Merge branch 'timeout' into usrp_e_mmap

6 files changed, 293 insertions, 320 deletions

diff --git a/host/lib/transport/CMakeLists.txt b/host/lib/transport/CMakeLists.txt
index e9e82932c..eafdd57d2 100644
--- a/host/lib/transport/CMakeLists.txt
+++ b/host/lib/transport/CMakeLists.txt
@@ -36,6 +36,8 @@ IF(LIBUSB_FOUND)
         INCLUDE_DIRECTORIES(${CMAKE_SOURCE_DIR}/lib/transport/msvc)
     ENDIF(MSVC)
     SET(HAVE_USB_SUPPORT TRUE)
+ELSE(LIBUSB_FOUND)
+    #TODO dummy usb
 ENDIF(LIBUSB_FOUND)
 
 IF(HAVE_USB_SUPPORT)
diff --git a/host/lib/transport/libusb1_base.cpp b/host/lib/transport/libusb1_base.cpp
index 49f524a32..910b04fc8 100644
--- a/host/lib/transport/libusb1_base.cpp
+++ b/host/lib/transport/libusb1_base.cpp
@@ -213,15 +213,21 @@ private:
 libusb::device_handle::sptr libusb::device_handle::get_cached_handle(device::sptr dev){
     static uhd::dict<libusb_device *, boost::weak_ptr<device_handle> > handles;
 
-    //not expired -> get existing session
+    //not expired -> get existing handle
     if (handles.has_key(dev->get()) and not handles[dev->get()].expired()){
         return handles[dev->get()].lock();
     }
 
-    //create a new global session
-    sptr new_handle(new libusb_device_handle_impl(dev));
-    handles[dev->get()] = new_handle;
-    return new_handle;
+    //create a new cached handle
+    try{
+        sptr new_handle(new libusb_device_handle_impl(dev));
+        handles[dev->get()] = new_handle;
+        return new_handle;
+    }
+    catch(const std::exception &e){
+        std::cerr << "USB open failed: see the application notes for your device." << std::endl;
+        throw std::runtime_error(e.what());
+    }
 }
 
 /***********************************************************************
diff --git a/host/lib/transport/libusb1_zero_copy.cpp b/host/lib/transport/libusb1_zero_copy.cpp
index f9beb0b4c..ab48e4fc4 100644
--- a/host/lib/transport/libusb1_zero_copy.cpp
+++ b/host/lib/transport/libusb1_zero_copy.cpp
@@ -24,12 +24,10 @@
 #include <boost/thread.hpp>
 #include <vector>
 #include <iostream>
-#include <iomanip>
 
 using namespace uhd::transport;
 
-const int libusb_timeout = 0;
-
+static const double CLEANUP_TIMEOUT   = 0.2;    //seconds
 static const size_t DEFAULT_NUM_XFERS = 16;     //num xfers
 static const size_t DEFAULT_XFER_SIZE = 32*512; //bytes
 
@@ -84,10 +82,10 @@ public:
      * Get an available transfer:
      * For inputs, this is a just filled transfer.
      * For outputs, this is a just emptied transfer.
-     * \param timeout_ms the timeout to wait for a lut
+     * \param timeout the timeout to wait for a lut
      * \return the transfer pointer or NULL if timeout
      */
-    libusb_transfer *get_lut_with_wait(size_t timeout_ms = 100);
+    libusb_transfer *get_lut_with_wait(double timeout);
 
     //Callback use only
     void callback_handle_transfer(libusb_transfer *lut);
@@ -97,21 +95,18 @@ private:
     int  _endpoint;
     bool _input;
 
-    size_t _transfer_size;
-    size_t _num_transfers;
-
     //! hold a bounded buffer of completed transfers
     typedef bounded_buffer<libusb_transfer *> lut_buff_type;
     lut_buff_type::sptr _completed_list;
 
     //! a list of all transfer structs we allocated
-    std::vector<libusb_transfer *>  _all_luts;
+    std::vector<libusb_transfer *> _all_luts;
 
-    //! a list of shared arrays for the transfer buffers
-    std::vector<boost::shared_array<boost::uint8_t> > _buffers;
+    //! a block of memory for the transfer buffers
+    boost::shared_array<char> _buffer;
 
     // Calls for processing asynchronous I/O
-    libusb_transfer *allocate_transfer(int buff_len);
+    libusb_transfer *allocate_transfer(void *mem, size_t len);
     void print_transfer_status(libusb_transfer *lut);
 };
 
@@ -156,14 +151,12 @@ usb_endpoint::usb_endpoint(
 ):
     _handle(handle),
     _endpoint(endpoint),
-    _input(input),
-    _transfer_size(transfer_size),
-    _num_transfers(num_transfers)
+    _input(input)
 {
     _completed_list = lut_buff_type::make(num_transfers);
-
-    for (size_t i = 0; i < _num_transfers; i++){
-        _all_luts.push_back(allocate_transfer(_transfer_size));
+    _buffer = boost::shared_array<char>(new char[num_transfers*transfer_size]);
+    for (size_t i = 0; i < num_transfers; i++){
+        _all_luts.push_back(allocate_transfer(_buffer.get() + i*transfer_size, transfer_size));
 
         //input luts are immediately submitted to be filled
         //output luts go into the completed list as free buffers
@@ -187,7 +180,7 @@ usb_endpoint::~usb_endpoint(void){
     }
 
