Added perfbd.py and block device performance sampling in bench-runner

Based loosely on Linux's perf tool, perfbd.py uses trace output with
backtraces to aggregate and show the block device usage of all functions
in a program, propagating block devices operation cost up the backtrace
for each operation.

This combined with --trace-period and --trace-freq for
sampling/filtering trace events allow the bench-runner to very
efficiently record the general cost of block device operations with very
little overhead.

Adopted this as the default side-effect of make bench, replacing
cycle-based performance measurements which are less important for
littlefs.

Makefile

@@ -50,6 +50,7 @@ TEST_CI	  := $(TEST_TAC:%.t.a.c=%.t.a.ci)
 TEST_GCNO  := $(TEST_TAC:%.t.a.c=%.t.a.gcno)
 TEST_GCDA  := $(TEST_TAC:%.t.a.c=%.t.a.gcda)
 TEST_PERF  := $(TEST_RUNNER:%=%.perf)
+TEST_TRACE := $(TEST_RUNNER:%=%.trace)
 
 BENCHES ?= $(wildcard benches/*.toml)
 BENCH_SRC ?= $(SRC) \
@@ -65,6 +66,7 @@ BENCH_CI   := $(BENCH_BAC:%.b.a.c=%.b.a.ci)
 BENCH_GCNO  := $(BENCH_BAC:%.b.a.c=%.b.a.gcno)
 BENCH_GCDA  := $(BENCH_BAC:%.b.a.c=%.b.a.gcda)
 BENCH_PERF  := $(BENCH_RUNNER:%=%.perf)
+BENCH_TRACE := $(BENCH_RUNNER:%=%.trace)
 
 ifdef DEBUG
 override CFLAGS += -O0
@@ -85,6 +87,9 @@ endif
 ifdef YES_PERF
 override CFLAGS += -fno-omit-frame-pointer
 endif
+ifdef YES_PERFBD
+override CFLAGS += -fno-omit-frame-pointer
+endif
 
 ifdef VERBOSE
 override CODEFLAGS   += -v
@@ -93,7 +98,11 @@ override STACKFLAGS  += -v
 override STRUCTFLAGS += -v
 override COVFLAGS    += -v
 override PERFFLAGS   += -v
+override PERFBDFLAGS += -v
 endif
+# forward -j flag
+override PERFFLAGS   += $(filter -j%,$(MAKEFLAGS))
+override PERFBDFLAGS += $(filter -j%,$(MAKEFLAGS))
 ifneq ($(NM),nm)
 override CODEFLAGS += --nm-tool="$(NM)"
 override DATAFLAGS += --nm-tool="$(NM)"
@@ -103,6 +112,7 @@ override CODEFLAGS   += --objdump-tool="$(OBJDUMP)"
 override DATAFLAGS   += --objdump-tool="$(OBJDUMP)"
 override STRUCTFLAGS += --objdump-tool="$(OBJDUMP)"
 override PERFFLAGS   += --objdump-tool="$(OBJDUMP)"
+override PERFBDFLAGS += --objdump-tool="$(OBJDUMP)"
 endif
 ifneq ($(PERF),perf)
 override PERFFLAGS += --perf-tool="$(PERF)"
@@ -114,10 +124,14 @@ override BENCHFLAGS += -b
 override TESTFLAGS  += $(filter -j%,$(MAKEFLAGS))
 override BENCHFLAGS += $(filter -j%,$(MAKEFLAGS))
 ifdef YES_PERF
-override TESTFLAGS += --perf=$(TEST_PERF)
+override TESTFLAGS  += -p$(TEST_PERF)
+override BENCHFLAGS += -p$(BENCH_PERF)
 endif
-ifndef NO_PERF
+ifdef YES_PERFBD
-override BENCHFLAGS += --perf=$(BENCH_PERF)
+override TESTFLAGS  += -t$(TEST_TRACE) --trace-backtrace --trace-freq=100
+endif
+ifndef NO_PERFBD
+override BENCHFLAGS  += -t$(BENCH_TRACE) --trace-backtrace --trace-freq=100
 endif
 ifdef VERBOSE
 override TESTFLAGS   += -v
@@ -165,6 +179,11 @@ endif
 ifdef YES_PERF
 bench-runner build-bench: override CFLAGS+=-fno-omit-frame-pointer
 endif
+ifdef YES_PERFBD
+bench-runner build-bench: override CFLAGS+=-fno-omit-frame-pointer
+endif
+# note we remove some binary dependent files during compilation,
+# otherwise it's way to easy to end up with outdated results
 test-runner build-test: $(TEST_RUNNER)
 ifndef NO_COV
 	rm -f $(TEST_GCDA)
@@ -172,6 +191,9 @@ endif
 ifdef YES_PERF
 	rm -f $(TEST_PERF)
 endif
+ifdef YES_PERFBD
+	rm -f $(TEST_TRACE)
+endif
 
 .PHONY: test
 test: test-runner
@@ -185,16 +207,24 @@ test-list: test-runner
 ifdef YES_COV
 bench-runner build-bench: override CFLAGS+=--coverage
 endif
-ifndef NO_PERF
+ifdef YES_PERF
 bench-runner build-bench: override CFLAGS+=-fno-omit-frame-pointer
 endif
+ifndef NO_PERFBD
+bench-runner build-bench: override CFLAGS+=-fno-omit-frame-pointer
+endif
+# note we remove some binary dependent files during compilation,
+# otherwise it's way to easy to end up with outdated results
 bench-runner build-bench: $(BENCH_RUNNER)
 ifdef YES_COV 
 	rm -f $(BENCH_GCDA)
 endif
-ifndef NO_PERF
+ifdef YES_PERF
 	rm -f $(BENCH_PERF)
 endif
+ifndef NO_PERFBD
+	rm -f $(BENCH_TRACE)
+endif
 
 .PHONY: bench
 bench: bench-runner
@@ -234,6 +264,13 @@ perf: $(BENCH_PERF)
 		-Scycles \
 		$(PERFFLAGS))
 
+.PHONY: perfbd
+perfbd: $(BENCH_TRACE)
+	$(strip ./scripts/perfbd.py \
+		$(BENCH_RUNNER) $^ $(patsubst %,-F%,$(SRC)) \
+		-Serased -Sproged -Sreaded \
+		$(PERFBDFLAGS))
+
 .PHONY: summary sizes
 summary sizes: $(BUILDDIR)lfs.csv
 	$(strip ./scripts/summary.py -Y $^ \
@@ -275,6 +312,11 @@ $(BUILDDIR)lfs.cov.csv: $(GCDA)
 $(BUILDDIR)lfs.perf.csv: $(BENCH_PERF)
 	./scripts/perf.py $^ $(patsubst %,-F%,$(SRC)) -q $(PERFFLAGS) -o $@
 
+$(BUILDDIR)lfs.perfbd.csv: $(BENCH_TRACE)
+	$(strip ./scripts/perfbd.py \
+		$(BENCH_RUNNER) $^ $(patsubst %,-F%,$(SRC)) \
+		-q $(PERFBDFLAGS) -o $@)
+
 $(BUILDDIR)lfs.csv: \
 		$(BUILDDIR)lfs.code.csv \
 		$(BUILDDIR)lfs.data.csv \
@@ -326,7 +368,8 @@ clean:
 		$(BUILDDIR)lfs.stack.csv \
 		$(BUILDDIR)lfs.struct.csv \
 		$(BUILDDIR)lfs.cov.csv \
-		$(BUILDDIR)lfs.perf.csv)
+		$(BUILDDIR)lfs.perf.csv \
+		$(BUILDDIR)lfs.perfbd.csv)
 	rm -f $(OBJ)
 	rm -f $(DEP)
 	rm -f $(ASM)
@@ -340,6 +383,7 @@ clean:
 	rm -f $(TEST_GCNO)
 	rm -f $(TEST_GCDA)
 	rm -f $(TEST_PERF)
+	rm -f $(TEST_TRACE)
 	rm -f $(BENCH_RUNNER)
 	rm -f $(BENCH_BC)
 	rm -f $(BENCH_BAC)
@@ -349,3 +393,4 @@ clean:
 	rm -f $(BENCH_GCNO)
 	rm -f $(BENCH_GCDA)
 	rm -f $(BENCH_PERF)
+	rm -f $(BENCH_TRACE)

bd/lfs_emubd.c

@@ -358,7 +358,8 @@ int lfs_emubd_prog(const struct lfs_config *cfg, lfs_block_t block,
 }
 
 int lfs_emubd_erase(const struct lfs_config *cfg, lfs_block_t block) {
-    LFS_EMUBD_TRACE("lfs_emubd_erase(%p, 0x%"PRIx32")", (void*)cfg, block);
+    LFS_EMUBD_TRACE("lfs_emubd_erase(%p, 0x%"PRIx32" (%"PRIu32"))",
+            (void*)cfg, block, cfg->block_size);
     lfs_emubd_t *bd = cfg->context;
 
     // check if erase is valid

bd/lfs_filebd.c

@@ -96,7 +96,8 @@ int lfs_filebd_read(const struct lfs_config *cfg, lfs_block_t block,
 
 int lfs_filebd_prog(const struct lfs_config *cfg, lfs_block_t block,
         lfs_off_t off, const void *buffer, lfs_size_t size) {
-    LFS_FILEBD_TRACE("lfs_filebd_prog(%p, 0x%"PRIx32", %"PRIu32", %p, %"PRIu32")",
+    LFS_FILEBD_TRACE("lfs_filebd_prog(%p, "
+                "0x%"PRIx32", %"PRIu32", %p, %"PRIu32")",
             (void*)cfg, block, off, buffer, size);
     lfs_filebd_t *bd = cfg->context;
 
@@ -127,7 +128,8 @@ int lfs_filebd_prog(const struct lfs_config *cfg, lfs_block_t block,
 }
 
 int lfs_filebd_erase(const struct lfs_config *cfg, lfs_block_t block) {
-    LFS_FILEBD_TRACE("lfs_filebd_erase(%p, 0x%"PRIx32")", (void*)cfg, block);
+    LFS_FILEBD_TRACE("lfs_filebd_erase(%p, 0x%"PRIx32" (%"PRIu32"))",
+            (void*)cfg, block, cfg->block_size);
 
     // check if erase is valid
     LFS_ASSERT(block < cfg->block_count);

bd/lfs_rambd.c

@@ -107,7 +107,8 @@ int lfs_rambd_prog(const struct lfs_config *cfg, lfs_block_t block,
 }
 
 int lfs_rambd_erase(const struct lfs_config *cfg, lfs_block_t block) {
-    LFS_RAMBD_TRACE("lfs_rambd_erase(%p, 0x%"PRIx32")", (void*)cfg, block);
+    LFS_RAMBD_TRACE("lfs_rambd_erase(%p, 0x%"PRIx32" (%"PRIu32"))",
+            (void*)cfg, block, cfg->block_size);
 
