sr_write does not propperly prevent from overwriting send-ethernet-buffer when writing too fast #258
Comments
|
@Kuurusch On which device? |
|
Also, is this on a custom RFNoC block? If so, did you set CMD_FIFO_SIZE to a custom value? |
|
If it is a custom RFNoC Block, try setting CMD_FIFO_SIZE to 8 (see noc_block_radio_core.v). |
|
(If that works, it's still a bug on our side, but it would narrow down the source). |
on a x310 with ubx-160 daughterboards. |
Yes, it is a custome RFNoC-block. We will try, thank you very much! |
|
The error we get is the following: |
|
So we've resized the fifo, but no difference! |
|
@Kuurusch Huh, I was not expecting that. We have a fix for a (maybe? hopefully?) related but in the pipeline, give us a few days to push it out and then we'll try again. |
|
@mbr0wn Thank you very much! |
|
@Kuurusch You can apply this in the meantime and see if it helps: From e9dc02d796e8a3206d3847c7d24fc39e78dff472 Mon Sep 17 00:00:00 2001
From: Sugandha Gupta <[email protected]>
Date: Tue, 12 Mar 2019 10:27:23 -0700
Subject: [PATCH] rfnoc: noc_shell: Change default address for the reg
readbacks
The rb_stb signal is used in the user logic (e.g. simple_spi_core
in the radio) as a means to back pressure and prevent spi setting
register writes during ongoing spi transactions. While in the
noc_shell register it is used more as a "valid" or "done" signal
for register reads. Without this fix, the default rb_addr is the
last used rb address and can affect the behavior of rb_stb/back
pressure in the spi core, which eventually results in spi reg
writes during a spi transaction. Changing the default to RB_USER,
allows the user to control the behavior of rb_stb.
---
usrp3/lib/rfnoc/cmd_pkt_proc.v | 6 +++++-
usrp3/lib/rfnoc/noc_shell.v | 1 +
2 files changed, 6 insertions(+), 1 deletion(-)
diff --git a/usrp3/lib/rfnoc/cmd_pkt_proc.v b/usrp3/lib/rfnoc/cmd_pkt_proc.v
index 1c1f422db..8506a89bf 100644
--- a/usrp3/lib/rfnoc/cmd_pkt_proc.v
+++ b/usrp3/lib/rfnoc/cmd_pkt_proc.v
@@ -52,6 +52,7 @@ module cmd_pkt_proc #(
// TODO: Eliminate extra readback address output once NoC Shell / user register readback address spaces are merged
parameter SR_RB_ADDR = 0, // Settings bus address to set NoC Shell readback address register
parameter SR_RB_ADDR_USER = 1, // Settings bus address to set user readback address register
+ parameter DEFAULT_RB_ADDR = 0, // Default address for the readback address bus
parameter FIFO_SIZE = 5 // Depth of command FIFO
)(
input clk, input reset, input clear,
@@ -150,7 +151,7 @@ module cmd_pkt_proc #(
resp_tvalid <= 1'b0;
set_stb <= 1'b0;
set_has_time <= 1'b0;
- rb_addr <= 'd0;
+ rb_addr <= DEFAULT_RB_ADDR;
rb_addr_user <= 'd0;
end else begin
case (state)
@@ -212,6 +213,9 @@ module cmd_pkt_proc #(
rb_addr <= int_tdata[RB_AWIDTH-1:0];
end else if (set_rb_addr_user) begin
rb_addr_user <= int_tdata[RB_USER_AWIDTH-1:0];
+ rb_addr <= DEFAULT_RB_ADDR;
+ end else begin
+ rb_addr <= DEFAULT_RB_ADDR;
end
set_stb <= 1'b1;
if (int_tlast) begin
diff --git a/usrp3/lib/rfnoc/noc_shell.v b/usrp3/lib/rfnoc/noc_shell.v
index b66688aed..aef26a3b2 100644
--- a/usrp3/lib/rfnoc/noc_shell.v
+++ b/usrp3/lib/rfnoc/noc_shell.v
@@ -249,6 +249,7 @@ module noc_shell
.USE_TIME(USE_TIMED_CMDS),
.SR_RB_ADDR(SR_RB_ADDR),
.SR_RB_ADDR_USER(SR_RB_ADDR_USER),
+ .DEFAULT_RB_ADDR(RB_USER_RB_DATA),
.FIFO_SIZE(CMD_FIFO_SIZE[8*k+7:8*k]))
cmd_pkt_proc (
.clk(clk), .reset(reset), .clear(1'b0),
--
2.20.1 |
|
@mbr0wn Hi, we hadn't time to test this out but we just moved from UHD 3.14.0.0-rc1 to the final release 3.14.0.0. But we saw, that nothing has changed in the FPGA-files. Where have you introduced your changes? |
|
Good question; they haven't been merged yet. I will ping the team and see what's up with that changeset. |