     //collect canceled transfers (drain the queue)
-    while (this->get_lut_with_wait() != NULL){};
+    while (this->get_lut_with_wait(CLEANUP_TIMEOUT) != NULL){};
 
     //free all transfers
     BOOST_FOREACH(libusb_transfer *lut, _all_luts){
@@ -200,23 +193,23 @@ usb_endpoint::~usb_endpoint(void){
  * Allocate a libusb transfer
  * The allocated transfer - and buffer it contains - is repeatedly
  * submitted, reaped, and reused and should not be freed until shutdown.
- * \param buff_len size of the individual buffer held by each transfer
+ * \param mem a pointer to the buffer memory
+ * \param len size of the individual buffer
  * \return pointer to an allocated libusb_transfer
  */
-libusb_transfer *usb_endpoint::allocate_transfer(int buff_len){
+libusb_transfer *usb_endpoint::allocate_transfer(void *mem, size_t len){
     libusb_transfer *lut = libusb_alloc_transfer(0);
-
-    boost::shared_array<boost::uint8_t> buff(new boost::uint8_t[buff_len]);
-    _buffers.push_back(buff); //store a reference to this shared array
+    UHD_ASSERT_THROW(lut != NULL);
 
     unsigned int endpoint = ((_endpoint & 0x7f) | (_input ? 0x80 : 0));
+    unsigned char *buff = reinterpret_cast<unsigned char *>(mem);
     libusb_transfer_cb_fn lut_callback = libusb_transfer_cb_fn(&callback);
 
     libusb_fill_bulk_transfer(lut,                // transfer
                               _handle->get(),     // dev_handle
                               endpoint,           // endpoint
-                              buff.get(),         // buffer
-                              buff_len,           // length
+                              buff,               // buffer
+                              len,                // length
                               lut_callback,       // callback
                               this,               // user_data
                               0);                 // timeout
@@ -239,6 +232,7 @@ void usb_endpoint::submit(libusb_transfer *lut){
  * \param lut pointer to an libusb_transfer
  */
 void usb_endpoint::print_transfer_status(libusb_transfer *lut){
+    std::cout << "here " << lut->status << std::endl;
     switch (lut->status) {
     case LIBUSB_TRANSFER_COMPLETED:
         if (lut->actual_length < lut->length) {
@@ -274,136 +268,52 @@ void usb_endpoint::print_transfer_status(libusb_transfer *lut){
     }
 }
 
-libusb_transfer *usb_endpoint::get_lut_with_wait(size_t timeout_ms){
+libusb_transfer *usb_endpoint::get_lut_with_wait(double timeout){
     boost::this_thread::disable_interruption di; //disable because the wait can throw
     libusb_transfer *lut;
-    if (_completed_list->pop_with_timed_wait(
-        lut, boost::posix_time::milliseconds(timeout_ms)
-    )) return lut;
+    if (_completed_list->pop_with_timed_wait(lut, timeout)) return lut;
     return NULL;
 }
 
 /***********************************************************************
- * Managed buffers
+ * USB zero_copy device class
  **********************************************************************/
-/*
- * Libusb managed receive buffer
- * Construct a recv buffer from a libusb transfer. The memory held by
- * the libusb transfer is exposed through the managed buffer interface.
- * Upon destruction, the transfer and buffer are resubmitted to the
- * endpoint for further use.
- */
-class libusb_managed_recv_buffer_impl : public managed_recv_buffer {
+class libusb_zero_copy_impl : public usb_zero_copy, public boost::enable_shared_from_this<libusb_zero_copy_impl> {
 public:
-    libusb_managed_recv_buffer_impl(libusb_transfer *lut,
-                                    usb_endpoint::sptr endpoint)
-        : _buff(lut->buffer, lut->length)
-    {
-        _lut = lut;
-        _endpoint = endpoint;
-    }
-
-    ~libusb_managed_recv_buffer_impl(void){
-        _endpoint->submit(_lut);
-    }
+    typedef boost::shared_ptr<libusb_zero_copy_impl> sptr;
 
-private:
-    const boost::asio::const_buffer &get(void) const{
-        return _buff;
-    }
+    libusb_zero_copy_impl(
+        libusb::device_handle::sptr handle,
+        unsigned int recv_endpoint, unsigned int send_endpoint,
+        size_t recv_xfer_size, size_t recv_num_xfers,
+        size_t send_xfer_size, size_t send_num_xfers
+    );
 
-    libusb_transfer *_lut;
-    usb_endpoint::sptr _endpoint;
-    const boost::asio::const_buffer _buff;
-};
+    managed_recv_buffer::sptr get_recv_buff(double);
+    managed_send_buffer::sptr get_send_buff(double);
 
-/*
- * Libusb managed send buffer
- * Construct a send buffer from a libusb transfer. The memory held by
- * the libusb transfer is exposed through the managed buffer interface.
- * Committing the buffer will set the data length and submit the buffer
- * to the endpoint. Submitting a buffer multiple times or destroying
- * the buffer before committing is an error. For the latter, the transfer
- * is returned to the endpoint with no data for reuse.
- */
-class libusb_managed_send_buffer_impl : public managed_send_buffer {
-public:
-    libusb_managed_send_buffer_impl(libusb_transfer *lut,
-                                    usb_endpoint::sptr endpoint)
-        : _buff(lut->buffer, lut->length), _committed(false)
-    {
-        _lut = lut;
-        _endpoint = endpoint;
-    }
+    size_t get_num_recv_frames(void) const { return _recv_num_frames; }
+    size_t get_num_send_frames(void) const { return _send_num_frames; }
 