     // check if erase is valid
     LFS_ASSERT(block < cfg->block_count);

runners/bench_runner.c

@@ -14,6 +14,8 @@
 #include <stdarg.h>
 #include <stdio.h>
 #include <unistd.h>
+#include <execinfo.h>
+#include <time.h>
 
 
 // some helpers
@@ -405,26 +407,63 @@ size_t bench_step_step = 1;
 
 const char *bench_disk_path = NULL;
 const char *bench_trace_path = NULL;
+bool bench_trace_backtrace = false;
+uint32_t bench_trace_period = 0;
+uint32_t bench_trace_freq = 0;
 FILE *bench_trace_file = NULL;
 uint32_t bench_trace_cycles = 0;
+uint64_t bench_trace_time = 0;
+uint64_t bench_trace_open_time = 0;
 lfs_emubd_sleep_t bench_read_sleep = 0.0;
 lfs_emubd_sleep_t bench_prog_sleep = 0.0;
 lfs_emubd_sleep_t bench_erase_sleep = 0.0;
 
+// this determines both the backtrace buffer and the trace printf buffer, if
+// trace ends up interleaved or truncated this may need to be increased
+#ifndef BENCH_TRACE_BACKTRACE_BUFFER_SIZE
+#define BENCH_TRACE_BACKTRACE_BUFFER_SIZE 8192
+#endif
+void *bench_trace_backtrace_buffer[
+    BENCH_TRACE_BACKTRACE_BUFFER_SIZE / sizeof(void*)];
 
 // trace printing
 void bench_trace(const char *fmt, ...) {
     if (bench_trace_path) {
-        if (!bench_trace_file) {
+        // sample at a specific period?
-            // Tracing output is heavy and trying to open every trace
+        if (bench_trace_period) {
-            // call is slow, so we only try to open the trace file every
+            if (bench_trace_cycles % bench_trace_period != 0) {
-            // so often. Note this doesn't affect successfully opened files
-            if (bench_trace_cycles % 128 != 0) {
                 bench_trace_cycles += 1;
                 return;
             }
             bench_trace_cycles += 1;
+        }
 
+        // sample at a specific frequency?
+        if (bench_trace_freq) {
+            struct timespec t;
+            clock_gettime(CLOCK_MONOTONIC, &t);
+            uint64_t now = (uint64_t)t.tv_sec*1000*1000*1000
+                    + (uint64_t)t.tv_nsec;
+            if (now - bench_trace_time < (1000*1000*1000) / bench_trace_freq) {
+                return;
+            }
+            bench_trace_time = now;
+        }
+
+        if (!bench_trace_file) {
+            // Tracing output is heavy and trying to open every trace
+            // call is slow, so we only try to open the trace file every
+            // so often. Note this doesn't affect successfully opened files
+            struct timespec t;
+            clock_gettime(CLOCK_MONOTONIC, &t);
+            uint64_t now = (uint64_t)t.tv_sec*1000*1000*1000
+                    + (uint64_t)t.tv_nsec;
+            if (now - bench_trace_open_time < 100*1000*1000) {
+                return;
+            }
+            bench_trace_open_time = now;
+
+            // try to open the trace file
             int fd;
             if (strcmp(bench_trace_path, "-") == 0) {
                 fd = dup(1);
@@ -445,19 +484,42 @@ void bench_trace(const char *fmt, ...) {
 
             FILE *f = fdopen(fd, "a");
             assert(f);
-            int err = setvbuf(f, NULL, _IOLBF, BUFSIZ);
+            int err = setvbuf(f, NULL, _IOFBF,
+                    BENCH_TRACE_BACKTRACE_BUFFER_SIZE);
             assert(!err);
             bench_trace_file = f;
         }
 
+        // print trace
         va_list va;
         va_start(va, fmt);
         int res = vfprintf(bench_trace_file, fmt, va);
+        va_end(va);
         if (res < 0) {
             fclose(bench_trace_file);
             bench_trace_file = NULL;
+            return;
         }
-        va_end(va);
+
+        if (bench_trace_backtrace) {
+            // print backtrace
+            size_t count = backtrace(
+                    bench_trace_backtrace_buffer,
+                    BENCH_TRACE_BACKTRACE_BUFFER_SIZE);
+            // note we skip our own stack frame
+            for (size_t i = 1; i < count; i++) {
+                res = fprintf(bench_trace_file, "\tat %p\n",
+                        bench_trace_backtrace_buffer[i]);
+                if (res < 0) {
+                    fclose(bench_trace_file);
+                    bench_trace_file = NULL;
+                    return;
+                }
+            }
+        }
+
+        // flush immediately
+        fflush(bench_trace_file);
     }
 }
 
@@ -1312,9 +1374,12 @@ enum opt_flags {
     OPT_STEP                     = 's',
     OPT_DISK                     = 'd',
     OPT_TRACE                    = 't',
-    OPT_READ_SLEEP               = 7,
+    OPT_TRACE_BACKTRACE          = 7,
-    OPT_PROG_SLEEP               = 8,
+    OPT_TRACE_PERIOD             = 8,
-    OPT_ERASE_SLEEP              = 9,
+    OPT_TRACE_FREQ               = 9,
+    OPT_READ_SLEEP               = 10,
+    OPT_PROG_SLEEP               = 11,
+    OPT_ERASE_SLEEP              = 12,
 };
 
 const char *short_opts = "hYlLD:G:s:d:t:";
@@ -1337,6 +1402,9 @@ const struct option long_opts[] = {
     {"step",             required_argument, NULL, OPT_STEP},
     {"disk",             required_argument, NULL, OPT_DISK},
     {"trace",            required_argument, NULL, OPT_TRACE},
+    {"trace-backtrace",  no_argument,       NULL, OPT_TRACE_BACKTRACE},
+    {"trace-period",     required_argument, NULL, OPT_TRACE_PERIOD},
+    {"trace-freq",       required_argument, NULL, OPT_TRACE_FREQ},
     {"read-sleep",       required_argument, NULL, OPT_READ_SLEEP},
     {"prog-sleep",       required_argument, NULL, OPT_PROG_SLEEP},
     {"erase-sleep",      required_argument, NULL, OPT_ERASE_SLEEP},
@@ -1357,8 +1425,11 @@ const char *const help_text[] = {
     "Override a bench define.",
     "Comma-separated list of disk geometries to bench.",
     "Comma-separated range of bench permutations to run (start,stop,step).",
-    "Redirect block device operations to this file.",
+    "Direct block device operations to this file.",
-    "Redirect trace output to this file.",
+    "Direct trace output to this file.",
+    "Include a backtrace with every trace statement.",
+    "Sample trace output at this period in cycles.",
+    "Sample trace output at this frequency in hz.",
     "Artificial read delay in seconds.",
     "Artificial prog delay in seconds.",
     "Artificial erase delay in seconds.",
@@ -1790,6 +1861,27 @@ step_unknown:
             case OPT_TRACE:
                 bench_trace_path = optarg;
                 break;
+            case OPT_TRACE_BACKTRACE:
+                bench_trace_backtrace = true;
+                break;
+            case OPT_TRACE_PERIOD: {
+                char *parsed = NULL;
+                bench_trace_period = strtoumax(optarg, &parsed, 0);
+                if (parsed == optarg) {
+                    fprintf(stderr, "error: invalid trace-period: %s\n", optarg);
+                    exit(-1);
+                }
+                break;
+            }
+            case OPT_TRACE_FREQ: {
+                char *parsed = NULL;
+                bench_trace_freq = strtoumax(optarg, &parsed, 0);
+                if (parsed == optarg) {
+                    fprintf(stderr, "error: invalid trace-freq: %s\n", optarg);
+                    exit(-1);
+                }
+                break;
+            }
             case OPT_READ_SLEEP: {
                 char *parsed = NULL;
                 double read_sleep = strtod(optarg, &parsed);

runners/test_runner.c

@@ -14,6 +14,8 @@
 #include <stdarg.h>
 #include <stdio.h>
 #include <unistd.h>
+#include <time.h>
+#include <execinfo.h>
 
 
 // some helpers
@@ -421,26 +423,63 @@ size_t test_step_step = 1;
 
 const char *test_disk_path = NULL;
 const char *test_trace_path = NULL;
+bool test_trace_backtrace = false;
+uint32_t test_trace_period = 0;
+uint32_t test_trace_freq = 0;
 FILE *test_trace_file = NULL;
 uint32_t test_trace_cycles = 0;
+uint64_t test_trace_time = 0;
+uint64_t test_trace_open_time = 0;
 lfs_emubd_sleep_t test_read_sleep = 0.0;
 lfs_emubd_sleep_t test_prog_sleep = 0.0;
 lfs_emubd_sleep_t test_erase_sleep = 0.0;
 
+// this determines both the backtrace buffer and the trace printf buffer, if
+// trace ends up interleaved or truncated this may need to be increased
+#ifndef TEST_TRACE_BACKTRACE_BUFFER_SIZE
+#define TEST_TRACE_BACKTRACE_BUFFER_SIZE 8192
+#endif
+void *test_trace_backtrace_buffer[
+    TEST_TRACE_BACKTRACE_BUFFER_SIZE / sizeof(void*)];
 
 // trace printing
 void test_trace(const char *fmt, ...) {
     if (test_trace_path) {
-        if (!test_trace_file) {
+        // sample at a specific period?
-            // Tracing output is heavy and trying to open every trace
+        if (test_trace_period) {
-            // call is slow, so we only try to open the trace file every
+            if (test_trace_cycles % test_trace_period != 0) {
-            // so often. Note this doesn't affect successfully opened files
-            if (test_trace_cycles % 128 != 0) {
                 test_trace_cycles += 1;
                 return;
             }
             test_trace_cycles += 1;
+        }
 
+        // sample at a specific frequency?
+        if (test_trace_freq) {
+            struct timespec t;
+            clock_gettime(CLOCK_MONOTONIC, &t);
+            uint64_t now = (uint64_t)t.tv_sec*1000*1000*1000
+                    + (uint64_t)t.tv_nsec;
+            if (now - test_trace_time < (1000*1000*1000) / test_trace_freq) {
+                return;
+            }
+            test_trace_time = now;
+        }
+
+        if (!test_trace_file) {
+            // Tracing output is heavy and trying to open every trace
+            // call is slow, so we only try to open the trace file every
+            // so often. Note this doesn't affect successfully opened files
+            struct timespec t;
+            clock_gettime(CLOCK_MONOTONIC, &t);
+            uint64_t now = (uint64_t)t.tv_sec*1000*1000*1000
+                    + (uint64_t)t.tv_nsec;
+            if (now - test_trace_open_time < 100*1000*1000) {
+                return;
+            }
+            test_trace_open_time = now;
+
+            // try to open the trace file
             int fd;
             if (strcmp(test_trace_path, "-") == 0) {
                 fd = dup(1);
@@ -461,19 +500,42 @@ void test_trace(const char *fmt, ...) {
 
