Fixed MTP to work with TWRP

tools/testing/selftests/powerpc/Makefile

@@ -0,0 +1,39 @@
+# Makefile for powerpc selftests
+
+# ARCH can be overridden by the user for cross compiling
+ARCH ?= $(shell uname -m)
+ARCH := $(shell echo $(ARCH) | sed -e s/ppc.*/powerpc/)
+
+ifeq ($(ARCH),powerpc)
+
+GIT_VERSION = $(shell git describe --always --long --dirty || echo "unknown")
+
+CC := $(CROSS_COMPILE)$(CC)
+CFLAGS := -Wall -O2 -flto -Wall -Werror -DGIT_VERSION='"$(GIT_VERSION)"' -I$(CURDIR) $(CFLAGS)
+
+export CC CFLAGS
+
+TARGETS = pmu copyloops mm tm primitives
+
+endif
+
+all: $(TARGETS)
+
+$(TARGETS):
+	$(MAKE) -k -C $@ all
+
+run_tests: all
+	@for TARGET in $(TARGETS); do \
+		$(MAKE) -C $$TARGET run_tests; \
+	done;
+
+clean:
+	@for TARGET in $(TARGETS); do \
+		$(MAKE) -C $$TARGET clean; \
+	done;
+	rm -f tags
+
+tags:
+	find . -name '*.c' -o -name '*.h' | xargs ctags
+
+.PHONY: all run_tests clean tags $(TARGETS)

tools/testing/selftests/powerpc/copyloops/Makefile

@@ -0,0 +1,29 @@
+# The loops are all 64-bit code
+CFLAGS += -m64
+CFLAGS += -I$(CURDIR)
+CFLAGS += -D SELFTEST
+
+# Use our CFLAGS for the implicit .S rule
+ASFLAGS = $(CFLAGS)
+
+PROGS := copyuser_64 copyuser_power7 memcpy_64 memcpy_power7
+EXTRA_SOURCES := validate.c ../harness.c
+
+all: $(PROGS)
+
+copyuser_64:     CPPFLAGS += -D COPY_LOOP=test___copy_tofrom_user_base
+copyuser_power7: CPPFLAGS += -D COPY_LOOP=test___copy_tofrom_user_power7
+memcpy_64:       CPPFLAGS += -D COPY_LOOP=test_memcpy
+memcpy_power7:   CPPFLAGS += -D COPY_LOOP=test_memcpy_power7
+
+$(PROGS): $(EXTRA_SOURCES)
+
+run_tests: all
+	@-for PROG in $(PROGS); do \
+		./$$PROG; \
+	done;
+
+clean:
+	rm -f $(PROGS) *.o
+
+.PHONY: all run_tests clean

tools/testing/selftests/powerpc/copyloops/asm/ppc_asm.h

@@ -0,0 +1,89 @@
+#include <ppc-asm.h>
+
+#define CONFIG_ALTIVEC
+
+#define r1	1
+
+#define vr0     0
+#define vr1     1
+#define vr2     2
+#define vr3     3
+#define vr4     4
+#define vr5     5
+#define vr6     6
+#define vr7     7
+#define vr8     8
+#define vr9     9
+#define vr10    10
+#define vr11    11
+#define vr12    12
+#define vr13    13
+#define vr14    14
+#define vr15    15
+#define vr16    16
+#define vr17    17
+#define vr18    18
+#define vr19    19
+#define vr20    20
+#define vr21    21
+#define vr22    22
+#define vr23    23
+#define vr24    24
+#define vr25    25
+#define vr26    26
+#define vr27    27
+#define vr28    28
+#define vr29    29
+#define vr30    30
+#define vr31    31
+
+#define R14 r14
+#define R15 r15
+#define R16 r16
+#define R17 r17
+#define R18 r18
+#define R19 r19
+#define R20 r20
+#define R21 r21
+#define R22 r22
+#define R29 r29
+#define R30 r30
+#define R31 r31
+
+#define STACKFRAMESIZE	256
+#define STK_REG(i)	(112 + ((i)-14)*8)
+
+#define _GLOBAL(A) FUNC_START(test_ ## A)
+#define _GLOBAL_TOC(A) _GLOBAL(A)
+
+#define PPC_MTOCRF(A, B)	mtocrf A, B
+
+FUNC_START(enter_vmx_usercopy)
+	li	r3,1
+	blr
+
+FUNC_START(exit_vmx_usercopy)
+	li	r3,0
+	blr
+
+FUNC_START(enter_vmx_copy)
+	li	r3,1
+	blr
+
+FUNC_START(exit_vmx_copy)
+	blr
+
+FUNC_START(memcpy_power7)
+	blr
+
+FUNC_START(__copy_tofrom_user_power7)
+	blr
+
+FUNC_START(__copy_tofrom_user_base)
+	blr
+
+#define BEGIN_FTR_SECTION
+#define FTR_SECTION_ELSE
+#define ALT_FTR_SECTION_END_IFCLR(x)
+#define ALT_FTR_SECTION_END(x, y)
+#define END_FTR_SECTION_IFCLR(x)

tools/testing/selftests/powerpc/copyloops/copyuser_64.S

@@ -0,0 +1 @@
+../../../../../arch/powerpc/lib/copyuser_64.S

tools/testing/selftests/powerpc/copyloops/copyuser_power7.S

@@ -0,0 +1 @@
+../../../../../arch/powerpc/lib/copyuser_power7.S

tools/testing/selftests/powerpc/copyloops/memcpy_64.S

@@ -0,0 +1 @@
+../../../../../arch/powerpc/lib/memcpy_64.S

tools/testing/selftests/powerpc/copyloops/memcpy_power7.S

@@ -0,0 +1 @@
+../../../../../arch/powerpc/lib/memcpy_power7.S

tools/testing/selftests/powerpc/copyloops/validate.c

@@ -0,0 +1,99 @@
+#include <malloc.h>
+#include <string.h>
+#include <stdlib.h>
+#include <stdbool.h>
+
+#include "../utils.h"
+
+#define MAX_LEN 8192
+#define MAX_OFFSET 16
+#define MIN_REDZONE 128
+#define BUFLEN (MAX_LEN+MAX_OFFSET+2*MIN_REDZONE)
+#define POISON 0xa5
+
+unsigned long COPY_LOOP(void *to, const void *from, unsigned long size);
+
+static void do_one(char *src, char *dst, unsigned long src_off,
+		   unsigned long dst_off, unsigned long len, void *redzone,
+		   void *fill)
+{
+	char *srcp, *dstp;
+	unsigned long ret;
+	unsigned long i;
+
+	srcp = src + MIN_REDZONE + src_off;
+	dstp = dst + MIN_REDZONE + dst_off;
+
+	memset(src, POISON, BUFLEN);
+	memset(dst, POISON, BUFLEN);
+	memcpy(srcp, fill, len);
+
+	ret = COPY_LOOP(dstp, srcp, len);
+	if (ret && ret != (unsigned long)dstp) {
+		printf("(%p,%p,%ld) returned %ld\n", dstp, srcp, len, ret);
+		abort();
+	}
+
+	if (memcmp(dstp, srcp, len)) {
+		printf("(%p,%p,%ld) miscompare\n", dstp, srcp, len);
+		printf("src: ");
+		for (i = 0; i < len; i++)
+			printf("%02x ", srcp[i]);
+		printf("\ndst: ");
+		for (i = 0; i < len; i++)
+			printf("%02x ", dstp[i]);
+		printf("\n");
+		abort();
+	}
+
+	if (memcmp(dst, redzone, dstp - dst)) {
+		printf("(%p,%p,%ld) redzone before corrupted\n",
+		       dstp, srcp, len);
+		abort();
+	}
+
+	if (memcmp(dstp+len, redzone, dst+BUFLEN-(dstp+len))) {
+		printf("(%p,%p,%ld) redzone after corrupted\n",
+		       dstp, srcp, len);
+		abort();
+	}
+}
+
+int test_copy_loop(void)
+{
+	char *src, *dst, *redzone, *fill;
+	unsigned long len, src_off, dst_off;
+	unsigned long i;
+
+	src = memalign(BUFLEN, BUFLEN);
+	dst = memalign(BUFLEN, BUFLEN);
+	redzone = malloc(BUFLEN);
+	fill = malloc(BUFLEN);
+
+	if (!src || !dst || !redzone || !fill) {
+		fprintf(stderr, "malloc failed\n");
+		exit(1);
+	}
+
+	memset(redzone, POISON, BUFLEN);
+
+	/* Fill with sequential bytes */
+	for (i = 0; i < BUFLEN; i++)
+		fill[i] = i & 0xff;
+
+	for (len = 1; len < MAX_LEN; len++) {
+		for (src_off = 0; src_off < MAX_OFFSET; src_off++) {
+			for (dst_off = 0; dst_off < MAX_OFFSET; dst_off++) {
+				do_one(src, dst, src_off, dst_off, len,
+				       redzone, fill);
+			}
+		}
+	}
+
+	return 0;
+}
+
+int main(void)
+{
+	return test_harness(test_copy_loop, str(COPY_LOOP));
+}

tools/testing/selftests/powerpc/harness.c

@@ -0,0 +1,114 @@
+/*
+ * Copyright 2013, Michael Ellerman, IBM Corp.
+ * Licensed under GPLv2.
+ */
+
+#include <errno.h>
+#include <signal.h>
+#include <stdbool.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <sys/types.h>
+#include <sys/wait.h>
+#include <unistd.h>
+
+#include "subunit.h"
+#include "utils.h"
+
+#define TIMEOUT		120
+#define KILL_TIMEOUT	5
+
+
+int run_test(int (test_function)(void), char *name)
+{
+	bool terminated;
+	int rc, status;
+	pid_t pid;
+
+	/* Make sure output is flushed before forking */
+	fflush(stdout);
+
+	pid = fork();
+	if (pid == 0) {
+		setpgid(0, 0);
+		exit(test_function());
+	} else if (pid == -1) {
+		perror("fork");
+		return 1;
+	}
+
+	setpgid(pid, pid);
+
+	/* Wake us up in timeout seconds */
+	alarm(TIMEOUT);
+	terminated = false;
+
+wait:
+	rc = waitpid(pid, &status, 0);
+	if (rc == -1) {
+		if (errno != EINTR) {
+			printf("unknown error from waitpid\n");
+			return 1;
+		}
+
+		if (terminated) {
+			printf("!! force killing %s\n", name);
+			kill(-pid, SIGKILL);
+			return 1;
+		} else {
+			printf("!! killing %s\n", name);
+			kill(-pid, SIGTERM);
+			terminated = true;
+			alarm(KILL_TIMEOUT);
+			goto wait;
+		}
+	}
+
+	/* Kill anything else in the process group that is still running */
+	kill(-pid, SIGTERM);
+
+	if (WIFEXITED(status))
+		status = WEXITSTATUS(status);
+	else {
+		if (WIFSIGNALED(status))
+			printf("!! child died by signal %d\n", WTERMSIG(status));
+		else
+			printf("!! child died by unknown cause\n");
+
+		status = 1; /* Signal or other */
+	}
+
+	return status;
+}
+
+static void alarm_handler(int signum)
+{
+	/* Jut wake us up from waitpid */
+}
+
+static struct sigaction alarm_action = {
+	.sa_handler = alarm_handler,
+};
+
+int test_harness(int (test_function)(void), char *name)
+{
+	int rc;
+
+	test_start(name);
+	test_set_git_version(GIT_VERSION);
+
+	if (sigaction(SIGALRM, &alarm_action, NULL)) {
+		perror("sigaction");
+		test_error(name);
+		return 1;
+	}
+
+	rc = run_test(test_function, name);
+
+	if (rc == MAGIC_SKIP_RETURN_VALUE)
+		test_skip(name);
+	else
+		test_finish(name, rc);
+
+	return rc;
+}

tools/testing/selftests/powerpc/mm/Makefile

@@ -0,0 +1,18 @@
+noarg:
+	$(MAKE) -C ../
+
+PROGS := hugetlb_vs_thp_test
+
+all: $(PROGS)
+
+$(PROGS): ../harness.c
+
+run_tests: all
+	@-for PROG in $(PROGS); do \
+		./$$PROG; \
+	done;
+
+clean:
+	rm -f $(PROGS)
+
+.PHONY: all run_tests clean

tools/testing/selftests/powerpc/mm/hugetlb_vs_thp_test.c

@@ -0,0 +1,72 @@
+#include <stdio.h>
+#include <sys/mman.h>
+#include <unistd.h>
+
+#include "utils.h"
+
+/* This must match the huge page & THP size */
+#define SIZE	(16 * 1024 * 1024)
+
+static int test_body(void)
+{
+	void *addr;
+	char *p;
+
+	addr = (void *)0xa0000000;
+
+	p = mmap(addr, SIZE, PROT_READ | PROT_WRITE,
+		 MAP_HUGETLB | MAP_ANONYMOUS | MAP_PRIVATE, -1, 0);
+	if (p != MAP_FAILED) {
+		/*
+		 * Typically the mmap will fail because no huge pages are
+		 * allocated on the system. But if there are huge pages
+		 * allocated the mmap will succeed. That's fine too, we just
+		 * munmap here before continuing.
+		 */
+		munmap(addr, SIZE);
+	}
+
+	p = mmap(addr, SIZE, PROT_READ | PROT_WRITE,
+		 MAP_ANONYMOUS | MAP_PRIVATE, -1, 0);
+	if (p == MAP_FAILED) {
+		printf("Mapping failed @ %p\n", addr);
+		perror("mmap");
+		return 1;
+	}
+
+	/*
+	 * Either a user or kernel access is sufficient to trigger the bug.
+	 * A kernel access is easier to spot & debug, as it will trigger the
+	 * softlockup or RCU stall detectors, and when the system is kicked
+	 * into xmon we get a backtrace in the kernel.
+	 *
+	 * A good option is:
+	 *  getcwd(p, SIZE);
+	 *
+	 * For the purposes of this testcase it's preferable to spin in
+	 * userspace, so the harness can kill us if we get stuck. That way we
+	 * see a test failure rather than a dead system.
+	 */
+	*p = 0xf;
+
+	munmap(addr, SIZE);
+
+	return 0;
+}
+
+static int test_main(void)
+{
+	int i;
+
+	/* 10,000 because it's a "bunch", and completes reasonably quickly */
+	for (i = 0; i < 10000; i++)
+		if (test_body())
+			return 1;
+
+	return 0;
+}
+
+int main(void)
+{
+	return test_harness(test_main, "hugetlb_vs_thp");
+}

tools/testing/selftests/powerpc/pmu/Makefile

@@ -0,0 +1,38 @@
+noarg:
+	$(MAKE) -C ../
+
+PROGS := count_instructions l3_bank_test per_event_excludes
+EXTRA_SOURCES := ../harness.c event.c lib.c
+
+SUB_TARGETS = ebb
+
+all: $(PROGS) $(SUB_TARGETS)
+
+$(PROGS): $(EXTRA_SOURCES)
+
+# loop.S can only be built 64-bit
+count_instructions: loop.S count_instructions.c $(EXTRA_SOURCES)
+	$(CC) $(CFLAGS) -m64 -o $@ $^
+
+run_tests: all sub_run_tests
+	@-for PROG in $(PROGS); do \
+		./$$PROG; \
+	done;
+
+clean: sub_clean
+	rm -f $(PROGS) loop.o
+
+$(SUB_TARGETS):
+	$(MAKE) -k -C $@ all
+
+sub_run_tests: all
+	@for TARGET in $(SUB_TARGETS); do \
+		$(MAKE) -C $$TARGET run_tests; \
+	done;
+
+sub_clean:
+	@for TARGET in $(SUB_TARGETS); do \
+		$(MAKE) -C $$TARGET clean; \
+	done;
+
+.PHONY: all run_tests clean sub_run_tests sub_clean $(SUB_TARGETS)

tools/testing/selftests/powerpc/pmu/count_instructions.c

@@ -0,0 +1,147 @@
+/*
+ * Copyright 2013, Michael Ellerman, IBM Corp.
+ * Licensed under GPLv2.
+ */
+
+#define _GNU_SOURCE
+
+#include <stdio.h>
+#include <stdbool.h>
+#include <string.h>
+#include <sys/prctl.h>
+
+#include "event.h"
+#include "utils.h"
+#include "lib.h"
+
+extern void thirty_two_instruction_loop(u64 loops);
+
+static void setup_event(struct event *e, u64 config, char *name)
+{
+	event_init_opts(e, config, PERF_TYPE_HARDWARE, name);
+
+	e->attr.disabled = 1;
+	e->attr.exclude_kernel = 1;
+	e->attr.exclude_hv = 1;
+	e->attr.exclude_idle = 1;
+}
+
+static int do_count_loop(struct event *events, u64 instructions,
+			 u64 overhead, bool report)
+{
+	s64 difference, expected;
+	double percentage;
+
+	prctl(PR_TASK_PERF_EVENTS_ENABLE);
+
+	/* Run for 1M instructions */
+	thirty_two_instruction_loop(instructions >> 5);
+
+	prctl(PR_TASK_PERF_EVENTS_DISABLE);
+
+	event_read(&events[0]);
+	event_read(&events[1]);
+
+	expected = instructions + overhead;
+	difference = events[0].result.value - expected;
+	percentage = (double)difference / events[0].result.value * 100;
+
+	if (report) {
+		event_report(&events[0]);
+		event_report(&events[1]);
+
+		printf("Looped for %llu instructions, overhead %llu\n", instructions, overhead);
+		printf("Expected %llu\n", expected);
+		printf("Actual   %llu\n", events[0].result.value);
+		printf("Delta    %lld, %f%%\n", difference, percentage);
+	}
+
+	event_reset(&events[0]);
+	event_reset(&events[1]);
+
+	if (difference < 0)
+		difference = -difference;
+
+	/* Tolerate a difference below 0.0001 % */
+	difference *= 10000 * 100;
+	if (difference / events[0].result.value)
+		return -1;
+
+	return 0;
+}
+
+/* Count how many instructions it takes to do a null loop */
+static u64 determine_overhead(struct event *events)
+{
+	u64 current, overhead;
+	int i;
+
+	do_count_loop(events, 0, 0, false);
+	overhead = events[0].result.value;
+
+	for (i = 0; i < 100; i++) {
+		do_count_loop(events, 0, 0, false);
+		current = events[0].result.value;
+		if (current < overhead) {
+			printf("Replacing overhead %llu with %llu\n", overhead, current);
+			overhead = current;
+		}
+	}
+
+	return overhead;
+}
+
+static int test_body(void)
+{
+	struct event events[2];
+	u64 overhead;
+
+	setup_event(&events[0], PERF_COUNT_HW_INSTRUCTIONS, "instructions");
+	setup_event(&events[1], PERF_COUNT_HW_CPU_CYCLES, "cycles");
+
+	if (event_open(&events[0])) {
+		perror("perf_event_open");
+		return -1;
+	}
+
+	if (event_open_with_group(&events[1], events[0].fd)) {
+		perror("perf_event_open");
+		return -1;
+	}
+
+	overhead = determine_overhead(events);
+	printf("Overhead of null loop: %llu instructions\n", overhead);
+
+	/* Run for 1Mi instructions */
+	FAIL_IF(do_count_loop(events, 1000000, overhead, true));
+
+	/* Run for 10Mi instructions */
+	FAIL_IF(do_count_loop(events, 10000000, overhead, true));
+
+	/* Run for 100Mi instructions */
+	FAIL_IF(do_count_loop(events, 100000000, overhead, true));
+
+	/* Run for 1Bi instructions */
+	FAIL_IF(do_count_loop(events, 1000000000, overhead, true));
+
+	/* Run for 16Bi instructions */
+	FAIL_IF(do_count_loop(events, 16000000000, overhead, true));
+
+	/* Run for 64Bi instructions */
+	FAIL_IF(do_count_loop(events, 64000000000, overhead, true));
+
+	event_close(&events[0]);
+	event_close(&events[1]);
+
+	return 0;
+}
+
+static int count_instructions(void)
+{
+	return eat_cpu(test_body);
+}
+
+int main(void)
+{
+	return test_harness(count_instructions, "count_instructions");
+}

tools/testing/selftests/powerpc/pmu/ebb/Makefile

@@ -0,0 +1,33 @@
+noarg:
+	$(MAKE) -C ../../
+
+# The EBB handler is 64-bit code and everything links against it
+CFLAGS += -m64
+
+PROGS := reg_access_test event_attributes_test cycles_test	\
+	 cycles_with_freeze_test pmc56_overflow_test		\
+	 ebb_vs_cpu_event_test cpu_event_vs_ebb_test		\
+	 cpu_event_pinned_vs_ebb_test task_event_vs_ebb_test	\
+	 task_event_pinned_vs_ebb_test multi_ebb_procs_test	\
+	 multi_counter_test pmae_handling_test			\
+	 close_clears_pmcc_test instruction_count_test		\
+	 fork_cleanup_test ebb_on_child_test			\
+	 ebb_on_willing_child_test back_to_back_ebbs_test	\
+	 lost_exception_test no_handler_test			\
+	 cycles_with_mmcr2_test
+
+all: $(PROGS)
+
+$(PROGS): ../../harness.c ../event.c ../lib.c ebb.c ebb_handler.S trace.c busy_loop.S
+
+instruction_count_test: ../loop.S
+
+lost_exception_test: ../lib.c
+
+run_tests: all
+	@-for PROG in $(PROGS); do \
+		./$$PROG; \
+	done;
+
+clean:
+	rm -f $(PROGS)