-    ~libusb_managed_send_buffer_impl(void){
-        if (!_committed) {
-            _lut->length = 0;
-            _lut->actual_length = 0;
-            _endpoint->submit(_lut);
-        }
+private:
+    void release(libusb_transfer *lut){
+        _recv_ep->submit(lut);
     }
 
-    ssize_t commit(size_t num_bytes)
-    {
-        if (_committed) {
-            std::cerr << "UHD: send buffer already committed" << std::endl;
-            return 0;
-        }
-
-        UHD_ASSERT_THROW(num_bytes <= boost::asio::buffer_size(_buff));
-
-        _lut->length = num_bytes;
-        _lut->actual_length = 0;
-
+    void commit(libusb_transfer *lut, size_t num_bytes){
+        lut->length = num_bytes;
         try{
-            _endpoint->submit(_lut);
-            _committed = true;
-            return num_bytes;
+            _send_ep->submit(lut);
         }
         catch(const std::exception &e){
             std::cerr << "Error in commit: " << e.what() << std::endl;
-            return -1;
         }
     }
 
-private:
-    const boost::asio::mutable_buffer &get(void) const{
-        return _buff;
-    }
-
-    libusb_transfer *_lut;
-    usb_endpoint::sptr _endpoint;
-    const boost::asio::mutable_buffer _buff;
-    bool _committed;
-};
-
-
-/***********************************************************************
- * USB zero_copy device class
- **********************************************************************/
-class libusb_zero_copy_impl : public usb_zero_copy
-{
-private:
     libusb::device_handle::sptr _handle;
+    size_t _recv_xfer_size, _send_xfer_size;
     size_t _recv_num_frames, _send_num_frames;
     usb_endpoint::sptr _recv_ep, _send_ep;
-
-public:
-    typedef boost::shared_ptr<libusb_zero_copy_impl> sptr;
-
-    libusb_zero_copy_impl(
-        libusb::device_handle::sptr handle,
-        unsigned int recv_endpoint, unsigned int send_endpoint,
-        size_t recv_xfer_size, size_t recv_num_xfers,
-        size_t send_xfer_size, size_t send_num_xfers
-    );
-
-    managed_recv_buffer::sptr get_recv_buff(size_t timeout_ms);
-    managed_send_buffer::sptr get_send_buff(void);
-
-    size_t get_num_recv_frames(void) const { return _recv_num_frames; }
-    size_t get_num_send_frames(void) const { return _send_num_frames; }
 };
 
 /*
@@ -430,7 +340,9 @@ libusb_zero_copy_impl::libusb_zero_copy_impl(
     UHD_ASSERT_THROW(send_xfer_size % 512 == 0);
 
     //store the num xfers for the num frames count
+    _recv_xfer_size = recv_xfer_size;
     _recv_num_frames = recv_num_xfers;
+    _send_xfer_size = send_xfer_size;
     _send_num_frames = send_num_xfers;
 
     _handle->claim_interface(2 /*in interface*/);
@@ -459,15 +371,16 @@ libusb_zero_copy_impl::libusb_zero_copy_impl(
  * Return empty pointer if no transfer is available (timeout or error).
  * \return pointer to a managed receive buffer
  */
-managed_recv_buffer::sptr libusb_zero_copy_impl::get_recv_buff(size_t timeout_ms){
-    libusb_transfer *lut = _recv_ep->get_lut_with_wait(timeout_ms);
+managed_recv_buffer::sptr libusb_zero_copy_impl::get_recv_buff(double timeout){
+    libusb_transfer *lut = _recv_ep->get_lut_with_wait(timeout);
     if (lut == NULL) {
         return managed_recv_buffer::sptr();
     }
     else {
-        return managed_recv_buffer::sptr(
-            new libusb_managed_recv_buffer_impl(lut,
-                                                _recv_ep));
+        return managed_recv_buffer::make_safe(
+            boost::asio::const_buffer(lut->buffer, lut->actual_length),
+            boost::bind(&libusb_zero_copy_impl::release, shared_from_this(), lut)
+        );
     }
 }
 
@@ -478,15 +391,16 @@ managed_recv_buffer::sptr libusb_zero_copy_impl::get_recv_buff(size_t timeout_ms
  * (timeout or error).
  * \return pointer to a managed send buffer
  */
-managed_send_buffer::sptr libusb_zero_copy_impl::get_send_buff(void){
-    libusb_transfer *lut = _send_ep->get_lut_with_wait(/* TODO timeout API */);
+managed_send_buffer::sptr libusb_zero_copy_impl::get_send_buff(double timeout){
+    libusb_transfer *lut = _send_ep->get_lut_with_wait(timeout);
     if (lut == NULL) {
         return managed_send_buffer::sptr();
     }
     else {
-        return managed_send_buffer::sptr(
-            new libusb_managed_send_buffer_impl(lut,
-                                                _send_ep));
+        return managed_send_buffer::make_safe(
+            boost::asio::mutable_buffer(lut->buffer, _send_xfer_size),
+            boost::bind(&libusb_zero_copy_impl::commit, shared_from_this(), lut, _1)
+        );
     }
 }
 
