Changes from v8:
* Probe PMU version based on ID_DFR0
* pmccntr_read() now returns 64bit and can handle both 32bit and 64bit
PMCCNTR based on PMU version.
* Add pmccntr_write() support
* Use a common printf format PRId64 to support 64bit variable smoothly in
test functions
* Add barriers to several PMU write functions
* Verfied on different execution modes
Note:
1) Current KVM code has bugs in handling PMCCFILTR write. A fix (see
below) is required for this unit testing code to work correctly under
KVM mode.
https://lists.cs.columbia.edu/pipermail/kvmarm/2016-November/022134.html.
Thanks,
-Wei
Wei Huang (3):
arm: Add PMU test
arm: pmu: Check cycle count increases
arm: pmu: Add CPI checking
arm/Makefile.common | 3 +-
arm/pmu.c | 339 ++++++++++++++++++++++++++++++++++++++++++++++++++++
arm/unittests.cfg | 19 +++
3 files changed, 360 insertions(+), 1 deletion(-)
create mode 100644 arm/pmu.c
--
1.8.3.1
From: Christopher Covington <[email protected]>
Ensure that reads of the PMCCNTR_EL0 are monotonically increasing,
even for the smallest delta of two subsequent reads.
Signed-off-by: Christopher Covington <[email protected]>
Signed-off-by: Wei Huang <[email protected]>
---
arm/pmu.c | 156 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
1 file changed, 156 insertions(+)
diff --git a/arm/pmu.c b/arm/pmu.c
index 9d9c53b..fa87de4 100644
--- a/arm/pmu.c
+++ b/arm/pmu.c
@@ -15,6 +15,9 @@
#include "libcflat.h"
#include "asm/barrier.h"
+#define PMU_PMCR_E (1 << 0)
+#define PMU_PMCR_C (1 << 2)
+#define PMU_PMCR_LC (1 << 6)
#define PMU_PMCR_N_SHIFT 11
#define PMU_PMCR_N_MASK 0x1f
#define PMU_PMCR_ID_SHIFT 16
@@ -22,6 +25,14 @@
#define PMU_PMCR_IMP_SHIFT 24
#define PMU_PMCR_IMP_MASK 0xff
+#define ID_DFR0_PERFMON_SHIFT 24
+#define ID_DFR0_PERFMON_MASK 0xf
+
+#define PMU_CYCLE_IDX 31
+
+#define NR_SAMPLES 10
+
+static unsigned int pmu_version;
#if defined(__arm__)
static inline uint32_t pmcr_read(void)
{
@@ -30,6 +41,69 @@ static inline uint32_t pmcr_read(void)
asm volatile("mrc p15, 0, %0, c9, c12, 0" : "=r" (ret));
return ret;
}
+
+static inline void pmcr_write(uint32_t value)
+{
+ asm volatile("mcr p15, 0, %0, c9, c12, 0" : : "r" (value));
+ isb();
+}
+
+static inline void pmselr_write(uint32_t value)
+{
+ asm volatile("mcr p15, 0, %0, c9, c12, 5" : : "r" (value));
+ isb();
+}
+
+static inline void pmxevtyper_write(uint32_t value)
+{
+ asm volatile("mcr p15, 0, %0, c9, c13, 1" : : "r" (value));
+}
+
+static inline uint64_t pmccntr_read(void)
+{
+ uint32_t lo, hi = 0;
+
+ if (pmu_version == 0x3)
+ asm volatile("mrrc p15, 0, %0, %1, c9" : "=r" (lo), "=r" (hi));
+ else
+ asm volatile("mrc p15, 0, %0, c9, c13, 0" : "=r" (lo));
+
+ return ((uint64_t)hi << 32) | lo;
+}
+
+static inline void pmccntr_write(uint64_t value)
+{
+ uint32_t lo, hi;
+
+ lo = value & 0xffffffff;
+ hi = (value >> 32) & 0xffffffff;
+
+ if (pmu_version == 0x3)
+ asm volatile("mcrr p15, 0, %0, %1, c9" : : "r" (lo), "r" (hi));
+ else
+ asm volatile("mcr p15, 0, %0, c9, c13, 0" : : "r" (lo));
+}
+
+static inline void pmcntenset_write(uint32_t value)
+{
+ asm volatile("mcr p15, 0, %0, c9, c12, 1" : : "r" (value));
+}
+
+/* PMCCFILTR is an obsolete name for PMXEVTYPER31 in ARMv7 */
+static inline void pmccfiltr_write(uint32_t value)
+{
+ pmselr_write(PMU_CYCLE_IDX);
+ pmxevtyper_write(value);