tools/testing/selftests/powerpc/pmu/ebb/back_to_back_ebbs_test.c

@@ -0,0 +1,106 @@
+/*
+ * Copyright 2014, Michael Ellerman, IBM Corp.
+ * Licensed under GPLv2.
+ */
+
+#include <stdbool.h>
+#include <stdio.h>
+#include <stdlib.h>
+
+#include "ebb.h"
+
+
+#define NUMBER_OF_EBBS	50
+
+/*
+ * Test that if we overflow the counter while in the EBB handler, we take
+ * another EBB on exiting from the handler.
+ *
+ * We do this by counting with a stupidly low sample period, causing us to
+ * overflow the PMU while we're still in the EBB handler, leading to another
+ * EBB.
+ *
+ * We get out of what would otherwise be an infinite loop by leaving the
+ * counter frozen once we've taken enough EBBs.
+ */
+
+static void ebb_callee(void)
+{
+	uint64_t siar, val;
+
+	val = mfspr(SPRN_BESCR);
+	if (!(val & BESCR_PMEO)) {
+		ebb_state.stats.spurious++;
+		goto out;
+	}
+
+	ebb_state.stats.ebb_count++;
+	trace_log_counter(ebb_state.trace, ebb_state.stats.ebb_count);
+
+	/* Resets the PMC */
+	count_pmc(1, sample_period);
+
+out:
+	if (ebb_state.stats.ebb_count == NUMBER_OF_EBBS)
+		/* Reset but leave counters frozen */
+		reset_ebb_with_clear_mask(MMCR0_PMAO);
+	else
+		/* Unfreezes */
+		reset_ebb();
+
+	/* Do some stuff to chew some cycles and pop the counter */
+	siar = mfspr(SPRN_SIAR);
+	trace_log_reg(ebb_state.trace, SPRN_SIAR, siar);
+
+	val = mfspr(SPRN_PMC1);
+	trace_log_reg(ebb_state.trace, SPRN_PMC1, val);
+
+	val = mfspr(SPRN_MMCR0);
+	trace_log_reg(ebb_state.trace, SPRN_MMCR0, val);
+}
+
+int back_to_back_ebbs(void)
+{
+	struct event event;
+
+	event_init_named(&event, 0x1001e, "cycles");
+	event_leader_ebb_init(&event);
+
+	event.attr.exclude_kernel = 1;
+	event.attr.exclude_hv = 1;
+	event.attr.exclude_idle = 1;
+
+	FAIL_IF(event_open(&event));
+
+	setup_ebb_handler(ebb_callee);
+
+	FAIL_IF(ebb_event_enable(&event));
+
+	sample_period = 5;
+
+	ebb_freeze_pmcs();
+	mtspr(SPRN_PMC1, pmc_sample_period(sample_period));
+	ebb_global_enable();
+	ebb_unfreeze_pmcs();
+
+	while (ebb_state.stats.ebb_count < NUMBER_OF_EBBS)
+		FAIL_IF(core_busy_loop());
+
+	ebb_global_disable();
+	ebb_freeze_pmcs();
+
+	count_pmc(1, sample_period);
+
+	dump_ebb_state();
+
+	event_close(&event);
+
+	FAIL_IF(ebb_state.stats.ebb_count != NUMBER_OF_EBBS);
+
+	return 0;
+}
+
+int main(void)
+{
+	return test_harness(back_to_back_ebbs, "back_to_back_ebbs");
+}

tools/testing/selftests/powerpc/pmu/ebb/busy_loop.S

@@ -0,0 +1,271 @@
+/*
+ * Copyright 2014, Michael Ellerman, IBM Corp.
+ * Licensed under GPLv2.
+ */
+
+#include <ppc-asm.h>
+
+	.text
+
+FUNC_START(core_busy_loop)
+	stdu	%r1, -168(%r1)
+	std	r14, 160(%r1)
+	std	r15, 152(%r1)
+	std	r16, 144(%r1)
+	std	r17, 136(%r1)
+	std	r18, 128(%r1)
+	std	r19, 120(%r1)
+	std	r20, 112(%r1)
+	std	r21, 104(%r1)
+	std	r22, 96(%r1)
+	std	r23, 88(%r1)
+	std	r24, 80(%r1)
+	std	r25, 72(%r1)
+	std	r26, 64(%r1)
+	std	r27, 56(%r1)
+	std	r28, 48(%r1)
+	std	r29, 40(%r1)
+	std	r30, 32(%r1)
+	std	r31, 24(%r1)
+
+	li	 r3, 0x3030
+	std	 r3, -96(%r1)
+	li	 r4, 0x4040
+	std	 r4, -104(%r1)
+	li	 r5, 0x5050
+	std	 r5, -112(%r1)
+	li	 r6, 0x6060
+	std	 r6, -120(%r1)
+	li	 r7, 0x7070
+	std	 r7, -128(%r1)
+	li	 r8, 0x0808
+	std	 r8, -136(%r1)
+	li	 r9, 0x0909
+	std	 r9, -144(%r1)
+	li	r10, 0x1010
+	std	r10, -152(%r1)
+	li	r11, 0x1111
+	std	r11, -160(%r1)
+	li	r14, 0x1414
+	std	r14, -168(%r1)
+	li	r15, 0x1515
+	std	r15, -176(%r1)
+	li	r16, 0x1616
+	std	r16, -184(%r1)
+	li	r17, 0x1717
+	std	r17, -192(%r1)
+	li	r18, 0x1818
+	std	r18, -200(%r1)
+	li	r19, 0x1919
+	std	r19, -208(%r1)
+	li	r20, 0x2020
+	std	r20, -216(%r1)
+	li	r21, 0x2121
+	std	r21, -224(%r1)
+	li	r22, 0x2222
+	std	r22, -232(%r1)
+	li	r23, 0x2323
+	std	r23, -240(%r1)
+	li	r24, 0x2424
+	std	r24, -248(%r1)
+	li	r25, 0x2525
+	std	r25, -256(%r1)
+	li	r26, 0x2626
+	std	r26, -264(%r1)
+	li	r27, 0x2727
+	std	r27, -272(%r1)
+	li	r28, 0x2828
+	std	r28, -280(%r1)
+	li	r29, 0x2929
+	std	r29, -288(%r1)
+	li	r30, 0x3030
+	li	r31, 0x3131
+
+	li	r3, 0
+0:	addi	r3, r3, 1
+	cmpwi	r3, 100
+	blt	0b
+
+	/* Return 1 (fail) unless we get through all the checks */
+	li	r3, 1
+
+	/* Check none of our registers have been corrupted */
+	cmpwi	r4,  0x4040
+	bne	1f
+	cmpwi	r5,  0x5050
+	bne	1f
+	cmpwi	r6,  0x6060
+	bne	1f
+	cmpwi	r7,  0x7070
+	bne	1f
+	cmpwi	r8,  0x0808
+	bne	1f
+	cmpwi	r9,  0x0909
+	bne	1f
+	cmpwi	r10, 0x1010
+	bne	1f
+	cmpwi	r11, 0x1111
+	bne	1f
+	cmpwi	r14, 0x1414
+	bne	1f
+	cmpwi	r15, 0x1515
+	bne	1f
+	cmpwi	r16, 0x1616
+	bne	1f
+	cmpwi	r17, 0x1717
+	bne	1f
+	cmpwi	r18, 0x1818
+	bne	1f
+	cmpwi	r19, 0x1919
+	bne	1f
+	cmpwi	r20, 0x2020
+	bne	1f
+	cmpwi	r21, 0x2121
+	bne	1f
+	cmpwi	r22, 0x2222
+	bne	1f
+	cmpwi	r23, 0x2323
+	bne	1f
+	cmpwi	r24, 0x2424
+	bne	1f
+	cmpwi	r25, 0x2525
+	bne	1f
+	cmpwi	r26, 0x2626
+	bne	1f
+	cmpwi	r27, 0x2727
+	bne	1f
+	cmpwi	r28, 0x2828
+	bne	1f
+	cmpwi	r29, 0x2929
+	bne	1f
+	cmpwi	r30, 0x3030
+	bne	1f
+	cmpwi	r31, 0x3131
+	bne	1f
+
+	/* Load junk into all our registers before we reload them from the stack. */
+	li	r3,  0xde
+	li	r4,  0xad
+	li	r5,  0xbe
+	li	r6,  0xef
+	li	r7,  0xde
+	li	r8,  0xad
+	li	r9,  0xbe
+	li	r10, 0xef
+	li	r11, 0xde
+	li	r14, 0xad
+	li	r15, 0xbe
+	li	r16, 0xef
+	li	r17, 0xde
+	li	r18, 0xad
+	li	r19, 0xbe
+	li	r20, 0xef
+	li	r21, 0xde
+	li	r22, 0xad
+	li	r23, 0xbe
+	li	r24, 0xef
+	li	r25, 0xde
+	li	r26, 0xad
+	li	r27, 0xbe
+	li	r28, 0xef
+	li	r29, 0xdd
+
+	ld	r3,	-96(%r1)
+	cmpwi	r3,  0x3030
+	bne	1f
+	ld	r4,	-104(%r1)
+	cmpwi	r4,  0x4040
+	bne	1f
+	ld	r5,	-112(%r1)
+	cmpwi	r5,  0x5050
+	bne	1f
+	ld	r6,	-120(%r1)
+	cmpwi	r6,  0x6060
+	bne	1f
+	ld	r7,	-128(%r1)
+	cmpwi	r7,  0x7070
+	bne	1f
+	ld	r8,	-136(%r1)
+	cmpwi	r8,  0x0808
+	bne	1f
+	ld	r9,	-144(%r1)
+	cmpwi	r9,  0x0909
+	bne	1f
+	ld	r10, -152(%r1)
+	cmpwi	r10, 0x1010
+	bne	1f
+	ld	r11, -160(%r1)
+	cmpwi	r11, 0x1111
+	bne	1f
+	ld	r14, -168(%r1)
+	cmpwi	r14, 0x1414
+	bne	1f
+	ld	r15, -176(%r1)
+	cmpwi	r15, 0x1515
+	bne	1f
+	ld	r16, -184(%r1)
+	cmpwi	r16, 0x1616
+	bne	1f
+	ld	r17, -192(%r1)
+	cmpwi	r17, 0x1717
+	bne	1f
+	ld	r18, -200(%r1)
+	cmpwi	r18, 0x1818
+	bne	1f
+	ld	r19, -208(%r1)
+	cmpwi	r19, 0x1919
+	bne	1f
+	ld	r20, -216(%r1)
+	cmpwi	r20, 0x2020
+	bne	1f
+	ld	r21, -224(%r1)
+	cmpwi	r21, 0x2121
+	bne	1f
+	ld	r22, -232(%r1)
+	cmpwi	r22, 0x2222
+	bne	1f
+	ld	r23, -240(%r1)
+	cmpwi	r23, 0x2323
+	bne	1f
+	ld	r24, -248(%r1)
+	cmpwi	r24, 0x2424
+	bne	1f
+	ld	r25, -256(%r1)
+	cmpwi	r25, 0x2525
+	bne	1f
+	ld	r26, -264(%r1)
+	cmpwi	r26, 0x2626
+	bne	1f
+	ld	r27, -272(%r1)
+	cmpwi	r27, 0x2727
+	bne	1f
+	ld	r28, -280(%r1)
+	cmpwi	r28, 0x2828
+	bne	1f
+	ld	r29, -288(%r1)
+	cmpwi	r29, 0x2929
+	bne	1f
+
+	/* Load 0 (success) to return */
+	li	r3, 0
+
+1:	ld	r14, 160(%r1)
+	ld	r15, 152(%r1)
+	ld	r16, 144(%r1)
+	ld	r17, 136(%r1)
+	ld	r18, 128(%r1)
+	ld	r19, 120(%r1)
+	ld	r20, 112(%r1)
+	ld	r21, 104(%r1)
+	ld	r22, 96(%r1)
+	ld	r23, 88(%r1)
+	ld	r24, 80(%r1)
+	ld	r25, 72(%r1)
+	ld	r26, 64(%r1)
+	ld	r27, 56(%r1)
+	ld	r28, 48(%r1)
+	ld	r29, 40(%r1)
+	ld	r30, 32(%r1)
+	ld	r31, 24(%r1)
+	addi	%r1, %r1, 168
+	blr

tools/testing/selftests/powerpc/pmu/ebb/close_clears_pmcc_test.c

@@ -0,0 +1,59 @@
+/*
+ * Copyright 2014, Michael Ellerman, IBM Corp.
+ * Licensed under GPLv2.
+ */
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <setjmp.h>
+#include <signal.h>
+
+#include "ebb.h"
+
+
+/*
+ * Test that closing the EBB event clears MMCR0_PMCC, preventing further access
+ * by userspace to the PMU hardware.
+ */
+
+int close_clears_pmcc(void)
+{
+	struct event event;
+
+	event_init_named(&event, 0x1001e, "cycles");
+	event_leader_ebb_init(&event);
+
+	FAIL_IF(event_open(&event));
+
+	ebb_enable_pmc_counting(1);
+	setup_ebb_handler(standard_ebb_callee);
+	ebb_global_enable();
+	FAIL_IF(ebb_event_enable(&event));
+
+	mtspr(SPRN_PMC1, pmc_sample_period(sample_period));
+
+	while (ebb_state.stats.ebb_count < 1)
+		FAIL_IF(core_busy_loop());
+
+	ebb_global_disable();
+	event_close(&event);
+
+	FAIL_IF(ebb_state.stats.ebb_count == 0);
+
+	/* The real test is here, do we take a SIGILL when writing PMU regs now
+	 * that we have closed the event. We expect that we will. */
+
+	FAIL_IF(catch_sigill(write_pmc1));
+
+	/* We should still be able to read EBB regs though */
+	mfspr(SPRN_EBBHR);
+	mfspr(SPRN_EBBRR);
+	mfspr(SPRN_BESCR);
+
+	return 0;
+}
+
+int main(void)
+{
+	return test_harness(close_clears_pmcc, "close_clears_pmcc");
+}

tools/testing/selftests/powerpc/pmu/ebb/cpu_event_pinned_vs_ebb_test.c

@@ -0,0 +1,93 @@
+/*
+ * Copyright 2014, Michael Ellerman, IBM Corp.
+ * Licensed under GPLv2.
+ */
+
+#include <signal.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <stdbool.h>
+#include <sys/types.h>
+#include <sys/wait.h>
+#include <unistd.h>
+
+#include "ebb.h"
+
+
+/*
+ * Tests a pinned cpu event vs an EBB - in that order. The pinned cpu event
+ * should remain and the EBB event should fail to enable.
+ */
+
+static int setup_cpu_event(struct event *event, int cpu)
+{
+	event_init_named(event, 0x400FA, "PM_RUN_INST_CMPL");
+
+	event->attr.pinned = 1;
+
+	event->attr.exclude_kernel = 1;
+	event->attr.exclude_hv = 1;
+	event->attr.exclude_idle = 1;
+
+	SKIP_IF(require_paranoia_below(1));
+	FAIL_IF(event_open_with_cpu(event, cpu));
+	FAIL_IF(event_enable(event));
+
+	return 0;
+}
+
+int cpu_event_pinned_vs_ebb(void)
+{
+	union pipe read_pipe, write_pipe;
+	struct event event;
+	int cpu, rc;
+	pid_t pid;
+
+	cpu = pick_online_cpu();
+	FAIL_IF(cpu < 0);
+	FAIL_IF(bind_to_cpu(cpu));
+
+	FAIL_IF(pipe(read_pipe.fds) == -1);
+	FAIL_IF(pipe(write_pipe.fds) == -1);
+
+	pid = fork();
+	if (pid == 0) {
+		/* NB order of pipes looks reversed */
+		exit(ebb_child(write_pipe, read_pipe));
+	}
+
+	/* We setup the cpu event first */
+	rc = setup_cpu_event(&event, cpu);
+	if (rc) {
+		kill_child_and_wait(pid);
+		return rc;
+	}
+
+	/* Signal the child to install its EBB event and wait */
+	if (sync_with_child(read_pipe, write_pipe))
+		/* If it fails, wait for it to exit */
+		goto wait;
+
+	/* Signal the child to run */
+	FAIL_IF(sync_with_child(read_pipe, write_pipe));
+
+wait:
+	/* We expect it to fail to read the event */
+	FAIL_IF(wait_for_child(pid) != 2);
+
+	FAIL_IF(event_disable(&event));
+	FAIL_IF(event_read(&event));
+
+	event_report(&event);
+
+	/* The cpu event should have run */
+	FAIL_IF(event.result.value == 0);
+	FAIL_IF(event.result.enabled != event.result.running);
+
+	return 0;
+}
+
+int main(void)
+{
+	return test_harness(cpu_event_pinned_vs_ebb, "cpu_event_pinned_vs_ebb");
+}

tools/testing/selftests/powerpc/pmu/ebb/cpu_event_vs_ebb_test.c

@@ -0,0 +1,89 @@
+/*
+ * Copyright 2014, Michael Ellerman, IBM Corp.
+ * Licensed under GPLv2.
+ */
+
+#include <signal.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <stdbool.h>
+#include <sys/types.h>
+#include <sys/wait.h>
+#include <unistd.h>
+
+#include "ebb.h"
+
+
+/*
+ * Tests a cpu event vs an EBB - in that order. The EBB should force the cpu
+ * event off the PMU.
+ */
+
+static int setup_cpu_event(struct event *event, int cpu)
+{
+	event_init_named(event, 0x400FA, "PM_RUN_INST_CMPL");
+
+	event->attr.exclude_kernel = 1;
+	event->attr.exclude_hv = 1;
+	event->attr.exclude_idle = 1;
+
+	SKIP_IF(require_paranoia_below(1));
+	FAIL_IF(event_open_with_cpu(event, cpu));
+	FAIL_IF(event_enable(event));
+
+	return 0;
+}
+
+int cpu_event_vs_ebb(void)
+{
+	union pipe read_pipe, write_pipe;
+	struct event event;
+	int cpu, rc;
+	pid_t pid;
+
+	cpu = pick_online_cpu();
+	FAIL_IF(cpu < 0);
+	FAIL_IF(bind_to_cpu(cpu));
+
+	FAIL_IF(pipe(read_pipe.fds) == -1);
+	FAIL_IF(pipe(write_pipe.fds) == -1);
+
+	pid = fork();
+	if (pid == 0) {
+		/* NB order of pipes looks reversed */
+		exit(ebb_child(write_pipe, read_pipe));
+	}
+
+	/* We setup the cpu event first */
+	rc = setup_cpu_event(&event, cpu);
+	if (rc) {
+		kill_child_and_wait(pid);
+		return rc;
+	}
+
+	/* Signal the child to install its EBB event and wait */
+	if (sync_with_child(read_pipe, write_pipe))
+		/* If it fails, wait for it to exit */
+		goto wait;
+
+	/* Signal the child to run */
+	FAIL_IF(sync_with_child(read_pipe, write_pipe));
+
+wait:
+	/* We expect the child to succeed */
+	FAIL_IF(wait_for_child(pid));
+
+	FAIL_IF(event_disable(&event));
+	FAIL_IF(event_read(&event));
+
+	event_report(&event);
+
+	/* The cpu event may have run */
+
+	return 0;
+}
+
+int main(void)
+{
+	return test_harness(cpu_event_vs_ebb, "cpu_event_vs_ebb");
+}

tools/testing/selftests/powerpc/pmu/ebb/cycles_test.c

@@ -0,0 +1,58 @@
+/*
+ * Copyright 2014, Michael Ellerman, IBM Corp.
+ * Licensed under GPLv2.
+ */
+
+#include <stdio.h>
+#include <stdlib.h>
+
+#include "ebb.h"
+
+
+/*
+ * Basic test that counts user cycles and takes EBBs.
+ */
+int cycles(void)
+{
+	struct event event;
+
+	event_init_named(&event, 0x1001e, "cycles");
+	event_leader_ebb_init(&event);
+
+	event.attr.exclude_kernel = 1;
+	event.attr.exclude_hv = 1;
+	event.attr.exclude_idle = 1;
+
+	FAIL_IF(event_open(&event));
+
+	ebb_enable_pmc_counting(1);
+	setup_ebb_handler(standard_ebb_callee);
+	ebb_global_enable();
+	FAIL_IF(ebb_event_enable(&event));
+
+	mtspr(SPRN_PMC1, pmc_sample_period(sample_period));
+
+	while (ebb_state.stats.ebb_count < 10) {
+		FAIL_IF(core_busy_loop());
+		FAIL_IF(ebb_check_mmcr0());
+	}
+
+	ebb_global_disable();
+	ebb_freeze_pmcs();
+
+	count_pmc(1, sample_period);
+
+	dump_ebb_state();
+
+	event_close(&event);
+
+	FAIL_IF(ebb_state.stats.ebb_count == 0);
+	FAIL_IF(!ebb_check_count(1, sample_period, 100));
+
+	return 0;
+}
+
+int main(void)
+{
+	return test_harness(cycles, "cycles");
+}

tools/testing/selftests/powerpc/pmu/ebb/cycles_with_freeze_test.c

@@ -0,0 +1,117 @@
+/*
+ * Copyright 2014, Michael Ellerman, IBM Corp.
+ * Licensed under GPLv2.
+ */
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <stdbool.h>
+
+#include "ebb.h"
+
+
+/*
+ * Test of counting cycles while using MMCR0_FC (freeze counters) to only count
+ * parts of the code. This is complicated by the fact that FC is set by the
+ * hardware when the event overflows. We may take the EBB after we have set FC,
+ * so we have to be careful about whether we clear FC at the end of the EBB
+ * handler or not.
+ */
+
+static bool counters_frozen = false;
+static int ebbs_while_frozen = 0;
+
+static void ebb_callee(void)
+{
+	uint64_t mask, val;
+
+	mask = MMCR0_PMAO | MMCR0_FC;
+
+	val = mfspr(SPRN_BESCR);
+	if (!(val & BESCR_PMEO)) {
+		ebb_state.stats.spurious++;
+		goto out;
+	}
+
+	ebb_state.stats.ebb_count++;
+	trace_log_counter(ebb_state.trace, ebb_state.stats.ebb_count);
+
+	val = mfspr(SPRN_MMCR0);
+	trace_log_reg(ebb_state.trace, SPRN_MMCR0, val);
+
+	if (counters_frozen) {
+		trace_log_string(ebb_state.trace, "frozen");
+		ebbs_while_frozen++;
+		mask &= ~MMCR0_FC;
+	}
+
+	count_pmc(1, sample_period);
+out:
+	reset_ebb_with_clear_mask(mask);
+}
+
+int cycles_with_freeze(void)
+{
+	struct event event;
+	uint64_t val;
+	bool fc_cleared;
+
+	event_init_named(&event, 0x1001e, "cycles");
+	event_leader_ebb_init(&event);
+
+	event.attr.exclude_kernel = 1;
+	event.attr.exclude_hv = 1;
+	event.attr.exclude_idle = 1;
+
+	FAIL_IF(event_open(&event));
+
+	setup_ebb_handler(ebb_callee);
+	ebb_global_enable();
+	FAIL_IF(ebb_event_enable(&event));
+
+	mtspr(SPRN_PMC1, pmc_sample_period(sample_period));
+
+	fc_cleared = false;
+
+	/* Make sure we loop until we take at least one EBB */
+	while ((ebb_state.stats.ebb_count < 20 && !fc_cleared) ||
+		ebb_state.stats.ebb_count < 1)
+	{
+		counters_frozen = false;
+		mb();
+		mtspr(SPRN_MMCR0, mfspr(SPRN_MMCR0) & ~MMCR0_FC);
+
+		FAIL_IF(core_busy_loop());
+
+		counters_frozen = true;
+		mb();
+		mtspr(SPRN_MMCR0, mfspr(SPRN_MMCR0) |  MMCR0_FC);
+
+		val = mfspr(SPRN_MMCR0);
+		if (! (val & MMCR0_FC)) {
+			printf("Outside of loop, FC NOT set MMCR0 0x%lx\n", val);
+			fc_cleared = true;
+		}
+	}
+
+	ebb_global_disable();
+	ebb_freeze_pmcs();
+
+	count_pmc(1, sample_period);
+
+	dump_ebb_state();
+
+	printf("EBBs while frozen %d\n", ebbs_while_frozen);
+
+	event_close(&event);
+
+	FAIL_IF(ebb_state.stats.ebb_count == 0);
+	FAIL_IF(fc_cleared);
+
+	return 0;
+}
+
+int main(void)
+{
+	return test_harness(cycles_with_freeze, "cycles_with_freeze");
+}

tools/testing/selftests/powerpc/pmu/ebb/cycles_with_mmcr2_test.c

@@ -0,0 +1,91 @@
+/*
+ * Copyright 2014, Michael Ellerman, IBM Corp.
+ * Licensed under GPLv2.
+ */
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <stdbool.h>
+
+#include "ebb.h"
+
+
+/*
+ * Test of counting cycles while manipulating the user accessible bits in MMCR2.
+ */
+
+/* We use two values because the first freezes PMC1 and so we would get no EBBs */
+#define MMCR2_EXPECTED_1 0x4020100804020000UL /* (FC1P|FC2P|FC3P|FC4P|FC5P|FC6P) */
+#define MMCR2_EXPECTED_2 0x0020100804020000UL /* (     FC2P|FC3P|FC4P|FC5P|FC6P) */
+
+
+int cycles_with_mmcr2(void)
+{
+	struct event event;
+	uint64_t val, expected[2], actual;
+	int i;
+	bool bad_mmcr2;
+
+	event_init_named(&event, 0x1001e, "cycles");
+	event_leader_ebb_init(&event);
+
+	event.attr.exclude_kernel = 1;
+	event.attr.exclude_hv = 1;
+	event.attr.exclude_idle = 1;
+
+	FAIL_IF(event_open(&event));
+
+	ebb_enable_pmc_counting(1);
+	setup_ebb_handler(standard_ebb_callee);
+	ebb_global_enable();
+
+	FAIL_IF(ebb_event_enable(&event));
+
+	mtspr(SPRN_PMC1, pmc_sample_period(sample_period));
+
+	/* XXX Set of MMCR2 must be after enable */
+	expected[0] = MMCR2_EXPECTED_1;
+	expected[1] = MMCR2_EXPECTED_2;
+	i = 0;
+	bad_mmcr2 = false;
+
+	/* Make sure we loop until we take at least one EBB */
+	while ((ebb_state.stats.ebb_count < 20 && !bad_mmcr2) ||
+		ebb_state.stats.ebb_count < 1)
+	{
+		mtspr(SPRN_MMCR2, expected[i % 2]);
+
+		FAIL_IF(core_busy_loop());
+
+		val = mfspr(SPRN_MMCR2);
+		if (val != expected[i % 2]) {
+			bad_mmcr2 = true;
+			actual = val;
+		}
+
+		i++;
+	}
+
+	ebb_global_disable();
+	ebb_freeze_pmcs();
+
+	count_pmc(1, sample_period);
+
+	dump_ebb_state();
+
+	event_close(&event);
+
+	FAIL_IF(ebb_state.stats.ebb_count == 0);
+
+	if (bad_mmcr2)
+		printf("Bad MMCR2 value seen is 0x%lx\n", actual);
+
+	FAIL_IF(bad_mmcr2);
+
+	return 0;
+}
+
+int main(void)
+{
+	return test_harness(cycles_with_mmcr2, "cycles_with_mmcr2");
+}

tools/testing/selftests/powerpc/pmu/ebb/ebb.c

@@ -0,0 +1,478 @@
+/*
+ * Copyright 2014, Michael Ellerman, IBM Corp.
+ * Licensed under GPLv2.
+ */
+
+#define _GNU_SOURCE	/* For CPU_ZERO etc. */
+
+#include <sched.h>
+#include <sys/wait.h>
+#include <setjmp.h>
+#include <signal.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <sys/ioctl.h>
+
+#include "trace.h"
+#include "reg.h"
+#include "ebb.h"
+
+
+void (*ebb_user_func)(void);
+
+void ebb_hook(void)
+{
+	if (ebb_user_func)
+		ebb_user_func();
+}
+
+struct ebb_state ebb_state;
+
+u64 sample_period = 0x40000000ull;
+
+void reset_ebb_with_clear_mask(unsigned long mmcr0_clear_mask)
+{
+	u64 val;
+
+	/* 2) clear MMCR0[PMAO] - docs say BESCR[PMEO] should do this */
+	/* 3) set MMCR0[PMAE]	- docs say BESCR[PME] should do this */
+	val = mfspr(SPRN_MMCR0);
+	mtspr(SPRN_MMCR0, (val & ~mmcr0_clear_mask) | MMCR0_PMAE);
+
+	/* 4) clear BESCR[PMEO] */
+	mtspr(SPRN_BESCRR, BESCR_PMEO);
+
+	/* 5) set BESCR[PME] */
+	mtspr(SPRN_BESCRS, BESCR_PME);
+
+	/* 6) rfebb 1 - done in our caller */
+}
+
+void reset_ebb(void)
+{
+	reset_ebb_with_clear_mask(MMCR0_PMAO | MMCR0_FC);
+}
+
+/* Called outside of the EBB handler to check MMCR0 is sane */
+int ebb_check_mmcr0(void)
+{
+	u64 val;
+
+	val = mfspr(SPRN_MMCR0);
+	if ((val & (MMCR0_FC | MMCR0_PMAO)) == MMCR0_FC) {
+		/* It's OK if we see FC & PMAO, but not FC by itself */
+		printf("Outside of loop, only FC set 0x%llx\n", val);
+		return 1;
+	}
+
+	return 0;
+}
+
+bool ebb_check_count(int pmc, u64 sample_period, int fudge)
+{
+	u64 count, upper, lower;
+
+	count = ebb_state.stats.pmc_count[PMC_INDEX(pmc)];
+
+	lower = ebb_state.stats.ebb_count * (sample_period - fudge);
+
+	if (count < lower) {
+		printf("PMC%d count (0x%llx) below lower limit 0x%llx (-0x%llx)\n",
+			pmc, count, lower, lower - count);
+		return false;
+	}
+
+	upper = ebb_state.stats.ebb_count * (sample_period + fudge);
+
+	if (count > upper) {
+		printf("PMC%d count (0x%llx) above upper limit 0x%llx (+0x%llx)\n",
+			pmc, count, upper, count - upper);
+		return false;
+	}
+
+	printf("PMC%d count (0x%llx) is between 0x%llx and 0x%llx delta +0x%llx/-0x%llx\n",
+		pmc, count, lower, upper, count - lower, upper - count);
+
+	return true;
+}
+
+void standard_ebb_callee(void)
+{
+	int found, i;
+	u64 val;
+
+	val = mfspr(SPRN_BESCR);
+	if (!(val & BESCR_PMEO)) {
+		ebb_state.stats.spurious++;
+		goto out;
+	}
+
+	ebb_state.stats.ebb_count++;
+	trace_log_counter(ebb_state.trace, ebb_state.stats.ebb_count);
+
+	val = mfspr(SPRN_MMCR0);
+	trace_log_reg(ebb_state.trace, SPRN_MMCR0, val);
+
+	found = 0;
+	for (i = 1; i <= 6; i++) {
+		if (ebb_state.pmc_enable[PMC_INDEX(i)])
+			found += count_pmc(i, sample_period);
+	}
+
+	if (!found)
+		ebb_state.stats.no_overflow++;
+
+out:
+	reset_ebb();
+}
+
+extern void ebb_handler(void);
+
+void setup_ebb_handler(void (*callee)(void))
+{
+	u64 entry;
+
+#if defined(_CALL_ELF) && _CALL_ELF == 2
+	entry = (u64)ebb_handler;
+#else
+	struct opd
+	{
+	    u64 entry;
+	    u64 toc;
+	} *opd;
+
+	opd = (struct opd *)ebb_handler;
+	entry = opd->entry;
+#endif
+	printf("EBB Handler is at %#llx\n", entry);
+
+	ebb_user_func = callee;
+
+	/* Ensure ebb_user_func is set before we set the handler */
+	mb();
+	mtspr(SPRN_EBBHR, entry);
+
+	/* Make sure the handler is set before we return */
+	mb();
+}
+
+void clear_ebb_stats(void)
+{
+	memset(&ebb_state.stats, 0, sizeof(ebb_state.stats));
+}
+
+void dump_summary_ebb_state(void)
+{
+	printf("ebb_state:\n"			\
+	       "  ebb_count    = %d\n"		\
+	       "  spurious     = %d\n"		\
+	       "  negative     = %d\n"		\
+	       "  no_overflow  = %d\n"		\
+	       "  pmc[1] count = 0x%llx\n"	\
+	       "  pmc[2] count = 0x%llx\n"	\
+	       "  pmc[3] count = 0x%llx\n"	\
+	       "  pmc[4] count = 0x%llx\n"	\
+	       "  pmc[5] count = 0x%llx\n"	\
+	       "  pmc[6] count = 0x%llx\n",
+		ebb_state.stats.ebb_count, ebb_state.stats.spurious,
+		ebb_state.stats.negative, ebb_state.stats.no_overflow,
+		ebb_state.stats.pmc_count[0], ebb_state.stats.pmc_count[1],
+		ebb_state.stats.pmc_count[2], ebb_state.stats.pmc_count[3],
+		ebb_state.stats.pmc_count[4], ebb_state.stats.pmc_count[5]);
+}
+
+static char *decode_mmcr0(u32 value)
+{
+	static char buf[16];
+
+	buf[0] = '\0';
+
+	if (value & (1 << 31))
+		strcat(buf, "FC ");
+	if (value & (1 << 26))
+		strcat(buf, "PMAE ");
+	if (value & (1 << 7))
+		strcat(buf, "PMAO ");
+
+	return buf;
+}
+
+static char *decode_bescr(u64 value)
+{
+	static char buf[16];
+
+	buf[0] = '\0';
+
+	if (value & (1ull << 63))
+		strcat(buf, "GE ");
+	if (value & (1ull << 32))
+		strcat(buf, "PMAE ");
+	if (value & 1)
+		strcat(buf, "PMAO ");
+
+	return buf;
+}
+
+void dump_ebb_hw_state(void)
+{
+	u64 bescr;
+	u32 mmcr0;
+
+	mmcr0 = mfspr(SPRN_MMCR0);
+	bescr = mfspr(SPRN_BESCR);
+
+	printf("HW state:\n"		\
+	       "MMCR0 0x%016x %s\n"	\
+	       "MMCR2 0x%016lx\n"	\
+	       "EBBHR 0x%016lx\n"	\
+	       "BESCR 0x%016llx %s\n"	\
+	       "PMC1  0x%016lx\n"	\
+	       "PMC2  0x%016lx\n"	\
+	       "PMC3  0x%016lx\n"	\
+	       "PMC4  0x%016lx\n"	\
+	       "PMC5  0x%016lx\n"	\
+	       "PMC6  0x%016lx\n"	\
+	       "SIAR  0x%016lx\n",
+	       mmcr0, decode_mmcr0(mmcr0), mfspr(SPRN_MMCR2),
+	       mfspr(SPRN_EBBHR), bescr, decode_bescr(bescr),
+	       mfspr(SPRN_PMC1), mfspr(SPRN_PMC2), mfspr(SPRN_PMC3),
+	       mfspr(SPRN_PMC4), mfspr(SPRN_PMC5), mfspr(SPRN_PMC6),
+	       mfspr(SPRN_SIAR));
+}
+
+void dump_ebb_state(void)
+{
+	dump_summary_ebb_state();
+
+	dump_ebb_hw_state();
+
+	trace_buffer_print(ebb_state.trace);
+}
+
+int count_pmc(int pmc, uint32_t sample_period)
+{
+	uint32_t start_value;
+	u64 val;
+
+	/* 0) Read PMC */
+	start_value = pmc_sample_period(sample_period);
+
+	val = read_pmc(pmc);
+	if (val < start_value)
+		ebb_state.stats.negative++;
+	else
+		ebb_state.stats.pmc_count[PMC_INDEX(pmc)] += val - start_value;
+
+	trace_log_reg(ebb_state.trace, SPRN_PMC1 + pmc - 1, val);
+
+	/* 1) Reset PMC */
+	write_pmc(pmc, start_value);
+
+	/* Report if we overflowed */
+	return val >= COUNTER_OVERFLOW;
+}
+
+int ebb_event_enable(struct event *e)
+{
+	int rc;
+
+	/* Ensure any SPR writes are ordered vs us */
+	mb();
+
+	rc = ioctl(e->fd, PERF_EVENT_IOC_ENABLE);
+	if (rc)
+		return rc;
+
+	rc = event_read(e);
+
+	/* Ditto */
+	mb();
+
+	return rc;
+}
+
+void ebb_freeze_pmcs(void)
+{
+	mtspr(SPRN_MMCR0, mfspr(SPRN_MMCR0) | MMCR0_FC);
+	mb();
+}
+
+void ebb_unfreeze_pmcs(void)
+{
+	/* Unfreeze counters */
+	mtspr(SPRN_MMCR0, mfspr(SPRN_MMCR0) & ~MMCR0_FC);
+	mb();
+}
+
+void ebb_global_enable(void)
+{
+	/* Enable EBBs globally and PMU EBBs */
+	mtspr(SPRN_BESCR, 0x8000000100000000ull);
+	mb();
+}
+
+void ebb_global_disable(void)
+{
+	/* Disable EBBs & freeze counters, events are still scheduled */
+	mtspr(SPRN_BESCRR, BESCR_PME);
+	mb();
+}
+
+void event_ebb_init(struct event *e)
+{
+	e->attr.config |= (1ull << 63);
+}
+
+void event_bhrb_init(struct event *e, unsigned ifm)
+{
+	e->attr.config |= (1ull << 62) | ((u64)ifm << 60);
+}
+
+void event_leader_ebb_init(struct event *e)
+{
+	event_ebb_init(e);
+
+	e->attr.exclusive = 1;
+	e->attr.pinned = 1;
+}
+
+int ebb_child(union pipe read_pipe, union pipe write_pipe)
+{
+	struct event event;
+	uint64_t val;
+
+	FAIL_IF(wait_for_parent(read_pipe));
+
+	event_init_named(&event, 0x1001e, "cycles");
+	event_leader_ebb_init(&event);
+
+	event.attr.exclude_kernel = 1;
+	event.attr.exclude_hv = 1;
+	event.attr.exclude_idle = 1;
+
+	FAIL_IF(event_open(&event));
+
+	ebb_enable_pmc_counting(1);
+	setup_ebb_handler(standard_ebb_callee);
+	ebb_global_enable();
+
+	FAIL_IF(event_enable(&event));
+
+	if (event_read(&event)) {
+		/*
+		 * Some tests expect to fail here, so don't report an error on
+		 * this line, and return a distinguisable error code. Tell the
+		 * parent an error happened.
+		 */
+		notify_parent_of_error(write_pipe);
+		return 2;
+	}
+
+	mtspr(SPRN_PMC1, pmc_sample_period(sample_period));
+
+	FAIL_IF(notify_parent(write_pipe));
+	FAIL_IF(wait_for_parent(read_pipe));
+	FAIL_IF(notify_parent(write_pipe));
+
+	while (ebb_state.stats.ebb_count < 20) {
+		FAIL_IF(core_busy_loop());
+
+		/* To try and hit SIGILL case */
+		val  = mfspr(SPRN_MMCRA);
+		val |= mfspr(SPRN_MMCR2);
+		val |= mfspr(SPRN_MMCR0);
+	}
+
+	ebb_global_disable();
+	ebb_freeze_pmcs();
+
+	count_pmc(1, sample_period);
+
+	dump_ebb_state();
+
+	event_close(&event);
+
+	FAIL_IF(ebb_state.stats.ebb_count == 0);
+
+	return 0;
+}
+
+static jmp_buf setjmp_env;
+
+static void sigill_handler(int signal)
+{
+	printf("Took sigill\n");
+	longjmp(setjmp_env, 1);
+}
+
+static struct sigaction sigill_action = {
+	.sa_handler = sigill_handler,
+};
+
+int catch_sigill(void (*func)(void))
+{
+	if (sigaction(SIGILL, &sigill_action, NULL)) {
+		perror("sigaction");
+		return 1;
+	}
+
+	if (setjmp(setjmp_env) == 0) {
+		func();
+		return 1;
+	}
+
+	return 0;
+}
+
+void write_pmc1(void)
+{
+	mtspr(SPRN_PMC1, 0);
+}
+
+void write_pmc(int pmc, u64 value)
+{
+	switch (pmc) {
+		case 1: mtspr(SPRN_PMC1, value); break;
+		case 2: mtspr(SPRN_PMC2, value); break;
+		case 3: mtspr(SPRN_PMC3, value); break;
+		case 4: mtspr(SPRN_PMC4, value); break;
+		case 5: mtspr(SPRN_PMC5, value); break;
+		case 6: mtspr(SPRN_PMC6, value); break;
+	}
+}
+
+u64 read_pmc(int pmc)
+{
+	switch (pmc) {
+		case 1: return mfspr(SPRN_PMC1);
+		case 2: return mfspr(SPRN_PMC2);
+		case 3: return mfspr(SPRN_PMC3);
+		case 4: return mfspr(SPRN_PMC4);
+		case 5: return mfspr(SPRN_PMC5);
+		case 6: return mfspr(SPRN_PMC6);
+	}
+
+	return 0;
+}
+
+static void term_handler(int signal)
+{
+	dump_summary_ebb_state();
+	dump_ebb_hw_state();
+	abort();
+}
+
+struct sigaction term_action = {
+	.sa_handler = term_handler,
+};
+
+static void __attribute__((constructor)) ebb_init(void)
+{
+	clear_ebb_stats();
+
+	if (sigaction(SIGTERM, &term_action, NULL))
+		perror("sigaction");
+
+	ebb_state.trace = trace_buffer_allocate(1 * 1024 * 1024);
+}

tools/testing/selftests/powerpc/pmu/ebb/ebb.h

@@ -0,0 +1,77 @@
+/*
+ * Copyright 2014, Michael Ellerman, IBM Corp.
+ * Licensed under GPLv2.
+ */
+
+#ifndef _SELFTESTS_POWERPC_PMU_EBB_EBB_H
+#define _SELFTESTS_POWERPC_PMU_EBB_EBB_H
+
+#include "../event.h"
+#include "../lib.h"
+#include "trace.h"
+#include "reg.h"
+
+#define PMC_INDEX(pmc)	((pmc)-1)
+
+#define NUM_PMC_VALUES	128
+
+struct ebb_state
+{
+	struct {
+		u64 pmc_count[6];
+		volatile int ebb_count;
+		int spurious;
+		int negative;
+		int no_overflow;
+	} stats;
+
+	bool pmc_enable[6];
+	struct trace_buffer *trace;
+};
+
+extern struct ebb_state ebb_state;
+
+#define COUNTER_OVERFLOW 0x80000000ull
+
+static inline uint32_t pmc_sample_period(uint32_t value)
+{
+	return COUNTER_OVERFLOW - value;
+}
+
+static inline void ebb_enable_pmc_counting(int pmc)
+{
+	ebb_state.pmc_enable[PMC_INDEX(pmc)] = true;
+}
+
+bool ebb_check_count(int pmc, u64 sample_period, int fudge);
+void event_leader_ebb_init(struct event *e);
+void event_ebb_init(struct event *e);
+void event_bhrb_init(struct event *e, unsigned ifm);
+void setup_ebb_handler(void (*callee)(void));
+void standard_ebb_callee(void);
+int ebb_event_enable(struct event *e);
+void ebb_global_enable(void);
+void ebb_global_disable(void);
+void ebb_freeze_pmcs(void);
+void ebb_unfreeze_pmcs(void);
+void event_ebb_init(struct event *e);
+void event_leader_ebb_init(struct event *e);
+int count_pmc(int pmc, uint32_t sample_period);
+void dump_ebb_state(void);
+void dump_summary_ebb_state(void);
+void dump_ebb_hw_state(void);
+void clear_ebb_stats(void);
+void write_pmc(int pmc, u64 value);
+u64 read_pmc(int pmc);
+void reset_ebb_with_clear_mask(unsigned long mmcr0_clear_mask);
+void reset_ebb(void);
+int ebb_check_mmcr0(void);
+
+extern u64 sample_period;
+
+int core_busy_loop(void);
+int ebb_child(union pipe read_pipe, union pipe write_pipe);
+int catch_sigill(void (*func)(void));
+void write_pmc1(void);
+
+#endif /* _SELFTESTS_POWERPC_PMU_EBB_EBB_H */

tools/testing/selftests/powerpc/pmu/ebb/ebb_handler.S

@@ -0,0 +1,365 @@
+/*
+ * Copyright 2014, Michael Ellerman, IBM Corp.
+ * Licensed under GPLv2.
+ */
+
+#include <ppc-asm.h>
+#include "reg.h"
+
+
+/* ppc-asm.h defines most of the reg aliases, but not r1/r2. */
+#define r1 1
+#define r2 2
+
+#define RFEBB   .long 0x4c000924
+
+/* Stack layout:
+ *
+ *                   ^
+ *  User stack       |
+ *  Back chain ------+	<- r1		<-------+
+ *  ...						|
+ *  Red zone / ABI Gap				|
+ *  ...						|
+ *  vr63	<+				|
+ *  vr0		 |				|
+ *  VSCR	 |				|
+ *  FSCR	 |				|
+ *  r31		 | Save area			|
+ *  r0		 |				|
+ *  XER		 |				|
+ *  CTR		 |				|
+ *  LR		 |				|
+ *  CCR		<+				|
+ *  ...		<+				|
+ *  LR		 | Caller frame			|
+ *  CCR		 |				|
+ *  Back chain	<+	<- updated r1	--------+
+ *
+ */
+
+#if defined(_CALL_ELF) && _CALL_ELF == 2
+#define ABIGAP		512
+#else
+#define ABIGAP		288
+#endif
+
+#define NR_GPR		32
+#define NR_SPR		6
+#define NR_VSR		64
+
+#define SAVE_AREA	((NR_GPR + NR_SPR) * 8 + (NR_VSR * 16))
+#define CALLER_FRAME	112
+
+#define STACK_FRAME	(ABIGAP + SAVE_AREA + CALLER_FRAME)
+
+#define CCR_SAVE	(CALLER_FRAME)
+#define LR_SAVE		(CCR_SAVE + 8)
+#define CTR_SAVE	(LR_SAVE  + 8)
+#define XER_SAVE	(CTR_SAVE + 8)
+#define GPR_SAVE(n)	(XER_SAVE + 8 + (8 * n))
+#define FSCR_SAVE	(GPR_SAVE(31) + 8)
+#define VSCR_SAVE	(FSCR_SAVE + 8)
+#define VSR_SAVE(n)	(VSCR_SAVE + 8 + (16 * n))
+
+#define SAVE_GPR(n)	std n,GPR_SAVE(n)(r1)
+#define REST_GPR(n)	ld  n,GPR_SAVE(n)(r1)
+#define TRASH_GPR(n)	lis n,0xaaaa
+
+#define SAVE_VSR(n, b)	li b, VSR_SAVE(n); stxvd2x n,b,r1
+#define LOAD_VSR(n, b)	li b, VSR_SAVE(n); lxvd2x  n,b,r1
+
+#define LOAD_REG_IMMEDIATE(reg,expr)	\
+	lis     reg,(expr)@highest;	\
+	ori     reg,reg,(expr)@higher;	\
+	rldicr  reg,reg,32,31;		\
+	oris    reg,reg,(expr)@h;	\
+	ori     reg,reg,(expr)@l;
+
+
+#if defined(_CALL_ELF) && _CALL_ELF == 2
+#define ENTRY_POINT(name) \
+	.type FUNC_NAME(name),@function; \
+	.globl FUNC_NAME(name); \
+	FUNC_NAME(name):
+
+#define RESTORE_TOC(name)	\
+	/* Restore our TOC pointer using our entry point */	\
+	LOAD_REG_IMMEDIATE(r12, name)				\
+0:	addis	r2,r12,(.TOC.-0b)@ha;				\
+	addi	r2,r2,(.TOC.-0b)@l;
+
+#else
+#define ENTRY_POINT(name) FUNC_START(name)
+#define RESTORE_TOC(name)	\
+	/* Restore our TOC pointer via our opd entry */	\
+	LOAD_REG_IMMEDIATE(r2, name)			\
+	ld      r2,8(r2);
+#endif
+
+    .text
+
+ENTRY_POINT(ebb_handler)
+    stdu    r1,-STACK_FRAME(r1)
+    SAVE_GPR(0)
+    mflr    r0
+    std     r0,LR_SAVE(r1)
+    mfcr    r0
+    std     r0,CCR_SAVE(r1)
+    mfctr   r0
+    std     r0,CTR_SAVE(r1)
+    mfxer   r0
+    std     r0,XER_SAVE(r1)
+    SAVE_GPR(2)
+    SAVE_GPR(3)
+    SAVE_GPR(4)
+    SAVE_GPR(5)
+    SAVE_GPR(6)
+    SAVE_GPR(7)
+    SAVE_GPR(8)
+    SAVE_GPR(9)
+    SAVE_GPR(10)
+    SAVE_GPR(11)
+    SAVE_GPR(12)
+    SAVE_GPR(13)
+    SAVE_GPR(14)
+    SAVE_GPR(15)
+    SAVE_GPR(16)
+    SAVE_GPR(17)
+    SAVE_GPR(18)
+    SAVE_GPR(19)
+    SAVE_GPR(20)
+    SAVE_GPR(21)
+    SAVE_GPR(22)
+    SAVE_GPR(23)
+    SAVE_GPR(24)
+    SAVE_GPR(25)
+    SAVE_GPR(26)
+    SAVE_GPR(27)
+    SAVE_GPR(28)
+    SAVE_GPR(29)
+    SAVE_GPR(30)
+    SAVE_GPR(31)
+    SAVE_VSR(0, r3)
+    mffs     f0
+    stfd     f0, FSCR_SAVE(r1)
+    mfvscr   f0
+    stfd     f0, VSCR_SAVE(r1)
+    SAVE_VSR(1,  r3)
+    SAVE_VSR(2,  r3)
+    SAVE_VSR(3,  r3)
+    SAVE_VSR(4,  r3)
+    SAVE_VSR(5,  r3)
+    SAVE_VSR(6,  r3)
+    SAVE_VSR(7,  r3)
+    SAVE_VSR(8,  r3)
+    SAVE_VSR(9,  r3)
+    SAVE_VSR(10, r3)
+    SAVE_VSR(11, r3)
+    SAVE_VSR(12, r3)
+    SAVE_VSR(13, r3)
+    SAVE_VSR(14, r3)
+    SAVE_VSR(15, r3)
+    SAVE_VSR(16, r3)
+    SAVE_VSR(17, r3)
+    SAVE_VSR(18, r3)
+    SAVE_VSR(19, r3)
+    SAVE_VSR(20, r3)
+    SAVE_VSR(21, r3)
+    SAVE_VSR(22, r3)
+    SAVE_VSR(23, r3)
+    SAVE_VSR(24, r3)
+    SAVE_VSR(25, r3)
+    SAVE_VSR(26, r3)
+    SAVE_VSR(27, r3)
+    SAVE_VSR(28, r3)
+    SAVE_VSR(29, r3)
+    SAVE_VSR(30, r3)
+    SAVE_VSR(31, r3)
+    SAVE_VSR(32, r3)
+    SAVE_VSR(33, r3)
+    SAVE_VSR(34, r3)
+    SAVE_VSR(35, r3)
+    SAVE_VSR(36, r3)
+    SAVE_VSR(37, r3)
+    SAVE_VSR(38, r3)
+    SAVE_VSR(39, r3)
+    SAVE_VSR(40, r3)
+    SAVE_VSR(41, r3)
+    SAVE_VSR(42, r3)
+    SAVE_VSR(43, r3)
+    SAVE_VSR(44, r3)
+    SAVE_VSR(45, r3)
+    SAVE_VSR(46, r3)
+    SAVE_VSR(47, r3)
+    SAVE_VSR(48, r3)
+    SAVE_VSR(49, r3)
+    SAVE_VSR(50, r3)
+    SAVE_VSR(51, r3)
+    SAVE_VSR(52, r3)
+    SAVE_VSR(53, r3)
+    SAVE_VSR(54, r3)
+    SAVE_VSR(55, r3)
+    SAVE_VSR(56, r3)
+    SAVE_VSR(57, r3)
+    SAVE_VSR(58, r3)
+    SAVE_VSR(59, r3)
+    SAVE_VSR(60, r3)
+    SAVE_VSR(61, r3)
+    SAVE_VSR(62, r3)
+    SAVE_VSR(63, r3)
+
+    TRASH_GPR(2)
+    TRASH_GPR(3)
+    TRASH_GPR(4)
+    TRASH_GPR(5)
+    TRASH_GPR(6)
+    TRASH_GPR(7)
+    TRASH_GPR(8)
+    TRASH_GPR(9)
+    TRASH_GPR(10)
+    TRASH_GPR(11)
+    TRASH_GPR(12)
+    TRASH_GPR(14)
+    TRASH_GPR(15)
+    TRASH_GPR(16)
+    TRASH_GPR(17)
+    TRASH_GPR(18)
+    TRASH_GPR(19)
+    TRASH_GPR(20)
+    TRASH_GPR(21)
+    TRASH_GPR(22)
+    TRASH_GPR(23)
+    TRASH_GPR(24)
+    TRASH_GPR(25)
+    TRASH_GPR(26)
+    TRASH_GPR(27)
+    TRASH_GPR(28)
+    TRASH_GPR(29)
+    TRASH_GPR(30)
+    TRASH_GPR(31)
+
+    RESTORE_TOC(ebb_handler)
+
+    /*
+     * r13 is our TLS pointer. We leave whatever value was in there when the
+     * EBB fired. That seems to be OK because once set the TLS pointer is not
+     * changed - but presumably that could change in future.
+     */
+
+    bl      ebb_hook
+    nop
+
+    /* r2 may be changed here but we don't care */
+
+    lfd      f0, FSCR_SAVE(r1)
+    mtfsf    0xff,f0
+    lfd      f0, VSCR_SAVE(r1)
+    mtvscr   f0
+    LOAD_VSR(0, r3)
+    LOAD_VSR(1,  r3)
+    LOAD_VSR(2,  r3)
+    LOAD_VSR(3,  r3)
+    LOAD_VSR(4,  r3)
+    LOAD_VSR(5,  r3)
+    LOAD_VSR(6,  r3)
+    LOAD_VSR(7,  r3)
+    LOAD_VSR(8,  r3)
+    LOAD_VSR(9,  r3)
+    LOAD_VSR(10, r3)
+    LOAD_VSR(11, r3)
+    LOAD_VSR(12, r3)
+    LOAD_VSR(13, r3)
+    LOAD_VSR(14, r3)
+    LOAD_VSR(15, r3)
+    LOAD_VSR(16, r3)
+    LOAD_VSR(17, r3)
+    LOAD_VSR(18, r3)
+    LOAD_VSR(19, r3)
+    LOAD_VSR(20, r3)
+    LOAD_VSR(21, r3)
+    LOAD_VSR(22, r3)
+    LOAD_VSR(23, r3)
+    LOAD_VSR(24, r3)
+    LOAD_VSR(25, r3)
+    LOAD_VSR(26, r3)
+    LOAD_VSR(27, r3)
+    LOAD_VSR(28, r3)
+    LOAD_VSR(29, r3)
+    LOAD_VSR(30, r3)
+    LOAD_VSR(31, r3)
+    LOAD_VSR(32, r3)
+    LOAD_VSR(33, r3)
+    LOAD_VSR(34, r3)
+    LOAD_VSR(35, r3)
+    LOAD_VSR(36, r3)
+    LOAD_VSR(37, r3)
+    LOAD_VSR(38, r3)
+    LOAD_VSR(39, r3)
+    LOAD_VSR(40, r3)
+    LOAD_VSR(41, r3)
+    LOAD_VSR(42, r3)
+    LOAD_VSR(43, r3)
+    LOAD_VSR(44, r3)
+    LOAD_VSR(45, r3)
+    LOAD_VSR(46, r3)
+    LOAD_VSR(47, r3)
+    LOAD_VSR(48, r3)
+    LOAD_VSR(49, r3)
+    LOAD_VSR(50, r3)
+    LOAD_VSR(51, r3)
+    LOAD_VSR(52, r3)
+    LOAD_VSR(53, r3)
+    LOAD_VSR(54, r3)
+    LOAD_VSR(55, r3)
+    LOAD_VSR(56, r3)
+    LOAD_VSR(57, r3)
+    LOAD_VSR(58, r3)
+    LOAD_VSR(59, r3)
+    LOAD_VSR(60, r3)
+    LOAD_VSR(61, r3)
+    LOAD_VSR(62, r3)
+    LOAD_VSR(63, r3)
+
+    ld      r0,XER_SAVE(r1)
+    mtxer   r0
+    ld      r0,CTR_SAVE(r1)
+    mtctr   r0
+    ld      r0,LR_SAVE(r1)
+    mtlr    r0
+    ld      r0,CCR_SAVE(r1)
+    mtcr    r0
+    REST_GPR(0)
+    REST_GPR(2)
+    REST_GPR(3)
+    REST_GPR(4)
+    REST_GPR(5)
+    REST_GPR(6)
+    REST_GPR(7)
+    REST_GPR(8)
+    REST_GPR(9)
+    REST_GPR(10)
+    REST_GPR(11)
+    REST_GPR(12)
+    REST_GPR(13)
+    REST_GPR(14)
+    REST_GPR(15)
+    REST_GPR(16)
+    REST_GPR(17)
+    REST_GPR(18)
+    REST_GPR(19)
+    REST_GPR(20)
+    REST_GPR(21)
+    REST_GPR(22)
+    REST_GPR(23)
+    REST_GPR(24)
+    REST_GPR(25)
+    REST_GPR(26)
+    REST_GPR(27)
+    REST_GPR(28)
+    REST_GPR(29)
+    REST_GPR(30)
+    REST_GPR(31)
+    addi    r1,r1,STACK_FRAME
+    RFEBB
+FUNC_END(ebb_handler)

tools/testing/selftests/powerpc/pmu/ebb/ebb_on_child_test.c

@@ -0,0 +1,86 @@
+/*
+ * Copyright 2014, Michael Ellerman, IBM Corp.
+ * Licensed under GPLv2.
+ */
+
+#include <signal.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <stdbool.h>
+#include <sys/types.h>
+#include <sys/wait.h>
+#include <unistd.h>
+
+#include "ebb.h"
+
+
+/*
+ * Tests we can setup an EBB on our child. Nothing interesting happens, because
+ * even though the event is enabled and running the child hasn't enabled the
+ * actual delivery of the EBBs.
+ */
+
+static int victim_child(union pipe read_pipe, union pipe write_pipe)
+{
+	int i;
+
+	FAIL_IF(wait_for_parent(read_pipe));
+	FAIL_IF(notify_parent(write_pipe));
+
+	/* Parent creates EBB event */
+
+	FAIL_IF(wait_for_parent(read_pipe));
+	FAIL_IF(notify_parent(write_pipe));
+
+	/* Check the EBB is enabled by writing PMC1 */
+	write_pmc1();
+
+	/* EBB event is enabled here */
+	for (i = 0; i < 1000000; i++) ;
+
+	return 0;
+}
+
+int ebb_on_child(void)
+{
+	union pipe read_pipe, write_pipe;
+	struct event event;
+	pid_t pid;
+
+	FAIL_IF(pipe(read_pipe.fds) == -1);
+	FAIL_IF(pipe(write_pipe.fds) == -1);
+
+	pid = fork();
+	if (pid == 0) {
+		/* NB order of pipes looks reversed */
+		exit(victim_child(write_pipe, read_pipe));
+	}
+
+	FAIL_IF(sync_with_child(read_pipe, write_pipe));
+
+	/* Child is running now */
+
+	event_init_named(&event, 0x1001e, "cycles");
+	event_leader_ebb_init(&event);
+
+	event.attr.exclude_kernel = 1;
+	event.attr.exclude_hv = 1;
+	event.attr.exclude_idle = 1;
+
+	FAIL_IF(event_open_with_pid(&event, pid));
+	FAIL_IF(ebb_event_enable(&event));
+
+	FAIL_IF(sync_with_child(read_pipe, write_pipe));
+
+	/* Child should just exit happily */
+	FAIL_IF(wait_for_child(pid));
+
+	event_close(&event);
+
+	return 0;
+}
+
+int main(void)
+{
+	return test_harness(ebb_on_child, "ebb_on_child");
+}

tools/testing/selftests/powerpc/pmu/ebb/ebb_on_willing_child_test.c

@@ -0,0 +1,92 @@
+/*
+ * Copyright 2014, Michael Ellerman, IBM Corp.
+ * Licensed under GPLv2.
+ */
+
+#include <signal.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <stdbool.h>
+#include <sys/types.h>
+#include <sys/wait.h>
+#include <unistd.h>
+
+#include "ebb.h"
+
+
+/*
+ * Tests we can setup an EBB on our child. The child expects this and enables
+ * EBBs, which are then delivered to the child, even though the event is
+ * created by the parent.
+ */
+
+static int victim_child(union pipe read_pipe, union pipe write_pipe)
+{
+	FAIL_IF(wait_for_parent(read_pipe));
+
+	/* Setup our EBB handler, before the EBB event is created */
+	ebb_enable_pmc_counting(1);
+	setup_ebb_handler(standard_ebb_callee);
+	ebb_global_enable();
+
+	FAIL_IF(notify_parent(write_pipe));
+
+	while (ebb_state.stats.ebb_count < 20) {
+		FAIL_IF(core_busy_loop());
+	}
+
+	ebb_global_disable();
+	ebb_freeze_pmcs();
+
+	count_pmc(1, sample_period);
+
+	dump_ebb_state();
+
+	FAIL_IF(ebb_state.stats.ebb_count == 0);
+
+	return 0;
+}
+
+/* Tests we can setup an EBB on our child - if it's expecting it */
+int ebb_on_willing_child(void)
+{
+	union pipe read_pipe, write_pipe;
+	struct event event;
+	pid_t pid;
+
+	FAIL_IF(pipe(read_pipe.fds) == -1);
+	FAIL_IF(pipe(write_pipe.fds) == -1);
+
+	pid = fork();
+	if (pid == 0) {
+		/* NB order of pipes looks reversed */
+		exit(victim_child(write_pipe, read_pipe));
+	}
+
+	/* Signal the child to setup its EBB handler */
+	FAIL_IF(sync_with_child(read_pipe, write_pipe));
+
+	/* Child is running now */
+
+	event_init_named(&event, 0x1001e, "cycles");
+	event_leader_ebb_init(&event);
+
+	event.attr.exclude_kernel = 1;
+	event.attr.exclude_hv = 1;
+	event.attr.exclude_idle = 1;
+
+	FAIL_IF(event_open_with_pid(&event, pid));
+	FAIL_IF(ebb_event_enable(&event));
+
+	/* Child show now take EBBs and then exit */
+	FAIL_IF(wait_for_child(pid));
+
+	event_close(&event);
+
+	return 0;
+}
+
+int main(void)
+{
+	return test_harness(ebb_on_willing_child, "ebb_on_willing_child");
+}