@@ -494,18 +408,18 @@ managed_send_buffer::sptr libusb_zero_copy_impl::get_send_buff(void){
  * USB zero_copy make functions
  **********************************************************************/
 usb_zero_copy::sptr usb_zero_copy::make(
-	usb_device_handle::sptr handle,
+    usb_device_handle::sptr handle,
     unsigned int recv_endpoint, unsigned int send_endpoint,
-	size_t recv_xfer_size, size_t recv_num_xfers,
-	size_t send_xfer_size, size_t send_num_xfers
+    size_t recv_xfer_size, size_t recv_num_xfers,
+    size_t send_xfer_size, size_t send_num_xfers
 ){
     libusb::device_handle::sptr dev_handle(libusb::device_handle::get_cached_handle(
         boost::static_pointer_cast<libusb::special_handle>(handle)->get_device()
     ));
     return sptr(new libusb_zero_copy_impl(
-		dev_handle,
-		recv_endpoint,  send_endpoint,
-		recv_xfer_size, recv_num_xfers,
-		send_xfer_size, send_num_xfers
+        dev_handle,
+        recv_endpoint,  send_endpoint,
+        recv_xfer_size, recv_num_xfers,
+        send_xfer_size, send_num_xfers
     ));
 }
diff --git a/host/lib/transport/udp_zero_copy_asio.cpp b/host/lib/transport/udp_zero_copy_asio.cpp
index 0a6c9f2af..2cf7bde18 100644
--- a/host/lib/transport/udp_zero_copy_asio.cpp
+++ b/host/lib/transport/udp_zero_copy_asio.cpp
@@ -17,25 +17,37 @@
 
 #include <uhd/transport/udp_zero_copy.hpp>
 #include <uhd/transport/udp_simple.hpp> //mtu
+#include <uhd/transport/bounded_buffer.hpp>
+#include <uhd/utils/thread_priority.hpp>
 #include <uhd/utils/assert.hpp>
 #include <uhd/utils/warning.hpp>
-#include <boost/cstdint.hpp>
+#include <boost/shared_array.hpp>
 #include <boost/asio.hpp>
 #include <boost/format.hpp>
+#include <boost/thread.hpp>
 #include <iostream>
 
+using namespace uhd;
 using namespace uhd::transport;
+namespace asio = boost::asio;
 
 /***********************************************************************
  * Constants
  **********************************************************************/
 //enough buffering for half a second of samples at full rate on usrp2
-static const size_t MIN_RECV_SOCK_BUFF_SIZE = size_t(sizeof(boost::uint32_t) * 25e6 * 0.5);
+static const size_t MIN_RECV_SOCK_BUFF_SIZE = size_t(4 * 25e6 * 0.5);
+
 //Large buffers cause more underflow at high rates.
 //Perhaps this is due to the kernel scheduling,
 //but may change with host-based flow control.
 static const size_t MIN_SEND_SOCK_BUFF_SIZE = size_t(10e3);
 
+//the number of async frames to allocate for each send and recv
+static const size_t DEFAULT_NUM_ASYNC_FRAMES = 32;
+
+//a single concurrent thread for io_service seems to be the fastest
+static const size_t CONCURRENCY_HINT = 1;
+
 /***********************************************************************
  * Zero Copy UDP implementation with ASIO:
  *   This is the portable zero copy implementation for systems
@@ -43,36 +55,57 @@ static const size_t MIN_SEND_SOCK_BUFF_SIZE = size_t(10e3);
  *   However, it is not a true zero copy implementation as each
  *   send and recv requires a copy operation to/from userspace.
  **********************************************************************/
-class udp_zero_copy_impl:
-    public phony_zero_copy_recv_if,
-    public phony_zero_copy_send_if,
-    public udp_zero_copy
-{
+class udp_zero_copy_asio_impl : public udp_zero_copy, public boost::enable_shared_from_this<udp_zero_copy_asio_impl> {
 public:
-    typedef boost::shared_ptr<udp_zero_copy_impl> sptr;
+    typedef boost::shared_ptr<udp_zero_copy_asio_impl> sptr;
 
-    udp_zero_copy_impl(
-        const std::string &addr,
-        const std::string &port
+    udp_zero_copy_asio_impl(
+        const std::string &addr, const std::string &port,
+        size_t recv_frame_size, size_t num_recv_frames,
+        size_t send_frame_size, size_t num_send_frames
     ):
-        phony_zero_copy_recv_if(udp_simple::mtu),
-        phony_zero_copy_send_if(udp_simple::mtu)
+        _io_service(CONCURRENCY_HINT),
+        _work(new asio::io_service::work(_io_service)),
+        _recv_frame_size(recv_frame_size), _num_recv_frames(num_recv_frames),
+        _send_frame_size(send_frame_size), _num_send_frames(num_send_frames)
     {
         //std::cout << boost::format("Creating udp transport for %s %s") % addr % port << std::endl;
 
-        // resolve the address
-        boost::asio::ip::udp::resolver resolver(_io_service);
-        boost::asio::ip::udp::resolver::query query(boost::asio::ip::udp::v4(), addr, port);
-        boost::asio::ip::udp::endpoint receiver_endpoint = *resolver.resolve(query);
+        //resolve the address
+        asio::ip::udp::resolver resolver(_io_service);
+        asio::ip::udp::resolver::query query(asio::ip::udp::v4(), addr, port);
+        asio::ip::udp::endpoint receiver_endpoint = *resolver.resolve(query);
 
