nghttpx: Rewrite TLS async handshake using memchunk buffers

src/memchunk.h

@@ -117,6 +117,31 @@ template <typename T> struct Pool {
 template <typename Memchunk> struct Memchunks {
   Memchunks(Pool<Memchunk> *pool)
       : pool(pool), head(nullptr), tail(nullptr), len(0) {}
+  Memchunks(const Memchunks &) = delete;
+  Memchunks(Memchunks &&other)
+      : pool(other.pool), head(other.head), tail(other.head), len(other.len) {
+    // keep other.pool
+    other.head = other.tail = nullptr;
+    other.len = 0;
+  }
+  Memchunks &operator=(const Memchunks &) = delete;
+  Memchunks &operator=(Memchunks &&other) {
+    if (this == &other) {
+      return *this;
+    }
+
+    reset();
+
+    pool = other.pool;
+    head = other.head;
+    tail = other.tail;
+    len = other.len;
+
+    other.head = other.tail = nullptr;
+    other.len = 0;
+
+    return *this;
+  }
   ~Memchunks() {
     if (!pool) {
       return;
@@ -223,15 +248,142 @@ template <typename Memchunk> struct Memchunks {
     return i;
   }
   size_t rleft() const { return len; }
+  void reset() {
+    for (auto m = head; m;) {
+      auto next = m->next;
+      pool->recycle(m);
+      m = next;
+    }
+    len = 0;
+    head = tail = nullptr;
+  }
 
   Pool<Memchunk> *pool;
   Memchunk *head, *tail;
   size_t len;
 };
 
+// Wrapper around Memchunks to offer "peeking" functionality.
+template <typename Memchunk> struct PeekMemchunks {
+  PeekMemchunks(Pool<Memchunk> *pool)
+      : memchunks(pool), cur(nullptr), cur_pos(nullptr), cur_last(nullptr),
+        len(0), peeking(true) {}
+  PeekMemchunks(const PeekMemchunks &) = delete;
+  PeekMemchunks(PeekMemchunks &&other)
+      : memchunks(std::move(other.memchunks)), cur(other.cur),
+        cur_pos(other.cur_pos), cur_last(other.cur_last), len(other.len),
+        peeking(other.peeking) {
+    other.reset();
+  }
+  PeekMemchunks &operator=(const PeekMemchunks &) = delete;
+  PeekMemchunks &operator=(PeekMemchunks &&other) {
+    if (this == &other) {
+      return *this;
+    }
+
+    memchunks = std::move(other.memchunks);
+    cur = other.cur;
+    cur_pos = other.cur_pos;
+    cur_last = other.cur_last;
+    len = other.len;
+    peeking = other.peeking;
+
+    other.reset();
+
+    return *this;
+  }
+  size_t append(const void *src, size_t count) {
+    count = memchunks.append(src, count);
+    len += count;
+    return count;
+  }
+  size_t remove(void *dest, size_t count) {
+    if (!peeking) {
+      count = memchunks.remove(dest, count);
+      len -= count;
+      return count;
+    }
+
+    if (count == 0 || len == 0) {
+      return 0;
+    }
+
+    if (!cur) {
+      cur = memchunks.head;
+      cur_pos = cur->pos;
+    }
+
+    // cur_last could be updated in append
+    cur_last = cur->last;
+
+    if (cur_pos == cur_last) {
+      assert(cur->next);
+      cur = cur->next;
+    }
+
+    auto first = static_cast<uint8_t *>(dest);
+    auto last = first + count;
+
+    for (;;) {
+      auto n = std::min(last - first, cur_last - cur_pos);
+
+      first = std::copy_n(cur_pos, n, first);
+      cur_pos += n;
+      len -= n;
+
+      if (first == last) {
+        break;
+      }
+      assert(cur_pos == cur_last);
+      if (!cur->next) {
+        break;
+      }
+      cur = cur->next;
+      cur_pos = cur->pos;
+      cur_last = cur->last;
+    }
+    return first - static_cast<uint8_t *>(dest);
+  }
+  size_t rleft() const { return len; }
+  size_t rleft_buffered() const { return memchunks.rleft(); }
+  void disable_peek(bool drain) {
+    if (!peeking) {
+      return;
+    }
+    if (drain) {
+      auto n = rleft_buffered() - rleft();
+      memchunks.drain(n);
+      assert(len == memchunks.rleft());
+    } else {
+      len = memchunks.rleft();
+    }
+    cur = nullptr;
+    cur_pos = cur_last = nullptr;
+    peeking = false;
+  }
+  void reset() {
+    memchunks.reset();
+    cur = nullptr;
+    cur_pos = cur_last = nullptr;
+    len = 0;
+    peeking = true;
+  }
+  Memchunks<Memchunk> memchunks;
+  // Pointer to the Memchunk currently we are reading/writing.
+  Memchunk *cur;
+  // Region inside cur, we have processed to cur_pos.
+  uint8_t *cur_pos, *cur_last;
+  // This is the length we have left unprocessed.  len <=
+  // memchunk.rleft() must hold.
+  size_t len;
+  // true if peeking is enabled.  Initially it is true.
+  bool peeking;
+};
+
 using Memchunk16K = Memchunk<16_k>;
 using MemchunkPool = Pool<Memchunk16K>;
 using DefaultMemchunks = Memchunks<Memchunk16K>;
+using DefaultPeekMemchunks = PeekMemchunks<Memchunk16K>;
 