+ isb();
+}
+
+static inline uint32_t id_dfr0_read(void)
+{
+ uint32_t val;
+
+ asm volatile("mrc p15, 0, %0, c0, c1, 2" : "=r" (val));
+ return val;
+}
#elif defined(__aarch64__)
static inline uint32_t pmcr_read(void)
{
@@ -38,6 +112,44 @@ static inline uint32_t pmcr_read(void)
asm volatile("mrs %0, pmcr_el0" : "=r" (ret));
return ret;
}
+
+static inline void pmcr_write(uint32_t value)
+{
+ asm volatile("msr pmcr_el0, %0" : : "r" (value));
+ isb();
+}
+
+static inline uint64_t pmccntr_read(void)
+{
+ uint64_t cycles;
+
+ asm volatile("mrs %0, pmccntr_el0" : "=r" (cycles));
+ return cycles;
+}
+
+static inline void pmccntr_write(uint64_t value)
+{
+ asm volatile("msr pmccntr_el0, %0" : : "r" (value));
+}
+
+static inline void pmcntenset_write(uint32_t value)
+{
+ asm volatile("msr pmcntenset_el0, %0" : : "r" (value));
+}
+
+static inline void pmccfiltr_write(uint32_t value)
+{
+ asm volatile("msr pmccfiltr_el0, %0" : : "r" (value));
+ isb();
+}
+
+static inline uint32_t id_dfr0_read(void)
+{
+ uint32_t id;
+
+ asm volatile("mrs %0, id_dfr0_el1" : "=r" (id));
+ return id;
+}
#endif
/*
@@ -64,11 +176,55 @@ static bool check_pmcr(void)
return ((pmcr >> PMU_PMCR_IMP_SHIFT) & PMU_PMCR_IMP_MASK) != 0;
}
+/*
+ * Ensure that the cycle counter progresses between back-to-back reads.
+ */
+static bool check_cycles_increase(void)
+{
+ bool success = true;
+
+ pmccntr_write(0);
+ pmcr_write(pmcr_read() | PMU_PMCR_LC | PMU_PMCR_C | PMU_PMCR_E);
+
+ for (int i = 0; i < NR_SAMPLES; i++) {
+ uint64_t a, b;
+
+ a = pmccntr_read();
+ b = pmccntr_read();
+
+ if (a >= b) {
+ printf("Read %"PRId64" then %"PRId64".\n", a, b);
+ success = false;
+ break;
+ }
+ }
+
+ pmcr_write(pmcr_read() & ~PMU_PMCR_E);
+
+ return success;
+}
+
+void pmu_init(void)
+{
+ uint32_t dfr0;
+
+ /* probe pmu version */
+ dfr0 = id_dfr0_read();
+ pmu_version = (dfr0 >> ID_DFR0_PERFMON_SHIFT) & ID_DFR0_PERFMON_MASK;
+ printf("PMU version: %d\n", pmu_version);
+
+ /* init for PMU event access, right now only care about cycle count */
+ pmcntenset_write(1 << PMU_CYCLE_IDX);
+ pmccfiltr_write(0); /* count cycles in EL0, EL1, but not EL2 */
+}
+
int main(void)
{
report_prefix_push("pmu");
+ pmu_init();
report("Control register", check_pmcr());
+ report("Monotonically increasing cycle count", check_cycles_increase());
return report_summary();
}
--
1.8.3.1
From: Christopher Covington <[email protected]>
Calculate the numbers of cycles per instruction (CPI) implied by ARM
PMU cycle counter values. The code includes a strict checking facility
intended for the -icount option in TCG mode in the configuration file.
Signed-off-by: Christopher Covington <[email protected]>
Signed-off-by: Wei Huang <[email protected]>
---
arm/pmu.c | 111 +++++++++++++++++++++++++++++++++++++++++++++++++++++-
arm/unittests.cfg | 14 +++++++
2 files changed, 124 insertions(+), 1 deletion(-)
diff --git a/arm/pmu.c b/arm/pmu.c
index fa87de4..b36c4fb 100644
--- a/arm/pmu.c
+++ b/arm/pmu.c
@@ -104,6 +104,25 @@ static inline uint32_t id_dfr0_read(void)
asm volatile("mrc p15, 0, %0, c0, c1, 2" : "=r" (val));
return val;
}
+
+/*
+ * Extra instructions inserted by the compiler would be difficult to compensate
+ * for, so hand assemble everything between, and including, the PMCR accesses
+ * to start and stop counting.