-        // create, open, and connect the socket
-        _socket = new boost::asio::ip::udp::socket(_io_service);
-        _socket->open(boost::asio::ip::udp::v4());
+        //create, open, and connect the socket
+        _socket = new asio::ip::udp::socket(_io_service);
+        _socket->open(asio::ip::udp::v4());
         _socket->connect(receiver_endpoint);
-        _sock_fd = _socket->native();
     }
 
-    ~udp_zero_copy_impl(void){
+    void init(void){
+        //allocate all recv frames and release them to begin xfers
+        _pending_recv_buffs = pending_buffs_type::make(_num_recv_frames);
+        _recv_buffer = boost::shared_array<char>(new char[_num_recv_frames*_recv_frame_size]);
+        for (size_t i = 0; i < _num_recv_frames; i++){
+            release(_recv_buffer.get() + i*_recv_frame_size);
+        }
+
+        //allocate all send frames and push them into the fifo
+        _pending_send_buffs = pending_buffs_type::make(_num_send_frames);
+        _send_buffer = boost::shared_array<char>(new char[_num_send_frames*_send_frame_size]);
+        for (size_t i = 0; i < _num_send_frames; i++){
+            handle_send(_send_buffer.get() + i*_send_frame_size);
+        }
+
+        //spawn the service threads that will run the io service
+        for (size_t i = 0; i < CONCURRENCY_HINT; i++) _thread_group.create_thread(
+            boost::bind(&udp_zero_copy_asio_impl::service, this)
+        );
+    }
+
+    ~udp_zero_copy_asio_impl(void){
+        delete _work; //allow io_service run to complete
+        _thread_group.join_all(); //wait for service threads to exit
         delete _socket;
     }
 
@@ -90,60 +123,106 @@ public:
         return get_buff_size<Opt>();
     }
 
+    //! pop a filled recv buffer off of the fifo and bind with the release callback
+    managed_recv_buffer::sptr get_recv_buff(double timeout){
+        boost::this_thread::disable_interruption di; //disable because the wait can throw
+        asio::mutable_buffer buff;
+        if (_pending_recv_buffs->pop_with_timed_wait(buff, timeout)){
+            return managed_recv_buffer::make_safe(
+                buff, boost::bind(
+                    &udp_zero_copy_asio_impl::release,
+                    shared_from_this(),
+                    asio::buffer_cast<void*>(buff)
+                )
+            );
+        }
+        return managed_recv_buffer::sptr();
+    }
 
-    //The number of frames is approximately the buffer size divided by the max datagram size.
-    //In reality, this is a phony zero-copy interface and the number of frames is infinite.
-    //However, its sensible to advertise a frame count that is approximate to buffer size.
-    //This way, the transport caller will have an idea about how much buffering to create.
+    size_t get_num_recv_frames(void) const {return _num_recv_frames;}
 
-    size_t get_num_recv_frames(void) const{
-        return this->get_buff_size<boost::asio::socket_base::receive_buffer_size>()/udp_simple::mtu;
+    //! pop an empty send buffer off of the fifo and bind with the commit callback
+    managed_send_buffer::sptr get_send_buff(double timeout){
+        boost::this_thread::disable_interruption di; //disable because the wait can throw
+        asio::mutable_buffer buff;
+        if (_pending_send_buffs->pop_with_timed_wait(buff, timeout)){
+            return managed_send_buffer::make_safe(
+                buff, boost::bind(
+                    &udp_zero_copy_asio_impl::commit,
+                    shared_from_this(),
+                    asio::buffer_cast<void*>(buff), _1
+                )
+            );
+        }
+        return managed_send_buffer::sptr();
     }
 
-    size_t get_num_send_frames(void) const{
-        return this->get_buff_size<boost::asio::socket_base::send_buffer_size>()/udp_simple::mtu;
-    }
+    size_t get_num_send_frames(void) const {return _num_send_frames;}
 
 private:
-    boost::asio::ip::udp::socket   *_socket;
-    boost::asio::io_service        _io_service;
-    int                            _sock_fd;
-
-    ssize_t recv(const boost::asio::mutable_buffer &buff, size_t timeout_ms){
-        //setup timeval for timeout
-        timeval tv;
-        tv.tv_sec = 0;
-        tv.tv_usec = timeout_ms*1000;
-
-        //setup rset for timeout
-        fd_set rset;
-        FD_ZERO(&rset);
-        FD_SET(_sock_fd, &rset);
-
-        //call select to perform timed wait
-        if (::select(_sock_fd+1, &rset, NULL, NULL, &tv) <= 0) return 0;
-
-        return ::recv(
-            _sock_fd,
-            boost::asio::buffer_cast<char *>(buff),
-            boost::asio::buffer_size(buff), 0
+    void service(void){
+        set_thread_priority_safe();
+        _io_service.run();
+    }
+
+    /*******************************************************************
+     * The async send and receive callbacks
+     ******************************************************************/
+
+    //! handle a recv callback -> push the filled memory into the fifo
+    void handle_recv(void *mem, size_t len){
+        boost::this_thread::disable_interruption di; //disable because the wait can throw
+        _pending_recv_buffs->push_with_wait(boost::asio::buffer(mem, len));
+    }
+
+    //! release a recv buffer -> start an async recv on the buffer
+    void release(void *mem){
+        _socket->async_receive(
+            boost::asio::buffer(mem, _recv_frame_size),
+            boost::bind(
+                &udp_zero_copy_asio_impl::handle_recv,
+                shared_from_this(), mem,
+                asio::placeholders::bytes_transferred
+            )
         );
     }
 
