Profiler WIP-ish

.vscode/settings.json

@@ -31,7 +31,12 @@
         "rmii_config.h": "c",
         "conn.h": "c",
         "ssi.h": "c",
-        "flash.h": "c"
+        "flash.h": "c",
+        "printf.h": "c",
+        "stdlib.h": "c",
+        "pico.h": "c",
+        "platform.h": "c",
+        "error.h": "c"
     },
     "cortex-debug.variableUseNaturalFormat": false,
     "cortex-debug.registerUseNaturalFormat": false

CMakeLists.txt

@@ -30,6 +30,8 @@ add_executable(pico-debug
     gdb.c gdb.h
     breakpoint.c breakpoint.h
 
+    prof.c
+
     cmdline.c cmdline.h
     utils.c utils.h
 
@@ -102,8 +104,11 @@ target_compile_definitions(pico-debug PRIVATE
   CMD_CDC=2
   CMD_TCP=3335
 
+  PROF_CDC=0
+  PROF_TCP=0
+
   # Debug on CDC2
-  # DEBUG_CDC=2
+  DEBUG_CDC=2
   DEBUG_BUF=1
 )
 

PROFILING.md

@@ -0,0 +1,55 @@
+# Profiling RP2040 Code
+
+This branch of `pico_debug` has a profiler interface instead of gdb on CDC port 0.
+(In the future it should be possible for the two to coexist, but that's not done yet.)
+Wifi support is forcibly disabled as well.
+
+This will only work with code compiled with `-fno-omit-frame-pointer`, because
+interpreting the ARM unwind tables is too expensive at runtime (and plus not implemented
+because it's super annoying). I may experiment with that in the future -- the unwind tables can be downloaded to the debug probe and evaluated over SWD -- but this is good enough for most
+use cases, even though it does make the `r7` register unavailable and thus does have a performance impact.
+
+The front end is based on [`samply`](https://github.com/mstange/samply).  Check out the `pico`
+branch and build (`cargo build`).
+
+## Setup
+
+Build this code and flash it to a Pico that's connected to your target in a "Picoprobe" setup.
+If you're targeting another rp2040-based device, you may need to modify some hardcoded values
+in `prof.c` (memory and flash regions). This will be runtime configurable in the future.
+
+Install `samply` (`pico` branch).
+
+You'll need your code compiled with `-fno-omit-frame-pointer` (`add_compile_options(-fno-omit-frame-pointer)`), and then you'll need to have the `program.elf` binary available. You should also download the [bootloader binary for your chip revision](https://github.com/raspberrypi/pico-bootrom/releases) for more complete symbols.
+
+## Running
+
+Connect the picoprobe to your computer via USB. Connecting to CDC port 3 with a terminal emulator
+should give you a `pico_debug>` prompt (you don't need to do anything here, but a good verification
+that things are working).
+
+Run:
+```
+samply --pico /dev/tty.usbXXXX1 --pico-bootloader b2.elf --pico-reset --rate 50 program.elf
+```
+
+the target will reset and sampling will start. Omit `--pico-reset` to profile an already-running device. Press `^C` when done to open the profiler front-end. (Note: `program.elf` should already be flashed onto the target pico. This program does not download the program, but it will do that in the future.)
+
+`--rate` takes an argument in Hz (not ms). Note that the default for `samply` is 1ms, which is too fast for this. (It will work, but your code won't run anywhere near full speed!)
+
+## Internals
+
+This uses the SWD protocol (the bulk of the code in this repo comes from https://github.com/essele/pico_debug -- a very fast rp2040 debug interface) to halt each core at a given sampling interval, and then read `pc`, `lr`, `fp`, and walk the frame pointer hierarchy back.
+
+In my measurements, it takes around 500µs to halt, sample (with shallow frame depth), and resume a core. It could be down to 300µs, but there's some weirdness with SWD I need to investigate where the cores aren't actually getting unhalted until I read their status. This isn't amazing, but is completely workable for a 20ms or so sampling interval.
+
+The protocol itself is simple. When a connection is made to CDC port 0, there are a few commands that can be sent:
+
+  * `r` -- reset both cores
+  * `0xa0` `0xvvvv` `0xff` -- start sampling
+    * `v` is the sampling interval, in ms (little endian)
+    * `f` is flags: `1 << 0` sample core 0 only instead of both cores
+  * `0xa1` -- stop sampling
+
+When sampling is running, a stream of data will be received with the samples. See `prof.c` for the (simple) format.
+

adi.c

@@ -1259,11 +1259,20 @@ int check_cores() {
 }
 
 
-
-
-int dp_init() {
+int dp_force_init() {
     CHECK_OK(dp_initialise());
     CHECK_OK(core_select(0));
     return SWD_OK;
 }
 
+int dp_init() {
+    static bool initialized = false;
+    if (!initialized) {
+        int rv = dp_force_init();
+        if (rv == SWD_OK) {
+            initialized = true;
+        }
+        return rv;
+    }
+    return SWD_OK;
+}

id_usb.c

@@ -38,6 +38,10 @@
 #define USBD_CDC_2_EP_OUT (0x04)
 #define USBD_CDC_2_EP_IN (0x86)
 
+#define USBD_CDC_3_EP_CMD (0x87)
+#define USBD_CDC_3_EP_OUT (0x08)
+#define USBD_CDC_3_EP_IN (0x89)
+
 #define USBD_CDC_CMD_MAX_SIZE (8)
 #define USBD_CDC_IN_OUT_MAX_SIZE (64)
 