 #define DEFAULT_WR_IOVCNT 16
 

src/memchunk_test.cc

@@ -84,6 +84,7 @@ void test_pool_recycle(void) {
 using Memchunk16 = Memchunk<16>;
 using MemchunkPool16 = Pool<Memchunk16>;
 using Memchunks16 = Memchunks<Memchunk16>;
+using PeekMemchunks16 = PeekMemchunks<Memchunk16>;
 
 void test_memchunks_append(void) {
   MemchunkPool16 pool;
@@ -196,4 +197,144 @@ void test_memchunks_recycle(void) {
   CU_ASSERT(nullptr == m->next);
 }
 
+void test_memchunks_reset(void) {
+  MemchunkPool16 pool;
+  Memchunks16 chunks(&pool);
+
+  std::array<uint8_t, 32> b{};
+
+  chunks.append(b.data(), b.size());
+
+  CU_ASSERT(32 == chunks.rleft());
+
+  chunks.reset();
+
+  CU_ASSERT(0 == chunks.rleft());
+  CU_ASSERT(nullptr == chunks.head);
+  CU_ASSERT(nullptr == chunks.tail);
+
+  auto m = pool.freelist;
+
+  CU_ASSERT(nullptr != m);
+  CU_ASSERT(nullptr != m->next);
+  CU_ASSERT(nullptr == m->next->next);
+}
+
+void test_peek_memchunks_append(void) {
+  MemchunkPool16 pool;
+  PeekMemchunks16 pchunks(&pool);
+
+  std::array<uint8_t, 32> b{{
+      '0', '1', '2', '3', '4', '5', '6', '7', '8', '9', '0', '1', '2', '3', '4',
+      '5', '6', '7', '8', '9', '0', '1',
+  }},
+      d;
+
+  pchunks.append(b.data(), b.size());
+
+  CU_ASSERT(32 == pchunks.rleft());
+  CU_ASSERT(32 == pchunks.rleft_buffered());
+
+  CU_ASSERT(0 == pchunks.remove(nullptr, 0));
+
+  CU_ASSERT(32 == pchunks.rleft());
+  CU_ASSERT(32 == pchunks.rleft_buffered());
+
+  CU_ASSERT(12 == pchunks.remove(d.data(), 12));
+
+  CU_ASSERT(std::equal(std::begin(b), std::begin(b) + 12, std::begin(d)));
+
+  CU_ASSERT(20 == pchunks.rleft());
+  CU_ASSERT(32 == pchunks.rleft_buffered());
+
+  CU_ASSERT(20 == pchunks.remove(d.data(), d.size()));
+
+  CU_ASSERT(std::equal(std::begin(b) + 12, std::end(b), std::begin(d)));
+
+  CU_ASSERT(0 == pchunks.rleft());
+  CU_ASSERT(32 == pchunks.rleft_buffered());
+}
+
+void test_peek_memchunks_disable_peek_drain(void) {
+  MemchunkPool16 pool;
+  PeekMemchunks16 pchunks(&pool);
+
+  std::array<uint8_t, 32> b{{
+      '0', '1', '2', '3', '4', '5', '6', '7', '8', '9', '0', '1', '2', '3', '4',
+      '5', '6', '7', '8', '9', '0', '1',
+  }},
+      d;
+
+  pchunks.append(b.data(), b.size());
+
+  CU_ASSERT(12 == pchunks.remove(d.data(), 12));
