1
0
Fork 0

runtime/kernel: store DMA and cache buffer on core0.

Closes #77.
This commit is contained in:
pca006132 2020-08-05 12:25:48 +08:00
parent 72427dbebb
commit 4b6c5d5679
5 changed files with 124 additions and 233 deletions

View File

@ -1,8 +1,7 @@
use core::fmt; use core::fmt;
use core::cell::RefCell; use core::cell::RefCell;
use core::str::Utf8Error; use core::str::Utf8Error;
use alloc::rc::Rc; use alloc::{vec, vec::Vec, string::String, collections::BTreeMap, rc::Rc};
use alloc::{vec, vec::Vec, string::String};
use log::{info, warn, error}; use log::{info, warn, error};
use num_derive::{FromPrimitive, ToPrimitive}; use num_derive::{FromPrimitive, ToPrimitive};
@ -18,7 +17,7 @@ use libboard_zynq::{
}, },
timer::GlobalTimer, timer::GlobalTimer,
}; };
use libcortex_a9::semaphore::Semaphore; use libcortex_a9::{semaphore::Semaphore, mutex::Mutex};
use futures::{select_biased, future::FutureExt}; use futures::{select_biased, future::FutureExt};
use libasync::{smoltcp::{Sockets, TcpStream}, task}; use libasync::{smoltcp::{Sockets, TcpStream}, task};
@ -83,6 +82,9 @@ enum Reply {
ClockFailure = 15, ClockFailure = 15,
} }
static CACHE_STORE: Mutex<BTreeMap<String, Vec<i32>>> = Mutex::new(BTreeMap::new());
static DMA_RECORD_STORE: Mutex<BTreeMap<String, (Vec<u8>, i64)>> = Mutex::new(BTreeMap::new());
async fn write_header(stream: &TcpStream, reply: Reply) -> Result<()> { async fn write_header(stream: &TcpStream, reply: Reply) -> Result<()> {
stream.send([0x5a, 0x5a, 0x5a, 0x5a, reply.to_u8().unwrap()].iter().copied()).await?; stream.send([0x5a, 0x5a, 0x5a, 0x5a, reply.to_u8().unwrap()].iter().copied()).await?;
Ok(()) Ok(())
@ -223,6 +225,25 @@ async fn handle_run_kernel(stream: Option<&TcpStream>, control: &Rc<RefCell<kern
} }
break; break;
} }
kernel::Message::CachePutRequest(key, value) => {
CACHE_STORE.lock().insert(key, value);
},
kernel::Message::CacheGetRequest(key) => {
const DEFAULT: Vec<i32> = Vec::new();
let value = CACHE_STORE.lock().get(&key).unwrap_or(&DEFAULT).clone();
control.borrow_mut().tx.async_send(kernel::Message::CacheGetReply(value)).await;
},
kernel::Message::DmaPutRequest(recorder) => {
DMA_RECORD_STORE.lock().insert(recorder.name, (recorder.buffer, recorder.duration));
},
kernel::Message::DmaEraseRequest(name) => {
// prevent possible OOM when we have large DMA record replacement.
DMA_RECORD_STORE.lock().remove(&name);
},
kernel::Message::DmaGetRequest(name) => {
let result = DMA_RECORD_STORE.lock().get(&name).map(|v| v.clone());
control.borrow_mut().tx.async_send(kernel::Message::DmaGetReply(result)).await;
},
_ => { _ => {
panic!("unexpected message from core1 while kernel was running: {:?}", reply); panic!("unexpected message from core1 while kernel was running: {:?}", reply);
} }