             FILE *f = fdopen(fd, "a");
             assert(f);
-            int err = setvbuf(f, NULL, _IOLBF, BUFSIZ);
+            int err = setvbuf(f, NULL, _IOFBF,
+                    TEST_TRACE_BACKTRACE_BUFFER_SIZE);
             assert(!err);
             test_trace_file = f;
         }
 
+        // print trace
         va_list va;
         va_start(va, fmt);
         int res = vfprintf(test_trace_file, fmt, va);
+        va_end(va);
         if (res < 0) {
             fclose(test_trace_file);
             test_trace_file = NULL;
+            return;
         }
-        va_end(va);
+
+        if (test_trace_backtrace) {
+            // print backtrace
+            size_t count = backtrace(
+                    test_trace_backtrace_buffer,
+                    TEST_TRACE_BACKTRACE_BUFFER_SIZE);
+            // note we skip our own stack frame
+            for (size_t i = 1; i < count; i++) {
+                res = fprintf(test_trace_file, "\tat %p\n",
+                        test_trace_backtrace_buffer[i]);
+                if (res < 0) {
+                    fclose(test_trace_file);
+                    test_trace_file = NULL;
+                    return;
+                }
+            }
+        }
+
+        // flush immediately
+        fflush(test_trace_file);
     }
 }
 
@@ -1838,9 +1900,12 @@ enum opt_flags {
     OPT_STEP                     = 's',
     OPT_DISK                     = 'd',
     OPT_TRACE                    = 't',
-    OPT_READ_SLEEP               = 8,
+    OPT_TRACE_BACKTRACE          = 8,
-    OPT_PROG_SLEEP               = 9,
+    OPT_TRACE_PERIOD             = 9,
-    OPT_ERASE_SLEEP              = 10,
+    OPT_TRACE_FREQ               = 10,
+    OPT_READ_SLEEP               = 11,
+    OPT_PROG_SLEEP               = 12,
+    OPT_ERASE_SLEEP              = 13,
 };
 
 const char *short_opts = "hYlLD:G:P:s:d:t:";
@@ -1865,6 +1930,9 @@ const struct option long_opts[] = {
     {"step",             required_argument, NULL, OPT_STEP},
     {"disk",             required_argument, NULL, OPT_DISK},
     {"trace",            required_argument, NULL, OPT_TRACE},
+    {"trace-backtrace",  no_argument,       NULL, OPT_TRACE_BACKTRACE},
+    {"trace-period",     required_argument, NULL, OPT_TRACE_PERIOD},
+    {"trace-freq",       required_argument, NULL, OPT_TRACE_FREQ},
     {"read-sleep",       required_argument, NULL, OPT_READ_SLEEP},
     {"prog-sleep",       required_argument, NULL, OPT_PROG_SLEEP},
     {"erase-sleep",      required_argument, NULL, OPT_ERASE_SLEEP},
@@ -1887,8 +1955,11 @@ const char *const help_text[] = {
     "Comma-separated list of disk geometries to test.",
     "Comma-separated list of power-loss scenarios to test.",
     "Comma-separated range of test permutations to run (start,stop,step).",
-    "Redirect block device operations to this file.",
+    "Direct block device operations to this file.",
-    "Redirect trace output to this file.",
+    "Direct trace output to this file.",
+    "Include a backtrace with every trace statement.",
+    "Sample trace output at this period in cycles.",
+    "Sample trace output at this frequency in hz.",
     "Artificial read delay in seconds.",
     "Artificial prog delay in seconds.",
     "Artificial erase delay in seconds.",
@@ -2460,6 +2531,27 @@ step_unknown:
             case OPT_TRACE:
                 test_trace_path = optarg;
                 break;
+            case OPT_TRACE_BACKTRACE:
+                test_trace_backtrace = true;
+                break;
+            case OPT_TRACE_PERIOD: {
+                char *parsed = NULL;
+                test_trace_period = strtoumax(optarg, &parsed, 0);
+                if (parsed == optarg) {
+                    fprintf(stderr, "error: invalid trace-period: %s\n", optarg);
+                    exit(-1);
+                }
+                break;
+            }
+            case OPT_TRACE_FREQ: {
+                char *parsed = NULL;
+                test_trace_freq = strtoumax(optarg, &parsed, 0);
+                if (parsed == optarg) {
+                    fprintf(stderr, "error: invalid trace-freq: %s\n", optarg);
+                    exit(-1);
+                }
+                break;
+            }
             case OPT_READ_SLEEP: {
                 char *parsed = NULL;
                 double read_sleep = strtod(optarg, &parsed);

scripts/bench.py

@@ -35,19 +35,12 @@ VALGRIND_TOOL = ['valgrind']
 PERF_SCRIPT = ['./scripts/perf.py']
 
 
-def openio(path, mode='r', buffering=-1, nb=False):
+def openio(path, mode='r', buffering=-1):
     if path == '-':
         if mode == 'r':
-            return os.fdopen(os.dup(sys.stdin.fileno()), 'r', buffering)
+            return os.fdopen(os.dup(sys.stdin.fileno()), mode, buffering)
         else:
-            return os.fdopen(os.dup(sys.stdout.fileno()), 'w', buffering)
+            return os.fdopen(os.dup(sys.stdout.fileno()), mode, buffering)
-    elif nb and 'a' in mode:
-        return os.fdopen(os.open(
-                path,
-                os.O_WRONLY | os.O_CREAT | os.O_APPEND | os.O_NONBLOCK,
-                0o666),
-            mode,
-            buffering)
     else:
         return open(path, mode, buffering)
 
@@ -533,6 +526,12 @@ def find_runner(runner, **args):
         cmd.append('-d%s' % args['disk'])
     if args.get('trace'):
         cmd.append('-t%s' % args['trace'])
+    if args.get('trace_backtrace'):
+        cmd.append('--trace-backtrace')
+    if args.get('trace_period'):
+        cmd.append('--trace-period=%s' % args['trace_period'])
+    if args.get('trace_freq'):
+        cmd.append('--trace-freq=%s' % args['trace_freq'])
     if args.get('read_sleep'):
         cmd.append('--read-sleep=%s' % args['read_sleep'])
     if args.get('prog_sleep'):
@@ -747,7 +746,7 @@ class BenchFailure(Exception):
         self.stdout = stdout
         self.assert_ = assert_
 
-def run_stage(name, runner_, ids, output_, **args):
+def run_stage(name, runner_, ids, stdout_, trace_, output_, **args):
     # get expected suite/case/perm counts
     (case_suites,
         expected_suite_perms,
@@ -795,7 +794,6 @@ def run_stage(name, runner_, ids, output_, **args):
         os.close(spty)
         children.add(proc)
         mpty = os.fdopen(mpty, 'r', 1)
-        stdout = None
 
         last_id = None
         last_stdout = []
@@ -812,18 +810,12 @@ def run_stage(name, runner_, ids, output_, **args):
                 if not line:
                     break
                 last_stdout.append(line)
-                if args.get('stdout'):
+                if stdout_:
                     try:
-                        if not stdout:
+                        stdout_.write(line)
-                            stdout = openio(args['stdout'], 'a', 1, nb=True)
+                        stdout_.flush()
-                        stdout.write(line)
+                    except BrokenPipeError:
-                    except OSError as e:
+                        pass
-                        if e.errno not in [
-                                errno.ENXIO,
-                                errno.EPIPE,
-                                errno.EAGAIN]:
-                            raise
-                        stdout = None
                 if args.get('verbose'):
                     sys.stdout.write(line)
 
@@ -1061,6 +1053,8 @@ def run(runner, bench_ids=[], **args):
                 by or 'benches',
                 runner_,
                 [by] if by is not None else bench_ids,
+                stdout,
+                trace,
                 output,
                 **args)
         # collect passes/failures
@@ -1076,9 +1070,15 @@ def run(runner, bench_ids=[], **args):
     stop = time.time()
 
     if stdout:
+        try:
             stdout.close()
+        except BrokenPipeError:
+            pass
     if trace:
+        try:
             trace.close()
+        except BrokenPipeError:
+            pass
     if output:
         output.close()
 
@@ -1276,6 +1276,16 @@ if __name__ == "__main__":
     bench_parser.add_argument(
         '-t', '--trace',
         help="Direct trace output to this file.")
+    bench_parser.add_argument(
+        '--trace-backtrace',
+        action='store_true',
+        help="Include a backtrace with every trace statement.")
+    bench_parser.add_argument(
+        '--trace-period',
+        help="Sample trace output at this period in cycles.")
+    bench_parser.add_argument(
+        '--trace-freq',
+        help="Sample trace output at this frequency in hz.")
     bench_parser.add_argument(
         '-O', '--stdout',
         help="Direct stdout to this file. Note stderr is already merged here.")
@@ -1353,7 +1363,7 @@ if __name__ == "__main__":
         default=VALGRIND_TOOL,
         help="Path to Valgrind tool to use. Defaults to %r." % VALGRIND_TOOL)
     bench_parser.add_argument(
-        '--perf',
+        '-p', '--perf',
         help="Run under Linux's perf to sample performance counters, writing "
             "samples to this file.")
     bench_parser.add_argument(

scripts/code.py

@@ -15,7 +15,6 @@
 import collections as co
 import csv
 import difflib
-import glob
 import itertools as it
 import math as m
 import os
@@ -24,7 +23,6 @@ import shlex
 import subprocess as sp
 
 
-OBJ_PATHS = ['*.o']
 NM_TOOL = ['nm']
 NM_TYPES = 'tTrRdD'
 OBJDUMP_TOOL = ['objdump']
@@ -126,16 +124,16 @@ class CodeResult(co.namedtuple('CodeResult', [
             self.size + other.size)
 
 
-def openio(path, mode='r'):
+def openio(path, mode='r', buffering=-1):
     if path == '-':
         if mode == 'r':
-            return os.fdopen(os.dup(sys.stdin.fileno()), 'r')
+            return os.fdopen(os.dup(sys.stdin.fileno()), mode, buffering)
         else:
-            return os.fdopen(os.dup(sys.stdout.fileno()), 'w')
+            return os.fdopen(os.dup(sys.stdout.fileno()), mode, buffering)
     else:
-        return open(path, mode)
+        return open(path, mode, buffering)
 