+
+  pchunks.disable_peek(true);
+
+  CU_ASSERT(!pchunks.peeking);
+  CU_ASSERT(20 == pchunks.rleft());
+  CU_ASSERT(20 == pchunks.rleft_buffered());
+
+  CU_ASSERT(20 == pchunks.remove(d.data(), d.size()));
+
+  CU_ASSERT(std::equal(std::begin(b) + 12, std::end(b), std::begin(d)));
+
+  CU_ASSERT(0 == pchunks.rleft());
+  CU_ASSERT(0 == pchunks.rleft_buffered());
+}
+
+void test_peek_memchunks_disable_peek_no_drain(void) {
+  MemchunkPool16 pool;
+  PeekMemchunks16 pchunks(&pool);
+
+  std::array<uint8_t, 32> b{{
+      '0', '1', '2', '3', '4', '5', '6', '7', '8', '9', '0', '1', '2', '3', '4',
+      '5', '6', '7', '8', '9', '0', '1',
+  }},
+      d;
+
+  pchunks.append(b.data(), b.size());
+
+  CU_ASSERT(12 == pchunks.remove(d.data(), 12));
+
+  pchunks.disable_peek(false);
+
+  CU_ASSERT(!pchunks.peeking);
+  CU_ASSERT(32 == pchunks.rleft());
+  CU_ASSERT(32 == pchunks.rleft_buffered());
+
+  CU_ASSERT(32 == pchunks.remove(d.data(), d.size()));
+
+  CU_ASSERT(std::equal(std::begin(b), std::end(b), std::begin(d)));
+
+  CU_ASSERT(0 == pchunks.rleft());
+  CU_ASSERT(0 == pchunks.rleft_buffered());
+}
+
+void test_peek_memchunks_reset(void) {
+  MemchunkPool16 pool;
+  PeekMemchunks16 pchunks(&pool);
+
+  std::array<uint8_t, 32> b{{
+      '0', '1', '2', '3', '4', '5', '6', '7', '8', '9', '0', '1', '2', '3', '4',
+      '5', '6', '7', '8', '9', '0', '1',
+  }},
+      d;
+
+  pchunks.append(b.data(), b.size());
+
+  CU_ASSERT(12 == pchunks.remove(d.data(), 12));
+
+  pchunks.disable_peek(true);
+  pchunks.reset();
+
+  CU_ASSERT(0 == pchunks.rleft());
+  CU_ASSERT(0 == pchunks.rleft_buffered());
+
+  CU_ASSERT(nullptr == pchunks.cur);
+  CU_ASSERT(nullptr == pchunks.cur_pos);
+  CU_ASSERT(nullptr == pchunks.cur_last);
+  CU_ASSERT(pchunks.peeking);
+}
+
 } // namespace nghttp2

src/memchunk_test.h

@@ -36,6 +36,11 @@ void test_memchunks_append(void);
 void test_memchunks_drain(void);
 void test_memchunks_riovec(void);
 void test_memchunks_recycle(void);
+void test_memchunks_reset(void);
+void test_peek_memchunks_append(void);
+void test_peek_memchunks_disable_peek_drain(void);
+void test_peek_memchunks_disable_peek_no_drain(void);
+void test_peek_memchunks_reset(void);
 