-    ssize_t send(const boost::asio::const_buffer &buff){
-        return ::send(
-            _sock_fd,
-            boost::asio::buffer_cast<const char *>(buff),
-            boost::asio::buffer_size(buff), 0
+    //! handle a send callback -> push the emptied memory into the fifo
+    void handle_send(void *mem){
+        boost::this_thread::disable_interruption di; //disable because the wait can throw
+        _pending_send_buffs->push_with_wait(boost::asio::buffer(mem, _send_frame_size));
+    }
+
+    //! commit a send buffer -> start an async send on the buffer
+    void commit(void *mem, size_t len){
+        _socket->async_send(
+            boost::asio::buffer(mem, len),
+            boost::bind(
+                &udp_zero_copy_asio_impl::handle_send,
+                shared_from_this(), mem
+            )
         );
     }
+
+    //asio guts -> socket and service
+    asio::ip::udp::socket   *_socket;
+    asio::io_service        _io_service;
+    asio::io_service::work  *_work;
+
+    //memory management -> buffers and fifos
+    boost::thread_group _thread_group;
+    boost::shared_array<char> _send_buffer, _recv_buffer;
+    typedef bounded_buffer<asio::mutable_buffer> pending_buffs_type;
+    pending_buffs_type::sptr _pending_recv_buffs, _pending_send_buffs;
+    const size_t _recv_frame_size, _num_recv_frames;
+    const size_t _send_frame_size, _num_send_frames;
 };
 
 /***********************************************************************
  * UDP zero copy make function
  **********************************************************************/
 template<typename Opt> static void resize_buff_helper(
-    udp_zero_copy_impl::sptr udp_trans,
+    udp_zero_copy_asio_impl::sptr udp_trans,
     size_t target_size,
     const std::string &name
 ){
@@ -183,11 +262,17 @@ udp_zero_copy::sptr udp_zero_copy::make(
     size_t recv_buff_size,
     size_t send_buff_size
 ){
-    udp_zero_copy_impl::sptr udp_trans(new udp_zero_copy_impl(addr, port));
+    udp_zero_copy_asio_impl::sptr udp_trans(new udp_zero_copy_asio_impl(
+        addr, port,
+        udp_simple::mtu, DEFAULT_NUM_ASYNC_FRAMES, //recv
+        udp_simple::mtu, DEFAULT_NUM_ASYNC_FRAMES  //send
+    ));
 
     //call the helper to resize send and recv buffers
-    resize_buff_helper<boost::asio::socket_base::receive_buffer_size>(udp_trans, recv_buff_size, "recv");
-    resize_buff_helper<boost::asio::socket_base::send_buffer_size>   (udp_trans, send_buff_size, "send");
+    resize_buff_helper<asio::socket_base::receive_buffer_size>(udp_trans, recv_buff_size, "recv");
+    resize_buff_helper<asio::socket_base::send_buffer_size>   (udp_trans, send_buff_size, "send");
+
+    udp_trans->init(); //buffers resized -> call init() to use
 
     return udp_trans;
 }
diff --git a/host/lib/transport/vrt_packet_handler.hpp b/host/lib/transport/vrt_packet_handler.hpp
index b603f1371..939517411 100644
--- a/host/lib/transport/vrt_packet_handler.hpp
+++ b/host/lib/transport/vrt_packet_handler.hpp
@@ -303,18 +303,18 @@ template <typename T> UHD_INLINE T get_context_code(
      * Pack a vrt header, copy-convert the data, and send it.
      *  - helper function for vrt_packet_handler::send
      ******************************************************************/
-    static UHD_INLINE void _send1(
+    static UHD_INLINE size_t _send1(
         send_state &state,
         const std::vector<const void *> &buffs,
-        size_t offset_bytes,
-        size_t num_samps,
+        const size_t offset_bytes,
+        const size_t num_samps,
         uhd::transport::vrt::if_packet_info_t &if_packet_info,
         const uhd::io_type_t &io_type,
         const uhd::otw_type_t &otw_type,
         const vrt_packer_t &vrt_packer,
         const get_send_buffs_t &get_send_buffs,
-        size_t vrt_header_offset_words32,
-        size_t chans_per_otw_buff
+        const size_t vrt_header_offset_words32,
+        const size_t chans_per_otw_buff
     ){
         //load the rest of the if_packet_info in here
         if_packet_info.num_payload_words32 = (num_samps*chans_per_otw_buff*otw_type.get_sample_size())/sizeof(boost::uint32_t);
@@ -322,7 +322,7 @@ template <typename T> UHD_INLINE T get_context_code(
 
         //get send buffers for each channel
         managed_send_buffs_t send_buffs(buffs.size()/chans_per_otw_buff);
-        UHD_ASSERT_THROW(get_send_buffs(send_buffs));
+        if (not get_send_buffs(send_buffs)) return 0;
 
         std::vector<const void *> io_buffs(chans_per_otw_buff);
         for (size_t i = 0; i < buffs.size(); i+=chans_per_otw_buff){
@@ -343,10 +343,9 @@ template <typename T> UHD_INLINE T get_context_code(
 