|
I'm having a similar issue: I changed CMD_FIFO_SIZE to 8 and applied the patch above but it did not have any effect. I noticed the radio block also sets RESP_FIFO_SIZE to 5. This buffers response packets coming out of the command packet processor. Could this help? I'm currently building this now.. |
|
hi current when I do this the sampeRate is 50e6, it's very easy to throw this exception when I recevie huge data. For testing this, I download gnuradio and use `/usr/local/bin/uhd_fft -f800e6 -s50e6' , it also will throw this exception when it's running for about 1minutes. When I change the sample rate to 20e6, it seems work well. void UsrpObj::ReceiveNSymbol(ComplexType *hostBuff, size_t nSymbol)
{
uhd::stream_cmd_t stream_cmd(uhd::stream_cmd_t::STREAM_MODE_START_CONTINUOUS);
stream_cmd.num_samps = nSymbol;
stream_cmd.stream_now = true;
stream_cmd.time_spec = uhd::time_spec_t();
//stream_cmd.time_spec = usrp_->get_time_now();
uhd::stream_args_t stream_args("fc32", "sc16");
std::vector<size_t> channel_nums;
channel_nums.push_back(boost::lexical_cast<size_t>(0));
stream_args.channels = channel_nums;
uhd::rx_streamer::sptr rx_stream = usrp_->get_rx_stream(stream_args);
rx_stream->issue_stream_cmd(stream_cmd);
uhd::rx_metadata_t md;
size_t samplePerBuff = 0;
size_t toWrite = 0;
size_t nSampsLeft = nSymbol;
const int kSAMPLE_PER_BUFF = rx_stream->get_max_num_samps();
cout << boost::format("sample_per_buff = %d\n") % kSAMPLE_PER_BUFF;
size_t numRxSamps;
const auto start = std::chrono::steady_clock::now();
while (nSampsLeft != 0)
{
if (nSampsLeft >= kSAMPLE_PER_BUFF)
samplePerBuff = kSAMPLE_PER_BUFF;
else
samplePerBuff = nSampsLeft;
// std::cout << nSampsLeft << std::endl;
try {
numRxSamps = rx_stream->recv(hostBuff + toWrite, samplePerBuff, md, 3.0, false);
}
catch (uhd::exception &e) {
cout << boost::format("uhd::io_error %s") % e.what() << endl;
continue;
}
if (md.error_code == uhd::rx_metadata_t::ERROR_CODE_TIMEOUT)
{
//printf("Timeout while streaming......\n");
break;
}
if (md.error_code == uhd::rx_metadata_t::ERROR_CODE_OVERFLOW)
{
//printf("Overflowing while stream......\n");
continue;
}
if (md.error_code != uhd::rx_metadata_t::ERROR_CODE_NONE)
{
//printf("Receive error: %s \n", md.strerror());
continue;
}
//std::copy(buff.begin(), buff.end(), hostBuff + toWrite);
toWrite += numRxSamps;
nSampsLeft -= numRxSamps;
}
const auto end = std::chrono::steady_clock::now();
const auto time_diff = std::chrono::duration_cast<std::chrono::milliseconds>(end - start);
cout << boost::format("recv cost %d milliseconds") % time_diff.count() << endl;
stream_cmd.stream_mode = uhd::stream_cmd_t::STREAM_MODE_STOP_CONTINUOUS;
stream_cmd.stream_now = false;
rx_stream->issue_stream_cmd(stream_cmd); //to end continuous
} |
|
Does someone tested this problem? And the fix? We hadent time so fare but run into the problem again at an other part of our CPP software. If not necessary, I would like to prevent to powering up all the environment to develop on the FPGA again to figure out, that the fix didn't work... |
|
We are no longer fixing issues on 3.15, so I'm closing this. As a workaround, you can try splitting command and data traffic, and reducing the poke speed. |
Issue Description
We have encountered, that when we write too many values too fast one after each other with sr_write() on a device3 (RFNoC), we get en error where it says, some packet-number is before an other. But this comes, because we overwrite a full buffer. When we write slower (with more time between each sr_write()-call, it works without problem.
But I would expect, that this is handled by the API and it should check, before writing to the ethernet-send-buffer, if it is not full!!
The text was updated successfully, but these errors were encountered: