[Cryptech-Commits] [sw/stm32] 08/08: Rebase branch 'profiling' from master

git at cryptech.is git at cryptech.is
Thu Sep 7 22:42:08 UTC 2017
Previous message (by thread): [Cryptech-Commits] [sw/stm32] 07/08: Add task_yield_maybe
Next message (by thread): [Cryptech-Commits] [sw/libhal] branch systolic_crt created (now 68019ff)
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]
This is an automated email from the git hooks/post-receive script.

paul at psgd.org pushed a commit to branch profiling
in repository sw/stm32.

commit 72227852729ed3125af58cff3f593340b3247fed
Merge: 2e1f880 5ff8c95
Author: Paul Selkirk <paul at psgd.org>
AuthorDate: Thu Sep 7 18:17:12 2017 -0400

    Rebase branch 'profiling' from master

 Makefile                     | 10 +++++++---
 libraries/libprof/README.txt |  2 +-
 libraries/libprof/profil.c   | 17 ++++++++++-------
 libraries/libtfm/Makefile    |  5 +++++
 projects/hsm/Makefile        |  6 ++++--
 projects/hsm/hsm.c           |  5 +++++
 projects/hsm/mgmt-task.c     |  9 ++++++---
 task.c                       |  8 ++++----
 task.h                       |  2 +-
 9 files changed, 43 insertions(+), 21 deletions(-)

diff --cc Makefile
index 12e7df8,68f4da3..e349ab1
--- a/Makefile
+++ b/Makefile
@@@ -117,6 -116,6 +117,9 @@@ CFLAGS += -I$(MBED_DIR)/targets/cmsis/T
  CFLAGS += -I$(MBED_DIR)/targets/cmsis/TARGET_STM/TARGET_STM32F4/$(BOARD)
  CFLAGS += -I$(MBED_DIR)/targets/hal/TARGET_STM/TARGET_STM32F4
  CFLAGS += -I$(MBED_DIR)/targets/hal/TARGET_STM/TARGET_STM32F4/$(BOARD)
++ifdef DO_TASK_METRICS
++CFLAGS += -DDO_TASK_METRICS
++endif
  
  %.o : %.c
  	$(CC) $(CFLAGS) -c -o $@ $<
@@@ -124,7 -123,7 +127,11 @@@
  %.o : %.S
  	$(CC) $(CFLAGS) -c -o $@ $<
  
++ifdef DO_PROFILING
++all: hsm
++else
  all: board-test cli-test libhal-test hsm bootloader
++endif
  
  $(MBED_DIR)/libstmf4.a: .FORCE
  	$(MAKE) -C $(MBED_DIR)
@@@ -152,9 -151,9 +159,6 @@@ libhal-test: $(BOARD_OBJS) $(LIBS) $(LI
  
  ifdef DO_PROFILING
  CFLAGS += -pg -DDO_PROFILING
--endif
--ifdef DO_TASK_METRICS
--CFLAGS += -DDO_TASK_METRICS
  hsm: $(BOARD_OBJS) $(LIBS) $(LIBHAL_BLD)/libhal.a $(LIBCLI_BLD)/libcli.a $(LIBPROF_BLD)/libprof.a .FORCE
  	$(MAKE) -C projects/hsm
  else
diff --cc libraries/libprof/README.txt
index 1fe378c,1fe378c..f0b8ee8
--- a/libraries/libprof/README.txt
+++ b/libraries/libprof/README.txt
@@@ -52,7 -52,7 +52,7 @@@ In the OpenOCD console, enable semihost
  
  In another window, start the debugger:
  
--    $ sw/stm32/bin/debug projects/hsm/hsm
++    $ ../../bin/debug hsm
  
  In the CLI, type `profile start`, then start the unit test or whatever
  will be exercising the hsm. Afterwards, in the CLI, type `profile stop`.
diff --cc libraries/libprof/profil.c
index 0654879,0654879..b0d8d55
--- a/libraries/libprof/profil.c
+++ b/libraries/libprof/profil.c
@@@ -25,27 -25,27 +25,30 @@@ static struct profinfo prof = 
    PROFILE_NOT_INIT, 0, 0, 0, 0
  };
  