-def collect(paths, *,
+def collect(obj_paths, *,
         nm_tool=NM_TOOL,
         nm_types=NM_TYPES,
         objdump_tool=OBJDUMP_TOOL,
@@ -147,17 +145,17 @@ def collect(paths, *,
         ' (?P<type>[%s])' % re.escape(nm_types) +
         ' (?P<func>.+?)$')
     line_pattern = re.compile(
-        '^\s+(?P<no>[0-9]+)\s+'
+        '^\s+(?P<no>[0-9]+)'
-            '(?:(?P<dir>[0-9]+)\s+)?'
+            '(?:\s+(?P<dir>[0-9]+))?'
-            '.*\s+'
+            '\s+.*'
-            '(?P<path>[^\s]+)$')
+            '\s+(?P<path>[^\s]+)$')
     info_pattern = re.compile(
         '^(?:.*(?P<tag>DW_TAG_[a-z_]+).*'
-            '|^.*DW_AT_name.*:\s*(?P<name>[^:\s]+)\s*'
+            '|.*DW_AT_name.*:\s*(?P<name>[^:\s]+)\s*'
-            '|^.*DW_AT_decl_file.*:\s*(?P<file>[0-9]+)\s*)$')
+            '|.*DW_AT_decl_file.*:\s*(?P<file>[0-9]+)\s*)$')
 
     results = []
-    for path in paths:
+    for path in obj_paths:
         # guess the source, if we have debug-info we'll replace this later
         file = re.sub('(\.o)?$', '.c', path, 1)
 
@@ -520,20 +518,7 @@ def main(obj_paths, *,
         **args):
     # find sizes
     if not args.get('use', None):
-        # find .o files
+        results = collect(obj_paths, **args)
-        paths = []
-        for path in obj_paths:
-            if os.path.isdir(path):
-                path = path + '/*.o'
-
-            for path in glob.glob(path):
-                paths.append(path)
-
-        if not paths:
-            print("error: no .o files found in %r?" % obj_paths)
-            sys.exit(-1)
-
-        results = collect(paths, **args)
     else:
         results = []
         with openio(args['use']) as f:
@@ -613,9 +598,7 @@ if __name__ == "__main__":
     parser.add_argument(
         'obj_paths',
         nargs='*',
-        default=OBJ_PATHS,
+        help="Input *.o files.")
-        help="Description of where to find *.o files. May be a directory "
-            "or a list of paths. Defaults to %r." % OBJ_PATHS)
     parser.add_argument(
         '-v', '--verbose',
         action='store_true',

scripts/cov.py

@@ -14,7 +14,6 @@
 
 import collections as co
 import csv
-import glob
 import itertools as it
 import json
 import math as m
@@ -26,8 +25,6 @@ import subprocess as sp
 # TODO use explode_asserts to avoid counting assert branches?
 # TODO use dwarf=info to find functions for inline functions?
 
-
-GCDA_PATHS = ['*.gcda']
 GCOV_TOOL = ['gcov']
 
 
@@ -202,14 +199,14 @@ class CovResult(co.namedtuple('CovResult', [
             self.branches + other.branches)
 
 
-def openio(path, mode='r'):
+def openio(path, mode='r', buffering=-1):
     if path == '-':
         if mode == 'r':
-            return os.fdopen(os.dup(sys.stdin.fileno()), 'r')
+            return os.fdopen(os.dup(sys.stdin.fileno()), mode, buffering)
         else:
-            return os.fdopen(os.dup(sys.stdout.fileno()), 'w')
+            return os.fdopen(os.dup(sys.stdout.fileno()), mode, buffering)
     else:
-        return open(path, mode)
+        return open(path, mode, buffering)
 
 def collect(gcda_paths, *,
         gcov_tool=GCOV_TOOL,
@@ -592,20 +589,7 @@ def main(gcda_paths, *,
 
     # find sizes
     if not args.get('use', None):
-        # find .gcda files
+        results = collect(gcda_paths, **args)
-        paths = []
-        for path in gcda_paths:
-            if os.path.isdir(path):
-                path = path + '/*.gcda'
-
-            for path in glob.glob(path):
-                paths.append(path)
-
-        if not paths:
-            print("error: no .gcda files found in %r?" % gcda_paths)
-            sys.exit(-1)
-
-        results = collect(paths, **args)
     else:
         results = []
         with openio(args['use']) as f:
@@ -707,9 +691,7 @@ if __name__ == "__main__":
     parser.add_argument(
         'gcda_paths',
         nargs='*',
-        default=GCDA_PATHS,
+        help="Input *.gcda files.")
-        help="Description of where to find *.gcda files. May be a directory "
-            "or a list of paths. Defaults to %r." % GCDA_PATHS)
     parser.add_argument(
         '-v', '--verbose',
         action='store_true',

scripts/data.py

@@ -15,7 +15,6 @@
 import collections as co
 import csv
 import difflib
-import glob
 import itertools as it
 import math as m
 import os
@@ -24,7 +23,6 @@ import shlex
 import subprocess as sp
 
 
-OBJ_PATHS = ['*.o']
 NM_TOOL = ['nm']
 NM_TYPES = 'dDbB'
 OBJDUMP_TOOL = ['objdump']
@@ -126,16 +124,16 @@ class DataResult(co.namedtuple('DataResult', [
             self.size + other.size)
 
 
-def openio(path, mode='r'):
+def openio(path, mode='r', buffering=-1):
     if path == '-':
         if mode == 'r':
-            return os.fdopen(os.dup(sys.stdin.fileno()), 'r')
+            return os.fdopen(os.dup(sys.stdin.fileno()), mode, buffering)
         else:
-            return os.fdopen(os.dup(sys.stdout.fileno()), 'w')
+            return os.fdopen(os.dup(sys.stdout.fileno()), mode, buffering)
     else:
-        return open(path, mode)
+        return open(path, mode, buffering)
 
-def collect(paths, *,
+def collect(obj_paths, *,
         nm_tool=NM_TOOL,
         nm_types=NM_TYPES,
         objdump_tool=OBJDUMP_TOOL,
@@ -147,17 +145,17 @@ def collect(paths, *,
         ' (?P<type>[%s])' % re.escape(nm_types) +
         ' (?P<func>.+?)$')
     line_pattern = re.compile(
-        '^\s+(?P<no>[0-9]+)\s+'
+        '^\s+(?P<no>[0-9]+)'
-            '(?:(?P<dir>[0-9]+)\s+)?'
+            '(?:\s+(?P<dir>[0-9]+))?'
-            '.*\s+'
+            '\s+.*'
-            '(?P<path>[^\s]+)$')
+            '\s+(?P<path>[^\s]+)$')
     info_pattern = re.compile(
         '^(?:.*(?P<tag>DW_TAG_[a-z_]+).*'
-            '|^.*DW_AT_name.*:\s*(?P<name>[^:\s]+)\s*'
+            '|.*DW_AT_name.*:\s*(?P<name>[^:\s]+)\s*'
-            '|^.*DW_AT_decl_file.*:\s*(?P<file>[0-9]+)\s*)$')
+            '|.*DW_AT_decl_file.*:\s*(?P<file>[0-9]+)\s*)$')
 
     results = []
-    for path in paths:
+    for path in obj_paths:
         # guess the source, if we have debug-info we'll replace this later
         file = re.sub('(\.o)?$', '.c', path, 1)
 
@@ -520,20 +518,7 @@ def main(obj_paths, *,
         **args):
     # find sizes
     if not args.get('use', None):
-        # find .o files
+        results = collect(obj_paths, **args)
-        paths = []
-        for path in obj_paths:
-            if os.path.isdir(path):
-                path = path + '/*.o'
-
-            for path in glob.glob(path):
-                paths.append(path)
-
-        if not paths:
-            print("error: no .o files found in %r?" % obj_paths)
-            sys.exit(-1)
-
-        results = collect(paths, **args)
     else:
         results = []
         with openio(args['use']) as f:
@@ -613,9 +598,7 @@ if __name__ == "__main__":
     parser.add_argument(
         'obj_paths',
         nargs='*',
-        default=OBJ_PATHS,
+        help="Input *.o files.")
-        help="Description of where to find *.o files. May be a directory "
-            "or a list of paths. Defaults to %r." % OBJ_PATHS)
     parser.add_argument(
         '-v', '--verbose',
         action='store_true',

scripts/perf.py

@@ -4,7 +4,7 @@
 #
 # Example:
 # ./scripts/perf.py -R -obench.perf ./runners/bench_runner
-# ./scripts/perf.py bench.perf -Flfs.c -Flfs_util.c -Scycles
+# ./scripts/perf.py bench.perf -j -Flfs.c -Flfs_util.c -Scycles
 #
 # Copyright (c) 2022, The littlefs authors.
 # SPDX-License-Identifier: BSD-3-Clause
@@ -16,7 +16,6 @@ import csv
 import errno
 import fcntl
 import functools as ft
-import glob
 import itertools as it
 import math as m
 import multiprocessing as mp
@@ -31,7 +30,6 @@ import zipfile
 # TODO support non-zip perf results?
 
