Device3: Change packet-based flow control to byte-based flow control

5 files changed, 47 insertions, 41 deletions

diff --git a/host/lib/rfnoc/ctrl_iface.cpp b/host/lib/rfnoc/ctrl_iface.cpp
index 11dfa7aaa..29e18fc3a 100644
--- a/host/lib/rfnoc/ctrl_iface.cpp
+++ b/host/lib/rfnoc/ctrl_iface.cpp
@@ -108,6 +108,7 @@ private:
         packet_info.tsf = timestamp;
         packet_info.sob = false;
         packet_info.eob = false;
+        packet_info.fc_ack = false;
         packet_info.sid = _xports.send_sid;
         packet_info.has_sid = true;
         packet_info.has_cid = false;
diff --git a/host/lib/rfnoc/graph_impl.cpp b/host/lib/rfnoc/graph_impl.cpp
index c361ea8f2..a2e0e64f4 100644
--- a/host/lib/rfnoc/graph_impl.cpp
+++ b/host/lib/rfnoc/graph_impl.cpp
@@ -122,9 +122,9 @@ void graph_impl::connect(
         UHD_LOGGER_WARNING("RFNOC") << "Assuming max packet size for " << src->get_block_id() ;
         pkt_size = uhd::rfnoc::MAX_PACKET_SIZE;
     }
-    // FC window (in packets) depends on FIFO size...          ...and packet size.
-    size_t buf_size_pkts = dst->get_fifo_size(dst_block_port) / pkt_size;
-    if (buf_size_pkts == 0) {
+    // FC window (in bytes) depends on FIFO size.
+    size_t buf_size_bytes = dst->get_fifo_size(dst_block_port);
+    if (buf_size_bytes < pkt_size) {
         throw uhd::runtime_error(str(
             boost::format("Input FIFO for block %s is too small (%d kiB) for packets of size %d kiB\n"
                           "coming from block %s.")
@@ -132,19 +132,20 @@ void graph_impl::connect(
             % (pkt_size / 1024) % src->get_block_id().get()
         ));
     }
-    src->configure_flow_control_out(buf_size_pkts, src_block_port);
-    // On the same crossbar, use lots of FC packets
-    size_t pkts_per_ack = std::min(
-            uhd::rfnoc::DEFAULT_FC_XBAR_PKTS_PER_ACK,
-            buf_size_pkts - 1
+    src->configure_flow_control_out(
+            true, /* enable output */
+            buf_size_bytes,
+            0, /* no packet limit. We need to revisit this at some point. */
+            src_block_port
     );
+    // On the same crossbar, use lots of FC packets
+    size_t bytes_per_response = std::ceil<size_t>(buf_size_bytes / uhd::rfnoc::DEFAULT_FC_XBAR_RESPONSE_FREQ);
     // Over the network, use less or we'd flood the transport
     if (sid.get_src_addr() != sid.get_dst_addr()) {
-        pkts_per_ack = std::max<size_t>(buf_size_pkts / uhd::rfnoc::DEFAULT_FC_TX_RESPONSE_FREQ, 1);
+        bytes_per_response = std::ceil<size_t>(buf_size_bytes / uhd::rfnoc::DEFAULT_FC_TX_RESPONSE_FREQ);
     }
     dst->configure_flow_control_in(
-            0, // Default to not use cycles
-            pkts_per_ack,
+            bytes_per_response,
             dst_block_port
     );
 
@@ -209,7 +210,7 @@ void graph_impl::connect_src(
 void graph_impl::connect_sink(
         const block_id_t &sink_block,
         const size_t dst_block_port,
-        const size_t pkts_per_ack
+        const size_t bytes_per_ack
 ) {
     device3::sptr device_ptr = _device_ptr.lock();
     if (not device_ptr) {
@@ -222,11 +223,7 @@ void graph_impl::connect_sink(
 
     uhd::rfnoc::sink_block_ctrl_base::sptr dst =
         device_ptr->get_block_ctrl<rfnoc::sink_block_ctrl_base>(sink_block);
-    dst->configure_flow_control_in(
-            0,
-            pkts_per_ack,
-            dst_block_port
-    );
+    dst->configure_flow_control_in(bytes_per_ack, dst_block_port);
 
     /********************************************************************
      * 5. Configure error policy
diff --git a/host/lib/rfnoc/legacy_compat.cpp b/host/lib/rfnoc/legacy_compat.cpp
index 7e9eec20e..c5bd7891a 100644
--- a/host/lib/rfnoc/legacy_compat.cpp
+++ b/host/lib/rfnoc/legacy_compat.cpp
@@ -158,10 +158,10 @@ public:
             UHD_LOGGER_WARNING("RFNOC") << "[legacy_compat] No DUCs detected. You will only be able to transmit at the radio frontend rate." ;
         }
         if (args.has_key("skip_dram")) {
-            UHD_LEGACY_LOG() << "[legacy_compat] Skipping DRAM by user request." << std::endl;
+            UHD_LEGACY_LOG() << "[legacy_compat] Skipping DRAM by user request." ;
         }
         if (args.has_key("skip_sram")) {
-            UHD_LEGACY_LOG() << "[legacy_compat] Skipping SRAM by user request." << std::endl;
+            UHD_LEGACY_LOG() << "[legacy_compat] Skipping SRAM by user request.";
         }
         if (not _has_dmafifo and not _has_sramfifo) {
             UHD_LOGGER_WARNING("RFNOC") << "[legacy_compat] No FIFO detected. Higher transmit rates may encounter errors.";
diff --git a/host/lib/rfnoc/sink_block_ctrl_base.cpp b/host/lib/rfnoc/sink_block_ctrl_base.cpp
index b620f917f..1562e134b 100644
--- a/host/lib/rfnoc/sink_block_ctrl_base.cpp
+++ b/host/lib/rfnoc/sink_block_ctrl_base.cpp
@@ -52,22 +52,17 @@ size_t sink_block_ctrl_base::get_fifo_size(size_t block_port) const {
 }
 
