = 0; i < queued_data.size(); ++i) {
+               char *data = reinterpret_cast<char *>(queued_data[i].iov_base);
+               delete[] data;
+       }

        queued_data.clear();
 
        // We have more data, so wake up all clients.
        queued_data.clear();
 
        // We have more data, so wake up all clients.

diff --git a/stream.h b/stream.h
index 97451f0c8afe7fe08dc57b7c44c18e64de4a0615..33fb55eeeab9180589d57275b182f806dcb3b9a3 100644 (file)
--- a/stream.h
+++ b/stream.h
@@ -7,6 +7,7 @@
 #include <stddef.h>
 #include <stdint.h>
 #include <sys/types.h>
 #include <stddef.h>
 #include <stdint.h>
 #include <sys/types.h>

+#include <sys/uio.h>

 #include <string>
 #include <vector>
 
 #include <string>
 #include <vector>
 


@@ -70,7 +71,8 @@ struct Stream {
        MarkPool *mark_pool;
 
        // Queued data, if any. Protected by the owning Server's <queued_data_mutex>.
        MarkPool *mark_pool;
 
        // Queued data, if any. Protected by the owning Server's <queued_data_mutex>.

-       std::string queued_data;
+       // The data pointers in the iovec are owned by us.
+       std::vector<iovec> queued_data;

 
        // Put client to sleep, since there is no more data for it; we will on
        // longer listen on POLLOUT until we get more data. Also, it will be put
 
        // Put client to sleep, since there is no more data for it; we will on
        // longer listen on POLLOUT until we get more data. Also, it will be put


@@ -78,6 +80,7 @@ struct Stream {
        void put_client_to_sleep(Client *client);
 
        // Add more data to <queued_data>, adding Metacube headers if needed.
        void put_client_to_sleep(Client *client);
 
        // Add more data to <queued_data>, adding Metacube headers if needed.

+       // Does not take ownership of <data>.

        // You should hold the owning Server's <queued_data_mutex>.
        void add_data_deferred(const char *data, size_t bytes);
 
        // You should hold the owning Server's <queued_data_mutex>.
        void add_data_deferred(const char *data, size_t bytes);
 


@@ -89,9 +92,9 @@ private:
        Stream(const Stream& other);
 
        // Adds data directly to the stream file descriptor, without adding headers or
        Stream(const Stream& other);
 
        // Adds data directly to the stream file descriptor, without adding headers or

-       // going through <queued_data>. You should hold the owning Server's
-       // <mutex>.
-       void add_data_raw(const char *data, ssize_t bytes);
+       // going through <queued_data>.
+       // You should hold the owning Server's <mutex>.
+       void add_data_raw(const std::vector<iovec> &data);

 };
 
 #endif  // !defined(_STREAM_H)
 };
 
 #endif  // !defined(_STREAM_H)