View File

@ -1,78 +1,27 @@
use alloc::{string::String, vec::Vec, collections::BTreeMap}; use alloc::string::String;
use libcortex_a9::mutex::Mutex;
use cslice::{CSlice, AsCSlice}; use cslice::{CSlice, AsCSlice};
use core::mem::transmute; use core::mem::transmute;
use core::str; use super::{KERNEL_CHANNEL_0TO1, KERNEL_CHANNEL_1TO0, Message};
use crate::artiq_raise;
#[derive(Debug)]
struct Entry {
data: Vec<i32>,
borrowed: bool
}
#[derive(Debug)]
struct Cache {
entries: BTreeMap<String, Entry>
}
impl Cache {
pub const fn new() -> Cache {
Cache { entries: BTreeMap::new() }
}
pub fn get(&mut self, key: &str) -> *const [i32] {
match self.entries.get_mut(key) {
None => &[],
Some(ref mut entry) => {
entry.borrowed = true;
&entry.data[..]
}
}
}
pub fn put(&mut self, key: &str, data: &[i32]) -> Result<(), ()> {
match self.entries.get_mut(key) {
None => (),
Some(ref mut entry) => {
if entry.borrowed { return Err(()) }
entry.data = Vec::from(data);
return Ok(())
}
}
self.entries.insert(String::from(key), Entry {
data: Vec::from(data),
borrowed: false
});
Ok(())
}
pub unsafe fn unborrow(&mut self) {
for (_key, entry) in self.entries.iter_mut() {
entry.borrowed = false;
}
}
}
static CACHE: Mutex<Cache> = Mutex::new(Cache::new());
pub extern fn get(key: CSlice<u8>) -> CSlice<'static, i32> { pub extern fn get(key: CSlice<u8>) -> CSlice<'static, i32> {
let value = CACHE.lock().get(str::from_utf8(key.as_ref()).unwrap()); let key = String::from_utf8(key.as_ref().to_vec()).unwrap();
KERNEL_CHANNEL_1TO0.lock().as_mut().unwrap().send(Message::CacheGetRequest(key));
let msg = KERNEL_CHANNEL_0TO1.lock().as_mut().unwrap().recv();
if let Message::CacheGetReply(v) = msg {
let slice = v.as_c_slice();
// we intentionally leak the memory here,
// which does not matter as core1 would restart
unsafe { unsafe {
transmute::<*const [i32], &'static [i32]>(value).as_c_slice() transmute(slice)
}
} else {
panic!("Expected CacheGetReply for CacheGetRequest");
} }
} }
pub extern fn put(key: CSlice<u8>, list: CSlice<i32>) { pub extern fn put(key: CSlice<u8>, list: CSlice<i32>) {
let result = CACHE.lock().put(str::from_utf8(key.as_ref()).unwrap(), list.as_ref()); let key = String::from_utf8(key.as_ref().to_vec()).unwrap();
if result.is_err() { let value = list.as_ref().to_vec();
artiq_raise!("CacheError", "cannot put into a busy cache row"); KERNEL_CHANNEL_1TO0.lock().as_mut().unwrap().send(Message::CachePutRequest(key, value));
}
} }
pub unsafe fn unborrow() {
CACHE.lock().unborrow();
}

View File

@ -22,7 +22,7 @@ use super::{
KERNEL_CHANNEL_0TO1, KERNEL_CHANNEL_1TO0, KERNEL_CHANNEL_0TO1, KERNEL_CHANNEL_1TO0,
KERNEL_IMAGE, KERNEL_IMAGE,
Message, Message,
cache dma,
}; };
// linker symbols // linker symbols
@ -149,6 +149,7 @@ pub fn main_core1() {
unsafe { unsafe {
core0_tx.reset(); core0_tx.reset();
core1_tx.reset(); core1_tx.reset();
dma::init_dma_recorder();
} }
*CHANNEL_0TO1.lock() = Some(core0_tx); *CHANNEL_0TO1.lock() = Some(core0_tx);
*CHANNEL_1TO0.lock() = Some(core0_rx); *CHANNEL_1TO0.lock() = Some(core0_rx);
@ -186,7 +187,6 @@ pub fn main_core1() {
core1_rx = core::mem::replace(&mut *KERNEL_CHANNEL_0TO1.lock(), None).unwrap(); core1_rx = core::mem::replace(&mut *KERNEL_CHANNEL_0TO1.lock(), None).unwrap();
core1_tx = core::mem::replace(&mut *KERNEL_CHANNEL_1TO0.lock(), None).unwrap(); core1_tx = core::mem::replace(&mut *KERNEL_CHANNEL_1TO0.lock(), None).unwrap();
} }
unsafe { cache::unborrow(); }
info!("kernel finished"); info!("kernel finished");
core1_tx.send(Message::KernelFinished); core1_tx.send(Message::KernelFinished);
} }
@ -197,8 +197,6 @@ pub fn main_core1() {
/// Called by eh_artiq /// Called by eh_artiq
pub fn terminate(exception: &'static eh_artiq::Exception<'static>, backtrace: &'static mut [usize]) -> ! { pub fn terminate(exception: &'static eh_artiq::Exception<'static>, backtrace: &'static mut [usize]) -> ! {
unsafe { cache::unborrow(); }
let load_addr = unsafe { let load_addr = unsafe {
KERNEL_IMAGE.as_ref().unwrap().get_load_addr() KERNEL_IMAGE.as_ref().unwrap().get_load_addr()
}; };