 void sink_block_ctrl_base::configure_flow_control_in(
-        size_t cycles,
-        size_t packets,
+        size_t bytes,
         size_t block_port
 ) {
-    UHD_RFNOC_BLOCK_TRACE() << boost::format("sink_block_ctrl_base::configure_flow_control_in(cycles=%d, packets=%d)") % cycles % packets ;
-    uint32_t cycles_word = 0;
-    if (cycles) {
-        cycles_word = (1<<31) | cycles;
-    }
-    sr_write(SR_FLOW_CTRL_CYCS_PER_ACK, cycles_word, block_port);
+    UHD_RFNOC_BLOCK_TRACE() << boost::format("sink_block_ctrl_base::configure_flow_control_in(bytes=%d)") % bytes;
 
-    uint32_t packets_word = 0;
-    if (packets) {
-        packets_word = (1<<31) | packets;
+    uint32_t bytes_word = 0;
+    if (bytes) {
+        // Bit 32 enables flow control
+        bytes_word = (1<<31) | bytes;
     }
-    sr_write(SR_FLOW_CTRL_PKTS_PER_ACK, packets_word, block_port);
+    sr_write(SR_FLOW_CTRL_BYTES_PER_ACK, bytes_word, block_port);
 }
 
 void sink_block_ctrl_base::set_error_policy(
diff --git a/host/lib/rfnoc/source_block_ctrl_base.cpp b/host/lib/rfnoc/source_block_ctrl_base.cpp
index 5ede899cb..afec6ba1b 100644
--- a/host/lib/rfnoc/source_block_ctrl_base.cpp
+++ b/host/lib/rfnoc/source_block_ctrl_base.cpp
@@ -77,25 +77,27 @@ void source_block_ctrl_base::set_destination(
 }
 
 void source_block_ctrl_base::configure_flow_control_out(
-            size_t buf_size_pkts,
+            bool enable_fc_output,
+            size_t buf_size_bytes,
+            size_t pkt_limit,
             size_t block_port,
             UHD_UNUSED(const uhd::sid_t &sid)
 ) {
-    UHD_RFNOC_BLOCK_TRACE() << "source_block_ctrl_base::configure_flow_control_out() buf_size_pkts==" << buf_size_pkts ;
-    if (buf_size_pkts < 2) {
+    UHD_RFNOC_BLOCK_TRACE() << "source_block_ctrl_base::configure_flow_control_out() buf_size_bytes==" << buf_size_bytes;
+    if (buf_size_bytes == 0) {
       throw uhd::runtime_error(str(
-              boost::format("Invalid window size %d for block %s. Window size must at least be 2.")
-              % buf_size_pkts % unique_id()
+              boost::format("Invalid window size %d for block %s. Window size cannot be 0 bytes.")
+              % buf_size_bytes % unique_id()
       ));
     }
 
-    //Disable the window and let all upstream data flush out
+    //Disable flow control entirely and let all upstream data flush out
     //We need to do this every time the window is changed because
     //a) We don't know what state the flow-control module was left in
     //   in the previous run (it should still be enabled)
     //b) Changing the window size where data is buffered upstream may
     //   result in stale packets entering the stream.
-    sr_write(SR_FLOW_CTRL_WINDOW_EN, 0, block_port);
+    sr_write(SR_FLOW_CTRL_EN, 0, block_port);
 
     //Wait for data to flush out.
     //In the FPGA we are guaranteed that all buffered packets are more-or-less consecutive.
@@ -107,13 +109,24 @@ void source_block_ctrl_base::configure_flow_control_out(
     //      module is done flushing.
     std::this_thread::sleep_for(std::chrono::milliseconds(1));
 
+    //Enable source flow control module and conditionally enable byte based and/or packet count
+    //based flow control
+    const bool enable_byte_fc = (buf_size_bytes != 0);
+    const bool enable_pkt_cnt_fc = (pkt_limit != 0);
+    const size_t config = enable_fc_output + (enable_byte_fc << 1) + (enable_pkt_cnt_fc << 2);
+
     //Resize the FC window.
     //Precondition: No data can be buffered upstream.
-    sr_write(SR_FLOW_CTRL_WINDOW_SIZE, buf_size_pkts, block_port);
+    if (enable_byte_fc) {
+        sr_write(SR_FLOW_CTRL_WINDOW_SIZE, buf_size_bytes, block_port);
+    }
+    if (enable_pkt_cnt_fc) {
+        sr_write(SR_FLOW_CTRL_PKT_LIMIT, pkt_limit, block_port);
+    }
 
     //Enable the FC window.
-    //Precondition: The window size must be set.
-    sr_write(SR_FLOW_CTRL_WINDOW_EN, (buf_size_pkts != 0), block_port);
+    //Precondition: The window size and/or packet limit must be set.
+    sr_write(SR_FLOW_CTRL_EN, config, block_port);
 }
 
 /***********************************************************************