tools/testing/selftests/powerpc/pmu/ebb/ebb_vs_cpu_event_test.c

@@ -0,0 +1,86 @@
+/*
+ * Copyright 2014, Michael Ellerman, IBM Corp.
+ * Licensed under GPLv2.
+ */
+
+#include <signal.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <stdbool.h>
+#include <sys/types.h>
+#include <sys/wait.h>
+#include <unistd.h>
+
+#include "ebb.h"
+
+
+/*
+ * Tests an EBB vs a cpu event - in that order. The EBB should force the cpu
+ * event off the PMU.
+ */
+
+static int setup_cpu_event(struct event *event, int cpu)
+{
+	event_init_named(event, 0x400FA, "PM_RUN_INST_CMPL");
+
+	event->attr.exclude_kernel = 1;
+	event->attr.exclude_hv = 1;
+	event->attr.exclude_idle = 1;
+
+	SKIP_IF(require_paranoia_below(1));
+	FAIL_IF(event_open_with_cpu(event, cpu));
+	FAIL_IF(event_enable(event));
+
+	return 0;
+}
+
+int ebb_vs_cpu_event(void)
+{
+	union pipe read_pipe, write_pipe;
+	struct event event;
+	int cpu, rc;
+	pid_t pid;
+
+	cpu = pick_online_cpu();
+	FAIL_IF(cpu < 0);
+	FAIL_IF(bind_to_cpu(cpu));
+
+	FAIL_IF(pipe(read_pipe.fds) == -1);
+	FAIL_IF(pipe(write_pipe.fds) == -1);
+
+	pid = fork();
+	if (pid == 0) {
+		/* NB order of pipes looks reversed */
+		exit(ebb_child(write_pipe, read_pipe));
+	}
+
+	/* Signal the child to install its EBB event and wait */
+	FAIL_IF(sync_with_child(read_pipe, write_pipe));
+
+	/* Now try to install our CPU event */
+	rc = setup_cpu_event(&event, cpu);
+	if (rc) {
+		kill_child_and_wait(pid);
+		return rc;
+	}
+
+	/* Signal the child to run */
+	FAIL_IF(sync_with_child(read_pipe, write_pipe));
+
+	/* .. and wait for it to complete */
+	FAIL_IF(wait_for_child(pid));
+	FAIL_IF(event_disable(&event));
+	FAIL_IF(event_read(&event));
+
+	event_report(&event);
+
+	/* The cpu event may have run, but we don't expect 100% */
+	FAIL_IF(event.result.enabled >= event.result.running);
+
+	return 0;
+}
+
+int main(void)
+{
+	return test_harness(ebb_vs_cpu_event, "ebb_vs_cpu_event");
+}

tools/testing/selftests/powerpc/pmu/ebb/event_attributes_test.c

@@ -0,0 +1,131 @@
+/*
+ * Copyright 2014, Michael Ellerman, IBM Corp.
+ * Licensed under GPLv2.
+ */
+
+#include <stdio.h>
+#include <stdlib.h>
+
+#include "ebb.h"
+
+
+/*
+ * Test various attributes of the EBB event are enforced.
+ */
+int event_attributes(void)
+{
+	struct event event, leader;
+
+	event_init(&event, 0x1001e);
+	event_leader_ebb_init(&event);
+	/* Expected to succeed */
+	FAIL_IF(event_open(&event));
+	event_close(&event);
+
+
+	event_init(&event, 0x001e); /* CYCLES - no PMC specified */
+	event_leader_ebb_init(&event);
+	/* Expected to fail, no PMC specified */
+	FAIL_IF(event_open(&event) == 0);
+
+
+	event_init(&event, 0x2001e);
+	event_leader_ebb_init(&event);
+	event.attr.exclusive = 0;
+	/* Expected to fail, not exclusive */
+	FAIL_IF(event_open(&event) == 0);
+
+
+	event_init(&event, 0x3001e);
+	event_leader_ebb_init(&event);
+	event.attr.freq = 1;
+	/* Expected to fail, sets freq */
+	FAIL_IF(event_open(&event) == 0);
+
+
+	event_init(&event, 0x4001e);
+	event_leader_ebb_init(&event);
+	event.attr.sample_period = 1;
+	/* Expected to fail, sets sample_period */
+	FAIL_IF(event_open(&event) == 0);
+
+
+	event_init(&event, 0x1001e);
+	event_leader_ebb_init(&event);
+	event.attr.enable_on_exec = 1;
+	/* Expected to fail, sets enable_on_exec */
+	FAIL_IF(event_open(&event) == 0);
+
+
+	event_init(&event, 0x1001e);
+	event_leader_ebb_init(&event);
+	event.attr.inherit = 1;
+	/* Expected to fail, sets inherit */
+	FAIL_IF(event_open(&event) == 0);
+
+
+	event_init(&leader, 0x1001e);
+	event_leader_ebb_init(&leader);
+	FAIL_IF(event_open(&leader));
+
+	event_init(&event, 0x20002);
+	event_ebb_init(&event);
+
+	/* Expected to succeed */
+	FAIL_IF(event_open_with_group(&event, leader.fd));
+	event_close(&leader);
+	event_close(&event);
+
+
+	event_init(&leader, 0x1001e);
+	event_leader_ebb_init(&leader);
+	FAIL_IF(event_open(&leader));
+
+	event_init(&event, 0x20002);
+
+	/* Expected to fail, event doesn't request EBB, leader does */
+	FAIL_IF(event_open_with_group(&event, leader.fd) == 0);
+	event_close(&leader);
+
+
+	event_init(&leader, 0x1001e);
+	event_leader_ebb_init(&leader);
+	/* Clear the EBB flag */
+	leader.attr.config &= ~(1ull << 63);
+
+	FAIL_IF(event_open(&leader));
+
+	event_init(&event, 0x20002);
+	event_ebb_init(&event);
+
+	/* Expected to fail, leader doesn't request EBB */
+	FAIL_IF(event_open_with_group(&event, leader.fd) == 0);
+	event_close(&leader);
+
+
+	event_init(&leader, 0x1001e);
+	event_leader_ebb_init(&leader);
+	leader.attr.exclusive = 0;
+	/* Expected to fail, leader isn't exclusive */
+	FAIL_IF(event_open(&leader) == 0);
+
+
+	event_init(&leader, 0x1001e);
+	event_leader_ebb_init(&leader);
+	leader.attr.pinned = 0;
+	/* Expected to fail, leader isn't pinned */
+	FAIL_IF(event_open(&leader) == 0);
+
+	event_init(&event, 0x1001e);
+	event_leader_ebb_init(&event);
+	/* Expected to fail, not a task event */
+	SKIP_IF(require_paranoia_below(1));
+	FAIL_IF(event_open_with_cpu(&event, 0) == 0);
+
+	return 0;
+}
+
+int main(void)
+{
+	return test_harness(event_attributes, "event_attributes");
+}

tools/testing/selftests/powerpc/pmu/ebb/fixed_instruction_loop.S

@@ -0,0 +1,43 @@
+/*
+ * Copyright 2014, Michael Ellerman, IBM Corp.
+ * Licensed under GPLv2.
+ */
+
+#include <ppc-asm.h>
+
+	.text
+
+FUNC_START(thirty_two_instruction_loop)
+	cmpwi	r3,0
+	beqlr
+	addi	r4,r3,1
+	addi	r4,r4,1
+	addi	r4,r4,1
+	addi	r4,r4,1
+	addi	r4,r4,1
+	addi	r4,r4,1
+	addi	r4,r4,1
+	addi	r4,r4,1
+	addi	r4,r4,1
+	addi	r4,r4,1
+	addi	r4,r4,1
+	addi	r4,r4,1
+	addi	r4,r4,1
+	addi	r4,r4,1
+	addi	r4,r4,1
+	addi	r4,r4,1
+	addi	r4,r4,1
+	addi	r4,r4,1
+	addi	r4,r4,1
+	addi	r4,r4,1
+	addi	r4,r4,1
+	addi	r4,r4,1
+	addi	r4,r4,1
+	addi	r4,r4,1
+	addi	r4,r4,1
+	addi	r4,r4,1
+	addi	r4,r4,1
+	addi	r4,r4,1	# 28 addi's
+	subi	r3,r3,1
+	b	FUNC_NAME(thirty_two_instruction_loop)
+FUNC_END(thirty_two_instruction_loop)

tools/testing/selftests/powerpc/pmu/ebb/fork_cleanup_test.c

@@ -0,0 +1,79 @@
+/*
+ * Copyright 2014, Michael Ellerman, IBM Corp.
+ * Licensed under GPLv2.
+ */
+
+#include <signal.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <stdbool.h>
+#include <sys/types.h>
+#include <sys/wait.h>
+#include <unistd.h>
+#include <setjmp.h>
+#include <signal.h>
+
+#include "ebb.h"
+
+
+/*
+ * Test that a fork clears the PMU state of the child. eg. BESCR/EBBHR/EBBRR
+ * are cleared, and MMCR0_PMCC is reset, preventing the child from accessing
+ * the PMU.
+ */
+
+static struct event event;
+
+static int child(void)
+{
+	/* Even though we have EBE=0 we can still see the EBB regs */
+	FAIL_IF(mfspr(SPRN_BESCR) != 0);
+	FAIL_IF(mfspr(SPRN_EBBHR) != 0);
+	FAIL_IF(mfspr(SPRN_EBBRR) != 0);
+
+	FAIL_IF(catch_sigill(write_pmc1));
+
+	/* We can still read from the event, though it is on our parent */
+	FAIL_IF(event_read(&event));
+
+	return 0;
+}
+
+/* Tests that fork clears EBB state */
+int fork_cleanup(void)
+{
+	pid_t pid;
+
+	event_init_named(&event, 0x1001e, "cycles");
+	event_leader_ebb_init(&event);
+
+	FAIL_IF(event_open(&event));
+
+	ebb_enable_pmc_counting(1);
+	setup_ebb_handler(standard_ebb_callee);
+	ebb_global_enable();
+
+	FAIL_IF(ebb_event_enable(&event));
+
+	mtspr(SPRN_MMCR0, MMCR0_FC);
+	mtspr(SPRN_PMC1, pmc_sample_period(sample_period));
+
+	/* Don't need to actually take any EBBs */
+
+	pid = fork();
+	if (pid == 0)
+		exit(child());
+
+	/* Child does the actual testing */
+	FAIL_IF(wait_for_child(pid));
+
+	/* After fork */
+	event_close(&event);
+
+	return 0;
+}
+
+int main(void)
+{
+	return test_harness(fork_cleanup, "fork_cleanup");
+}

tools/testing/selftests/powerpc/pmu/ebb/instruction_count_test.c

@@ -0,0 +1,164 @@
+/*
+ * Copyright 2014, Michael Ellerman, IBM Corp.
+ * Licensed under GPLv2.
+ */
+
+#define _GNU_SOURCE
+
+#include <stdio.h>
+#include <stdbool.h>
+#include <string.h>
+#include <sys/prctl.h>
+
+#include "ebb.h"
+
+
+/*
+ * Run a calibrated instruction loop and count instructions executed using
+ * EBBs. Make sure the counts look right.
+ */
+
+extern void thirty_two_instruction_loop(uint64_t loops);
+
+static bool counters_frozen = true;
+
+static int do_count_loop(struct event *event, uint64_t instructions,
+			 uint64_t overhead, bool report)
+{
+	int64_t difference, expected;
+	double percentage;
+
+	clear_ebb_stats();
+
+	counters_frozen = false;
+	mb();
+	mtspr(SPRN_MMCR0, mfspr(SPRN_MMCR0) & ~MMCR0_FC);
+
+	thirty_two_instruction_loop(instructions >> 5);
+
+	counters_frozen = true;
+	mb();
+	mtspr(SPRN_MMCR0, mfspr(SPRN_MMCR0) | MMCR0_FC);
+
+	count_pmc(4, sample_period);
+
+	event->result.value = ebb_state.stats.pmc_count[4-1];
+	expected = instructions + overhead;
+	difference = event->result.value - expected;
+	percentage = (double)difference / event->result.value * 100;
+
+	if (report) {
+		printf("Looped for %lu instructions, overhead %lu\n", instructions, overhead);
+		printf("Expected %lu\n", expected);
+		printf("Actual   %llu\n", event->result.value);
+		printf("Error    %ld, %f%%\n", difference, percentage);
+		printf("Took %d EBBs\n", ebb_state.stats.ebb_count);
+	}
+
+	if (difference < 0)
+		difference = -difference;
+
+	/* Tolerate a difference of up to 0.0001 % */
+	difference *= 10000 * 100;
+	if (difference / event->result.value)
+		return -1;
+
+	return 0;
+}
+
+/* Count how many instructions it takes to do a null loop */
+static uint64_t determine_overhead(struct event *event)
+{
+	uint64_t current, overhead;
+	int i;
+
+	do_count_loop(event, 0, 0, false);
+	overhead = event->result.value;
+
+	for (i = 0; i < 100; i++) {
+		do_count_loop(event, 0, 0, false);
+		current = event->result.value;
+		if (current < overhead) {
+			printf("Replacing overhead %lu with %lu\n", overhead, current);
+			overhead = current;
+		}
+	}
+
+	return overhead;
+}
+
+static void pmc4_ebb_callee(void)
+{
+	uint64_t val;
+
+	val = mfspr(SPRN_BESCR);
+	if (!(val & BESCR_PMEO)) {
+		ebb_state.stats.spurious++;
+		goto out;
+	}
+
+	ebb_state.stats.ebb_count++;
+	count_pmc(4, sample_period);
+out:
+	if (counters_frozen)
+		reset_ebb_with_clear_mask(MMCR0_PMAO);
+	else
+		reset_ebb();
+}
+
+int instruction_count(void)
+{
+	struct event event;
+	uint64_t overhead;
+
+	event_init_named(&event, 0x400FA, "PM_RUN_INST_CMPL");
+	event_leader_ebb_init(&event);
+	event.attr.exclude_kernel = 1;
+	event.attr.exclude_hv = 1;
+	event.attr.exclude_idle = 1;
+
+	FAIL_IF(event_open(&event));
+	FAIL_IF(ebb_event_enable(&event));
+
+	sample_period = COUNTER_OVERFLOW;
+
+	setup_ebb_handler(pmc4_ebb_callee);
+	mtspr(SPRN_MMCR0, mfspr(SPRN_MMCR0) & ~MMCR0_FC);
+	ebb_global_enable();
+
+	overhead = determine_overhead(&event);
+	printf("Overhead of null loop: %lu instructions\n", overhead);
+
+	/* Run for 1M instructions */
+	FAIL_IF(do_count_loop(&event, 0x100000, overhead, true));
+
+	/* Run for 10M instructions */
+	FAIL_IF(do_count_loop(&event, 0xa00000, overhead, true));
+
+	/* Run for 100M instructions */
+	FAIL_IF(do_count_loop(&event, 0x6400000, overhead, true));
+
+	/* Run for 1G instructions */
+	FAIL_IF(do_count_loop(&event, 0x40000000, overhead, true));
+
+	/* Run for 16G instructions */
+	FAIL_IF(do_count_loop(&event, 0x400000000, overhead, true));
+
+	/* Run for 64G instructions */
+	FAIL_IF(do_count_loop(&event, 0x1000000000, overhead, true));
+
+	/* Run for 128G instructions */
+	FAIL_IF(do_count_loop(&event, 0x2000000000, overhead, true));
+
+	ebb_global_disable();
+	event_close(&event);
+
+	printf("Finished OK\n");
+
+	return 0;
+}
+
+int main(void)
+{
+	return test_harness(instruction_count, "instruction_count");
+}

tools/testing/selftests/powerpc/pmu/ebb/lost_exception_test.c

@@ -0,0 +1,100 @@
+/*
+ * Copyright 2014, Michael Ellerman, IBM Corp.
+ * Licensed under GPLv2.
+ */
+
+#include <sched.h>
+#include <signal.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <sys/mman.h>
+
+#include "ebb.h"
+
+
+/*
+ * Test that tries to trigger CPU_FTR_PMAO_BUG. Which is a hardware defect
+ * where an exception triggers but we context switch before it is delivered and
+ * lose the exception.
+ */
+
+static int test_body(void)
+{
+	int i, orig_period, max_period;
+	struct event event;
+
+	/* We use PMC4 to make sure the kernel switches all counters correctly */
+	event_init_named(&event, 0x40002, "instructions");
+	event_leader_ebb_init(&event);
+
+	event.attr.exclude_kernel = 1;
+	event.attr.exclude_hv = 1;
+	event.attr.exclude_idle = 1;
+
+	FAIL_IF(event_open(&event));
+
+	ebb_enable_pmc_counting(4);
+	setup_ebb_handler(standard_ebb_callee);
+	ebb_global_enable();
+	FAIL_IF(ebb_event_enable(&event));
+
+	/*
+	 * We want a low sample period, but we also want to get out of the EBB
+	 * handler without tripping up again.
+	 *
+	 * This value picked after much experimentation.
+	 */
+	orig_period = max_period = sample_period = 400;
+
+	mtspr(SPRN_PMC4, pmc_sample_period(sample_period));
+
+	while (ebb_state.stats.ebb_count < 1000000) {
+		/*
+		 * We are trying to get the EBB exception to race exactly with
+		 * us entering the kernel to do the syscall. We then need the
+		 * kernel to decide our timeslice is up and context switch to
+		 * the other thread. When we come back our EBB will have been
+		 * lost and we'll spin in this while loop forever.
+		 */
+
+		for (i = 0; i < 100000; i++)
+			sched_yield();
+
+		/* Change the sample period slightly to try and hit the race */
+		if (sample_period >= (orig_period + 200))
+			sample_period = orig_period;
+		else
+			sample_period++;
+
+		if (sample_period > max_period)
+			max_period = sample_period;
+	}
+
+	ebb_freeze_pmcs();
+	ebb_global_disable();
+
+	count_pmc(4, sample_period);
+	mtspr(SPRN_PMC4, 0xdead);
+
+	dump_summary_ebb_state();
+	dump_ebb_hw_state();
+
+	event_close(&event);
+
+	FAIL_IF(ebb_state.stats.ebb_count == 0);
+
+	/* We vary our sample period so we need extra fudge here */
+	FAIL_IF(!ebb_check_count(4, orig_period, 2 * (max_period - orig_period)));
+
+	return 0;
+}
+
+static int lost_exception(void)
+{
+	return eat_cpu(test_body);
+}
+
+int main(void)
+{
+	return test_harness(lost_exception, "lost_exception");
+}