             //commit the samples to the zero-copy interface
             size_t num_bytes_total = (vrt_header_offset_words32+if_packet_info.num_packet_words32)*sizeof(boost::uint32_t);
-            if (send_buffs[i]->commit(num_bytes_total) < ssize_t(num_bytes_total)){
-                std::cerr << "commit to send buffer returned less than commit size" << std::endl;
-            }
+            send_buffs[i]->commit(num_bytes_total);
         }
+        return num_samps;
     }
 
     /*******************************************************************
@@ -381,7 +380,6 @@ template <typename T> UHD_INLINE T get_context_code(
         ////////////////////////////////////////////////////////////////
         case uhd::device::SEND_MODE_ONE_PACKET:{
         ////////////////////////////////////////////////////////////////
-            size_t num_samps = std::min(total_num_samps, max_samples_per_packet);
 
             //fill in parts of the packet info overwrote in full buff mode
             if_packet_info.has_tsi = metadata.has_time_spec;
@@ -389,10 +387,10 @@ template <typename T> UHD_INLINE T get_context_code(
             if_packet_info.sob = metadata.start_of_burst;
             if_packet_info.eob = metadata.end_of_burst;
 
-            _send1(
+            return _send1(
                 state,
                 buffs, 0,
-                num_samps,
+                std::min(total_num_samps, max_samples_per_packet),
                 if_packet_info,
                 io_type, otw_type,
                 vrt_packer,
@@ -400,31 +398,32 @@ template <typename T> UHD_INLINE T get_context_code(
                 vrt_header_offset_words32,
                 chans_per_otw_buff
             );
-            return num_samps;
         }
 
         ////////////////////////////////////////////////////////////////
         case uhd::device::SEND_MODE_FULL_BUFF:{
         ////////////////////////////////////////////////////////////////
-            //calculate constants for fragmentation
-            const size_t num_fragments = (total_num_samps+max_samples_per_packet-1)/max_samples_per_packet;
-            static const size_t first_fragment_index = 0;
-            const size_t final_fragment_index = num_fragments-1;
+            size_t total_num_samps_sent = 0;
 
             //loop through the following fragment indexes
-            for (size_t n = first_fragment_index; n <= final_fragment_index; n++){
+            while(total_num_samps_sent < total_num_samps){
+
+                //calculate per-loop-iteration variables
+                const size_t total_num_samps_unsent = total_num_samps - total_num_samps_sent;
+                const bool first_fragment = (total_num_samps_sent == 0);
+                const bool final_fragment = (total_num_samps_unsent <= max_samples_per_packet);
 
                 //calculate new flags for the fragments
-                if_packet_info.has_tsi = metadata.has_time_spec  and (n == first_fragment_index);
-                if_packet_info.has_tsf = metadata.has_time_spec  and (n == first_fragment_index);
-                if_packet_info.sob     = metadata.start_of_burst and (n == first_fragment_index);
-                if_packet_info.eob     = metadata.end_of_burst   and (n == final_fragment_index);
+                if_packet_info.has_tsi = metadata.has_time_spec  and first_fragment;
+                if_packet_info.has_tsf = if_packet_info.has_tsi;
+                if_packet_info.sob     = metadata.start_of_burst and first_fragment;
+                if_packet_info.eob     = metadata.end_of_burst   and final_fragment;
 
                 //send the fragment with the helper function
-                _send1(
+                const size_t num_samps_sent = _send1(
                     state,
-                    buffs, n*max_samples_per_packet*io_type.size,
-                    (n == final_fragment_index)?(total_num_samps%max_samples_per_packet):max_samples_per_packet,
+                    buffs, total_num_samps_sent*io_type.size,
+                    std::min(total_num_samps_unsent, max_samples_per_packet),
                     if_packet_info,
                     io_type, otw_type,
                     vrt_packer,
@@ -432,8 +431,10 @@ template <typename T> UHD_INLINE T get_context_code(
                     vrt_header_offset_words32,
                     chans_per_otw_buff
                 );
+                total_num_samps_sent += num_samps_sent;
+                if (num_samps_sent == 0) return total_num_samps_sent;
             }
-            return total_num_samps;
+            return total_num_samps_sent;
         }
 
         default: throw std::runtime_error("unknown send mode");
diff --git a/host/lib/transport/zero_copy.cpp b/host/lib/transport/zero_copy.cpp
index 1fcf846a0..a5a864a04 100644
--- a/host/lib/transport/zero_copy.cpp
+++ b/host/lib/transport/zero_copy.cpp
@@ -16,32 +16,35 @@
 //
 
 #include <uhd/transport/zero_copy.hpp>
-#include <boost/cstdint.hpp>
-#include <boost/function.hpp>
-#include <boost/bind.hpp>
 
 using namespace uhd::transport;
 
 /***********************************************************************
- * The pure-virtual deconstructor needs an implementation to be happy
+ * Safe managed receive buffer
  **********************************************************************/
-managed_recv_buffer::~managed_recv_buffer(void){
+static void release_nop(void){
     /* NOP */
 }
 