View File

@ -3,110 +3,15 @@ use crate::{
artiq_raise, artiq_raise,
rtio, rtio,
}; };
use alloc::{vec::Vec, string::String, collections::BTreeMap, str}; use alloc::{vec::Vec, string::String, boxed::Box};
use cslice::CSlice; use cslice::CSlice;
use super::KERNEL_IMAGE; use super::{KERNEL_IMAGE, KERNEL_CHANNEL_0TO1, KERNEL_CHANNEL_1TO0, Message};
use core::mem; use core::mem;
use log::debug; use log::debug;
use libcortex_a9::{ use libcortex_a9::cache::dcci_slice;
cache::dcci_slice,
asm::dsb,
};
const ALIGNMENT: usize = 16 * 8; const ALIGNMENT: usize = 16 * 8;
const DMA_BUFFER_SIZE: usize = 16 * 8 * 1024;
struct DmaRecorder {
active: bool,
data_len: usize,
buffer: [u8; DMA_BUFFER_SIZE],
}
static mut DMA_RECORDER: DmaRecorder = DmaRecorder {
active: false,
data_len: 0,
buffer: [0; DMA_BUFFER_SIZE],
};
#[derive(Debug)]
struct Entry {
trace: Vec<u8>,
padding_len: usize,
duration: u64
}
#[derive(Debug)]
pub struct Manager {
entries: BTreeMap<String, Entry>,
recording_name: String,
recording_trace: Vec<u8>
}
// Copied from https://github.com/m-labs/artiq/blob/master/artiq/firmware/runtime/rtio_dma.rs
// basically without modification except removing some warnings.
impl Manager {
pub const fn new() -> Manager {
Manager {
entries: BTreeMap::new(),
recording_name: String::new(),
recording_trace: Vec::new(),
}
}
pub fn record_start(&mut self, name: &str) {
self.recording_name = String::from(name);
self.recording_trace = Vec::new();
// or we could needlessly OOM replacing a large trace
self.entries.remove(name);
}
pub fn record_append(&mut self, data: &[u8]) {
self.recording_trace.extend_from_slice(data);
}
pub fn record_stop(&mut self, duration: u64) {
let mut trace = Vec::new();
mem::swap(&mut self.recording_trace, &mut trace);
trace.push(0);
let data_len = trace.len();
// Realign.
trace.reserve(ALIGNMENT - 1);
let padding = ALIGNMENT - trace.as_ptr() as usize % ALIGNMENT;
let padding = if padding == ALIGNMENT { 0 } else { padding };
for _ in 0..padding {
// Vec guarantees that this will not reallocate
trace.push(0)
}
for i in 1..data_len + 1 {
trace[data_len + padding - i] = trace[data_len - i]
}
let mut name = String::new();
mem::swap(&mut self.recording_name, &mut name);
self.entries.insert(name, Entry {
trace, duration,
padding_len: padding,
});
}
pub fn erase(&mut self, name: &str) {
self.entries.remove(name);
}
pub fn with_trace<F, R>(&self, name: &str, f: F) -> R
where F: FnOnce(Option<&[u8]>, u64) -> R {
match self.entries.get(name) {
Some(entry) => f(Some(&entry.trace[entry.padding_len..]), entry.duration),
None => f(None, 0)
}
}
}
static mut DMA_MANAGER: Manager = Manager::new();
#[repr(C)] #[repr(C)]
pub struct DmaTrace { pub struct DmaTrace {
@ -114,18 +19,26 @@ pub struct DmaTrace {
address: i32, address: i32,
} }
fn dma_record_flush() { #[derive(Clone, Debug)]
unsafe { pub struct DmaRecorder {
DMA_MANAGER.record_append(&DMA_RECORDER.buffer[..DMA_RECORDER.data_len]); pub name: String,
DMA_RECORDER.data_len = 0; pub buffer: Vec<u8>,
} pub duration: i64,
}
static mut RECORDER: Option<DmaRecorder> = None;
pub unsafe fn init_dma_recorder() {
// as static would remain after restart, we have to reset it,
// without running its destructor.
mem::forget(mem::replace(&mut RECORDER, None));
} }
pub extern fn dma_record_start(name: CSlice<u8>) { pub extern fn dma_record_start(name: CSlice<u8>) {
let name = str::from_utf8(name.as_ref()).unwrap(); let name = String::from_utf8(name.as_ref().to_vec()).unwrap();
KERNEL_CHANNEL_1TO0.lock().as_mut().unwrap().send(Message::DmaEraseRequest(name.clone()));
unsafe { unsafe {
if DMA_RECORDER.active { if RECORDER.is_some() {