tools/testing/selftests/powerpc/pmu/ebb/multi_counter_test.c

@@ -0,0 +1,91 @@
+/*
+ * Copyright 2014, Michael Ellerman, IBM Corp.
+ * Licensed under GPLv2.
+ */
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <sys/ioctl.h>
+
+#include "ebb.h"
+
+
+/*
+ * Test counting multiple events using EBBs.
+ */
+int multi_counter(void)
+{
+	struct event events[6];
+	int i, group_fd;
+
+	event_init_named(&events[0], 0x1001C, "PM_CMPLU_STALL_THRD");
+	event_init_named(&events[1], 0x2D016, "PM_CMPLU_STALL_FXU");
+	event_init_named(&events[2], 0x30006, "PM_CMPLU_STALL_OTHER_CMPL");
+	event_init_named(&events[3], 0x4000A, "PM_CMPLU_STALL");
+	event_init_named(&events[4], 0x600f4, "PM_RUN_CYC");
+	event_init_named(&events[5], 0x500fa, "PM_RUN_INST_CMPL");
+
+	event_leader_ebb_init(&events[0]);
+	for (i = 1; i < 6; i++)
+		event_ebb_init(&events[i]);
+
+	group_fd = -1;
+	for (i = 0; i < 6; i++) {
+		events[i].attr.exclude_kernel = 1;
+		events[i].attr.exclude_hv = 1;
+		events[i].attr.exclude_idle = 1;
+
+		FAIL_IF(event_open_with_group(&events[i], group_fd));
+		if (group_fd == -1)
+			group_fd = events[0].fd;
+	}
+
+	ebb_enable_pmc_counting(1);
+	ebb_enable_pmc_counting(2);
+	ebb_enable_pmc_counting(3);
+	ebb_enable_pmc_counting(4);
+	ebb_enable_pmc_counting(5);
+	ebb_enable_pmc_counting(6);
+	setup_ebb_handler(standard_ebb_callee);
+
+	FAIL_IF(ioctl(events[0].fd, PERF_EVENT_IOC_ENABLE, PERF_IOC_FLAG_GROUP));
+	FAIL_IF(event_read(&events[0]));
+
+	ebb_global_enable();
+
+	mtspr(SPRN_PMC1, pmc_sample_period(sample_period));
+	mtspr(SPRN_PMC2, pmc_sample_period(sample_period));
+	mtspr(SPRN_PMC3, pmc_sample_period(sample_period));
+	mtspr(SPRN_PMC4, pmc_sample_period(sample_period));
+	mtspr(SPRN_PMC5, pmc_sample_period(sample_period));
+	mtspr(SPRN_PMC6, pmc_sample_period(sample_period));
+
+	while (ebb_state.stats.ebb_count < 50) {
+		FAIL_IF(core_busy_loop());
+		FAIL_IF(ebb_check_mmcr0());
+	}
+
+	ebb_global_disable();
+	ebb_freeze_pmcs();
+
+	count_pmc(1, sample_period);
+	count_pmc(2, sample_period);
+	count_pmc(3, sample_period);
+	count_pmc(4, sample_period);
+	count_pmc(5, sample_period);
+	count_pmc(6, sample_period);
+
+	dump_ebb_state();
+
+	for (i = 0; i < 6; i++)
+		event_close(&events[i]);
+
+	FAIL_IF(ebb_state.stats.ebb_count == 0);
+
+	return 0;
+}
+
+int main(void)
+{
+	return test_harness(multi_counter, "multi_counter");
+}

tools/testing/selftests/powerpc/pmu/ebb/multi_ebb_procs_test.c

@@ -0,0 +1,109 @@
+/*
+ * Copyright 2014, Michael Ellerman, IBM Corp.
+ * Licensed under GPLv2.
+ */
+
+#include <stdbool.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <signal.h>
+
+#include "ebb.h"
+
+
+/*
+ * Test running multiple EBB using processes at once on a single CPU. They
+ * should all run happily without interfering with each other.
+ */
+
+static bool child_should_exit;
+
+static void sigint_handler(int signal)
+{
+	child_should_exit = true;
+}
+
+struct sigaction sigint_action = {
+	.sa_handler = sigint_handler,
+};
+
+static int cycles_child(void)
+{
+	struct event event;
+
+	if (sigaction(SIGINT, &sigint_action, NULL)) {
+		perror("sigaction");
+		return 1;
+	}
+
+	event_init_named(&event, 0x1001e, "cycles");
+	event_leader_ebb_init(&event);
+
+	event.attr.exclude_kernel = 1;
+	event.attr.exclude_hv = 1;
+	event.attr.exclude_idle = 1;
+
+	FAIL_IF(event_open(&event));
+
+	ebb_enable_pmc_counting(1);
+	setup_ebb_handler(standard_ebb_callee);
+	ebb_global_enable();
+
+	FAIL_IF(ebb_event_enable(&event));
+
+	mtspr(SPRN_PMC1, pmc_sample_period(sample_period));
+
+	while (!child_should_exit) {
+		FAIL_IF(core_busy_loop());
+		FAIL_IF(ebb_check_mmcr0());
+	}
+
+	ebb_global_disable();
+	ebb_freeze_pmcs();
+
+	count_pmc(1, sample_period);
+
+	dump_summary_ebb_state();
+
+	event_close(&event);
+
+	FAIL_IF(ebb_state.stats.ebb_count == 0);
+
+	return 0;
+}
+
+#define NR_CHILDREN	4
+
+int multi_ebb_procs(void)
+{
+	pid_t pids[NR_CHILDREN];
+	int cpu, rc, i;
+
+	cpu = pick_online_cpu();
+	FAIL_IF(cpu < 0);
+	FAIL_IF(bind_to_cpu(cpu));
+
+	for (i = 0; i < NR_CHILDREN; i++) {
+		pids[i] = fork();
+		if (pids[i] == 0)
+			exit(cycles_child());
+	}
+
+	/* Have them all run for "a while" */
+	sleep(10);
+
+	rc = 0;
+	for (i = 0; i < NR_CHILDREN; i++) {
+		/* Tell them to stop */
+		kill(pids[i], SIGINT);
+		/* And wait */
+		rc |= wait_for_child(pids[i]);
+	}
+
+	return rc;
+}
+
+int main(void)
+{
+	return test_harness(multi_ebb_procs, "multi_ebb_procs");
+}

tools/testing/selftests/powerpc/pmu/ebb/no_handler_test.c

@@ -0,0 +1,61 @@
+/*
+ * Copyright 2014, Michael Ellerman, IBM Corp.
+ * Licensed under GPLv2.
+ */
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <setjmp.h>
+#include <signal.h>
+
+#include "ebb.h"
+
+
+/* Test that things work sanely if we have no handler */
+
+static int no_handler_test(void)
+{
+	struct event event;
+	u64 val;
+	int i;
+
+	event_init_named(&event, 0x1001e, "cycles");
+	event_leader_ebb_init(&event);
+
+	event.attr.exclude_kernel = 1;
+	event.attr.exclude_hv = 1;
+	event.attr.exclude_idle = 1;
+
+	FAIL_IF(event_open(&event));
+	FAIL_IF(ebb_event_enable(&event));
+
+	val = mfspr(SPRN_EBBHR);
+	FAIL_IF(val != 0);
+
+	/* Make sure it overflows quickly */
+	sample_period = 1000;
+	mtspr(SPRN_PMC1, pmc_sample_period(sample_period));
+
+	/* Spin to make sure the event has time to overflow */
+	for (i = 0; i < 1000; i++)
+		mb();
+
+	dump_ebb_state();
+
+	/* We expect to see the PMU frozen & PMAO set */
+	val = mfspr(SPRN_MMCR0);
+	FAIL_IF(val != 0x0000000080000080);
+
+	event_close(&event);
+
+	dump_ebb_state();
+
+	/* The real test is that we never took an EBB at 0x0 */
+
+	return 0;
+}
+
+int main(void)
+{
+	return test_harness(no_handler_test,"no_handler_test");
+}

tools/testing/selftests/powerpc/pmu/ebb/pmae_handling_test.c

@@ -0,0 +1,106 @@
+/*
+ * Copyright 2014, Michael Ellerman, IBM Corp.
+ * Licensed under GPLv2.
+ */
+
+#include <sched.h>
+#include <signal.h>
+#include <stdbool.h>
+#include <stdio.h>
+#include <stdlib.h>
+
+#include "ebb.h"
+
+
+/*
+ * Test that the kernel properly handles PMAE across context switches.
+ *
+ * We test this by calling into the kernel inside our EBB handler, where PMAE
+ * is clear. A cpu eater companion thread is running on the same CPU as us to
+ * encourage the scheduler to switch us.
+ *
+ * The kernel must make sure that when it context switches us back in, it
+ * honours the fact that we had PMAE clear.
+ *
+ * Observed to hit the failing case on the first EBB with a broken kernel.
+ */
+
+static bool mmcr0_mismatch;
+static uint64_t before, after;
+
+static void syscall_ebb_callee(void)
+{
+	uint64_t val;
+
+	val = mfspr(SPRN_BESCR);
+	if (!(val & BESCR_PMEO)) {
+		ebb_state.stats.spurious++;
+		goto out;
+	}
+
+	ebb_state.stats.ebb_count++;
+	count_pmc(1, sample_period);
+
+	before = mfspr(SPRN_MMCR0);
+
+	/* Try and get ourselves scheduled, to force a PMU context switch */
+	sched_yield();
+
+	after = mfspr(SPRN_MMCR0);
+	if (before != after)
+		mmcr0_mismatch = true;
+
+out:
+	reset_ebb();
+}
+
+static int test_body(void)
+{
+	struct event event;
+
+	event_init_named(&event, 0x1001e, "cycles");
+	event_leader_ebb_init(&event);
+
+	event.attr.exclude_kernel = 1;
+	event.attr.exclude_hv = 1;
+	event.attr.exclude_idle = 1;
+
+	FAIL_IF(event_open(&event));
+
+	setup_ebb_handler(syscall_ebb_callee);
+	ebb_global_enable();
+
+	FAIL_IF(ebb_event_enable(&event));
+
+	mtspr(SPRN_PMC1, pmc_sample_period(sample_period));
+
+	while (ebb_state.stats.ebb_count < 20 && !mmcr0_mismatch)
+		FAIL_IF(core_busy_loop());
+
+	ebb_global_disable();
+	ebb_freeze_pmcs();
+
+	count_pmc(1, sample_period);
+
+	dump_ebb_state();
+
+	if (mmcr0_mismatch)
+		printf("Saw MMCR0 before 0x%lx after 0x%lx\n", before, after);
+
+	event_close(&event);
+
+	FAIL_IF(ebb_state.stats.ebb_count == 0);
+	FAIL_IF(mmcr0_mismatch);
+
+	return 0;
+}
+
+int pmae_handling(void)
+{
+	return eat_cpu(test_body);
+}
+
+int main(void)
+{
+	return test_harness(pmae_handling, "pmae_handling");
+}

tools/testing/selftests/powerpc/pmu/ebb/pmc56_overflow_test.c

@@ -0,0 +1,93 @@
+/*
+ * Copyright 2014, Michael Ellerman, IBM Corp.
+ * Licensed under GPLv2.
+ */
+
+#include <stdio.h>
+#include <stdlib.h>
+
+#include "ebb.h"
+
+
+/*
+ * Test that PMC5 & 6 are frozen (ie. don't overflow) when they are not being
+ * used. Tests the MMCR0_FC56 logic in the kernel.
+ */
+
+static int pmc56_overflowed;
+
+static void ebb_callee(void)
+{
+	uint64_t val;
+
+	val = mfspr(SPRN_BESCR);
+	if (!(val & BESCR_PMEO)) {
+		ebb_state.stats.spurious++;
+		goto out;
+	}
+
+	ebb_state.stats.ebb_count++;
+	count_pmc(2, sample_period);
+
+	val = mfspr(SPRN_PMC5);
+	if (val >= COUNTER_OVERFLOW)
+		pmc56_overflowed++;
+
+	count_pmc(5, COUNTER_OVERFLOW);
+
+	val = mfspr(SPRN_PMC6);
+	if (val >= COUNTER_OVERFLOW)
+		pmc56_overflowed++;
+
+	count_pmc(6, COUNTER_OVERFLOW);
+
+out:
+	reset_ebb();
+}
+
+int pmc56_overflow(void)
+{
+	struct event event;
+
+	/* Use PMC2 so we set PMCjCE, which enables PMC5/6 */
+	event_init(&event, 0x2001e);
+	event_leader_ebb_init(&event);
+
+	event.attr.exclude_kernel = 1;
+	event.attr.exclude_hv = 1;
+	event.attr.exclude_idle = 1;
+
+	FAIL_IF(event_open(&event));
+
+	setup_ebb_handler(ebb_callee);
+	ebb_global_enable();
+
+	FAIL_IF(ebb_event_enable(&event));
+
+	mtspr(SPRN_PMC1, pmc_sample_period(sample_period));
+	mtspr(SPRN_PMC5, 0);
+	mtspr(SPRN_PMC6, 0);
+
+	while (ebb_state.stats.ebb_count < 10)
+		FAIL_IF(core_busy_loop());
+
+	ebb_global_disable();
+	ebb_freeze_pmcs();
+
+	count_pmc(2, sample_period);
+
+	dump_ebb_state();
+
+	printf("PMC5/6 overflow %d\n", pmc56_overflowed);
+
+	event_close(&event);
+
+	FAIL_IF(ebb_state.stats.ebb_count == 0 || pmc56_overflowed != 0);
+
+	return 0;
+}
+
+int main(void)
+{
+	return test_harness(pmc56_overflow, "pmc56_overflow");
+}

tools/testing/selftests/powerpc/pmu/ebb/reg.h

@@ -0,0 +1,49 @@
+/*
+ * Copyright 2014, Michael Ellerman, IBM Corp.
+ * Licensed under GPLv2.
+ */
+
+#ifndef _SELFTESTS_POWERPC_REG_H
+#define _SELFTESTS_POWERPC_REG_H
+
+#define __stringify_1(x)        #x
+#define __stringify(x)          __stringify_1(x)
+
+#define mfspr(rn)       ({unsigned long rval; \
+                         asm volatile("mfspr %0," __stringify(rn) \
+                                 : "=r" (rval)); rval; })
+#define mtspr(rn, v)    asm volatile("mtspr " __stringify(rn) ",%0" : \
+                                    : "r" ((unsigned long)(v)) \
+                                    : "memory")
+
+#define mb()		asm volatile("sync" : : : "memory");
+
+#define SPRN_MMCR2     769
+#define SPRN_MMCRA     770
+#define SPRN_MMCR0     779
+#define   MMCR0_PMAO   0x00000080
+#define   MMCR0_PMAE   0x04000000
+#define   MMCR0_FC     0x80000000
+#define SPRN_EBBHR     804
+#define SPRN_EBBRR     805
+#define SPRN_BESCR     806     /* Branch event status & control register */
+#define SPRN_BESCRS    800     /* Branch event status & control set (1 bits set to 1) */
+#define SPRN_BESCRSU   801     /* Branch event status & control set upper */
+#define SPRN_BESCRR    802     /* Branch event status & control REset (1 bits set to 0) */
+#define SPRN_BESCRRU   803     /* Branch event status & control REset upper */
+
+#define BESCR_PMEO     0x1     /* PMU Event-based exception Occurred */
+#define BESCR_PME      (0x1ul << 32) /* PMU Event-based exception Enable */
+
+#define SPRN_PMC1      771
+#define SPRN_PMC2      772
+#define SPRN_PMC3      773
+#define SPRN_PMC4      774
+#define SPRN_PMC5      775
+#define SPRN_PMC6      776
+
+#define SPRN_SIAR      780
+#define SPRN_SDAR      781
+#define SPRN_SIER      768
+
+#endif /* _SELFTESTS_POWERPC_REG_H */

tools/testing/selftests/powerpc/pmu/ebb/reg_access_test.c

@@ -0,0 +1,39 @@
+/*
+ * Copyright 2014, Michael Ellerman, IBM Corp.
+ * Licensed under GPLv2.
+ */
+
+#include <stdio.h>
+#include <stdlib.h>
+
+#include "ebb.h"
+#include "reg.h"
+
+
+/*
+ * Test basic access to the EBB regs, they should be user accessible with no
+ * kernel interaction required.
+ */
+int reg_access(void)
+{
+	uint64_t val, expected;
+
+	expected = 0x8000000100000000ull;
+	mtspr(SPRN_BESCR, expected);
+	val = mfspr(SPRN_BESCR);
+
+	FAIL_IF(val != expected);
+
+	expected = 0x0000000001000000ull;
+	mtspr(SPRN_EBBHR, expected);
+	val = mfspr(SPRN_EBBHR);
+
+	FAIL_IF(val != expected);
+
+	return 0;
+}
+
+int main(void)
+{
+	return test_harness(reg_access, "reg_access");
+}

tools/testing/selftests/powerpc/pmu/ebb/task_event_pinned_vs_ebb_test.c

@@ -0,0 +1,91 @@
+/*
+ * Copyright 2014, Michael Ellerman, IBM Corp.
+ * Licensed under GPLv2.
+ */
+
+#include <signal.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <stdbool.h>
+#include <sys/types.h>
+#include <sys/wait.h>
+#include <unistd.h>
+
+#include "ebb.h"
+
+
+/*
+ * Tests a pinned per-task event vs an EBB - in that order. The pinned per-task
+ * event should prevent the EBB event from being enabled.
+ */
+
+static int setup_child_event(struct event *event, pid_t child_pid)
+{
+	event_init_named(event, 0x400FA, "PM_RUN_INST_CMPL");
+
+	event->attr.pinned = 1;
+
+	event->attr.exclude_kernel = 1;
+	event->attr.exclude_hv = 1;
+	event->attr.exclude_idle = 1;
+
+	FAIL_IF(event_open_with_pid(event, child_pid));
+	FAIL_IF(event_enable(event));
+
+	return 0;
+}
+
+int task_event_pinned_vs_ebb(void)
+{
+	union pipe read_pipe, write_pipe;
+	struct event event;
+	pid_t pid;
+	int rc;
+
+	FAIL_IF(pipe(read_pipe.fds) == -1);
+	FAIL_IF(pipe(write_pipe.fds) == -1);
+
+	pid = fork();
+	if (pid == 0) {
+		/* NB order of pipes looks reversed */
+		exit(ebb_child(write_pipe, read_pipe));
+	}
+
+	/* We setup the task event first */
+	rc = setup_child_event(&event, pid);
+	if (rc) {
+		kill_child_and_wait(pid);
+		return rc;
+	}
+
+	/* Signal the child to install its EBB event and wait */
+	if (sync_with_child(read_pipe, write_pipe))
+		/* If it fails, wait for it to exit */
+		goto wait;
+
+	/* Signal the child to run */
+	FAIL_IF(sync_with_child(read_pipe, write_pipe));
+
+wait:
+	/* We expect it to fail to read the event */
+	FAIL_IF(wait_for_child(pid) != 2);
+	FAIL_IF(event_disable(&event));
+	FAIL_IF(event_read(&event));
+
+	event_report(&event);
+
+	FAIL_IF(event.result.value == 0);
+	/*
+	 * For reasons I don't understand enabled is usually just slightly
+	 * lower than running. Would be good to confirm why.
+	 */
+	FAIL_IF(event.result.enabled == 0);
+	FAIL_IF(event.result.running == 0);
+
+	return 0;
+}
+
+int main(void)
+{
+	return test_harness(task_event_pinned_vs_ebb, "task_event_pinned_vs_ebb");
+}

tools/testing/selftests/powerpc/pmu/ebb/task_event_vs_ebb_test.c

@@ -0,0 +1,83 @@
+/*
+ * Copyright 2014, Michael Ellerman, IBM Corp.
+ * Licensed under GPLv2.
+ */
+
+#include <signal.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <stdbool.h>
+#include <sys/types.h>
+#include <sys/wait.h>
+#include <unistd.h>
+
+#include "ebb.h"
+
+
+/*
+ * Tests a per-task event vs an EBB - in that order. The EBB should push the
+ * per-task event off the PMU.
+ */
+
+static int setup_child_event(struct event *event, pid_t child_pid)
+{
+	event_init_named(event, 0x400FA, "PM_RUN_INST_CMPL");
+
+	event->attr.exclude_kernel = 1;
+	event->attr.exclude_hv = 1;
+	event->attr.exclude_idle = 1;
+
+	FAIL_IF(event_open_with_pid(event, child_pid));
+	FAIL_IF(event_enable(event));
+
+	return 0;
+}
+
+int task_event_vs_ebb(void)
+{
+	union pipe read_pipe, write_pipe;
+	struct event event;
+	pid_t pid;
+	int rc;
+
+	FAIL_IF(pipe(read_pipe.fds) == -1);
+	FAIL_IF(pipe(write_pipe.fds) == -1);
+
+	pid = fork();
+	if (pid == 0) {
+		/* NB order of pipes looks reversed */
+		exit(ebb_child(write_pipe, read_pipe));
+	}
+
+	/* We setup the task event first */
+	rc = setup_child_event(&event, pid);
+	if (rc) {
+		kill_child_and_wait(pid);
+		return rc;
+	}
+
+	/* Signal the child to install its EBB event and wait */
+	if (sync_with_child(read_pipe, write_pipe))
+		/* If it fails, wait for it to exit */
+		goto wait;
+
+	/* Signal the child to run */
+	FAIL_IF(sync_with_child(read_pipe, write_pipe));
+
+wait:
+	/* The EBB event should push the task event off so the child should succeed */
+	FAIL_IF(wait_for_child(pid));
+	FAIL_IF(event_disable(&event));
+	FAIL_IF(event_read(&event));
+
+	event_report(&event);
+
+	/* The task event may have run, or not so we can't assert anything about it */
+
+	return 0;
+}
+
+int main(void)
+{
+	return test_harness(task_event_vs_ebb, "task_event_vs_ebb");
+}