--extern void set_SysTick_hook(void (*hook)(void));
--
  /* sample the current program counter */
--static void SysTick_hook(void) {
--  size_t pc = (size_t)((uint32_t *)__get_MSP())[8];
--  if (pc >= prof.lowpc && pc < prof.highpc) {
++void profil_callback(void) {
++  if (prof.state == PROFILE_ON) {
++    /* The interrupt mechanism pushes xPSR, PC, LR, R12, and R3-R0 onto the
++     * stack, so PC is the 6th word from the top at that point. However, the
++     * normal function entry code pushes registers as well, so the stack
++     * offset right now depends on the call tree that got us here.
++     */
++    size_t pc = (size_t)((uint32_t *)__get_MSP())[6 + 6];
++    if (pc >= prof.lowpc && pc < prof.highpc) {
        size_t idx = PROFIDX (pc, prof.lowpc, prof.scale);
        prof.counter[idx]++;
++    }
    }
  }
  
  /* Stop profiling to the profiling buffer pointed to by p. */
  static int profile_off (struct profinfo *p) {
--  set_SysTick_hook(NULL);
    p->state = PROFILE_OFF;
    return 0;
  }
  
  /* Create a timer thread and pass it a pointer P to the profiling buffer. */
  static int profile_on (struct profinfo *p) {
--  set_SysTick_hook(SysTick_hook);
    p->state = PROFILE_ON;
    return 0; /* ok */
  }
diff --cc libraries/libtfm/Makefile
index ceb8541,6da552d..5be45f4
--- a/libraries/libtfm/Makefile
+++ b/libraries/libtfm/Makefile
@@@ -10,33 -10,10 +10,38 @@@ BITS := 819
  HDR  := ${LIBTFM_SRC}/tomsfastmath/src/headers/tfm.h
  LIB  := tomsfastmath/libtfm.a
  
 -#CFLAGS	+= -DTFM_X86
 -#CFLAGS	+= -DTFM_NO_ASM
 +# See sw/thirdparty/libtfm/Makefile for compilation options.  Note
 +# that libtfm platform-specific assembly code has opinions on the
 +# optimization level (and appears to be best tested with -O3).
  
 -CFLAGS	+= -fPIC -Wall -W -Wshadow -I${LIBTFM_SRC}/tomsfastmath/src/headers -g3 -DFP_MAX_SIZE="(${BITS}*2+(8*DIGIT_BIT))"
 +# Using $(subst...) here is a kludge.  A cleaner approach might be for
 +# sw/stm32/Makefile to build up the non-variant parts of CFLAGS in a
 +# different variable before merging the variant and non-variant parts
 +# into CFLAGS, which would give us a clean copy of the non-variant
 +# parts to use when constructing our own CFLAGS.  Later.
 +
 +# The ARM assembly code in libtfm still generates a lot of warnings of the form:
 +#
 +#   warning: matching constraint does not allow a register [enabled by default]
 +#
 +# This is just a warning, the resulting library appears to work
 +# correctly, and the fix appears to require a nasty intervention in
 +# the guts of the libtfm assembly code, so we live with the warning
 +# for now, at least until we confirm that it hasn't already been fixed
 +# in a newer version of libtfm.
 +
++ifdef DO_PROFILING
++# arm-none-eabi-gcc: error: -pg and -fomit-frame-pointer are incompatible
++STM32_LIBTFM_CFLAGS_OPTIMIZATION := -O3 -funroll-loops
++else
 +STM32_LIBTFM_CFLAGS_OPTIMIZATION := -O3 -funroll-loops -fomit-frame-pointer
++endif
 +
 +CFLAGS := $(subst ${STM32_CFLAGS_OPTIMIZATION},${STM32_LIBTFM_CFLAGS_OPTIMIZATION},${CFLAGS})
 +CFLAGS += -DTFM_ARM -Dasm=__asm__ -Wa,-mimplicit-it=thumb
 +CFLAGS += -I${LIBTFM_SRC}/tomsfastmath/src/headers
 +CFLAGS += -DFP_MAX_SIZE="(${BITS}*2+(8*DIGIT_BIT))"
 +CFLAGS += -Wall -W -Wshadow
  
  TARGETS	:= $(notdir ${HDR} ${LIB})
  
diff --cc projects/hsm/hsm.c
index b6b8820,33342c0..5ef2ccc
--- a/projects/hsm/hsm.c
+++ b/projects/hsm/hsm.c
@@@ -244,38 -244,25 +244,43 @@@ static void RxCallback(uint8_t c
      }
  }
  