@@ -48,6 +52,7 @@
 #define USBD_STR_CDC0 (0x04)
 #define USBD_STR_CDC1 (0x05)
 #define USBD_STR_CDC2 (0x06)
+#define USBD_STR_CDC3 (0x07)
 
 //
 // Officially, according to the USB specs, we shoul duse TUSB_CLASS_MISC, SUBCLASS_COMMON, and PROTOCOL_IAD
@@ -85,6 +90,8 @@ enum
   ITF_NUM_CDC_1_DATA,
   ITF_NUM_CDC_2,
   ITF_NUM_CDC_2_DATA,
+  ITF_NUM_CDC_3,
+  ITF_NUM_CDC_3_DATA,
   ITF_NUM_TOTAL
 };
 
@@ -104,6 +111,9 @@ uint8_t const usbd_desc_cfg[] =
 
     TUD_CDC_DESCRIPTOR(ITF_NUM_CDC_2, USBD_STR_CDC2, USBD_CDC_2_EP_CMD,
         USBD_CDC_CMD_MAX_SIZE, USBD_CDC_2_EP_OUT, USBD_CDC_2_EP_IN, USBD_CDC_IN_OUT_MAX_SIZE),
+
+    TUD_CDC_DESCRIPTOR(ITF_NUM_CDC_3, USBD_STR_CDC3, USBD_CDC_3_EP_CMD,
+        USBD_CDC_CMD_MAX_SIZE, USBD_CDC_3_EP_OUT, USBD_CDC_3_EP_IN, USBD_CDC_IN_OUT_MAX_SIZE),
 };
 
 
@@ -159,7 +169,7 @@ const uint16_t *tud_descriptor_string_cb(uint8_t index, __unused uint16_t langid
             break;
 
         case USBD_STR_SERIAL:
-            len = string_to_descriptor("11223344", desc_str);
+            len = string_to_descriptor("pdebug", desc_str);
             break;
 
         case USBD_STR_CDC0:
@@ -174,6 +184,10 @@ const uint16_t *tud_descriptor_string_cb(uint8_t index, __unused uint16_t langid
             len = string_to_descriptor("debug-debug", desc_str);
             break;
 
+        case USBD_STR_CDC3:
+            len = string_to_descriptor("debug-prof", desc_str);
+            break;
+
         default:
             return NULL;
     }

main.c

@@ -46,6 +46,7 @@ void main_poll() {
     debug_poll();
 }
 
+void prof_init();
 
 int main() {
     // Take us to 150Mhz (for future rmii support)
@@ -70,15 +71,15 @@ int main() {
     dbg_uart_init();
 
     // Create the GDB server task..
-    gdb_init();
+    //gdb_init();
+
+    prof_init();
 
     // See if we have enough information to start the wifi...
     // This is horrible here, needs to be a separate func/file with wifi/net related stuff.
-    if (*cf->main->wifi.ssid && *cf->main->wifi.creds) {
-        cyw43_arch_wifi_connect_async(cf->main->wifi.ssid, cf->main->wifi.creds, CYW43_AUTH_WPA2_AES_PSK);
-    }
-
-
+    //if (*cf->main->wifi.ssid && *cf->main->wifi.creds) {
+    //    cyw43_arch_wifi_connect_async(cf->main->wifi.ssid, cf->main->wifi.creds, CYW43_AUTH_WPA2_AES_PSK);
+    //}
 
     // And start the scheduler...
     leos_init(main_poll);

modules/lerp_io/include/lerp/io.h

@@ -44,7 +44,9 @@ struct io *io_init(int cdc_port, int tcp_port, int buf_size);
 
 void io_poll();
 int io_get_byte(struct io *io);
+int io_get_byte_nb(struct io *io);
 int io_put_byte(struct io *io, uint8_t ch);
+int io_put_bytes(struct io *io, uint8_t *data, int len);
 int io_put_hexbyte(struct io *io, uint8_t b);
 int io_printf(struct io *io, char *format, ...);
 int io_aprintf(struct io *io, char *format, va_list args);

modules/lerp_io/io.c

@@ -337,6 +337,14 @@ int io_is_connected(struct io *io) {
     return io->usb_is_connected || (io->pcb != NULL);
 }
 
+int io_get_byte_nb(struct io *io) {
+    if (circ_is_empty(io->input)) {
+        return 0;
+    }
+
+    return circ_get_byte(io->input);
+}
+
 int io_get_byte(struct io *io) {
     int reason;
 
@@ -369,6 +377,19 @@ int io_put_byte(struct io *io, uint8_t ch) {
     return 0;
 }
 
+int io_put_bytes(struct io *io, uint8_t *data, int len) {
+    int reason;
+
+    if (circ_is_full(io->output)) {
+        io->waiting_on_output = current_task();
+        reason = task_block();
+        if (reason < 0) return reason;
+    }
+    //debug_putch(ch);
+    circ_add_bytes(io->output, data, len);
+    return 0;
+}
+
 int io_read_flush(struct io *io) {
     if (io->usb_is_connected) tud_cdc_n_read_flush(io->cdc_port);
     circ_clean(io->input);
@@ -489,6 +510,8 @@ struct io *io_init(int cdc_port, int tcp_port, int buf_size) {
     struct io *io = malloc(sizeof(struct io));
     if (!io) return NULL;
 
+    tcp_port = 0;
+
     io->cdc_port = cdc_port;
     io->tcp_port = tcp_port;
     io->pcb = NULL;