artiq_raise!("DMAError", "DMA is already recording") artiq_raise!("DMAError", "DMA is already recording")
} }
@ -135,16 +48,17 @@ pub extern fn dma_record_start(name: CSlice<u8>) {
library.rebind(b"rtio_output_wide", library.rebind(b"rtio_output_wide",
dma_record_output_wide as *const ()).unwrap(); dma_record_output_wide as *const ()).unwrap();
DMA_RECORDER.active = true; RECORDER = Some(DmaRecorder {
DMA_MANAGER.record_start(name); name,
buffer: Vec::new(),
duration: 0,
});
} }
} }
pub extern fn dma_record_stop(duration: i64) { pub extern fn dma_record_stop(_: i64) {
unsafe { unsafe {
dma_record_flush(); if RECORDER.is_none() {
if !DMA_RECORDER.active {
artiq_raise!("DMAError", "DMA is not recording") artiq_raise!("DMAError", "DMA is not recording")
} }
@ -154,29 +68,22 @@ pub extern fn dma_record_stop(duration: i64) {
library.rebind(b"rtio_output_wide", library.rebind(b"rtio_output_wide",
rtio::output_wide as *const ()).unwrap(); rtio::output_wide as *const ()).unwrap();
DMA_RECORDER.active = false; KERNEL_CHANNEL_1TO0.lock().as_mut().unwrap().send(
DMA_MANAGER.record_stop(duration as u64); Message::DmaPutRequest(RECORDER.take().unwrap())
);
} }
} }
#[inline(always)] #[inline(always)]
unsafe fn dma_record_output_prepare(timestamp: i64, target: i32, unsafe fn dma_record_output_prepare(timestamp: i64, target: i32,
words: usize) -> &'static mut [u8] { words: usize) {
// See gateware/rtio/dma.py. // See gateware/rtio/dma.py.
const HEADER_LENGTH: usize = /*length*/1 + /*channel*/3 + /*timestamp*/8 + /*address*/1; const HEADER_LENGTH: usize = /*length*/1 + /*channel*/3 + /*timestamp*/8 + /*address*/1;
let length = HEADER_LENGTH + /*data*/words * 4; let length = HEADER_LENGTH + /*data*/words * 4;
if DMA_RECORDER.buffer.len() - DMA_RECORDER.data_len < length { let buffer = &mut RECORDER.as_mut().unwrap().buffer;
dma_record_flush() buffer.reserve(length);
} buffer.extend_from_slice(&[
let record = &mut DMA_RECORDER.buffer[DMA_RECORDER.data_len..
DMA_RECORDER.data_len + length];
DMA_RECORDER.data_len += length;
let (header, data) = record.split_at_mut(HEADER_LENGTH);
header.copy_from_slice(&[
(length >> 0) as u8, (length >> 0) as u8,
(target >> 8) as u8, (target >> 8) as u8,
(target >> 16) as u8, (target >> 16) as u8,
@ -191,15 +98,13 @@ unsafe fn dma_record_output_prepare(timestamp: i64, target: i32,
(timestamp >> 56) as u8, (timestamp >> 56) as u8,
(target >> 0) as u8, (target >> 0) as u8,
]); ]);
data
} }
pub extern fn dma_record_output(target: i32, word: i32) { pub extern fn dma_record_output(target: i32, word: i32) {
unsafe { unsafe {
let timestamp = rtio::now_mu(); let timestamp = rtio::now_mu();
let data = dma_record_output_prepare(timestamp, target, 1); dma_record_output_prepare(timestamp, target, 1);
data.copy_from_slice(&[ RECORDER.as_mut().unwrap().buffer.extend_from_slice(&[
(word >> 0) as u8, (word >> 0) as u8,
(word >> 8) as u8, (word >> 8) as u8,
(word >> 16) as u8, (word >> 16) as u8,
@ -213,53 +118,60 @@ pub extern fn dma_record_output_wide(target: i32, words: CSlice<i32>) {
unsafe { unsafe {
let timestamp = rtio::now_mu(); let timestamp = rtio::now_mu();
let mut data = dma_record_output_prepare(timestamp, target, words.len()); dma_record_output_prepare(timestamp, target, words.len());
let buffer = &mut RECORDER.as_mut().unwrap().buffer;
for word in words.as_ref().iter() { for word in words.as_ref().iter() {
data[..4].copy_from_slice(&[ buffer.extend_from_slice(&[
(word >> 0) as u8, (word >> 0) as u8,
(word >> 8) as u8, (word >> 8) as u8,
(word >> 16) as u8, (word >> 16) as u8,
(word >> 24) as u8, (word >> 24) as u8,
]); ]);