 
-PERF_PATHS = ['*.perf']
 PERF_TOOL = ['perf']
 PERF_EVENTS = 'cycles,branch-misses,branches,cache-misses,cache-references'
 PERF_FREQ = 100
@@ -147,14 +145,14 @@ class PerfResult(co.namedtuple('PerfResult', [
             self.children + other.children)
 
 
-def openio(path, mode='r'):
+def openio(path, mode='r', buffering=-1):
     if path == '-':
         if mode == 'r':
-            return os.fdopen(os.dup(sys.stdin.fileno()), 'r')
+            return os.fdopen(os.dup(sys.stdin.fileno()), mode, buffering)
         else:
-            return os.fdopen(os.dup(sys.stdout.fileno()), 'w')
+            return os.fdopen(os.dup(sys.stdout.fileno()), mode, buffering)
     else:
-        return open(path, mode)
+        return open(path, mode, buffering)
 
 # run perf as a subprocess, storing measurements into a zip file
 def record(command, *,
@@ -164,14 +162,6 @@ def record(command, *,
         perf_events=PERF_EVENTS,
         perf_tool=PERF_TOOL,
         **args):
-    if not command:
-        print('error: no command specified?')
-        sys.exit(-1)
-
-    if not output:
-        print('error: no output file specified?')
-        sys.exit(-1)
-
     # create a temporary file for perf to write to, as far as I can tell
     # this is strictly needed because perf's pipe-mode only works with stdout
     with tempfile.NamedTemporaryFile('rb') as f:
@@ -214,148 +204,49 @@ def record(command, *,
     return err
 
 
-def collect_decompressed(path, *,
+# try to only process each dso onceS
-        perf_tool=PERF_TOOL,
+#
-        everything=False,
+# note this only caches with the non-keyword arguments
-        propagate=0,
+def multiprocessing_cache(f):
-        depth=1,
+    local_cache = {}
-        **args):
+    manager = mp.Manager()
-    sample_pattern = re.compile(
+    global_cache = manager.dict()
-        '(?P<comm>\w+)'
+    lock = mp.Lock()
-        '\s+(?P<pid>\w+)'
-        '\s+(?P<time>[\w.]+):'
-        '\s*(?P<period>\w+)'
-        '\s+(?P<event>[^:]+):')
-    frame_pattern = re.compile(
-        '\s+(?P<addr>\w+)'
-        '\s+(?P<sym>[^\s]+)'
-        '\s+\((?P<dso>[^\)]+)\)')
-    events = {
-        'cycles':           'cycles',
-        'branch-misses':    'bmisses',
-        'branches':         'branches',
-        'cache-misses':     'cmisses',
-        'cache-references': 'caches'}
 
-    # note perf_tool may contain extra args
+    def multiprocessing_cache(*args, **kwargs):
-    cmd = perf_tool + [
+        # check local cache?
-        'script',
+        if args in local_cache:
-        '-i%s' % path]
+            return local_cache[args]
-    if args.get('verbose'):
+        # check global cache?
-        print(' '.join(shlex.quote(c) for c in cmd))
+        with lock:
-    proc = sp.Popen(cmd,
+            if args in global_cache:
-        stdout=sp.PIPE,
+                v = global_cache[args]
-        stderr=sp.PIPE if not args.get('verbose') else None,
+                local_cache[args] = v
-        universal_newlines=True,
+                return v
-        errors='replace',
+            # fall back to calling the function
-        close_fds=False)
+            v = f(*args, **kwargs)
+            global_cache[args] = v
+            local_cache[args] = v
+            return v
 
-    last_filtered = False
+    return multiprocessing_cache
-    last_event = ''
-    last_period = 0
-    last_stack = []
-    results = {}
 
-    def commit():
+@multiprocessing_cache
-        # tail-recursively propagate measurements
+def collect_syms_and_lines(obj_path, *,
-        for i in range(len(last_stack)):
-            results_ = results
-            for j in reversed(range(i+1)):
-                if i+1-j > depth:
-                    break
-
-                # propagate
-                name = last_stack[j]
-                if name not in results_:
-                    results_[name] = (co.defaultdict(lambda: 0), {})
-                results_[name][0][last_event] += last_period
-
-                # recurse
-                results_ = results_[name][1]
-
-    for line in proc.stdout:
-        # we need to process a lot of data, so wait to use regex as late
-        # as possible
-        if not line:
-            continue
-        if not line.startswith('\t'):
-            m = sample_pattern.match(line)
-            if m:
-                if last_stack:
-                    commit()
-                last_event = m.group('event')
-                last_filtered = last_event in events
-                last_period = int(m.group('period'), 0)
-                last_stack = []
-        elif last_filtered:
-            m = frame_pattern.match(line)
-            if m:
-                # filter out internal/kernel functions
-                if not everything and (
-                        m.group('sym').startswith('__')
-                        or m.group('dso').startswith('/usr/lib')
-                        or not m.group('sym')[:1].isalpha()):
-                    continue
-
-                last_stack.append((
-                    m.group('dso'),
-                    m.group('sym'),
-                    int(m.group('addr'), 16)))
-
-                # stop propogating?
-                if propagate and len(last_stack) >= propagate:
-                    last_filtered = False
-    if last_stack:
-        commit()
-
-    proc.wait()
-    if proc.returncode != 0:
-        if not args.get('verbose'):
-            for line in proc.stderr:
-                sys.stdout.write(line)
-        sys.exit(-1)
-
-    # rearrange results into result type
-    def to_results(results):
-        results_ = []
-        for name, (r, children) in results.items():
-            results_.append(PerfResult(*name,
-                **{events[k]: v for k, v in r.items()},
-                children=to_results(children)))
-        return results_
-
-    return to_results(results)
-
-def collect_job(path, i, **args):
-    # decompress into a temporary file, this is to work around
-    # some limitations of perf
-    with zipfile.ZipFile(path) as z:
-        with z.open(i) as f:
-            with tempfile.NamedTemporaryFile('wb') as g:
-                shutil.copyfileobj(f, g)
-                g.flush()
-
-                return collect_decompressed(g.name, **args)
-
-def starapply(args):
-    f, args, kwargs = args
-    return f(*args, **kwargs)
-
-def collect(paths, *,
-        jobs=None,
         objdump_tool=None,
-        sources=None,
-        everything=False,
         **args):
     symbol_pattern = re.compile(
-        '^(?P<addr>[0-9a-fA-F]+)\s.*\s(?P<name>[^\s]+)\s*$')
+        '^(?P<addr>[0-9a-fA-F]+)'
+            '\s+.*'
+            '\s+(?P<size>[0-9a-fA-F]+)'
+            '\s+(?P<name>[^\s]+)\s*$')
     line_pattern = re.compile(
         '^\s+(?:'
             # matches dir/file table
-            '(?P<no>[0-9]+)\s+'
+            '(?P<no>[0-9]+)'
-                '(?:(?P<dir>[0-9]+)\s+)?'
+                '(?:\s+(?P<dir>[0-9]+))?'
-                '.*\s+'
+                '\s+.*'
-                '(?P<path>[^\s]+)'
+                '\s+(?P<path>[^\s]+)'
             # matches line opcodes
             '|' '\[[^\]]*\]\s+'
                 '(?:'
@@ -368,41 +259,10 @@ def collect(paths, *,
                     '|' '.' ')*'
             ')$', re.IGNORECASE)
 
-    records = []
+    # figure out symbol addresses and file+line ranges
-    for path in paths:
-        # each .perf file is actually a zip file containing perf files from
-        # multiple runs
-        with zipfile.ZipFile(path) as z:
-            records.extend((path, i) for i in z.infolist())
-
-    # we're dealing with a lot of data but also surprisingly
-    # parallelizable
-    dsos = {}
-    results = []
-    with mp.Pool(jobs or len(os.sched_getaffinity(0))) as p:
-        for results_ in p.imap_unordered(
-                starapply,
-                ((collect_job, (path, i), dict(
-                    everything=everything,
-                    **args))
-                    for path, i in records)):
-
-            # organize by dso
-            results__ = {}
-            for r in results_:
-                if r.file not in results__:
-                    results__[r.file] = []
-                results__[r.file].append(r)
-            results_ = results__
-
-            for dso, results_ in results_.items():
-                if dso not in dsos:
-                    # find file+line ranges for dsos
-                    #
-                    # do this here so we only process each dso once
     syms = {}
     sym_at = []
-                    cmd = objdump_tool + ['-t', dso]
+    cmd = objdump_tool + ['-t', obj_path]
     if args.get('verbose'):
         print(' '.join(shlex.quote(c) for c in cmd))
     proc = sp.Popen(cmd,
@@ -416,11 +276,15 @@ def collect(paths, *,
         if m:
             name = m.group('name')
             addr = int(m.group('addr'), 16)
+            size = int(m.group('size'), 16)
+            # ignore zero-sized symbols
+            if not size:
+                continue
             # note multiple symbols can share a name
             if name not in syms:
                 syms[name] = set()
-                            syms[name].add(addr)
+            syms[name].add((addr, size))
-                            sym_at.append((addr, name))
+            sym_at.append((addr, name, size))
     proc.wait()
     if proc.returncode != 0:
         if not args.get('verbose'):
@@ -429,24 +293,25 @@ def collect(paths, *,
         # assume no debug-info on failure
         pass
 
-                    # sort and keep first when duplicates
+    # sort and keep largest/first when duplicates
-                    sym_at.sort()
+    sym_at.sort(key=lambda x: (x[0], -x[2], x[1]))
     sym_at_ = []
-                    for addr, name in sym_at:
+    for addr, name, size in sym_at:
         if len(sym_at_) == 0 or sym_at_[-1][0] != addr:
-                            sym_at_.append((addr, name))
+            sym_at_.append((addr, name, size))
     sym_at = sym_at_
 
     # state machine for dwarf line numbers, note that objdump's
     # decodedline seems to have issues with multiple dir/file
     # tables, which is why we need this
+    lines = []
     line_at = []
     dirs = {}
     files = {}
     op_file = 1
     op_line = 1
     op_addr = 0
-                    cmd = objdump_tool + ['--dwarf=rawline', dso]
+    cmd = objdump_tool + ['--dwarf=rawline', obj_path]
     if args.get('verbose'):
         print(' '.join(shlex.quote(c) for c in cmd))
     proc = sp.Popen(cmd,
@@ -482,10 +347,9 @@ def collect(paths, *,
                 if (m.group('op_special')
                         or m.group('op_copy')
                         or m.group('op_end')):
-                                    line_at.append((
+                    file = os.path.abspath(files.get(op_file, '?'))
-                                        op_addr,
+                    lines.append((file, op_line, op_addr))
-                                        files.get(op_file, '?'),
+                    line_at.append((op_addr, file, op_line))
-                                        op_line))
 
                 if m.group('op_end'):
                     op_file = 1
@@ -500,8 +364,14 @@ def collect(paths, *,
         pass
 
     # sort and keep first when duplicates
-                    #
+    lines.sort()
-                    # I think dwarf requires this to be sorted but just in case
+    lines_ = []
+    for file, line, addr in lines:
+        if len(lines_) == 0 or lines_[-1][0] != file or lines[-1][1] != line:
+            lines_.append((file, line, addr))
+    lines = lines_
+
+    # sort and keep first when duplicates
     line_at.sort()
     line_at_ = []
     for addr, file, line in line_at:
@@ -509,76 +379,165 @@ def collect(paths, *,
             line_at_.append((addr, file, line))
     line_at = line_at_
 
-                    # discard lines outside of the range of the containing
+    return syms, sym_at, lines, line_at
-                    # function, these are introduced by dwarf for inlined
-                    # functions but don't map to elf-level symbols
-                    sym_at_ = []
-                    for addr, sym in sym_at:
-                        i = bisect.bisect(line_at, addr, key=lambda x: x[0])
-                        if i > 0:
-                            _, file, line = line_at[i-1]
-                            sym_at_.append((file, line, sym))
-                    sym_at_.sort()
 