+ */
+static inline void loop(int i, uint32_t pmcr)
+{
+ asm volatile(
+ " mcr p15, 0, %[pmcr], c9, c12, 0\n"
+ " isb\n"
+ "1: subs %[i], %[i], #1\n"
+ " bgt 1b\n"
+ " mcr p15, 0, %[z], c9, c12, 0\n"
+ " isb\n"
+ : [i] "+r" (i)
+ : [pmcr] "r" (pmcr), [z] "r" (0)
+ : "cc");
+}
#elif defined(__aarch64__)
static inline uint32_t pmcr_read(void)
{
@@ -150,6 +169,25 @@ static inline uint32_t id_dfr0_read(void)
asm volatile("mrs %0, id_dfr0_el1" : "=r" (id));
return id;
}
+
+/*
+ * Extra instructions inserted by the compiler would be difficult to compensate
+ * for, so hand assemble everything between, and including, the PMCR accesses
+ * to start and stop counting.
+ */
+static inline void loop(int i, uint32_t pmcr)
+{
+ asm volatile(
+ " msr pmcr_el0, %[pmcr]\n"
+ " isb\n"
+ "1: subs %[i], %[i], #1\n"
+ " b.gt 1b\n"
+ " msr pmcr_el0, xzr\n"
+ " isb\n"
+ : [i] "+r" (i)
+ : [pmcr] "r" (pmcr)
+ : "cc");
+}
#endif
/*
@@ -204,6 +242,71 @@ static bool check_cycles_increase(void)
return success;
}
+/*
+ * Execute a known number of guest instructions. Only odd instruction counts
+ * greater than or equal to 3 are supported by the in-line assembly code. The
+ * control register (PMCR_EL0) is initialized with the provided value (allowing
+ * for example for the cycle counter or event counters to be reset). At the end
+ * of the exact instruction loop, zero is written to PMCR_EL0 to disable
+ * counting, allowing the cycle counter or event counters to be read at the
+ * leisure of the calling code.
+ */
+static void measure_instrs(int num, uint32_t pmcr)
+{
+ int i = (num - 1) / 2;
+
+ assert(num >= 3 && ((num - 1) % 2 == 0));
+ loop(i, pmcr);
+}
+
+/*
+ * Measure cycle counts for various known instruction counts. Ensure that the
+ * cycle counter progresses (similar to check_cycles_increase() but with more
+ * instructions and using reset and stop controls). If supplied a positive,
+ * nonzero CPI parameter, also strictly check that every measurement matches
+ * it. Strict CPI checking is used to test -icount mode.
+ */
+static bool check_cpi(int cpi)
+{
+ uint32_t pmcr = pmcr_read() | PMU_PMCR_LC | PMU_PMCR_C | PMU_PMCR_E;
+
+ if (cpi > 0)
+ printf("Checking for CPI=%d.\n", cpi);
+ printf("instrs : cycles0 cycles1 ...\n");
+
+ for (unsigned int i = 3; i < 300; i += 32) {
+ uint64_t avg, sum = 0;
+
+ printf("%d :", i);
+ for (int j = 0; j < NR_SAMPLES; j++) {
+ uint64_t cycles;
+
+ pmccntr_write(0);
+ measure_instrs(i, pmcr);
+ cycles = pmccntr_read();
+ printf(" %"PRId64"", cycles);
+
+ /*
+ * The cycles taken by the loop above should fit in
+ * 32 bits easily. We check the upper 32 bits of the
+ * cycle counter to make sure there is no supprise.
+ */
+ if (!cycles || (cpi > 0 && cycles != i * cpi) ||
+ (cycles & 0xffffffff00000000)) {
+ printf("\n");
+ return false;
+ }
+
+ sum += cycles;
+ }
+ avg = sum / NR_SAMPLES;
+ printf(" sum=%"PRId64" avg=%"PRId64" avg_ipc=%"PRId64" "
+ "avg_cpi=%"PRId64"\n", sum, avg, i / avg, avg / i);
+ }
+
+ return true;
+}
+
void pmu_init(void)
{
uint32_t dfr0;
@@ -218,13 +321,19 @@ void pmu_init(void)
pmccfiltr_write(0); /* count cycles in EL0, EL1, but not EL2 */
}
-int main(void)
+int main(int argc, char *argv[])
{
+ int cpi = 0;
+
+ if (argc >= 1)
+ cpi = atol(argv[0]);
+
report_prefix_push("pmu");
pmu_init();
report("Control register", check_pmcr());
report("Monotonically increasing cycle count", check_cycles_increase());
+ report("Cycle/instruction ratio", check_cpi(cpi));
return report_summary();
}
diff --git a/arm/unittests.cfg b/arm/unittests.cfg
index 7645180..2050dc8 100644
--- a/arm/unittests.cfg
+++ b/arm/unittests.cfg
@@ -59,3 +59,17 @@ groups = selftest
[pmu]
file = pmu.flat
groups = pmu
+
+# Test PMU support (TCG) with -icount IPC=1
+[pmu-tcg-icount-1]
+file = pmu.flat
+extra_params = -icount 0 -append '1'
+groups = pmu
+accel = tcg
+
+# Test PMU support (TCG) with -icount IPC=256
+[pmu-tcg-icount-256]
+file = pmu.flat
+extra_params = -icount 8 -append '256'
+groups = pmu
+accel = tcg
--
1.8.3.1
From: Christopher Covington <[email protected]>
Beginning with a simple sanity check of the control register, add
a unit test for the ARM Performance Monitors Unit (PMU).