data = &mut data[4..];
} }
} }
} }
pub extern fn dma_erase(name: CSlice<u8>) { pub extern fn dma_erase(name: CSlice<u8>) {
let name = str::from_utf8(name.as_ref()).unwrap(); let name = String::from_utf8(name.as_ref().to_vec()).unwrap();
KERNEL_CHANNEL_1TO0.lock().as_mut().unwrap().send(Message::DmaEraseRequest(name));
unsafe {
DMA_MANAGER.erase(name);
};
} }
pub extern fn dma_retrieve(name: CSlice<u8>) -> DmaTrace { pub extern fn dma_retrieve(name: CSlice<u8>) -> DmaTrace {
let name = str::from_utf8(name.as_ref()).unwrap(); let name = String::from_utf8(name.as_ref().to_vec()).unwrap();
KERNEL_CHANNEL_1TO0.lock().as_mut().unwrap().send(Message::DmaGetRequest(name));
let (trace, duration) = unsafe { match KERNEL_CHANNEL_0TO1.lock().as_mut().unwrap().recv() {
DMA_MANAGER.with_trace(name, |trace, duration| (trace.map(|v| { Message::DmaGetReply(None) => (),
dcci_slice(v); Message::DmaGetReply(Some((mut v, duration))) => {
dsb(); v.reserve(ALIGNMENT - 1);
v.as_ptr() let original_length = v.len();
}), duration)) let padding = ALIGNMENT - v.as_ptr() as usize % ALIGNMENT;
let padding = if padding == ALIGNMENT { 0 } else { padding };
for _ in 0..padding {
v.push(0);
}
v.copy_within(0..original_length, padding);
let v = Box::new(v);
let address = Box::into_raw(v) as *mut Vec<u8> as i32;
return DmaTrace {
address,
duration,
}; };
match trace { },
Some(ptr) => Ok(DmaTrace { _ => panic!("Expected DmaGetReply after DmaGetRequest!"),
address: ptr as i32, }
duration: duration as i64, // we have to defer raising error as we have to drop the message first...
}),
None => Err(())
}.unwrap_or_else(|_| {
artiq_raise!("DMAError", "DMA trace not found"); artiq_raise!("DMAError", "DMA trace not found");
})
} }
pub extern fn dma_playback(timestamp: i64, ptr: i32) { pub extern fn dma_playback(timestamp: i64, ptr: i32) {
assert!(ptr % ALIGNMENT as i32 == 0);
debug!("DMA playback started"); debug!("DMA playback started");
unsafe { unsafe {
let v = Box::from_raw(ptr as *mut Vec<u8>);
let padding = ALIGNMENT - v.as_ptr() as usize % ALIGNMENT;
let padding = if padding == ALIGNMENT { 0 } else { padding };
dcci_slice(&v[padding..]);
let ptr = v.as_ptr().add(padding) as i32;
csr::rtio_dma::base_address_write(ptr as u32); csr::rtio_dma::base_address_write(ptr as u32);
csr::rtio_dma::time_offset_write(timestamp as u64); csr::rtio_dma::time_offset_write(timestamp as u64);
@ -268,6 +180,7 @@ pub extern fn dma_playback(timestamp: i64, ptr: i32) {
while csr::rtio_dma::enable_read() != 0 {} while csr::rtio_dma::enable_read() != 0 {}
csr::cri_con::selected_write(0); csr::cri_con::selected_write(0);
mem::drop(v);
debug!("DMA playback finished"); debug!("DMA playback finished");
let error = csr::rtio_dma::error_read(); let error = csr::rtio_dma::error_read();

View File

@ -10,6 +10,7 @@ pub mod core1;
mod api; mod api;
mod rpc; mod rpc;
mod dma; mod dma;
pub use dma::DmaRecorder;
mod cache; mod cache;
#[derive(Debug, Clone)] #[derive(Debug, Clone)]
@ -34,6 +35,15 @@ pub enum Message {
RpcSend { is_async: bool, data: Vec<u8> }, RpcSend { is_async: bool, data: Vec<u8> },
RpcRecvRequest(*mut ()), RpcRecvRequest(*mut ()),
RpcRecvReply(Result<usize, RPCException>), RpcRecvReply(Result<usize, RPCException>),
CacheGetRequest(String),
CacheGetReply(Vec<i32>),
CachePutRequest(String, Vec<i32>),
DmaPutRequest(DmaRecorder),
DmaEraseRequest(String),
DmaGetRequest(String),
DmaGetReply(Option<(Vec<u8>, i64)>),
} }
static CHANNEL_0TO1: Mutex<Option<sync_channel::Sender<'static, Message>>> = Mutex::new(None); static CHANNEL_0TO1: Mutex<Option<sync_channel::Sender<'static, Message>>> = Mutex::new(None);