 } // namespace nghttp2
 

src/shrpx-unittest.cc

@@ -172,7 +172,16 @@ int main(int argc, char *argv[]) {
       !CU_add_test(pSuite, "memchunk_drain", nghttp2::test_memchunks_drain) ||
       !CU_add_test(pSuite, "memchunk_riovec", nghttp2::test_memchunks_riovec) ||
       !CU_add_test(pSuite, "memchunk_recycle",
-                   nghttp2::test_memchunks_recycle)) {
+                   nghttp2::test_memchunks_recycle) ||
+      !CU_add_test(pSuite, "memchunk_reset", nghttp2::test_memchunks_reset) ||
+      !CU_add_test(pSuite, "peek_memchunk_append",
+                   nghttp2::test_peek_memchunks_append) ||
+      !CU_add_test(pSuite, "peek_memchunk_disable_peek_drain",
+                   nghttp2::test_peek_memchunks_disable_peek_drain) ||
+      !CU_add_test(pSuite, "peek_memchunk_disable_peek_no_drain",
+                   nghttp2::test_peek_memchunks_disable_peek_no_drain) ||
+      !CU_add_test(pSuite, "peek_memchunk_reset",
+                   nghttp2::test_peek_memchunks_reset)) {
     CU_cleanup_registry();
     return CU_get_error();
   }

src/shrpx_client_handler.cc

@@ -358,7 +358,8 @@ int ClientHandler::upstream_http1_connhd_read() {
 
 ClientHandler::ClientHandler(Worker *worker, int fd, SSL *ssl,
                              const char *ipaddr, const char *port)
-    : conn_(worker->get_loop(), fd, ssl, get_config()->upstream_write_timeout,
+    : conn_(worker->get_loop(), fd, ssl, worker->get_mcpool(),
+            get_config()->upstream_write_timeout,
             get_config()->upstream_read_timeout, get_config()->write_rate,
             get_config()->write_burst, get_config()->read_rate,
             get_config()->read_burst, writecb, readcb, timeoutcb, this),

src/shrpx_connection.cc

@@ -40,12 +40,13 @@ using namespace nghttp2;
 
 namespace shrpx {
 Connection::Connection(struct ev_loop *loop, int fd, SSL *ssl,
-                       ev_tstamp write_timeout, ev_tstamp read_timeout,
-                       size_t write_rate, size_t write_burst, size_t read_rate,
-                       size_t read_burst, IOCb writecb, IOCb readcb,
-                       TimerCb timeoutcb, void *data)
-    : tls{}, wlimit(loop, &wev, write_rate, write_burst),
-      rlimit(loop, &rev, read_rate, read_burst, ssl), writecb(writecb),
+                       MemchunkPool *mcpool, ev_tstamp write_timeout,
+                       ev_tstamp read_timeout, size_t write_rate,
+                       size_t write_burst, size_t read_rate, size_t read_burst,
+                       IOCb writecb, IOCb readcb, TimerCb timeoutcb, void *data)
+    : tls{DefaultMemchunks(mcpool), DefaultPeekMemchunks(mcpool)},
+      wlimit(loop, &wev, write_rate, write_burst),
+      rlimit(loop, &rev, read_rate, read_burst, this), writecb(writecb),
       readcb(readcb), timeoutcb(timeoutcb), loop(loop), data(data), fd(fd) {
 
   ev_io_init(&wev, writecb, fd, EV_WRITE);
@@ -83,10 +84,12 @@ void Connection::disconnect() {
 
     if (tls.cached_session) {
       SSL_SESSION_free(tls.cached_session);
+      tls.cached_session = nullptr;
     }
 
     if (tls.cached_session_lookup_req) {
       tls.cached_session_lookup_req->canceled = true;
+      tls.cached_session_lookup_req = nullptr;
     }
 
     // To reuse SSL/TLS session, we have to shutdown, and don't free
@@ -96,7 +99,15 @@ void Connection::disconnect() {
       tls.ssl = nullptr;
     }
 
-    tls = {tls.ssl};
+    tls.wbuf.reset();
+    tls.rbuf.reset();
+    tls.last_write_idle = 0.;
+    tls.warmup_writelen = 0;
+    tls.last_writelen = 0;
+    tls.last_readlen = 0;
+    tls.handshake_state = 0;
+    tls.initial_handshake_done = false;
+    tls.reneg_started = false;
   }
 
   if (fd != -1) {
@@ -114,22 +125,9 @@ void Connection::disconnect() {
   wlimit.stopw();
 }
 