-/***********************************************************************
- * Phony zero-copy recv interface implementation
- **********************************************************************/
-
-//! phony zero-copy recv buffer implementation
-class managed_recv_buffer_impl : public managed_recv_buffer{
+class safe_managed_receive_buffer : public managed_recv_buffer{
 public:
-    managed_recv_buffer_impl(const boost::asio::const_buffer &buff) : _buff(buff){
+    safe_managed_receive_buffer(
+        const boost::asio::const_buffer &buff,
+        const release_fcn_t &release_fcn
+    ):
+        _buff(buff), _release_fcn(release_fcn)
+    {
         /* NOP */
     }
 
-    ~managed_recv_buffer_impl(void){
-        delete [] this->cast<const boost::uint8_t *>();
+    ~safe_managed_receive_buffer(void){
+        _release_fcn();
+    }
+
+    void release(void){
+        release_fcn_t release_fcn = _release_fcn;
+        _release_fcn = &release_nop;
+        return release_fcn();
     }
 
 private:
@@ -50,64 +53,42 @@ private:
     }
 
     const boost::asio::const_buffer _buff;
+    release_fcn_t _release_fcn;
 };
 
-//! phony zero-copy recv interface implementation
-struct phony_zero_copy_recv_if::impl{
-    impl(size_t max_buff_size) : max_buff_size(max_buff_size){
-        /* NOP */
-    }
-    size_t max_buff_size;
-};
-
-phony_zero_copy_recv_if::phony_zero_copy_recv_if(size_t max_buff_size){
-    _impl = UHD_PIMPL_MAKE(impl, (max_buff_size));
-}
-
-phony_zero_copy_recv_if::~phony_zero_copy_recv_if(void){
-    /* NOP */
-}
-
-managed_recv_buffer::sptr phony_zero_copy_recv_if::get_recv_buff(size_t timeout_ms){
-    //allocate memory
-    boost::uint8_t *recv_mem = new boost::uint8_t[_impl->max_buff_size];
-
-    //call recv() with timeout option
-    ssize_t num_bytes = this->recv(boost::asio::buffer(recv_mem, _impl->max_buff_size), timeout_ms);
-
-    if (num_bytes <= 0) return managed_recv_buffer::sptr(); //NULL sptr
-
-    //create a new managed buffer to house the data
-    return managed_recv_buffer::sptr(
-        new managed_recv_buffer_impl(boost::asio::buffer(recv_mem, num_bytes))
-    );
+managed_recv_buffer::sptr managed_recv_buffer::make_safe(
+    const boost::asio::const_buffer &buff,
+    const release_fcn_t &release_fcn
+){
+    return sptr(new safe_managed_receive_buffer(buff, release_fcn));
 }
 
 /***********************************************************************
- * Phony zero-copy send interface implementation
+ * Safe managed send buffer
  **********************************************************************/
+static void commit_nop(size_t){
+    /* NOP */
+}
 
-//! phony zero-copy send buffer implementation
-class managed_send_buffer_impl : public managed_send_buffer{
+class safe_managed_send_buffer : public managed_send_buffer{
 public:
-    typedef boost::function<ssize_t(const boost::asio::const_buffer &)> send_fcn_t;
-
-    managed_send_buffer_impl(
+    safe_managed_send_buffer(
         const boost::asio::mutable_buffer &buff,
-        const send_fcn_t &send_fcn
+        const commit_fcn_t &commit_fcn
     ):
-        _buff(buff),
-        _send_fcn(send_fcn)
+        _buff(buff), _commit_fcn(commit_fcn)
     {
         /* NOP */
     }
 
-    ~managed_send_buffer_impl(void){
-        /* NOP */
+    ~safe_managed_send_buffer(void){
+        _commit_fcn(0);
     }
 
-    ssize_t commit(size_t num_bytes){
-        return _send_fcn(boost::asio::buffer(_buff, num_bytes));
+    void commit(size_t num_bytes){
+        commit_fcn_t commit_fcn = _commit_fcn;
+        _commit_fcn = &commit_nop;
+        return commit_fcn(num_bytes);
     }
 
 private:
@@ -116,28 +97,12 @@ private:
     }
 
     const boost::asio::mutable_buffer _buff;
-    const send_fcn_t                  _send_fcn;
-};
-
-//! phony zero-copy send interface implementation
-struct phony_zero_copy_send_if::impl{
-    boost::uint8_t *send_mem;
-    managed_send_buffer::sptr send_buff;
+    commit_fcn_t _commit_fcn;
 };
 
-phony_zero_copy_send_if::phony_zero_copy_send_if(size_t max_buff_size){
-    _impl = UHD_PIMPL_MAKE(impl, ());
-    _impl->send_mem = new boost::uint8_t[max_buff_size];
-    _impl->send_buff = managed_send_buffer::sptr(new managed_send_buffer_impl(
-        boost::asio::buffer(_impl->send_mem, max_buff_size),
-        boost::bind(&phony_zero_copy_send_if::send, this, _1)
-    ));
-}
-
-phony_zero_copy_send_if::~phony_zero_copy_send_if(void){
-    delete [] _impl->send_mem;
-}
-
-managed_send_buffer::sptr phony_zero_copy_send_if::get_send_buff(void){
-    return _impl->send_buff; //FIXME there is only ever one send buff, we assume that the caller doesnt hang onto these
+safe_managed_send_buffer::sptr managed_send_buffer::make_safe(
+    const boost::asio::mutable_buffer &buff,
+    const commit_fcn_t &commit_fcn
+){
+    return sptr(new safe_managed_send_buffer(buff, commit_fcn));
 }