-                    line_at_ = []
-                    for addr, file, line in line_at:
-                        # only keep if sym-at-addr and sym-at-line match
-                        i = bisect.bisect(
-                            sym_at, addr, key=lambda x: x[0])
-                        j = bisect.bisect(
-                            sym_at_, (file, line), key=lambda x: (x[0], x[1]))
-                        if i > 0 and j > 0 and (
-                                sym_at[i-1][1] == sym_at_[j-1][2]):
-                            line_at_.append((addr, file, line))
-                    line_at = line_at_
 
-                    dsos[dso] = (syms, sym_at, line_at)
+def collect_decompressed(path, *,
+        perf_tool=PERF_TOOL,
+        sources=None,
+        everything=False,
+        propagate=0,
+        depth=1,
+        **args):
+    sample_pattern = re.compile(
+        '(?P<comm>\w+)'
+        '\s+(?P<pid>\w+)'
+        '\s+(?P<time>[\w.]+):'
+        '\s*(?P<period>\w+)'
+        '\s+(?P<event>[^:]+):')
+    frame_pattern = re.compile(
+        '\s+(?P<addr>\w+)'
+        '\s+(?P<sym>[^\s\+]+)(?:\+(?P<off>\w+))?'
+        '\s+\((?P<dso>[^\)]+)\)')
+    events = {
+        'cycles':           'cycles',
+        'branch-misses':    'bmisses',
+        'branches':         'branches',
+        'cache-misses':     'cmisses',
+        'cache-references': 'caches'}
 
-                syms, _, line_at = dsos[dso]
+    # note perf_tool may contain extra args
+    cmd = perf_tool + [
+        'script',
+        '-i%s' % path]
+    if args.get('verbose'):
+        print(' '.join(shlex.quote(c) for c in cmd))
+    proc = sp.Popen(cmd,
+        stdout=sp.PIPE,
+        stderr=sp.PIPE if not args.get('verbose') else None,
+        universal_newlines=True,
+        errors='replace',
+        close_fds=False)
 
-                # first try to reverse ASLR
+    last_filtered = False
-                def deltas(r, d):
+    last_event = ''
-                    if '+' in r.function:
+    last_period = 0
-                        sym, off = r.function.split('+', 1)
+    last_stack = []
-                        off = int(off, 0)
+    deltas = co.defaultdict(lambda: {})
-                    else:
+    syms_ = co.defaultdict(lambda: {})
-                        sym, off = r.function, 0
+    at_cache = {}
-                    addr = r.line - off + d
+    results = {}
 
-                    for addr_ in syms.get(sym, []):
+    def commit():
-                        yield addr_ - addr
+        # tail-recursively propagate measurements
+        for i in range(len(last_stack)):
+            results_ = results
+            for j in reversed(range(i+1)):
+                if i+1-j > depth:
+                    break
 
-                delta = min(
+                # propagate
-                    it.chain.from_iterable(
+                name = last_stack[j]
-                        deltas(r, 0) for r in results_),
+                if name not in results_:
-                    key=lambda d: sum(it.chain.from_iterable(
+                    results_[name] = (co.defaultdict(lambda: 0), {})
-                        deltas(r, d) for r in results_)),
+                results_[name][0][last_event] += last_period
-                    default=0)
 
-                # then try to map addrs -> file+line
+                # recurse
+                results_ = results_[name][1]
+
+    for line in proc.stdout:
+        # we need to process a lot of data, so wait to use regex as late
+        # as possible
+        if not line.startswith('\t'):
+            if last_filtered:
+                commit()
+            last_filtered = False
+
+            if line:
+                m = sample_pattern.match(line)
+                if m and m.group('event') in events:
+                    last_filtered = True
+                    last_event = m.group('event')
+                    last_period = int(m.group('period'), 0)
+                    last_stack = []
+
+        elif last_filtered:
+            m = frame_pattern.match(line)
+            if m:
+                # filter out internal/kernel functions
+                if not everything and (
+                        m.group('sym').startswith('__')
+                        or m.group('sym').startswith('0')
+                        or m.group('sym').startswith('-')
+                        or m.group('sym').startswith('[')
+                        or m.group('dso').startswith('/usr/lib')):
+                    continue
+
+                dso = m.group('dso')
+                sym = m.group('sym')
+                off = int(m.group('off'), 0) if m.group('off') else 0
+                addr_ = int(m.group('addr'), 16)
+
+                # get the syms/lines for the dso, this is cached
+                syms, sym_at, lines, line_at = collect_syms_and_lines(
+                    dso,
+                    **args)
+
+                # ASLR is tricky, we have symbols+offsets, but static symbols
+                # means we may have multiple options for each symbol.
                 #
-                # note we need to do this recursively
+                # To try to solve this, we use previous seen symbols to build
-                def remap(results):
+                # confidence for the correct ASLR delta. This means we may
-                    results_ = []
+                # guess incorrectly for early symbols, but this will only affect
-                    for r in results:
+                # a few samples.
-                        addr = r.line + delta
+                if sym in syms:
+                    sym_addr_ = addr_ - off
+
+                    # track possible deltas?
+                    for sym_addr, size in syms[sym]:
+                        delta = sym_addr - sym_addr_
+                        if delta not in deltas[dso]:
+                            deltas[dso][delta] = sum(
+                                abs(a_+delta - a)
+                                for s, (a_, _) in syms_[dso].items()
+                                for a, _ in syms[s])
+                    for delta in deltas[dso].keys():
+                        deltas[dso][delta] += abs(sym_addr_+delta - sym_addr)
+                    syms_[dso][sym] = sym_addr_, size
+
+                    # guess the best delta
+                    delta, _ = min(deltas[dso].items(),
+                        key=lambda x: (x[1], x[0]))
+                    addr = addr_ + delta
+
+                    # cached?
+                    if (dso,addr) in at_cache:
+                        cached = at_cache[(dso,addr)]
+                        if cached is None:
+                            # cache says to skip
+                            continue
+                        file, line = cached
+                    else:
+                        # find file+line
                         i = bisect.bisect(line_at, addr, key=lambda x: x[0])
                         if i > 0:
                             _, file, line = line_at[i-1]
                         else:
-                            file, line = re.sub('(\.o)?$', '.c', r.file, 1), 0
+                            file, line = re.sub('(\.o)?$', '.c', dso, 1), 0
 
                         # ignore filtered sources
                         if sources is not None:
                             if not any(
                                     os.path.abspath(file) == os.path.abspath(s)
                                     for s in sources):
+                                at_cache[(dso,addr)] = None
                                 continue
                         else:
                             # default to only cwd
                             if not everything and not os.path.commonpath([
                                     os.getcwd(),
                                     os.path.abspath(file)]) == os.getcwd():
+                                at_cache[(dso,addr)] = None
                                 continue
 
                         # simplify path
@@ -589,13 +548,79 @@ def collect(paths, *,
                         else:
                             file = os.path.abspath(file)
 
-                        function, *_ = r.function.split('+', 1)
+                        at_cache[(dso,addr)] = file, line
-                        results_.append(r._replace(
+                else:
-                            file=file, function=function, line=line,
+                    file, line = re.sub('(\.o)?$', '.c', dso, 1), 0
-                            children=remap(r.children)))
+
+                last_stack.append((file, sym, line))
+
+                # stop propogating?
+                if propagate and len(last_stack) >= propagate:
+                    commit()
+                    last_filtered = False
+    if last_filtered:
+        commit()
+
+    proc.wait()
+    if proc.returncode != 0:
+        if not args.get('verbose'):
+            for line in proc.stderr:
+                sys.stdout.write(line)
+        sys.exit(-1)
+
+    # rearrange results into result type
+    def to_results(results):
+        results_ = []
+        for name, (r, children) in results.items():
+            results_.append(PerfResult(*name,
+                **{events[k]: v for k, v in r.items()},
+                children=to_results(children)))
         return results_
 
-                results.extend(remap(results_))
+    return to_results(results)
+
+def collect_job(path, i, **args):
+    # decompress into a temporary file, this is to work around
+    # some limitations of perf
+    with zipfile.ZipFile(path) as z:
+        with z.open(i) as f:
+            with tempfile.NamedTemporaryFile('wb') as g:
+                shutil.copyfileobj(f, g)
+                g.flush()
+
+                return collect_decompressed(g.name, **args)
+
+def starapply(args):
+    f, args, kwargs = args
+    return f(*args, **kwargs)
+
+def collect(perf_paths, *,
+        jobs=None,
+        **args):
+    # automatic job detection?
+    if jobs == 0:
+        jobs = len(os.sched_getaffinity(0))
+
+    records = []
+    for path in perf_paths:
+        # each .perf file is actually a zip file containing perf files from
+        # multiple runs
+        with zipfile.ZipFile(path) as z:
+            records.extend((path, i) for i in z.infolist())
+
+    # we're dealing with a lot of data but also surprisingly
+    # parallelizable
+    if jobs is not None:
+        results = []
+        with mp.Pool(jobs) as p:
+            for results_ in p.imap_unordered(
+                    starapply,
+                    ((collect_job, (path, i), args) for path, i in records)):
+                results.extend(results_)
+    else:
+        results = []
+        for path, i in records:
+            results.extend(collect_job(path, i, **args))
 
     return results
 
@@ -983,7 +1008,6 @@ def report(perf_paths, *,
         fields=None,
         defines=None,
         sort=None,
-        self=False,
         branches=False,
         caches=False,
         **args):
@@ -1001,20 +1025,7 @@ def report(perf_paths, *,
 
     # find sizes
     if not args.get('use', None):
-        # find .o files
+        results = collect(perf_paths, **args)
-        paths = []
-        for path in perf_paths:
-            if os.path.isdir(path):
-                path = path + '/*.perf'
-
-            for path in glob.glob(path):
-                paths.append(path)
-
-        if not paths:
-            print("error: no .perf files found in %r?" % perf_paths)
-            sys.exit(-1)
-
-        results = collect(paths, **args)
     else:
         results = []
         with openio(args['use']) as f:
@@ -1124,8 +1135,7 @@ if __name__ == "__main__":
     parser.add_argument(
         'perf_paths',
         nargs=nargs,
-        help="Description of where to find *.perf files. May be a directory "
+        help="Input *.perf files.")
-            "or a list of paths. Defaults to %r." % PERF_PATHS)
     parser.add_argument(
         '-v', '--verbose',
         action='store_true',
@@ -1224,7 +1234,7 @@ if __name__ == "__main__":
         nargs='?',
         type=lambda x: tuple(float(x) for x in x.split(',')),
         const=THRESHOLD,
-        help="Show lines wth samples above this threshold as a percent of "
+        help="Show lines with samples above this threshold as a percent of "
             "all lines. Defaults to %s." % ','.join(str(t) for t in THRESHOLD))
     parser.add_argument(
         '-c', '--context',
@@ -1295,7 +1305,13 @@ if __name__ == "__main__":
 