-namespace {
-void allocate_buffer(Connection *conn) {
-  conn->tls.rb = make_unique<Buffer<16_k>>();
-  conn->tls.wb = make_unique<Buffer<16_k>>();
-}
-} // namespace
-
-void Connection::prepare_client_handshake() {
-  SSL_set_connect_state(tls.ssl);
-  allocate_buffer(this);
-}
+void Connection::prepare_client_handshake() { SSL_set_connect_state(tls.ssl); }
 
-void Connection::prepare_server_handshake() {
-  SSL_set_accept_state(tls.ssl);
-  allocate_buffer(this);
-}
+void Connection::prepare_server_handshake() { SSL_set_accept_state(tls.ssl); }
 
 // BIO implementation is inspired by openldap implementation:
 // http://www.openldap.org/devel/cvsweb.cgi/~checkout~/libraries/libldap/tls_o.c
@@ -140,27 +138,26 @@ int shrpx_bio_write(BIO *b, const char *buf, int len) {
   }
 
   auto conn = static_cast<Connection *>(b->ptr);
-  auto &wb = conn->tls.wb;
+  auto &wbuf = conn->tls.wbuf;
 
   BIO_clear_retry_flags(b);
 
   if (conn->tls.initial_handshake_done) {
     // After handshake finished, send |buf| of length |len| to the
     // socket directly.
-    if (wb && wb->rleft()) {
-      auto nwrite = conn->write_clear(wb->pos, wb->rleft());
+    if (wbuf.rleft()) {
+      std::array<struct iovec, 4> iov;
+      auto iovcnt = wbuf.riovec(iov.data(), iov.size());
+      auto nwrite = conn->writev_clear(iov.data(), iovcnt);
       if (nwrite < 0) {
         return -1;
       }
 
-      wb->drain(nwrite);
-      if (wb->rleft()) {
+      wbuf.drain(nwrite);
+      if (wbuf.rleft()) {
         BIO_set_retry_write(b);
         return -1;
       }
-
-      // Here delete TLS write buffer
-      wb.reset();
     }
     auto nwrite = conn->write_clear(buf, len);
     if (nwrite < 0) {
@@ -175,16 +172,9 @@ int shrpx_bio_write(BIO *b, const char *buf, int len) {
     return nwrite;
   }
 
-  auto nwrite = std::min(static_cast<size_t>(len), wb->wleft());
-
-  if (nwrite == 0) {
-    BIO_set_retry_write(b);
-    return -1;
-  }
-
-  wb->write(buf, nwrite);
+  wbuf.append(buf, len);
 
-  return nwrite;
+  return len;
 }
 } // namespace
 
@@ -195,11 +185,11 @@ int shrpx_bio_read(BIO *b, char *buf, int len) {
   }
 
   auto conn = static_cast<Connection *>(b->ptr);
-  auto &rb = conn->tls.rb;
+  auto &rbuf = conn->tls.rbuf;
 
   BIO_clear_retry_flags(b);
 
-  if (conn->tls.initial_handshake_done && !rb) {
+  if (conn->tls.initial_handshake_done && rbuf.rleft() == 0) {
     auto nread = conn->read_clear(buf, len);
     if (nread < 0) {
       return -1;
@@ -211,22 +201,12 @@ int shrpx_bio_read(BIO *b, char *buf, int len) {
     return nread;
   }
 
-  auto nread = std::min(static_cast<size_t>(len), rb->rleft());
-
-  if (nread == 0) {
-    if (conn->tls.initial_handshake_done) {
-      rb.reset();
-    }
-
+  if (rbuf.rleft() == 0) {
     BIO_set_retry_read(b);
     return -1;
   }
 
-  std::copy_n(rb->pos, nread, buf);
-
-  rb->drain(nread);
-
-  return nread;
+  return rbuf.remove(buf, len);
 }
 } // namespace
 
@@ -289,51 +269,47 @@ void Connection::set_ssl(SSL *ssl) {
   bio->ptr = this;
   SSL_set_bio(tls.ssl, bio, bio);
   SSL_set_app_data(tls.ssl, this);
-  rlimit.set_ssl(tls.ssl);
 }
 
+namespace {
+// We should buffer at least full encrypted TLS record here.
+// Theoretically, peer can send client hello in several TLS records,
+// which could exeed this limit, but it is not portable, and we don't
+// have to handle such exotic behaviour.
+bool read_buffer_full(DefaultPeekMemchunks &rbuf) {
+  return rbuf.rleft_buffered() >= 20_k;
+}
+} // namespace
+
 int Connection::tls_handshake() {
   wlimit.stopw();
   ev_timer_stop(loop, &wt);
 
