Merge branch 'maint'

4 files changed, 61 insertions, 38 deletions

diff --git a/host/lib/usrp/device3/device3_io_impl.cpp b/host/lib/usrp/device3/device3_io_impl.cpp
index 8a42fe148..46cc0ee5a 100644
--- a/host/lib/usrp/device3/device3_io_impl.cpp
+++ b/host/lib/usrp/device3/device3_io_impl.cpp
@@ -31,6 +31,8 @@
 #include "../../rfnoc/tx_stream_terminator.hpp"
 #include <uhd/rfnoc/rate_node_ctrl.hpp>
 #include <uhd/rfnoc/radio_ctrl.hpp>
+#include <uhd/transport/zero_copy_flow_ctrl.hpp>
+#include <boost/atomic.hpp>
 
 #define UHD_TX_STREAMER_LOG() UHD_LOGGER_TRACE("STREAMER")
 #define UHD_RX_STREAMER_LOG() UHD_LOGGER_TRACE("STREAMER")
@@ -305,12 +307,13 @@ struct tx_fc_cache_t
         device_channel(0),
         last_seq_out(0),
         last_seq_ack(0),
-        seq_queue(1){}
+        last_seq_ack_cache(0) {}
+
     size_t stream_channel;
     size_t device_channel;
     size_t last_seq_out;
-    size_t last_seq_ack;
-    uhd::transport::bounded_buffer<size_t> seq_queue;
+    boost::atomic_size_t last_seq_ack;
+    size_t last_seq_ack_cache;
     boost::shared_ptr<device3_impl::async_md_type> async_queue;
     boost::shared_ptr<device3_impl::async_md_type> old_async_queue;
 };
@@ -338,33 +341,43 @@ static size_t get_tx_flow_control_window(
     return window_in_pkts;
 }
 
-static managed_send_buffer::sptr get_tx_buff_with_flowctrl(
-    task::sptr /*holds ref*/,
-    boost::shared_ptr<tx_fc_cache_t> fc_cache,
+// TODO: Remove this function
+// This function only exists to make sure the transport is not destroyed
+// until it is no longer needed.
+static managed_send_buffer::sptr get_tx_buff(
     zero_copy_if::sptr xport,
-    size_t fc_window,
     const double timeout
 ){
+    return xport->get_send_buff(timeout);
+}
+
+static bool tx_flow_ctrl(
+    task::sptr /*holds ref*/,
+    boost::shared_ptr<tx_fc_cache_t> fc_cache,
+    size_t fc_window,
+    managed_buffer::sptr
+) {
+    // Busy loop waiting for flow control update.  This is necessary because
+    // at this point there is data trying to be sent and it must be sent as
+    // quickly as possible when the flow control update arrives to avoid
+    // underruns at high rates.  This is also OK because it only occurs when
+    // data needs to be sent and flow control is holding it back.
     while (true)
     {
         // delta is the amount of FC credit we've used up
-        const size_t delta = (fc_cache->last_seq_out & HW_SEQ_NUM_MASK) - (fc_cache->last_seq_ack & HW_SEQ_NUM_MASK);
+        const size_t delta = (fc_cache->last_seq_out & HW_SEQ_NUM_MASK) -
+            (fc_cache->last_seq_ack_cache & HW_SEQ_NUM_MASK);
         // If we want to send another packet, we must have FC credit left
         if ((delta & HW_SEQ_NUM_MASK) < fc_window)
-            break;
-
-        // If credit is all used up, we check seq_queue for more.
-        const bool ok = fc_cache->seq_queue.pop_with_timed_wait(fc_cache->last_seq_ack, timeout);
-        if (not ok) {
-            return managed_send_buffer::sptr(); //timeout waiting for flow control
+        {
+            // Packet will be sent
+            fc_cache->last_seq_out++; //update seq
+            return true;
         }
+        // update the cached value from the atomic
+        fc_cache->last_seq_ack_cache = fc_cache->last_seq_ack;
     }
-
-    managed_send_buffer::sptr buff = xport->get_send_buff(timeout);
-    if (buff) {
-        fc_cache->last_seq_out++; //update seq, this will actually be a send
-    }
-    return buff;
+    return false;
 }
 
 #define DEVICE3_ASYNC_EVENT_CODE_FLOW_CTRL 0
@@ -381,7 +394,9 @@ static void handle_tx_async_msgs(
 ) {
     managed_recv_buffer::sptr buff = xport->get_recv_buff();
     if (not buff)
+    {
         return;
+    }
 
     //extract packet info
     vrt::if_packet_info_t if_packet_info;
@@ -430,8 +445,7 @@ static void handle_tx_async_msgs(
     //The FC response and the burst ack are two indicators that the radio
     //consumed packets. Use them to update the FC metadata
     if (metadata.event_code == DEVICE3_ASYNC_EVENT_CODE_FLOW_CTRL) {
-        const size_t seq = metadata.user_payload[0];
-        fc_cache->seq_queue.push_with_pop_on_full(seq);
+        fc_cache->last_seq_ack = metadata.user_payload[0];
     }
 
     //FC responses don't propagate up to the user so filter them here
@@ -842,13 +856,19 @@ tx_streamer::sptr device3_impl::get_tx_stream(const uhd::stream_args_t &args_)
             }
         }
 