     # perf_paths/command overlap, so need to do some munging here
     args.command = args.perf_paths
-    args.perf_paths = args.perf_paths or PERF_PATHS
+    if args.record:
+        if not args.command:
+            print('error: no command specified?')
+            sys.exit(-1)
+        if not args.output:
+            print('error: no output file specified?')
+            sys.exit(-1)
 
     sys.exit(main(**{k: v
         for k, v in vars(args).items()

scripts/plot.py

@@ -11,7 +11,6 @@
 
 import collections as co
 import csv
-import glob
 import io
 import itertools as it
 import math as m
@@ -19,7 +18,6 @@ import os
 import shutil
 import time
 
-CSV_PATHS = ['*.csv']
 COLORS = [
     '1;34', # bold blue
     '1;31', # bold red
@@ -80,14 +78,14 @@ def si(x, w=4):
         s = s.rstrip('.')
     return '%s%s%s' % ('-' if x < 0 else '', s, SI_PREFIXES[p])
 
-def openio(path, mode='r'):
+def openio(path, mode='r', buffering=-1):
     if path == '-':
         if mode == 'r':
-            return os.fdopen(os.dup(sys.stdin.fileno()), 'r')
+            return os.fdopen(os.dup(sys.stdin.fileno()), mode, buffering)
         else:
-            return os.fdopen(os.dup(sys.stdout.fileno()), 'w')
+            return os.fdopen(os.dup(sys.stdout.fileno()), mode, buffering)
     else:
-        return open(path, mode)
+        return open(path, mode, buffering)
 
 class LinesIO:
     def __init__(self, maxlen=None):
@@ -356,16 +354,8 @@ class Plot:
 
 def collect(csv_paths, renames=[]):
     # collect results from CSV files
-    paths = []
-    for path in csv_paths:
-        if os.path.isdir(path):
-            path = path + '/*.csv'
-
-        for path in glob.glob(path):
-            paths.append(path)
-
     results = []
-    for path in paths:
+    for path in csv_paths:
         try:
             with openio(path) as f:
                 reader = csv.DictReader(f, restval='')
@@ -732,9 +722,7 @@ if __name__ == "__main__":
     parser.add_argument(
         'csv_paths',
         nargs='*',
-        default=CSV_PATHS,
+        help="Input *.csv files.")
-        help="Description of where to find *.csv files. May be a directory "
-            "or list of paths. Defaults to %r." % CSV_PATHS)
     parser.add_argument(
         '-b', '--by',
         action='append',

scripts/prettyasserts.py

@@ -42,14 +42,14 @@ LEXEMES = {
 }
 
 
-def openio(path, mode='r'):
+def openio(path, mode='r', buffering=-1):
     if path == '-':
         if mode == 'r':
-            return os.fdopen(os.dup(sys.stdin.fileno()), 'r')
+            return os.fdopen(os.dup(sys.stdin.fileno()), mode, buffering)
         else:
-            return os.fdopen(os.dup(sys.stdout.fileno()), 'w')
+            return os.fdopen(os.dup(sys.stdout.fileno()), mode, buffering)
     else:
-        return open(path, mode)
+        return open(path, mode, buffering)
 
 def write_header(f, limit=LIMIT):
     f.writeln("// Generated by %s:" % sys.argv[0])

scripts/stack.py

@@ -12,15 +12,12 @@
 
 import collections as co
 import csv
-import glob
 import itertools as it
 import math as m
 import os
 import re
 
 
-CI_PATHS = ['*.ci']
-
 
 # integer fields
 class Int(co.namedtuple('Int', 'x')):
@@ -121,16 +118,16 @@ class StackResult(co.namedtuple('StackResult', [
             self.children | other.children)
 
 
-def openio(path, mode='r'):
+def openio(path, mode='r', buffering=-1):
     if path == '-':
         if mode == 'r':
-            return os.fdopen(os.dup(sys.stdin.fileno()), 'r')
+            return os.fdopen(os.dup(sys.stdin.fileno()), mode, buffering)
         else:
-            return os.fdopen(os.dup(sys.stdout.fileno()), 'w')
+            return os.fdopen(os.dup(sys.stdout.fileno()), mode, buffering)
     else:
-        return open(path, mode)
+        return open(path, mode, buffering)
 
-def collect(paths, *,
+def collect(ci_paths, *,
         sources=None,
         everything=False,
         **args):
@@ -167,7 +164,7 @@ def collect(paths, *,
     callgraph = co.defaultdict(lambda: (None, None, 0, set()))
     f_pattern = re.compile(
         r'([^\\]*)\\n([^:]*)[^\\]*\\n([0-9]+) bytes \((.*)\)')
-    for path in paths:
+    for path in ci_paths:
         with open(path) as f:
             vcg = parse_vcg(f.read())
         for k, graph in vcg:
@@ -546,20 +543,7 @@ def main(ci_paths,
 
     # find sizes
     if not args.get('use', None):
-        # find .ci files
+        results = collect(ci_paths, **args)
-        paths = []
-        for path in ci_paths:
-            if os.path.isdir(path):
-                path = path + '/*.ci'
-
-            for path in glob.glob(path):
-                paths.append(path)
-
-        if not paths:
-            print("error: no .ci files found in %r?" % ci_paths)
-            sys.exit(-1)
-
-        results = collect(paths, **args)
     else:
         results = []
         with openio(args['use']) as f:
@@ -644,9 +628,7 @@ if __name__ == "__main__":
     parser.add_argument(
         'ci_paths',
         nargs='*',
-        default=CI_PATHS,
+        help="Input *.ci files.")
-        help="Description of where to find *.ci files. May be a directory "
-            "or a list of paths. Defaults to %r." % CI_PATHS)
     parser.add_argument(
         '-v', '--verbose',
         action='store_true',

scripts/struct_.py

@@ -12,7 +12,6 @@
 import collections as co
 import csv
 import difflib
-import glob
 import itertools as it
 import math as m
 import os
@@ -21,7 +20,6 @@ import shlex
 import subprocess as sp
 
 
-OBJ_PATHS = ['*.o']
 OBJDUMP_TOOL = ['objdump']
 
 
@@ -120,14 +118,14 @@ class StructResult(co.namedtuple('StructResult', ['file', 'struct', 'size'])):
             self.size + other.size)
 
 
-def openio(path, mode='r'):
+def openio(path, mode='r', buffering=-1):
     if path == '-':
         if mode == 'r':
-            return os.fdopen(os.dup(sys.stdin.fileno()), 'r')
+            return os.fdopen(os.dup(sys.stdin.fileno()), mode, buffering)
         else:
-            return os.fdopen(os.dup(sys.stdout.fileno()), 'w')
+            return os.fdopen(os.dup(sys.stdout.fileno()), mode, buffering)
     else:
-        return open(path, mode)
+        return open(path, mode, buffering)
 
 def collect(obj_paths, *,
         objdump_tool=OBJDUMP_TOOL,
@@ -136,15 +134,15 @@ def collect(obj_paths, *,
         internal=False,
         **args):
     line_pattern = re.compile(
-        '^\s+(?P<no>[0-9]+)\s+'
+        '^\s+(?P<no>[0-9]+)'
-            '(?:(?P<dir>[0-9]+)\s+)?'
+            '(?:\s+(?P<dir>[0-9]+))?'
-            '.*\s+'
+            '\s+.*'
-            '(?P<path>[^\s]+)$')
+            '\s+(?P<path>[^\s]+)$')
     info_pattern = re.compile(
         '^(?:.*(?P<tag>DW_TAG_[a-z_]+).*'
-            '|^.*DW_AT_name.*:\s*(?P<name>[^:\s]+)\s*'
+            '|.*DW_AT_name.*:\s*(?P<name>[^:\s]+)\s*'
-            '|^.*DW_AT_decl_file.*:\s*(?P<file>[0-9]+)\s*'
+            '|.*DW_AT_decl_file.*:\s*(?P<file>[0-9]+)\s*'
-            '|^.*DW_AT_byte_size.*:\s*(?P<size>[0-9]+)\s*)$')
+            '|.*DW_AT_byte_size.*:\s*(?P<size>[0-9]+)\s*)$')
 
     results = []
     for path in obj_paths:
@@ -468,20 +466,7 @@ def main(obj_paths, *,
         **args):
     # find sizes
     if not args.get('use', None):
-        # find .o files
+        results = collect(obj_paths, **args)
-        paths = []
-        for path in obj_paths:
-            if os.path.isdir(path):
-                path = path + '/*.o'
-
-            for path in glob.glob(path):
-                paths.append(path)
-
-        if not paths:
-            print("error: no .o files found in %r?" % obj_paths)
-            sys.exit(-1)
-
-        results = collect(paths, **args)
     else:
         results = []
         with openio(args['use']) as f:
@@ -565,9 +550,7 @@ if __name__ == "__main__":
     parser.add_argument(
         'obj_paths',
         nargs='*',
-        default=OBJ_PATHS,
+        help="Input *.o files.")
-        help="Description of where to find *.o files. May be a directory "
-            "or a list of paths. Defaults to %r." % OBJ_PATHS)
     parser.add_argument(
         '-v', '--verbose',
         action='store_true',

scripts/summary.py

@@ -15,15 +15,12 @@
 import collections as co
 import csv
 import functools as ft
-import glob
 import itertools as it
 import math as m
 import os
 import re
 
 
-CSV_PATHS = ['*.csv']
-
 # supported merge operations
 #
 # this is a terrible way to express these
@@ -548,14 +545,14 @@ def table(Result, results, diff_results=None, *,
             line[-1]))
 
 
-def openio(path, mode='r'):
+def openio(path, mode='r', buffering=-1):
     if path == '-':
         if mode == 'r':
-            return os.fdopen(os.dup(sys.stdin.fileno()), 'r')
+            return os.fdopen(os.dup(sys.stdin.fileno()), mode, buffering)
         else:
-            return os.fdopen(os.dup(sys.stdout.fileno()), 'w')
+            return os.fdopen(os.dup(sys.stdout.fileno()), mode, buffering)
     else:
-        return open(path, mode)
+        return open(path, mode, buffering)
 
 def main(csv_paths, *,
         by=None,
@@ -605,20 +602,8 @@ def main(csv_paths, *,
         ops.update(ops_)
 