-  auto nread = read_clear(tls.rb->last, tls.rb->wleft());
-  if (nread < 0) {
-    if (LOG_ENABLED(INFO)) {
-      LOG(INFO) << "tls: handshake read error";
-    }
-    return -1;
-  }
-  tls.rb->write(nread);
-
-  // We have limited space for read buffer, so stop reading if it
-  // filled up.
-  if (tls.rb->wleft() == 0) {
-    if (tls.handshake_state != TLS_CONN_WRITE_STARTED) {
-      // Reading 16KiB before writing server hello is unlikely for
-      // ordinary client.
+  if (ev_is_active(&rev)) {
+    std::array<uint8_t, 8_k> buf;
+    auto nread = read_clear(buf.data(), buf.size());
+    if (nread < 0) {
       if (LOG_ENABLED(INFO)) {
-        LOG(INFO) << "tls: client hello is too large";
+        LOG(INFO) << "tls: handshake read error";
       }
       return -1;
     }
-
-    rlimit.stopw();
-    ev_timer_stop(loop, &rt);
+    tls.rbuf.append(buf.data(), nread);
+    if (read_buffer_full(tls.rbuf)) {
+      rlimit.stopw();
+    }
   }
 
   switch (tls.handshake_state) {
   case TLS_CONN_WAIT_FOR_SESSION_CACHE:
-    if (nread > 0) {
-      if (LOG_ENABLED(INFO)) {
-        LOG(INFO) << "tls: client sent addtional data after client hello";
-      }
-      return -1;
-    }
     return SHRPX_ERR_INPROGRESS;
   case TLS_CONN_GOT_SESSION_CACHE: {
-    // Use the same trick invented by @kazuho in h2o project
-    tls.wb->reset();
-    tls.rb->pos = tls.rb->begin();
+    // Use the same trick invented by @kazuho in h2o project.
+
+    // Discard all outgoing data.
+    tls.wbuf.reset();
+    // Rewind buffered incoming data to replay client hello.
+    tls.rbuf.disable_peek(false);
 
     auto ssl_ctx = SSL_get_SSL_CTX(tls.ssl);
     auto ssl_opts = SSL_get_options(tls.ssl);
@@ -382,32 +358,33 @@ int Connection::tls_handshake() {
     return SHRPX_ERR_INPROGRESS;
   }
 
-  if (tls.wb->rleft()) {
+  if (tls.wbuf.rleft()) {
     // First write indicates that resumption stuff has done.
-    tls.handshake_state = TLS_CONN_WRITE_STARTED;
-    auto nwrite = write_clear(tls.wb->pos, tls.wb->rleft());
+    if (tls.handshake_state != TLS_CONN_WRITE_STARTED) {
+      tls.handshake_state = TLS_CONN_WRITE_STARTED;
+      // If peek has already disabled, this is noop.
+      tls.rbuf.disable_peek(true);
+    }
+    std::array<struct iovec, 4> iov;
+    auto iovcnt = tls.wbuf.riovec(iov.data(), iov.size());
+    auto nwrite = writev_clear(iov.data(), iovcnt);
     if (nwrite < 0) {
       if (LOG_ENABLED(INFO)) {
         LOG(INFO) << "tls: handshake write error";
       }
       return -1;
     }
-    tls.wb->drain(nwrite);
-  }
+    tls.wbuf.drain(nwrite);
 
-  if (tls.wb->rleft()) {
-    wlimit.startw();
-    ev_timer_again(loop, &wt);
-  } else {
-    tls.wb->reset();
+    if (tls.wbuf.rleft()) {
+      wlimit.startw();
+      ev_timer_again(loop, &wt);
+    }
   }
 
-  if (tls.handshake_state == TLS_CONN_WRITE_STARTED && tls.rb->rleft() == 0) {
-    tls.rb->reset();
-
+  if (!read_buffer_full(tls.rbuf)) {
     // We may have stopped reading
     rlimit.startw();
-    ev_timer_again(loop, &rt);
   }
 
   if (rv != 1) {
@@ -419,13 +396,14 @@ int Connection::tls_handshake() {
 
   tls.initial_handshake_done = true;
 
