From 924ae7300026b1f28a29b1dd7c24ffcb0a867876 Mon Sep 17 00:00:00 2001
From: whitequark <whitequark@whitequark.org>
Date: Fri, 3 Mar 2017 09:14:43 +0000
Subject: [PATCH] firmware: fix DMA trace alignment and flush caches.

---
 artiq/firmware/ksupport/lib.rs     |  7 ++++---
 artiq/firmware/runtime/rtio_dma.rs | 28 ++++++++++++++++++++++------
 artiq/firmware/runtime/session.rs  |  2 ++
 3 files changed, 28 insertions(+), 9 deletions(-)

diff --git a/artiq/firmware/ksupport/lib.rs b/artiq/firmware/ksupport/lib.rs
index 8b54b8111..8a4be01da 100644
--- a/artiq/firmware/ksupport/lib.rs
+++ b/artiq/firmware/ksupport/lib.rs
@@ -320,9 +320,10 @@ extern fn dma_playback(timestamp: i64, name: CSlice<u8>) {
     let succeeded = recv!(&DmaPlaybackReply(data) => unsafe {
         match data {
             Some(bytes) => {
-                // Here, we take advantage of the fact that DmaPlaybackReply always refers
-                // to an entire heap allocation, which is 4-byte-aligned.
-                csr::rtio_dma::base_address_write(bytes.as_ptr() as u64);
+                let ptr = bytes.as_ptr() as usize;
+                assert!(ptr % 64 == 0);
+
+                csr::rtio_dma::base_address_write(ptr as u64);
                 csr::rtio_dma::time_offset_write(timestamp as u64);
 
                 rtio_arb_dma();
diff --git a/artiq/firmware/runtime/rtio_dma.rs b/artiq/firmware/runtime/rtio_dma.rs
index 48a34bd87..a8d624462 100644
--- a/artiq/firmware/runtime/rtio_dma.rs
+++ b/artiq/firmware/runtime/rtio_dma.rs
@@ -4,9 +4,12 @@ use std::string::String;
 use std::btree_map::BTreeMap;
 use std::io::Write;
 
+const ALIGNMENT: usize = 64;
+
 #[derive(Debug)]
 struct Entry {
-    data: Vec<u8>
+    data: Vec<u8>,
+    padding: usize
 }
 
 #[derive(Debug)]
@@ -24,7 +27,7 @@ impl Manager {
     }
 
     pub fn record_start(&mut self) {
-        self.recording.clear();
+        self.recording = Vec::new();
     }
 
     pub fn record_append(&mut self, timestamp: u64, channel: u32,
@@ -63,12 +66,23 @@ impl Manager {
         let mut recorded = Vec::new();
         mem::swap(&mut self.recording, &mut recorded);
         recorded.push(0);
-        recorded.shrink_to_fit();
+        let data_len = recorded.len();
 
-        info!("recorded DMA data: {:?}", recorded);
+        // Realign.
+        recorded.reserve(ALIGNMENT - 1);
+        let padding = ALIGNMENT - recorded.as_ptr() as usize % ALIGNMENT;
+        let padding = if padding == ALIGNMENT { 0 } else { padding };
+        for _ in 0..padding {
+            // Vec guarantees that this will not reallocate
+            recorded.push(0)
+        }
+        for i in 1..data_len + 1 {
+            recorded[data_len + padding - i] = recorded[data_len - i]
+        }
 
         self.entries.insert(String::from(name), Entry {
-            data: recorded
+            data: recorded,
+            padding: padding,
         });
     }
 
@@ -77,6 +91,8 @@ impl Manager {
     }
 
     pub fn with_trace<F: FnOnce(Option<&[u8]>) -> R, R>(&self, name: &str, f: F) -> R {
-        f(self.entries.get(name).map(|entry| &entry.data[..]))
+        f(self.entries
+              .get(name)
+              .map(|entry| &entry.data[entry.padding..]))
     }
 }
diff --git a/artiq/firmware/runtime/session.rs b/artiq/firmware/runtime/session.rs
index f46ea6ed1..ad9b6577e 100644
--- a/artiq/firmware/runtime/session.rs
+++ b/artiq/firmware/runtime/session.rs
@@ -410,6 +410,8 @@ fn process_kern_message(io: &Io, mut stream: Option<&mut TcpStream>,
             }
             &kern::DmaRecordStop(name) => {
                 session.congress.dma_manager.record_stop(name);
+                board::cache::flush_cpu_dcache();
+                board::cache::flush_l2_cache();
                 kern_acknowledge()
             }
             &kern::DmaEraseRequest(name) => {