 -static uint8_t uart_rx[2];      /* current character received from UART */
 -static uint32_t uart_rx_idx = 0;
 -
 -/* UART DMA half-complete and complete callbacks. With a 2-character DMA
 - * buffer, one or the other of these will fire on each incoming character.
 - * Under heavy load, these will sometimes fire in the wrong order, but the
 - * data are in the right order in the DMA buffer, so we have a flip-flop
 - * buffer index that doesn't depend on the order of the callbacks.
 +/* A ring buffer for the UART DMA receiver. In theory, it should get at most
 + * 92 characters per 1ms tick, but we're going to up-size it for safety.
   */
 -void HAL_UART2_RxHalfCpltCallback(UART_HandleTypeDef *huart)
 -{
 -    RxCallback(uart_rx[uart_rx_idx]);
 -    uart_rx_idx ^= 1;
 -}
 +#ifndef RPC_UART_RECVBUF_SIZE
 +#define RPC_UART_RECVBUF_SIZE  1024  /* must be a power of 2 */
 +#endif
 +#define RPC_UART_RECVBUF_MASK  (RPC_UART_RECVBUF_SIZE - 1)
 +
 +typedef struct {
 +    uint32_t ridx;
 +    uint8_t buf[RPC_UART_RECVBUF_SIZE];
 +} uart_ringbuf_t;
 +
 +volatile uart_ringbuf_t uart_ringbuf = {0, {0}};
  
 -void HAL_UART2_RxCpltCallback(UART_HandleTypeDef *huart)
 +#define RINGBUF_RIDX(rb)       (rb.ridx & RPC_UART_RECVBUF_MASK)
 +#define RINGBUF_WIDX(rb)       (sizeof(rb.buf) - __HAL_DMA_GET_COUNTER(huart_user.hdmarx))
 +#define RINGBUF_COUNT(rb)      ((RINGBUF_WIDX(rb) - RINGBUF_RIDX(rb)) & RPC_UART_RECVBUF_MASK)
 +#define RINGBUF_READ(rb, dst)  {dst = rb.buf[RINGBUF_RIDX(rb)]; rb.ridx++;}
 +
 +size_t uart_rx_max = 0;
 +
 +void HAL_SYSTICK_Callback(void)
  {
 -    RxCallback(uart_rx[uart_rx_idx]);
 -    uart_rx_idx ^= 1;
++#ifdef DO_PROFILING
++    extern void profil_callback(void);
++    profil_callback();
++#endif
++
 +    size_t count = RINGBUF_COUNT(uart_ringbuf);
 +    if (uart_rx_max < count) uart_rx_max = count;
 +
 +    while (RINGBUF_COUNT(uart_ringbuf)) {
 +        uint8_t c;
 +        RINGBUF_READ(uart_ringbuf, c);
 +        RxCallback(c);
 +    }
  }
  
  /* Send one character over the UART. This is called from
diff --cc projects/hsm/mgmt-task.c
index 9f6a908,1658a80..4668585
--- a/projects/hsm/mgmt-task.c
+++ b/projects/hsm/mgmt-task.c
@@@ -99,9 -95,9 +99,12 @@@ void configure_cli_task(struct cli_def 
      struct cli_command *c = cli_register_command(cli, NULL, "task", NULL, 0, 0, NULL);
  
      /* task show */
--    struct cli_command *c_show = cli_register_command(cli, c, "show", cmd_task_show, 0, 0, "Show the active tasks");
++#ifdef DO_TASK_METRICS
++    struct cli_command *c_show =
++#endif
++    cli_register_command(cli, c, "show", cmd_task_show, 0, 0, "Show the active tasks");
  
- #ifdef TASK_METRICS
+ #ifdef DO_TASK_METRICS
      /* task show metrics */
      cli_register_command(cli, c_show, "metrics", cmd_task_show_metrics, 0, 0, "Show task metrics");
Previous message (by thread): [Cryptech-Commits] [sw/stm32] 07/08: Add task_yield_maybe
Next message (by thread): [Cryptech-Commits] [sw/libhal] branch systolic_crt created (now 68019ff)
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]
More information about the Commits mailing list