experiments: add bandwidth tester

experiments/src/main.rs

@@ -33,7 +33,7 @@ use libregister::RegisterR;
 use libsupport_zynq::{
     boot, ram,
 };
-use log::info;
+use log::{info, warn};
 
 const HWADDR: [u8; 6] = [0, 0x23, 0xde, 0xea, 0xbe, 0xef];
 
@@ -180,10 +180,10 @@ pub fn main_core0() {
     let eth = zynq::eth::Eth::default(HWADDR.clone());
     println!("Eth on");
 
-    const RX_LEN: usize = 8;
+    const RX_LEN: usize = 4096;
     // Number of transmission buffers (minimum is two because with
     // one, duplicate packet transmission occurs)
-    const TX_LEN: usize = 8;
+    const TX_LEN: usize = 4096;
     let eth = eth.start_rx(RX_LEN);
     let mut eth = eth.start_tx(TX_LEN);
 
@@ -238,18 +238,42 @@ pub fn main_core0() {
         Ok(())
     }
 
-    let counter = alloc::rc::Rc::new(core::cell::RefCell::new(0));
+    // (rx, tx)
+    let stats = alloc::rc::Rc::new(core::cell::RefCell::new((0, 0)));
+    let stats_tx = stats.clone();
     task::spawn(async move {
-        while let Ok(stream) = TcpStream::accept(TCP_PORT, 2048, 2408).await {
+        while let Ok(stream) = TcpStream::accept(TCP_PORT, 0x10_0000, 0x10_0000).await {
-            let counter = counter.clone();
+            let stats_tx = stats_tx.clone();
             task::spawn(async move {
-                *counter.borrow_mut() += 1;
+                let tx_data = (0..=255).take(4096).collect::<alloc::vec::Vec<u8>>();
-                println!("Serving {} connections", *counter.borrow());
+                loop {
-                handle_connection(stream)
+                    // const CHUNK_SIZE: usize = 65536;
-                    .await
+                    // match stream.send((0..=255).cycle().take(CHUNK_SIZE)).await {
-                    .unwrap_or_else(|e| println!("Connection: {:?}", e));
+                    match stream.send_slice(&tx_data[..]).await {
-                *counter.borrow_mut() -= 1;
+                        Ok(len) => stats_tx.borrow_mut().1 += tx_data.len(), //CHUNK_SIZE,
-                println!("Now serving {} connections", *counter.borrow());
+                        Err(e) => {
+                            warn!("tx: {:?}", e);
+                            break
+                        }
+                    }
+                }
+            });
+        }
+    });
+    let stats_rx = stats.clone();
+    task::spawn(async move {
+        while let Ok(stream) = TcpStream::accept(TCP_PORT+1, 0x10_0000, 0x10_0000).await {
+            let stats_rx = stats_rx.clone();
+            task::spawn(async move {
+                loop {
+                    match stream.recv(|buf| Poll::Ready((buf.len(), buf.len()))).await {
+                        Ok(len) => stats_rx.borrow_mut().0 += len,
+                        Err(e) => {
+                            warn!("rx: {:?}", e);
+                            break
+                        }
+                    }
+                }
             });
         }
     });
@@ -262,7 +286,13 @@ pub fn main_core0() {
             let timestamp = timer.get_us();
             let seconds = timestamp / 1_000_000;
             let micros = timestamp % 1_000_000;
-            info!("time: {:6}.{:06}s", seconds, micros);
+            let (rx, tx) = {
+                let mut stats = stats.borrow_mut();
+                let result = *stats;
+                *stats = (0, 0);
+                result
+            };
+            info!("time: {:6}.{:06}s, rx: {}k/s, tx: {}k/s", seconds, micros, rx / 1024, tx / 1024);
         }
     });
 

libcortex_a9/src/cache.rs

@@ -44,6 +44,15 @@ pub fn dcisw(setway: u32) {
     }
 }
 
+/// Data cache clean by set/way
+#[inline(always)]
+pub fn dccisw(setway: u32) {
+    unsafe {
+        llvm_asm!("mcr p15, 0, $0, c7, c14, 2" :: "r" (setway) :: "volatile");
+    }
+}
+
+
 /// A made-up "instruction": invalidate all of the L1 D-Cache
 #[inline(always)]
 pub fn dciall() {
@@ -71,6 +80,33 @@ pub fn dciall() {
     }
 }
 
+/// A made-up "instruction": flush and invalidate all of the L1 D-Cache
+#[inline(always)]
+pub fn dcciall() {
+    // the cache associativity could be read from a register, but will
+    // always be 4 in L1 data cache of a cortex a9
+    let ways = 4;
+    let bit_pos_of_way = 30; // 32 - log2(ways)
+
+    // the cache sets could be read from a register, but are always
+    // 256 for the cores in the zync-7000; in general, 128 or 512 are
+    // also possible.
+    let sets = 256;
+    let bit_pos_of_set = 5; // for a line size of 8 words = 2^5 bytes
+
+    // select L1 data cache
+    unsafe {
+        llvm_asm!("mcr p15, 2, $0, c0, c0, 0" :: "r" (0) :: "volatile");
+    }
+
+    // Invalidate entire D-Cache by iterating every set and every way
+    for set in 0..sets {
+        for way in 0..ways {
+            dccisw((set << bit_pos_of_set) | (way << bit_pos_of_way));
+        }
+    }
+}
+
 
 const CACHE_LINE: usize = 0x20;
 const CACHE_LINE_MASK: usize = CACHE_LINE - 1;

libcortex_a9/src/mmu.rs

@@ -1,5 +1,5 @@
 use bit_field::BitField;
-use super::{regs::*, asm, cache};
+use super::{regs::*, asm::*, cache::*};
 use libregister::RegisterW;
 
 #[derive(Copy, Clone)]
@@ -368,10 +368,19 @@ impl L1Table {
         let result = f(&mut section);
         entry.set_section(section);
 
-        asm::dmb();
+        // Flush L1Dcache
-        cache::tlbiall();
+        dcciall();
-        asm::dsb();
+        // // TODO: L2?
-        asm::isb();
+
+        // Invalidate TLB
+        tlbiall();
+        // Invalidate all branch predictors
+        bpiall();
+
+        // ensure completion of the BP and TLB invalidation
+        dsb();
+        // synchronize context on this processor
+        isb();
 
         result
     }
@@ -406,9 +415,9 @@ pub fn with_mmu<F: FnMut() -> !>(l1table: &L1Table, mut f: F) -> ! {
 
     // Synchronization barriers
     // Allows MMU to start
-    asm::dsb();
+    dsb();
     // Flushes pre-fetch buffer
-    asm::isb();
+    isb();
 
     f();
 }

libcortex_a9/src/uncached.rs

@@ -25,6 +25,7 @@ impl<T> UncachedSlice<T> {
         for page_start in (start..(start + size)).step_by(L1_PAGE_SIZE) {
             L1Table::get()
                 .update(page_start as *const (), |l1_section| {
+                    l1_section.tex = 0b100;
                     l1_section.cacheable = false;
                     l1_section.bufferable = false;
                 });