+        // Add flow control
+        xport.send = zero_copy_flow_ctrl::make(
+            xport.send,
+            boost::bind(&tx_flow_ctrl, task, fc_cache, fc_window, _1),
+            NULL);
+
         //Give the streamer a functor to get the send buffer
-        //get_tx_buff_with_flowctrl is static so bind has no lifetime issues
+        //get_tx_buff is static so bind has no lifetime issues
         //xport.send (sptr) is required to add streamer->data-transport lifetime dependency
         //task (sptr) is required to add  a streamer->async-handler lifetime dependency
         my_streamer->set_xport_chan_get_buff(
             stream_i,
-            boost::bind(&get_tx_buff_with_flowctrl, task, fc_cache, xport.send, fc_window, _1)
+            boost::bind(&get_tx_buff, xport.send, _1)
         );
         //Give the streamer a functor handled received async messages
         my_streamer->set_async_receiver(
diff --git a/host/lib/usrp/x300/x300_fw_ctrl.cpp b/host/lib/usrp/x300/x300_fw_ctrl.cpp
index 1df0fa611..5ff40c966 100644
--- a/host/lib/usrp/x300/x300_fw_ctrl.cpp
+++ b/host/lib/usrp/x300/x300_fw_ctrl.cpp
@@ -292,7 +292,7 @@ protected:
 
 private:
     niriok_proxy::sptr _drv_proxy;
-    static const uint32_t READ_TIMEOUT_IN_MS = 10;
+    static const uint32_t READ_TIMEOUT_IN_MS = 100;
     static const uint32_t INIT_TIMEOUT_IN_MS = 5000;
 };
 
diff --git a/host/lib/usrp/x300/x300_impl.cpp b/host/lib/usrp/x300/x300_impl.cpp
index 785f7b4a3..ac08cf565 100644
--- a/host/lib/usrp/x300/x300_impl.cpp
+++ b/host/lib/usrp/x300/x300_impl.cpp
@@ -1128,14 +1128,14 @@ uhd::both_xports_t x300_impl::make_transport(
                 ? X300_PCIE_RX_DATA_FRAME_SIZE
                 : X300_PCIE_MSG_FRAME_SIZE;
 
-            default_buff_args.num_send_frames =
-                (xport_type == TX_DATA)
-                ? X300_PCIE_DATA_NUM_FRAMES
+			default_buff_args.num_send_frames =
+				(xport_type == TX_DATA)
+                ? X300_PCIE_TX_DATA_NUM_FRAMES
                 : X300_PCIE_MSG_NUM_FRAMES;
 
             default_buff_args.num_recv_frames =
                 (xport_type == RX_DATA)
-                ? X300_PCIE_DATA_NUM_FRAMES
+                ? X300_PCIE_RX_DATA_NUM_FRAMES
                 : X300_PCIE_MSG_NUM_FRAMES;
 
             xports.recv = nirio_zero_copy::make(
diff --git a/host/lib/usrp/x300/x300_impl.hpp b/host/lib/usrp/x300/x300_impl.hpp
index 2de295bd9..27f3f130e 100644
--- a/host/lib/usrp/x300/x300_impl.hpp
+++ b/host/lib/usrp/x300/x300_impl.hpp
@@ -52,15 +52,18 @@ static const size_t X300_RX_SW_BUFF_SIZE_ETH        = 0x2000000;//32MiB    For a
 static const size_t X300_RX_SW_BUFF_SIZE_ETH_MACOS  = 0x100000; //1Mib
 
 //The FIFO closest to the DMA controller is 1023 elements deep for RX and 1029 elements deep for TX
-//where an element is 8 bytes. For best throughput ensure that the data frame fits in these buffers.
-//Also ensure that the kernel has enough frames to hold buffered TX and RX data
-static const size_t X300_PCIE_RX_DATA_FRAME_SIZE    = 8184;     //bytes
-static const size_t X300_PCIE_TX_DATA_FRAME_SIZE    = 8184;     //bytes
-static const size_t X300_PCIE_DATA_NUM_FRAMES       = 2048;
-static const size_t X300_PCIE_MSG_FRAME_SIZE        = 256;      //bytes
-static const size_t X300_PCIE_MSG_NUM_FRAMES        = 64;
-static const size_t X300_PCIE_MAX_CHANNELS          = 6;
-static const size_t X300_PCIE_MAX_MUXED_XPORTS      = 32;
+//where an element is 8 bytes. The buffers (number of frames * frame size) must be aligned to the
+//memory page size.  For the control, we are getting lucky because 64 frames * 256 bytes each aligns
+//with the typical page size of 4096 bytes.  Since most page sizes are 4096 bytes or some multiple of
+//that, keep the number of frames * frame size aligned to it.
+static const size_t X300_PCIE_RX_DATA_FRAME_SIZE        = 4096;     //bytes
+static const size_t X300_PCIE_RX_DATA_NUM_FRAMES        = 4096;
+static const size_t X300_PCIE_TX_DATA_FRAME_SIZE        = 4096;     //bytes
+static const size_t X300_PCIE_TX_DATA_NUM_FRAMES	    = 4096;
+static const size_t X300_PCIE_MSG_FRAME_SIZE            = 256;      //bytes
+static const size_t X300_PCIE_MSG_NUM_FRAMES            = 64;
+static const size_t X300_PCIE_MAX_CHANNELS              = 6;
+static const size_t X300_PCIE_MAX_MUXED_XPORTS          = 32;
 
 static const size_t X300_10GE_DATA_FRAME_MAX_SIZE   = 8000;     // CHDR packet size in bytes
 static const size_t X300_1GE_DATA_FRAME_MAX_SIZE    = 1472;     // CHDR packet size in bytes