-  if (tls.rb->rleft()) {
-    ev_feed_event(loop, &rev, EV_READ);
-  }
-
-  // We may have stopped reading
+  // We have to start read watcher, since later stage of code expects
+  // this.
   rlimit.startw();
-  ev_timer_again(loop, &rt);
+
+  // We may have whole request in tls.rbuf.  This means that we don't
+  // get notified further read event.  This is especially true for
+  // HTTP/1.1.
+  handle_tls_pending_read();
 
   if (LOG_ENABLED(INFO)) {
     LOG(INFO) << "SSL/TLS handshake completed";
@@ -506,8 +484,8 @@ ssize_t Connection::write_tls(const void *data, size_t len) {
       return SHRPX_ERR_NETWORK;
     case SSL_ERROR_WANT_WRITE:
       tls.last_writelen = len;
-      wlimit.startw();
-      ev_timer_again(loop, &wt);
+      // starting write watcher and timer is done in write_clear via
+      // bio.
       return 0;
     default:
       if (LOG_ENABLED(INFO)) {

src/shrpx_connection.h

@@ -35,7 +35,7 @@
 
 #include "shrpx_rate_limit.h"
 #include "shrpx_error.h"
-#include "buffer.h"
+#include "memchunk.h"
 
 namespace shrpx {
 
@@ -50,6 +50,8 @@ enum {
 };
 
 struct TLSConnection {
+  DefaultMemchunks wbuf;
+  DefaultPeekMemchunks rbuf;
   SSL *ssl;
   SSL_SESSION *cached_session;
   MemcachedRequest *cached_session_lookup_req;
@@ -62,8 +64,6 @@ struct TLSConnection {
   int handshake_state;
   bool initial_handshake_done;
   bool reneg_started;
-  std::unique_ptr<Buffer<16_k>> rb;
-  std::unique_ptr<Buffer<16_k>> wb;
 };
 
 template <typename T> using EVCb = void (*)(struct ev_loop *, T *, int);
@@ -72,10 +72,10 @@ using IOCb = EVCb<ev_io>;
 using TimerCb = EVCb<ev_timer>;
 
 struct Connection {
-  Connection(struct ev_loop *loop, int fd, SSL *ssl, ev_tstamp write_timeout,
-             ev_tstamp read_timeout, size_t write_rate, size_t write_burst,
-             size_t read_rate, size_t read_burst, IOCb writecb, IOCb readcb,
-             TimerCb timeoutcb, void *data);
+  Connection(struct ev_loop *loop, int fd, SSL *ssl, MemchunkPool *mcpool,
+             ev_tstamp write_timeout, ev_tstamp read_timeout, size_t write_rate,
+             size_t write_burst, size_t read_rate, size_t read_burst,
+             IOCb writecb, IOCb readcb, TimerCb timeoutcb, void *data);
   ~Connection();
 
   void disconnect();

src/shrpx_http2_session.cc

@@ -144,7 +144,8 @@ void writecb(struct ev_loop *loop, ev_io *w, int revents) {
 Http2Session::Http2Session(struct ev_loop *loop, SSL_CTX *ssl_ctx,
                            ConnectBlocker *connect_blocker, Worker *worker,
                            size_t group, size_t idx)
-    : conn_(loop, -1, nullptr, get_config()->downstream_write_timeout,
+    : conn_(loop, -1, nullptr, worker->get_mcpool(),
+            get_config()->downstream_write_timeout,
             get_config()->downstream_read_timeout, 0, 0, 0, 0, writecb, readcb,
             timeoutcb, this),
       worker_(worker), connect_blocker_(connect_blocker), ssl_ctx_(ssl_ctx),

src/shrpx_http_downstream_connection.cc

@@ -112,7 +112,7 @@ void connectcb(struct ev_loop *loop, ev_io *w, int revents) {
 HttpDownstreamConnection::HttpDownstreamConnection(
     DownstreamConnectionPool *dconn_pool, size_t group, struct ev_loop *loop)
     : DownstreamConnection(dconn_pool),
-      conn_(loop, -1, nullptr, get_config()->downstream_write_timeout,
+      conn_(loop, -1, nullptr, nullptr, get_config()->downstream_write_timeout,
             get_config()->downstream_read_timeout, 0, 0, 0, 0, connectcb,
             readcb, timeoutcb, this),
       ioctrl_(&conn_.rlimit), response_htp_{0}, group_(group), addr_idx_(0),

src/shrpx_memcached_connection.cc

@@ -92,7 +92,7 @@ constexpr ev_tstamp read_timeout = 10.;
 