tools/testing/selftests/powerpc/pmu/ebb/trace.c

@@ -0,0 +1,300 @@
+/*
+ * Copyright 2014, Michael Ellerman, IBM Corp.
+ * Licensed under GPLv2.
+ */
+
+#include <errno.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <sys/mman.h>
+
+#include "trace.h"
+
+
+struct trace_buffer *trace_buffer_allocate(u64 size)
+{
+	struct trace_buffer *tb;
+
+	if (size < sizeof(*tb)) {
+		fprintf(stderr, "Error: trace buffer too small\n");
+		return NULL;
+	}
+
+	tb = mmap(NULL, size, PROT_READ | PROT_WRITE,
+		  MAP_ANONYMOUS | MAP_PRIVATE, -1, 0);
+	if (tb == MAP_FAILED) {
+		perror("mmap");
+		return NULL;
+	}
+
+	tb->size = size;
+	tb->tail = tb->data;
+	tb->overflow = false;
+
+	return tb;
+}
+
+static bool trace_check_bounds(struct trace_buffer *tb, void *p)
+{
+	return p < ((void *)tb + tb->size);
+}
+
+static bool trace_check_alloc(struct trace_buffer *tb, void *p)
+{
+	/*
+	 * If we ever overflowed don't allow any more input. This prevents us
+	 * from dropping a large item and then later logging a small one. The
+	 * buffer should just stop when overflow happened, not be patchy. If
+	 * you're overflowing, make your buffer bigger.
+	 */
+	if (tb->overflow)
+		return false;
+
+	if (!trace_check_bounds(tb, p)) {
+		tb->overflow = true;
+		return false;
+	}
+
+	return true;
+}
+
+static void *trace_alloc(struct trace_buffer *tb, int bytes)
+{
+	void *p, *newtail;
+
+	p = tb->tail;
+	newtail = tb->tail + bytes;
+	if (!trace_check_alloc(tb, newtail))
+		return NULL;
+
+	tb->tail = newtail;
+
+	return p;
+}
+
+static struct trace_entry *trace_alloc_entry(struct trace_buffer *tb, int payload_size)
+{
+	struct trace_entry *e;
+
+	e = trace_alloc(tb, sizeof(*e) + payload_size);
+	if (e)
+		e->length = payload_size;
+
+	return e;
+}
+
+int trace_log_reg(struct trace_buffer *tb, u64 reg, u64 value)
+{
+	struct trace_entry *e;
+	u64 *p;
+
+	e = trace_alloc_entry(tb, sizeof(reg) + sizeof(value));
+	if (!e)
+		return -ENOSPC;
+
+	e->type = TRACE_TYPE_REG;
+	p = (u64 *)e->data;
+	*p++ = reg;
+	*p++ = value;
+
+	return 0;
+}
+
+int trace_log_counter(struct trace_buffer *tb, u64 value)
+{
+	struct trace_entry *e;
+	u64 *p;
+
+	e = trace_alloc_entry(tb, sizeof(value));
+	if (!e)
+		return -ENOSPC;
+
+	e->type = TRACE_TYPE_COUNTER;
+	p = (u64 *)e->data;
+	*p++ = value;
+
+	return 0;
+}
+
+int trace_log_string(struct trace_buffer *tb, char *str)
+{
+	struct trace_entry *e;
+	char *p;
+	int len;
+
+	len = strlen(str);
+
+	/* We NULL terminate to make printing easier */
+	e = trace_alloc_entry(tb, len + 1);
+	if (!e)
+		return -ENOSPC;
+
+	e->type = TRACE_TYPE_STRING;
+	p = (char *)e->data;
+	memcpy(p, str, len);
+	p += len;
+	*p = '\0';
+
+	return 0;
+}
+
+int trace_log_indent(struct trace_buffer *tb)
+{
+	struct trace_entry *e;
+
+	e = trace_alloc_entry(tb, 0);
+	if (!e)
+		return -ENOSPC;
+
+	e->type = TRACE_TYPE_INDENT;
+
+	return 0;
+}
+
+int trace_log_outdent(struct trace_buffer *tb)
+{
+	struct trace_entry *e;
+
+	e = trace_alloc_entry(tb, 0);
+	if (!e)
+		return -ENOSPC;
+
+	e->type = TRACE_TYPE_OUTDENT;
+
+	return 0;
+}
+
+static void trace_print_header(int seq, int prefix)
+{
+	printf("%*s[%d]: ", prefix, "", seq);
+}
+
+static char *trace_decode_reg(int reg)
+{
+	switch (reg) {
+		case 769: return "SPRN_MMCR2"; break;
+		case 770: return "SPRN_MMCRA"; break;
+		case 779: return "SPRN_MMCR0"; break;
+		case 804: return "SPRN_EBBHR"; break;
+		case 805: return "SPRN_EBBRR"; break;
+		case 806: return "SPRN_BESCR"; break;
+		case 800: return "SPRN_BESCRS"; break;
+		case 801: return "SPRN_BESCRSU"; break;
+		case 802: return "SPRN_BESCRR"; break;
+		case 803: return "SPRN_BESCRRU"; break;
+		case 771: return "SPRN_PMC1"; break;
+		case 772: return "SPRN_PMC2"; break;
+		case 773: return "SPRN_PMC3"; break;
+		case 774: return "SPRN_PMC4"; break;
+		case 775: return "SPRN_PMC5"; break;
+		case 776: return "SPRN_PMC6"; break;
+		case 780: return "SPRN_SIAR"; break;
+		case 781: return "SPRN_SDAR"; break;
+		case 768: return "SPRN_SIER"; break;
+	}
+
+	return NULL;
+}
+
+static void trace_print_reg(struct trace_entry *e)
+{
+	u64 *p, *reg, *value;
+	char *name;
+
+	p = (u64 *)e->data;
+	reg = p++;
+	value = p;
+
+	name = trace_decode_reg(*reg);
+	if (name)
+		printf("register %-10s = 0x%016llx\n", name, *value);
+	else
+		printf("register %lld = 0x%016llx\n", *reg, *value);
+}
+
+static void trace_print_counter(struct trace_entry *e)
+{
+	u64 *value;
+
+	value = (u64 *)e->data;
+	printf("counter = %lld\n", *value);
+}
+
+static void trace_print_string(struct trace_entry *e)
+{
+	char *str;
+
+	str = (char *)e->data;
+	puts(str);
+}
+
+#define BASE_PREFIX	2
+#define PREFIX_DELTA	8
+
+static void trace_print_entry(struct trace_entry *e, int seq, int *prefix)
+{
+	switch (e->type) {
+	case TRACE_TYPE_REG:
+		trace_print_header(seq, *prefix);
+		trace_print_reg(e);
+		break;
+	case TRACE_TYPE_COUNTER:
+		trace_print_header(seq, *prefix);
+		trace_print_counter(e);
+		break;
+	case TRACE_TYPE_STRING:
+		trace_print_header(seq, *prefix);
+		trace_print_string(e);
+		break;
+	case TRACE_TYPE_INDENT:
+		trace_print_header(seq, *prefix);
+		puts("{");
+		*prefix += PREFIX_DELTA;
+		break;
+	case TRACE_TYPE_OUTDENT:
+		*prefix -= PREFIX_DELTA;
+		if (*prefix < BASE_PREFIX)
+			*prefix = BASE_PREFIX;
+		trace_print_header(seq, *prefix);
+		puts("}");
+		break;
+	default:
+		trace_print_header(seq, *prefix);
+		printf("entry @ %p type %d\n", e, e->type);
+		break;
+	}
+}
+
+void trace_buffer_print(struct trace_buffer *tb)
+{
+	struct trace_entry *e;
+	int i, prefix;
+	void *p;
+
+	printf("Trace buffer dump:\n");
+	printf("  address  %p \n", tb);
+	printf("  tail     %p\n", tb->tail);
+	printf("  size     %llu\n", tb->size);
+	printf("  overflow %s\n", tb->overflow ? "TRUE" : "false");
+	printf("  Content:\n");
+
+	p = tb->data;
+
+	i = 0;
+	prefix = BASE_PREFIX;
+
+	while (trace_check_bounds(tb, p) && p < tb->tail) {
+		e = p;
+
+		trace_print_entry(e, i, &prefix);
+
+		i++;
+		p = (void *)e + sizeof(*e) + e->length;
+	}
+}
+
+void trace_print_location(struct trace_buffer *tb)
+{
+	printf("Trace buffer 0x%llx bytes @ %p\n", tb->size, tb);
+}

tools/testing/selftests/powerpc/pmu/ebb/trace.h

@@ -0,0 +1,41 @@
+/*
+ * Copyright 2014, Michael Ellerman, IBM Corp.
+ * Licensed under GPLv2.
+ */
+
+#ifndef _SELFTESTS_POWERPC_PMU_EBB_TRACE_H
+#define _SELFTESTS_POWERPC_PMU_EBB_TRACE_H
+
+#include "utils.h"
+
+#define TRACE_TYPE_REG		1
+#define TRACE_TYPE_COUNTER	2
+#define TRACE_TYPE_STRING	3
+#define TRACE_TYPE_INDENT	4
+#define TRACE_TYPE_OUTDENT	5
+
+struct trace_entry
+{
+	u8 type;
+	u8 length;
+	u8 data[0];
+};
+
+struct trace_buffer
+{
+	u64  size;
+	bool overflow;
+	void *tail;
+	u8   data[0];
+};
+
+struct trace_buffer *trace_buffer_allocate(u64 size);
+int trace_log_reg(struct trace_buffer *tb, u64 reg, u64 value);
+int trace_log_counter(struct trace_buffer *tb, u64 value);
+int trace_log_string(struct trace_buffer *tb, char *str);
+int trace_log_indent(struct trace_buffer *tb);
+int trace_log_outdent(struct trace_buffer *tb);
+void trace_buffer_print(struct trace_buffer *tb);
+void trace_print_location(struct trace_buffer *tb);
+
+#endif /* _SELFTESTS_POWERPC_PMU_EBB_TRACE_H */

tools/testing/selftests/powerpc/pmu/event.c

@@ -0,0 +1,131 @@
+/*
+ * Copyright 2013, Michael Ellerman, IBM Corp.
+ * Licensed under GPLv2.
+ */
+
+#define _GNU_SOURCE
+#include <unistd.h>
+#include <sys/syscall.h>
+#include <string.h>
+#include <stdio.h>
+#include <sys/ioctl.h>
+
+#include "event.h"
+
+
+int perf_event_open(struct perf_event_attr *attr, pid_t pid, int cpu,
+		int group_fd, unsigned long flags)
+{
+	return syscall(__NR_perf_event_open, attr, pid, cpu,
+			   group_fd, flags);
+}
+
+void event_init_opts(struct event *e, u64 config, int type, char *name)
+{
+	memset(e, 0, sizeof(*e));
+
+	e->name = name;
+
+	e->attr.type = type;
+	e->attr.config = config;
+	e->attr.size = sizeof(e->attr);
+	/* This has to match the structure layout in the header */
+	e->attr.read_format = PERF_FORMAT_TOTAL_TIME_ENABLED | \
+				  PERF_FORMAT_TOTAL_TIME_RUNNING;
+}
+
+void event_init_named(struct event *e, u64 config, char *name)
+{
+	event_init_opts(e, config, PERF_TYPE_RAW, name);
+}
+
+void event_init(struct event *e, u64 config)
+{
+	event_init_opts(e, config, PERF_TYPE_RAW, "event");
+}
+
+#define PERF_CURRENT_PID	0
+#define PERF_NO_PID		-1
+#define PERF_NO_CPU		-1
+#define PERF_NO_GROUP		-1
+
+int event_open_with_options(struct event *e, pid_t pid, int cpu, int group_fd)
+{
+	e->fd = perf_event_open(&e->attr, pid, cpu, group_fd, 0);
+	if (e->fd == -1) {
+		perror("perf_event_open");
+		return -1;
+	}
+
+	return 0;
+}
+
+int event_open_with_group(struct event *e, int group_fd)
+{
+	return event_open_with_options(e, PERF_CURRENT_PID, PERF_NO_CPU, group_fd);
+}
+
+int event_open_with_pid(struct event *e, pid_t pid)
+{
+	return event_open_with_options(e, pid, PERF_NO_CPU, PERF_NO_GROUP);
+}
+
+int event_open_with_cpu(struct event *e, int cpu)
+{
+	return event_open_with_options(e, PERF_NO_PID, cpu, PERF_NO_GROUP);
+}
+
+int event_open(struct event *e)
+{
+	return event_open_with_options(e, PERF_CURRENT_PID, PERF_NO_CPU, PERF_NO_GROUP);
+}
+
+void event_close(struct event *e)
+{
+	close(e->fd);
+}
+
+int event_enable(struct event *e)
+{
+	return ioctl(e->fd, PERF_EVENT_IOC_ENABLE);
+}
+
+int event_disable(struct event *e)
+{
+	return ioctl(e->fd, PERF_EVENT_IOC_DISABLE);
+}
+
+int event_reset(struct event *e)
+{
+	return ioctl(e->fd, PERF_EVENT_IOC_RESET);
+}
+
+int event_read(struct event *e)
+{
+	int rc;
+
+	rc = read(e->fd, &e->result, sizeof(e->result));
+	if (rc != sizeof(e->result)) {
+		fprintf(stderr, "read error on event %p!\n", e);
+		return -1;
+	}
+
+	return 0;
+}
+
+void event_report_justified(struct event *e, int name_width, int result_width)
+{
+	printf("%*s: result %*llu ", name_width, e->name, result_width,
+	       e->result.value);
+
+	if (e->result.running == e->result.enabled)
+		printf("running/enabled %llu\n", e->result.running);
+	else
+		printf("running %llu enabled %llu\n", e->result.running,
+			e->result.enabled);
+}
+
+void event_report(struct event *e)
+{
+	event_report_justified(e, 0, 0);
+}

tools/testing/selftests/powerpc/pmu/event.h

@@ -0,0 +1,43 @@
+/*
+ * Copyright 2013, Michael Ellerman, IBM Corp.
+ * Licensed under GPLv2.
+ */
+
+#ifndef _SELFTESTS_POWERPC_PMU_EVENT_H
+#define _SELFTESTS_POWERPC_PMU_EVENT_H
+
+#include <unistd.h>
+#include <linux/perf_event.h>
+
+#include "utils.h"
+
+
+struct event {
+	struct perf_event_attr attr;
+	char *name;
+	int fd;
+	/* This must match the read_format we use */
+	struct {
+		u64 value;
+		u64 running;
+		u64 enabled;
+	} result;
+};
+
+void event_init(struct event *e, u64 config);
+void event_init_named(struct event *e, u64 config, char *name);
+void event_init_opts(struct event *e, u64 config, int type, char *name);
+int event_open_with_options(struct event *e, pid_t pid, int cpu, int group_fd);
+int event_open_with_group(struct event *e, int group_fd);
+int event_open_with_pid(struct event *e, pid_t pid);
+int event_open_with_cpu(struct event *e, int cpu);
+int event_open(struct event *e);
+void event_close(struct event *e);
+int event_enable(struct event *e);
+int event_disable(struct event *e);
+int event_reset(struct event *e);
+int event_read(struct event *e);
+void event_report_justified(struct event *e, int name_width, int result_width);
+void event_report(struct event *e);
+
+#endif /* _SELFTESTS_POWERPC_PMU_EVENT_H */

tools/testing/selftests/powerpc/pmu/l3_bank_test.c

@@ -0,0 +1,48 @@
+/*
+ * Copyright 2014, Michael Ellerman, IBM Corp.
+ * Licensed under GPLv2.
+ */
+
+#include <stdio.h>
+#include <stdlib.h>
+
+#include "event.h"
+#include "utils.h"
+
+#define MALLOC_SIZE     (0x10000 * 10)  /* Ought to be enough .. */
+
+/*
+ * Tests that the L3 bank handling is correct. We fixed it in commit e9aaac1.
+ */
+static int l3_bank_test(void)
+{
+	struct event event;
+	char *p;
+	int i;
+
+	p = malloc(MALLOC_SIZE);
+	FAIL_IF(!p);
+
+	event_init(&event, 0x84918F);
+
+	FAIL_IF(event_open(&event));
+
+	for (i = 0; i < MALLOC_SIZE; i += 0x10000)
+		p[i] = i;
+
+	event_read(&event);
+	event_report(&event);
+
+	FAIL_IF(event.result.running == 0);
+	FAIL_IF(event.result.enabled == 0);
+
+	event_close(&event);
+	free(p);
+
+	return 0;
+}
+
+int main(void)
+{
+	return test_harness(l3_bank_test, "l3_bank_test");
+}

tools/testing/selftests/powerpc/pmu/lib.c

@@ -0,0 +1,300 @@
+/*
+ * Copyright 2014, Michael Ellerman, IBM Corp.
+ * Licensed under GPLv2.
+ */
+
+#define _GNU_SOURCE	/* For CPU_ZERO etc. */
+
+#include <elf.h>
+#include <errno.h>
+#include <fcntl.h>
+#include <link.h>
+#include <sched.h>
+#include <setjmp.h>
+#include <stdlib.h>
+#include <sys/stat.h>
+#include <sys/types.h>
+#include <sys/wait.h>
+
+#include "utils.h"
+#include "lib.h"
+
+
+int pick_online_cpu(void)
+{
+	cpu_set_t mask;
+	int cpu;
+
+	CPU_ZERO(&mask);
+
+	if (sched_getaffinity(0, sizeof(mask), &mask)) {
+		perror("sched_getaffinity");
+		return -1;
+	}
+
+	/* We prefer a primary thread, but skip 0 */
+	for (cpu = 8; cpu < CPU_SETSIZE; cpu += 8)
+		if (CPU_ISSET(cpu, &mask))
+			return cpu;
+
+	/* Search for anything, but in reverse */
+	for (cpu = CPU_SETSIZE - 1; cpu >= 0; cpu--)
+		if (CPU_ISSET(cpu, &mask))
+			return cpu;
+
+	printf("No cpus in affinity mask?!\n");
+	return -1;
+}
+
+int bind_to_cpu(int cpu)
+{
+	cpu_set_t mask;
+
+	printf("Binding to cpu %d\n", cpu);
+
+	CPU_ZERO(&mask);
+	CPU_SET(cpu, &mask);
+
+	return sched_setaffinity(0, sizeof(mask), &mask);
+}
+
+#define PARENT_TOKEN	0xAA
+#define CHILD_TOKEN	0x55
+
+int sync_with_child(union pipe read_pipe, union pipe write_pipe)
+{
+	char c = PARENT_TOKEN;
+
+	FAIL_IF(write(write_pipe.write_fd, &c, 1) != 1);
+	FAIL_IF(read(read_pipe.read_fd, &c, 1) != 1);
+	if (c != CHILD_TOKEN) /* sometimes expected */
+		return 1;
+
+	return 0;
+}
+
+int wait_for_parent(union pipe read_pipe)
+{
+	char c;
+
+	FAIL_IF(read(read_pipe.read_fd, &c, 1) != 1);
+	FAIL_IF(c != PARENT_TOKEN);
+
+	return 0;
+}
+
+int notify_parent(union pipe write_pipe)
+{
+	char c = CHILD_TOKEN;
+
+	FAIL_IF(write(write_pipe.write_fd, &c, 1) != 1);
+
+	return 0;
+}
+
+int notify_parent_of_error(union pipe write_pipe)
+{
+	char c = ~CHILD_TOKEN;
+
+	FAIL_IF(write(write_pipe.write_fd, &c, 1) != 1);
+
+	return 0;
+}
+
+int wait_for_child(pid_t child_pid)
+{
+	int rc;
+
+	if (waitpid(child_pid, &rc, 0) == -1) {
+		perror("waitpid");
+		return 1;
+	}
+
+	if (WIFEXITED(rc))
+		rc = WEXITSTATUS(rc);
+	else
+		rc = 1; /* Signal or other */
+
+	return rc;
+}
+
+int kill_child_and_wait(pid_t child_pid)
+{
+	kill(child_pid, SIGTERM);
+
+	return wait_for_child(child_pid);
+}
+
+static int eat_cpu_child(union pipe read_pipe, union pipe write_pipe)
+{
+	volatile int i = 0;
+
+	/*
+	 * We are just here to eat cpu and die. So make sure we can be killed,
+	 * and also don't do any custom SIGTERM handling.
+	 */
+	signal(SIGTERM, SIG_DFL);
+
+	notify_parent(write_pipe);
+	wait_for_parent(read_pipe);
+
+	/* Soak up cpu forever */
+	while (1) i++;
+
+	return 0;
+}
+
+pid_t eat_cpu(int (test_function)(void))
+{
+	union pipe read_pipe, write_pipe;
+	int cpu, rc;
+	pid_t pid;
+
+	cpu = pick_online_cpu();
+	FAIL_IF(cpu < 0);
+	FAIL_IF(bind_to_cpu(cpu));
+
+	if (pipe(read_pipe.fds) == -1)
+		return -1;
+
+	if (pipe(write_pipe.fds) == -1)
+		return -1;
+
+	pid = fork();
+	if (pid == 0)
+		exit(eat_cpu_child(write_pipe, read_pipe));
+
+	if (sync_with_child(read_pipe, write_pipe)) {
+		rc = -1;
+		goto out;
+	}
+
+	printf("main test running as pid %d\n", getpid());
+
+	rc = test_function();
+out:
+	kill(pid, SIGKILL);
+
+	return rc;
+}
+
+struct addr_range libc, vdso;
+
+int parse_proc_maps(void)
+{
+	unsigned long start, end;
+	char execute, name[128];
+	FILE *f;
+	int rc;
+
+	f = fopen("/proc/self/maps", "r");
+	if (!f) {
+		perror("fopen");
+		return -1;
+	}
+
+	do {
+		/* This skips line with no executable which is what we want */
+		rc = fscanf(f, "%lx-%lx %*c%*c%c%*c %*x %*d:%*d %*d %127s\n",
+			    &start, &end, &execute, name);
+		if (rc <= 0)
+			break;
+
+		if (execute != 'x')
+			continue;
+
+		if (strstr(name, "libc")) {
+			libc.first = start;
+			libc.last = end - 1;
+		} else if (strstr(name, "[vdso]")) {
+			vdso.first = start;
+			vdso.last = end - 1;
+		}
+	} while(1);
+
+	fclose(f);
+
+	return 0;
+}
+
+#define PARANOID_PATH	"/proc/sys/kernel/perf_event_paranoid"
+
+bool require_paranoia_below(int level)
+{
+	unsigned long current;
+	char *end, buf[16];
+	FILE *f;
+	int rc;
+
+	rc = -1;
+
+	f = fopen(PARANOID_PATH, "r");
+	if (!f) {
+		perror("fopen");
+		goto out;
+	}
+
+	if (!fgets(buf, sizeof(buf), f)) {
+		printf("Couldn't read " PARANOID_PATH "?\n");
+		goto out_close;
+	}
+
+	current = strtoul(buf, &end, 10);
+
+	if (end == buf) {
+		printf("Couldn't parse " PARANOID_PATH "?\n");
+		goto out_close;
+	}
+
+	if (current >= level)
+		goto out;
+
+	rc = 0;
+out_close:
+	fclose(f);
+out:
+	return rc;
+}
+
+static char auxv[4096];
+
+void *get_auxv_entry(int type)
+{
+	ElfW(auxv_t) *p;
+	void *result;
+	ssize_t num;
+	int fd;
+
+	fd = open("/proc/self/auxv", O_RDONLY);
+	if (fd == -1) {
+		perror("open");
+		return NULL;
+	}
+
+	result = NULL;
+
+	num = read(fd, auxv, sizeof(auxv));
+	if (num < 0) {
+		perror("read");
+		goto out;
+	}
+
+	if (num > sizeof(auxv)) {
+		printf("Overflowed auxv buffer\n");
+		goto out;
+	}
+
+	p = (ElfW(auxv_t) *)auxv;
+
+	while (p->a_type != AT_NULL) {
+		if (p->a_type == type) {
+			result = (void *)p->a_un.a_val;
+			break;
+		}
+
+		p++;
+	}
+out:
+	close(fd);
+	return result;
+}