     # find CSV files
-    paths = []
-    for path in csv_paths:
-        if os.path.isdir(path):
-            path = path + '/*.csv'
-
-        for path in glob.glob(path):
-            paths.append(path)
-
-    if not paths:
-        print("error: no .csv files found in %r?" % csv_paths)
-        sys.exit(-1)
-
     results = []
-    for path in paths:
+    for path in csv_paths:
         try:
             with openio(path) as f:
                 reader = csv.DictReader(f, restval='')
@@ -721,9 +706,7 @@ if __name__ == "__main__":
     parser.add_argument(
         'csv_paths',
         nargs='*',
-        default=CSV_PATHS,
+        help="Input *.csv files.")
-        help="Description of where to find *.csv files. May be a directory "
-            "or list of paths. Defaults to %r." % CSV_PATHS)
     parser.add_argument(
         '-q', '--quiet',
         action='store_true',

scripts/tailpipe.py

@@ -17,14 +17,14 @@ import sys
 import time
 
 
-def openio(path, mode='r'):
+def openio(path, mode='r', buffering=-1):
     if path == '-':
         if mode == 'r':
-            return os.fdopen(os.dup(sys.stdin.fileno()), 'r')
+            return os.fdopen(os.dup(sys.stdin.fileno()), mode, buffering)
         else:
-            return os.fdopen(os.dup(sys.stdout.fileno()), 'w')
+            return os.fdopen(os.dup(sys.stdout.fileno()), mode, buffering)
     else:
-        return open(path, mode)
+        return open(path, mode, buffering)
 
 class LinesIO:
     def __init__(self, maxlen=None):
@@ -104,6 +104,9 @@ def main(path='-', *, lines=5, cat=False, sleep=0.01, keep_open=False):
                 break
             # don't just flood open calls
             time.sleep(sleep or 0.1)
+    except FileNotFoundError as e:
+        print("error: file not found %r" % path)
+        sys.exit(-1)
     except KeyboardInterrupt:
         pass
 

scripts/test.py

@@ -35,19 +35,12 @@ VALGRIND_TOOL = ['valgrind']
 PERF_SCRIPT = ['./scripts/perf.py']
 
 
-def openio(path, mode='r', buffering=-1, nb=False):
+def openio(path, mode='r', buffering=-1):
     if path == '-':
         if mode == 'r':
-            return os.fdopen(os.dup(sys.stdin.fileno()), 'r', buffering)
+            return os.fdopen(os.dup(sys.stdin.fileno()), mode, buffering)
         else:
-            return os.fdopen(os.dup(sys.stdout.fileno()), 'w', buffering)
+            return os.fdopen(os.dup(sys.stdout.fileno()), mode, buffering)
-    elif nb and 'a' in mode:
-        return os.fdopen(os.open(
-                path,
-                os.O_WRONLY | os.O_CREAT | os.O_APPEND | os.O_NONBLOCK,
-                0o666),
-            mode,
-            buffering)
     else:
         return open(path, mode, buffering)
 
@@ -549,6 +542,12 @@ def find_runner(runner, **args):
         cmd.append('-d%s' % args['disk'])
     if args.get('trace'):
         cmd.append('-t%s' % args['trace'])
+    if args.get('trace_backtrace'):
+        cmd.append('--trace-backtrace')
+    if args.get('trace_period'):
+        cmd.append('--trace-period=%s' % args['trace_period'])
+    if args.get('trace_freq'):
+        cmd.append('--trace-freq=%s' % args['trace_freq'])
     if args.get('read_sleep'):
         cmd.append('--read-sleep=%s' % args['read_sleep'])
     if args.get('prog_sleep'):
@@ -764,7 +763,7 @@ class TestFailure(Exception):
         self.stdout = stdout
         self.assert_ = assert_
 
-def run_stage(name, runner_, ids, output_, **args):
+def run_stage(name, runner_, ids, stdout_, trace_, output_, **args):
     # get expected suite/case/perm counts
     (case_suites,
         expected_suite_perms,
@@ -805,7 +804,6 @@ def run_stage(name, runner_, ids, output_, **args):
         os.close(spty)
         children.add(proc)
         mpty = os.fdopen(mpty, 'r', 1)
-        stdout = None
 
         last_id = None
         last_stdout = []
@@ -822,18 +820,12 @@ def run_stage(name, runner_, ids, output_, **args):
                 if not line:
                     break
                 last_stdout.append(line)
-                if args.get('stdout'):
+                if stdout_:
                     try:
-                        if not stdout:
+                        stdout_.write(line)
-                            stdout = openio(args['stdout'], 'a', 1, nb=True)
+                        stdout_.flush()
-                        stdout.write(line)
+                    except BrokenPipeError:
-                    except OSError as e:
+                        pass
-                        if e.errno not in [
-                                errno.ENXIO,
-                                errno.EPIPE,
-                                errno.EAGAIN]:
-                            raise
-                        stdout = None
                 if args.get('verbose'):
                     sys.stdout.write(line)
 
@@ -1063,6 +1055,8 @@ def run(runner, test_ids=[], **args):
                 by or 'tests',
                 runner_,
                 [by] if by is not None else test_ids,
+                stdout,
+                trace,
                 output,
                 **args)
         # collect passes/failures
@@ -1076,9 +1070,15 @@ def run(runner, test_ids=[], **args):
     stop = time.time()
 
     if stdout:
+        try:
             stdout.close()
+        except BrokenPipeError:
+            pass
     if trace:
+        try:
             trace.close()
+        except BrokenPipeError:
+            pass
     if output:
         output.close()
 
@@ -1284,6 +1284,16 @@ if __name__ == "__main__":
     test_parser.add_argument(
         '-t', '--trace',
         help="Direct trace output to this file.")
+    test_parser.add_argument(
+        '--trace-backtrace',
+        action='store_true',
+        help="Include a backtrace with every trace statement.")
+    test_parser.add_argument(
+        '--trace-period',
+        help="Sample trace output at this period in cycles.")
+    test_parser.add_argument(
+        '--trace-freq',
+        help="Sample trace output at this frequency in hz.")
     test_parser.add_argument(
         '-O', '--stdout',
         help="Direct stdout to this file. Note stderr is already merged here.")
@@ -1361,7 +1371,7 @@ if __name__ == "__main__":
         default=VALGRIND_TOOL,
         help="Path to Valgrind tool to use. Defaults to %r." % VALGRIND_TOOL)
     test_parser.add_argument(
-        '--perf',
+        '-p', '--perf',
         help="Run under Linux's perf to sample performance counters, writing "
             "samples to this file.")
     test_parser.add_argument(

scripts/tracebd.py

@@ -41,14 +41,14 @@ CHARS_BRAILLE = (
     '⠋⢋⡋⣋⠫⢫⡫⣫⠏⢏⡏⣏⠯⢯⡯⣯' '⠛⢛⡛⣛⠻⢻⡻⣻⠟⢟⡟⣟⠿⢿⡿⣿')
 
 
-def openio(path, mode='r'):
+def openio(path, mode='r', buffering=-1):
     if path == '-':
         if mode == 'r':
-            return os.fdopen(os.dup(sys.stdin.fileno()), 'r')
+            return os.fdopen(os.dup(sys.stdin.fileno()), mode, buffering)
         else:
-            return os.fdopen(os.dup(sys.stdout.fileno()), 'w')
+            return os.fdopen(os.dup(sys.stdout.fileno()), mode, buffering)
     else:
-        return open(path, mode)
+        return open(path, mode, buffering)
 
 class LinesIO:
     def __init__(self, maxlen=None):
@@ -663,29 +663,30 @@ def main(path='-', *,
 
     # parse a line of trace output
     pattern = re.compile(
-        'trace.*?bd_(?:'
+        '^(?P<file>[^:]*):(?P<line>[0-9]+):trace:.*?bd_(?:'
             '(?P<create>create\w*)\('
                 '(?:'
                     'block_size=(?P<block_size>\w+)'
                     '|' 'block_count=(?P<block_count>\w+)'
                     '|' '.*?' ')*' '\)'
             '|' '(?P<read>read)\('
-                '\s*(?P<read_ctx>\w+)\s*' ','
+                '\s*(?P<read_ctx>\w+)' '\s*,'
-                '\s*(?P<read_block>\w+)\s*' ','
+                '\s*(?P<read_block>\w+)' '\s*,'
-                '\s*(?P<read_off>\w+)\s*' ','
+                '\s*(?P<read_off>\w+)' '\s*,'
-                '\s*(?P<read_buffer>\w+)\s*' ','
+                '\s*(?P<read_buffer>\w+)' '\s*,'
-                '\s*(?P<read_size>\w+)\s*' '\)'
+                '\s*(?P<read_size>\w+)' '\s*\)'
             '|' '(?P<prog>prog)\('
-                '\s*(?P<prog_ctx>\w+)\s*' ','
+                '\s*(?P<prog_ctx>\w+)' '\s*,'
-                '\s*(?P<prog_block>\w+)\s*' ','
+                '\s*(?P<prog_block>\w+)' '\s*,'
-                '\s*(?P<prog_off>\w+)\s*' ','
+                '\s*(?P<prog_off>\w+)' '\s*,'
-                '\s*(?P<prog_buffer>\w+)\s*' ','
+                '\s*(?P<prog_buffer>\w+)' '\s*,'
-                '\s*(?P<prog_size>\w+)\s*' '\)'
+                '\s*(?P<prog_size>\w+)' '\s*\)'
             '|' '(?P<erase>erase)\('
-                '\s*(?P<erase_ctx>\w+)\s*' ','
+                '\s*(?P<erase_ctx>\w+)' '\s*,'
-                '\s*(?P<erase_block>\w+)\s*' '\)'
+                '\s*(?P<erase_block>\w+)'
+                '\s*\(\s*(?P<erase_size>\w+)\s*\)' '\s*\)'
             '|' '(?P<sync>sync)\('
-                '\s*(?P<sync_ctx>\w+)\s*' '\)' ')')
+                '\s*(?P<sync_ctx>\w+)' '\s*\)' ')\s*$')
     def parse(line):
         nonlocal bd
 
@@ -694,7 +695,7 @@ def main(path='-', *,
         # through here
         if 'trace' not in line or 'bd' not in line:
             return False
-        m = pattern.search(line)
+        m = pattern.match(line)
         if not m:
             return False
 
@@ -748,12 +749,16 @@ def main(path='-', *,
 
         elif m.group('erase') and (erase or wear):
             block = int(m.group('erase_block'), 0)
+            size = int(m.group('erase_size'), 0)
 
             if block_stop is not None and block >= block_stop:
                 return False
             block -= block_start
+            if off_stop is not None:
+                size = min(size, off_stop)
+            off = -off_start
 
-            bd.erase(block)
+            bd.erase(block, off, size)
             return True
 
         else:
@@ -818,6 +823,9 @@ def main(path='-', *,
                 break
             # don't just flood open calls
             time.sleep(sleep or 0.1)
+    except FileNotFoundError as e:
+        print("error: file not found %r" % path)
+        sys.exit(-1)
     except KeyboardInterrupt:
         pass