 MemcachedConnection::MemcachedConnection(const Address *addr,
                                          struct ev_loop *loop)
-    : conn_(loop, -1, nullptr, write_timeout, read_timeout, 0, 0, 0, 0,
+    : conn_(loop, -1, nullptr, nullptr, write_timeout, read_timeout, 0, 0, 0, 0,
             connectcb, readcb, timeoutcb, this),
       parse_state_{}, addr_(addr), sendsum_(0), connected_(false) {}
 
@@ -403,6 +403,7 @@ int MemcachedConnection::parse_packet() {
   return 0;
 }
 
+#undef DEFAULT_WR_IOVCNT
 #define DEFAULT_WR_IOVCNT 128
 
 #if defined(IOV_MAX) && IOV_MAX < DEFAULT_WR_IOVCNT

src/shrpx_rate_limit.cc

@@ -26,6 +26,8 @@
 
 #include <limits>
 
+#include "shrpx_connection.h"
+
 namespace shrpx {
 
 namespace {
@@ -36,9 +38,9 @@ void regencb(struct ev_loop *loop, ev_timer *w, int revents) {
 } // namespace
 
 RateLimit::RateLimit(struct ev_loop *loop, ev_io *w, size_t rate, size_t burst,
-                     SSL *ssl)
-    : w_(w), loop_(loop), ssl_(ssl), rate_(rate), burst_(burst), avail_(burst),
-      startw_req_(false) {
+                     Connection *conn)
+    : w_(w), loop_(loop), conn_(conn), rate_(rate), burst_(burst),
+      avail_(burst), startw_req_(false) {
   ev_timer_init(&t_, regencb, 0., 1.);
   t_.data = this;
   if (rate_ > 0) {
@@ -97,7 +99,8 @@ void RateLimit::stopw() {
 }
 
 void RateLimit::handle_tls_pending_read() {
-  if (!ssl_ || SSL_pending(ssl_) == 0) {
+  if (!conn_ || !conn_->tls.ssl ||
+      (SSL_pending(conn_->tls.ssl) == 0 && conn_->tls.rbuf.rleft() == 0)) {
     return;
   }
 
@@ -106,6 +109,4 @@ void RateLimit::handle_tls_pending_read() {
   ev_feed_event(loop_, w_, EV_READ);
 }
 
-void RateLimit::set_ssl(SSL *ssl) { ssl_ = ssl; }
-
 } // namespace shrpx

src/shrpx_rate_limit.h

@@ -33,28 +33,30 @@
 
 namespace shrpx {
 
+struct Connection;
+
 class RateLimit {
 public:
-  // We need |ssl| object to check that it has unread decrypted bytes.
+  // We need |conn| object to check that it has unread bytes for TLS
+  // connection.
   RateLimit(struct ev_loop *loop, ev_io *w, size_t rate, size_t burst,
-            SSL *ssl = nullptr);
+            Connection *conn = nullptr);
   ~RateLimit();
   size_t avail() const;
   void drain(size_t n);
   void regen();
   void startw();
   void stopw();
-  // Feeds event if ssl_ object has unread decrypted bytes.  This is
-  // required since it is buffered in ssl_ object, io event is not
-  // generated unless new incoming data is received.
+  // Feeds event if conn_->tls object has unread bytes.  This is
+  // required since it is buffered in conn_->tls object, io event is
+  // not generated unless new incoming data is received.
   void handle_tls_pending_read();
-  void set_ssl(SSL *ssl);
 
 private:
   ev_timer t_;
   ev_io *w_;
   struct ev_loop *loop_;
-  SSL *ssl_;
+  Connection *conn_;
   size_t rate_;
   size_t burst_;
   size_t avail_;