tools/testing/selftests/powerpc/pmu/lib.h

@@ -0,0 +1,42 @@
+/*
+ * Copyright 2014, Michael Ellerman, IBM Corp.
+ * Licensed under GPLv2.
+ */
+
+#ifndef __SELFTESTS_POWERPC_PMU_LIB_H
+#define __SELFTESTS_POWERPC_PMU_LIB_H
+
+#include <stdio.h>
+#include <stdint.h>
+#include <string.h>
+#include <unistd.h>
+
+union pipe {
+	struct {
+		int read_fd;
+		int write_fd;
+	};
+	int fds[2];
+};
+
+extern int pick_online_cpu(void);
+extern int bind_to_cpu(int cpu);
+extern int kill_child_and_wait(pid_t child_pid);
+extern int wait_for_child(pid_t child_pid);
+extern int sync_with_child(union pipe read_pipe, union pipe write_pipe);
+extern int wait_for_parent(union pipe read_pipe);
+extern int notify_parent(union pipe write_pipe);
+extern int notify_parent_of_error(union pipe write_pipe);
+extern pid_t eat_cpu(int (test_function)(void));
+extern bool require_paranoia_below(int level);
+extern void *get_auxv_entry(int type);
+
+struct addr_range {
+	uint64_t first, last;
+};
+
+extern struct addr_range libc, vdso;
+
+int parse_proc_maps(void);
+
+#endif /* __SELFTESTS_POWERPC_PMU_LIB_H */

tools/testing/selftests/powerpc/pmu/loop.S

@@ -0,0 +1,43 @@
+/*
+ * Copyright 2013, Michael Ellerman, IBM Corp.
+ * Licensed under GPLv2.
+ */
+
+#include <ppc-asm.h>
+
+	.text
+
+FUNC_START(thirty_two_instruction_loop)
+	cmpdi	r3,0
+	beqlr
+	addi	r4,r3,1
+	addi	r4,r4,1
+	addi	r4,r4,1
+	addi	r4,r4,1
+	addi	r4,r4,1
+	addi	r4,r4,1
+	addi	r4,r4,1
+	addi	r4,r4,1
+	addi	r4,r4,1
+	addi	r4,r4,1
+	addi	r4,r4,1
+	addi	r4,r4,1
+	addi	r4,r4,1
+	addi	r4,r4,1
+	addi	r4,r4,1
+	addi	r4,r4,1
+	addi	r4,r4,1
+	addi	r4,r4,1
+	addi	r4,r4,1
+	addi	r4,r4,1
+	addi	r4,r4,1
+	addi	r4,r4,1
+	addi	r4,r4,1
+	addi	r4,r4,1
+	addi	r4,r4,1
+	addi	r4,r4,1
+	addi	r4,r4,1
+	addi	r4,r4,1	# 28 addi's
+	subi	r3,r3,1
+	b	FUNC_NAME(thirty_two_instruction_loop)
+FUNC_END(thirty_two_instruction_loop)

tools/testing/selftests/powerpc/pmu/per_event_excludes.c

@@ -0,0 +1,114 @@
+/*
+ * Copyright 2014, Michael Ellerman, IBM Corp.
+ * Licensed under GPLv2.
+ */
+
+#define _GNU_SOURCE
+
+#include <elf.h>
+#include <limits.h>
+#include <stdio.h>
+#include <stdbool.h>
+#include <string.h>
+#include <sys/prctl.h>
+
+#include "event.h"
+#include "lib.h"
+#include "utils.h"
+
+/*
+ * Test that per-event excludes work.
+ */
+
+static int per_event_excludes(void)
+{
+	struct event *e, events[4];
+	char *platform;
+	int i;
+
+	platform = (char *)get_auxv_entry(AT_BASE_PLATFORM);
+	FAIL_IF(!platform);
+	SKIP_IF(strcmp(platform, "power8") != 0);
+
+	/*
+	 * We need to create the events disabled, otherwise the running/enabled
+	 * counts don't match up.
+	 */
+	e = &events[0];
+	event_init_opts(e, PERF_COUNT_HW_INSTRUCTIONS,
+			PERF_TYPE_HARDWARE, "instructions");
+	e->attr.disabled = 1;
+
+	e = &events[1];
+	event_init_opts(e, PERF_COUNT_HW_INSTRUCTIONS,
+			PERF_TYPE_HARDWARE, "instructions(k)");
+	e->attr.disabled = 1;
+	e->attr.exclude_user = 1;
+	e->attr.exclude_hv = 1;
+
+	e = &events[2];
+	event_init_opts(e, PERF_COUNT_HW_INSTRUCTIONS,
+			PERF_TYPE_HARDWARE, "instructions(h)");
+	e->attr.disabled = 1;
+	e->attr.exclude_user = 1;
+	e->attr.exclude_kernel = 1;
+
+	e = &events[3];
+	event_init_opts(e, PERF_COUNT_HW_INSTRUCTIONS,
+			PERF_TYPE_HARDWARE, "instructions(u)");
+	e->attr.disabled = 1;
+	e->attr.exclude_hv = 1;
+	e->attr.exclude_kernel = 1;
+
+	FAIL_IF(event_open(&events[0]));
+
+	/*
+	 * The open here will fail if we don't have per event exclude support,
+	 * because the second event has an incompatible set of exclude settings
+	 * and we're asking for the events to be in a group.
+	 */
+	for (i = 1; i < 4; i++)
+		FAIL_IF(event_open_with_group(&events[i], events[0].fd));
+
+	/*
+	 * Even though the above will fail without per-event excludes we keep
+	 * testing in order to be thorough.
+	 */
+	prctl(PR_TASK_PERF_EVENTS_ENABLE);
+
+	/* Spin for a while */
+	for (i = 0; i < INT_MAX; i++)
+		asm volatile("" : : : "memory");
+
+	prctl(PR_TASK_PERF_EVENTS_DISABLE);
+
+	for (i = 0; i < 4; i++) {
+		FAIL_IF(event_read(&events[i]));
+		event_report(&events[i]);
+	}
+
+	/*
+	 * We should see that all events have enabled == running. That
+	 * shows that they were all on the PMU at once.
+	 */
+	for (i = 0; i < 4; i++)
+		FAIL_IF(events[i].result.running != events[i].result.enabled);
+
+	/*
+	 * We can also check that the result for instructions is >= all the
+	 * other counts. That's because it is counting all instructions while
+	 * the others are counting a subset.
+	 */
+	for (i = 1; i < 4; i++)
+		FAIL_IF(events[0].result.value < events[i].result.value);
+
+	for (i = 0; i < 4; i++)
+		event_close(&events[i]);
+
+	return 0;
+}
+
+int main(void)
+{
+	return test_harness(per_event_excludes, "per_event_excludes");
+}

tools/testing/selftests/powerpc/primitives/Makefile

@@ -0,0 +1,17 @@
+CFLAGS += -I$(CURDIR)
+
+PROGS := load_unaligned_zeropad
+
+all: $(PROGS)
+
+$(PROGS): ../harness.c
+
+run_tests: all
+	@-for PROG in $(PROGS); do \
+		./$$PROG; \
+	done;
+
+clean:
+	rm -f $(PROGS) *.o
+
+.PHONY: all run_tests clean

tools/testing/selftests/powerpc/primitives/asm/asm-compat.h

@@ -0,0 +1 @@
+../.././../../../../arch/powerpc/include/asm/asm-compat.h

tools/testing/selftests/powerpc/primitives/load_unaligned_zeropad.c

@@ -0,0 +1,147 @@
+/*
+ * Userspace test harness for load_unaligned_zeropad. Creates two
+ * pages and uses mprotect to prevent access to the second page and
+ * a SEGV handler that walks the exception tables and runs the fixup
+ * routine.
+ *
+ * The results are compared against a normal load that is that is
+ * performed while access to the second page is enabled via mprotect.
+ *
+ * Copyright (C) 2014 Anton Blanchard <anton@au.ibm.com>, IBM
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License
+ * as published by the Free Software Foundation; either version
+ * 2 of the License, or (at your option) any later version.
+ */
+
+#include <stdlib.h>
+#include <string.h>
+#include <stdio.h>
+#include <stdbool.h>
+#include <signal.h>
+#include <unistd.h>
+#include <sys/mman.h>
+
+#define FIXUP_SECTION ".ex_fixup"
+
+#include "word-at-a-time.h"
+
+#include "utils.h"
+
+
+static int page_size;
+static char *mem_region;
+
+static int protect_region(void)
+{
+	if (mprotect(mem_region + page_size, page_size, PROT_NONE)) {
+		perror("mprotect");
+		return 1;
+	}
+
+	return 0;
+}
+
+static int unprotect_region(void)
+{
+	if (mprotect(mem_region + page_size, page_size, PROT_READ|PROT_WRITE)) {
+		perror("mprotect");
+		return 1;
+	}
+
+	return 0;
+}
+
+extern char __start___ex_table[];
+extern char __stop___ex_table[];
+
+#if defined(__powerpc64__)
+#define UCONTEXT_NIA(UC)	(UC)->uc_mcontext.gp_regs[PT_NIP]
+#elif defined(__powerpc__)
+#define UCONTEXT_NIA(UC)	(UC)->uc_mcontext.uc_regs->gregs[PT_NIP]
+#else
+#error implement UCONTEXT_NIA
+#endif
+
+static int segv_error;
+
+static void segv_handler(int signr, siginfo_t *info, void *ptr)
+{
+	ucontext_t *uc = (ucontext_t *)ptr;
+	unsigned long addr = (unsigned long)info->si_addr;
+	unsigned long *ip = &UCONTEXT_NIA(uc);
+	unsigned long *ex_p = (unsigned long *)__start___ex_table;
+
+	while (ex_p < (unsigned long *)__stop___ex_table) {
+		unsigned long insn, fixup;
+
+		insn = *ex_p++;
+		fixup = *ex_p++;
+
+		if (insn == *ip) {
+			*ip = fixup;
+			return;
+		}
+	}
+
+	printf("No exception table match for NIA %lx ADDR %lx\n", *ip, addr);
+	segv_error++;
+}
+
+static void setup_segv_handler(void)
+{
+	struct sigaction action;
+
+	memset(&action, 0, sizeof(action));
+	action.sa_sigaction = segv_handler;
+	action.sa_flags = SA_SIGINFO;
+	sigaction(SIGSEGV, &action, NULL);
+}
+
+static int do_one_test(char *p, int page_offset)
+{
+	unsigned long should;
+	unsigned long got;
+
+	FAIL_IF(unprotect_region());
+	should = *(unsigned long *)p;
+	FAIL_IF(protect_region());
+
+	got = load_unaligned_zeropad(p);
+
+	if (should != got)
+		printf("offset %u load_unaligned_zeropad returned 0x%lx, should be 0x%lx\n", page_offset, got, should);
+
+	return 0;
+}
+
+static int test_body(void)
+{
+	unsigned long i;
+
+	page_size = getpagesize();
+	mem_region = mmap(NULL, page_size * 2, PROT_READ|PROT_WRITE,
+		MAP_PRIVATE|MAP_ANONYMOUS, -1, 0);
+
+	FAIL_IF(mem_region == MAP_FAILED);
+
+	for (i = 0; i < page_size; i++)
+		mem_region[i] = i;
+
+	memset(mem_region+page_size, 0, page_size);
+
+	setup_segv_handler();
+
+	for (i = 0; i < page_size; i++)
+		FAIL_IF(do_one_test(mem_region+i, i));
+
+	FAIL_IF(segv_error);
+
+	return 0;
+}
+
+int main(void)
+{
+	return test_harness(test_body, "load_unaligned_zeropad");
+}

tools/testing/selftests/powerpc/primitives/word-at-a-time.h

@@ -0,0 +1 @@
+../../../../../arch/powerpc/include/asm/word-at-a-time.h

tools/testing/selftests/powerpc/subunit.h

@@ -0,0 +1,52 @@
+/*
+ * Copyright 2013, Michael Ellerman, IBM Corp.
+ * Licensed under GPLv2.
+ */
+
+#ifndef _SELFTESTS_POWERPC_SUBUNIT_H
+#define _SELFTESTS_POWERPC_SUBUNIT_H
+
+static inline void test_start(char *name)
+{
+	printf("test: %s\n", name);
+}
+
+static inline void test_failure_detail(char *name, char *detail)
+{
+	printf("failure: %s [%s]\n", name, detail);
+}
+
+static inline void test_failure(char *name)
+{
+	printf("failure: %s\n", name);
+}
+
+static inline void test_error(char *name)
+{
+	printf("error: %s\n", name);
+}
+
+static inline void test_skip(char *name)
+{
+	printf("skip: %s\n", name);
+}
+
+static inline void test_success(char *name)
+{
+	printf("success: %s\n", name);
+}
+
+static inline void test_finish(char *name, int status)
+{
+	if (status)
+		test_failure(name);
+	else
+		test_success(name);
+}
+
+static inline void test_set_git_version(char *value)
+{
+	printf("tags: git_version:%s\n", value);
+}
+
+#endif /* _SELFTESTS_POWERPC_SUBUNIT_H */

tools/testing/selftests/powerpc/tm/Makefile

@@ -0,0 +1,15 @@
+PROGS := tm-resched-dscr
+
+all: $(PROGS)
+
+$(PROGS): ../harness.c
+
+run_tests: all
+	@-for PROG in $(PROGS); do \
+		./$$PROG; \
+	done;
+
+clean:
+	rm -f $(PROGS) *.o
+
+.PHONY: all run_tests clean

tools/testing/selftests/powerpc/tm/tm-resched-dscr.c

@@ -0,0 +1,98 @@
+/* Test context switching to see if the DSCR SPR is correctly preserved
+ * when within a transaction.
+ *
+ * Note: We assume that the DSCR has been left at the default value (0)
+ * for all CPUs.
+ *
+ * Method:
+ *
+ * Set a value into the DSCR.
+ *
+ * Start a transaction, and suspend it (*).
+ *
+ * Hard loop checking to see if the transaction has become doomed.
+ *
+ * Now that we *may* have been preempted, record the DSCR and TEXASR SPRS.
+ *
+ * If the abort was because of a context switch, check the DSCR value.
+ * Otherwise, try again.
+ *
+ * (*) If the transaction is not suspended we can't see the problem because
+ * the transaction abort handler will restore the DSCR to it's checkpointed
+ * value before we regain control.
+ */
+
+#include <inttypes.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <assert.h>
+#include <asm/tm.h>
+
+#include "utils.h"
+
+#define TBEGIN          ".long 0x7C00051D ;"
+#define TEND            ".long 0x7C00055D ;"
+#define TCHECK          ".long 0x7C00059C ;"
+#define TSUSPEND        ".long 0x7C0005DD ;"
+#define TRESUME         ".long 0x7C2005DD ;"
+#define SPRN_TEXASR     0x82
+#define SPRN_DSCR       0x03
+
+int test_body(void)
+{
+	uint64_t rv, dscr1 = 1, dscr2, texasr;
+
+	printf("Check DSCR TM context switch: ");
+	fflush(stdout);
+	for (;;) {
+		rv = 1;
+		asm __volatile__ (
+			/* set a known value into the DSCR */
+			"ld      3, %[dscr1];"
+			"mtspr   %[sprn_dscr], 3;"
+
+			/* start and suspend a transaction */
+			TBEGIN
+			"beq     1f;"
+			TSUSPEND
+
+			/* hard loop until the transaction becomes doomed */
+			"2: ;"
+			TCHECK
+			"bc      4, 0, 2b;"
+
+			/* record DSCR and TEXASR */
+			"mfspr   3, %[sprn_dscr];"
+			"std     3, %[dscr2];"
+			"mfspr   3, %[sprn_texasr];"
+			"std     3, %[texasr];"
+
+			TRESUME
+			TEND
+			"li      %[rv], 0;"
+			"1: ;"
+			: [rv]"=r"(rv), [dscr2]"=m"(dscr2), [texasr]"=m"(texasr)
+			: [dscr1]"m"(dscr1)
+			, [sprn_dscr]"i"(SPRN_DSCR), [sprn_texasr]"i"(SPRN_TEXASR)
+			: "memory", "r3"
+		);
+		assert(rv); /* make sure the transaction aborted */
+		if ((texasr >> 56) != TM_CAUSE_RESCHED) {
+			putchar('.');
+			fflush(stdout);
+			continue;
+		}
+		if (dscr2 != dscr1) {
+			printf(" FAIL\n");
+			return 1;
+		} else {
+			printf(" OK\n");
+			return 0;
+		}
+	}
+}
+
+int main(void)
+{
+	return test_harness(test_body, "tm_resched_dscr");
+}

tools/testing/selftests/powerpc/utils.h

@@ -0,0 +1,49 @@
+/*
+ * Copyright 2013, Michael Ellerman, IBM Corp.
+ * Licensed under GPLv2.
+ */
+
+#ifndef _SELFTESTS_POWERPC_UTILS_H
+#define _SELFTESTS_POWERPC_UTILS_H
+
+#include <stdint.h>
+#include <stdbool.h>
+
+/* Avoid headaches with PRI?64 - just use %ll? always */
+typedef unsigned long long u64;
+typedef   signed long long s64;
+
+/* Just for familiarity */
+typedef uint32_t u32;
+typedef uint8_t u8;
+
+
+int test_harness(int (test_function)(void), char *name);
+
+
+/* Yes, this is evil */
+#define FAIL_IF(x)						\
+do {								\
+	if ((x)) {						\
+		fprintf(stderr,					\
+		"[FAIL] Test FAILED on line %d\n", __LINE__);	\
+		return 1;					\
+	}							\
+} while (0)
+
+/* The test harness uses this, yes it's gross */
+#define MAGIC_SKIP_RETURN_VALUE	99
+
+#define SKIP_IF(x)						\
+do {								\
+	if ((x)) {						\
+		fprintf(stderr,					\
+		"[SKIP] Test skipped on line %d\n", __LINE__);	\
+		return MAGIC_SKIP_RETURN_VALUE;			\
+	}							\
+} while (0)
+
+#define _str(s) #s
+#define str(s) _str(s)
+
+#endif /* _SELFTESTS_POWERPC_UTILS_H */