Signed-off-by: Christopher Covington <[email protected]>
Signed-off-by: Wei Huang <[email protected]>
Reviewed-by: Andrew Jones <[email protected]>
---
arm/Makefile.common | 3 ++-
arm/pmu.c | 74 +++++++++++++++++++++++++++++++++++++++++++++++++++++
arm/unittests.cfg | 5 ++++
3 files changed, 81 insertions(+), 1 deletion(-)
create mode 100644 arm/pmu.c
diff --git a/arm/Makefile.common b/arm/Makefile.common
index ccb554d..f98f422 100644
--- a/arm/Makefile.common
+++ b/arm/Makefile.common
@@ -11,7 +11,8 @@ endif
tests-common = \
$(TEST_DIR)/selftest.flat \
- $(TEST_DIR)/spinlock-test.flat
+ $(TEST_DIR)/spinlock-test.flat \
+ $(TEST_DIR)/pmu.flat
all: test_cases
diff --git a/arm/pmu.c b/arm/pmu.c
new file mode 100644
index 0000000..9d9c53b
--- /dev/null
+++ b/arm/pmu.c
@@ -0,0 +1,74 @@
+/*
+ * Test the ARM Performance Monitors Unit (PMU).
+ *
+ * Copyright (c) 2015-2016, The Linux Foundation. All rights reserved.
+ *
+ * This program is free software; you can redistribute it and/or modify it
+ * under the terms of the GNU Lesser General Public License version 2.1 and
+ * only version 2.1 as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE. See the GNU Lesser General Public License
+ * for more details.
+ */
+#include "libcflat.h"
+#include "asm/barrier.h"
+
+#define PMU_PMCR_N_SHIFT 11
+#define PMU_PMCR_N_MASK 0x1f
+#define PMU_PMCR_ID_SHIFT 16
+#define PMU_PMCR_ID_MASK 0xff
+#define PMU_PMCR_IMP_SHIFT 24
+#define PMU_PMCR_IMP_MASK 0xff
+
+#if defined(__arm__)
+static inline uint32_t pmcr_read(void)
+{
+ uint32_t ret;
+
+ asm volatile("mrc p15, 0, %0, c9, c12, 0" : "=r" (ret));
+ return ret;
+}
+#elif defined(__aarch64__)
+static inline uint32_t pmcr_read(void)
+{
+ uint32_t ret;
+
+ asm volatile("mrs %0, pmcr_el0" : "=r" (ret));
+ return ret;
+}
+#endif
+
+/*
+ * As a simple sanity check on the PMCR_EL0, ensure the implementer field isn't
+ * null. Also print out a couple other interesting fields for diagnostic
+ * purposes. For example, as of fall 2016, QEMU TCG mode doesn't implement
+ * event counters and therefore reports zero event counters, but hopefully
+ * support for at least the instructions event will be added in the future and
+ * the reported number of event counters will become nonzero.
+ */
+static bool check_pmcr(void)
+{
+ uint32_t pmcr;
+
+ pmcr = pmcr_read();
+
+ printf("PMU implementer: %c\n",
+ (pmcr >> PMU_PMCR_IMP_SHIFT) & PMU_PMCR_IMP_MASK);
+ printf("Identification code: 0x%x\n",
+ (pmcr >> PMU_PMCR_ID_SHIFT) & PMU_PMCR_ID_MASK);
+ printf("Event counters: %d\n",
+ (pmcr >> PMU_PMCR_N_SHIFT) & PMU_PMCR_N_MASK);
+
+ return ((pmcr >> PMU_PMCR_IMP_SHIFT) & PMU_PMCR_IMP_MASK) != 0;
+}
+
+int main(void)
+{
+ report_prefix_push("pmu");
+
+ report("Control register", check_pmcr());
+
+ return report_summary();
+}
diff --git a/arm/unittests.cfg b/arm/unittests.cfg
index 3f6fa45..7645180 100644
--- a/arm/unittests.cfg
+++ b/arm/unittests.cfg
@@ -54,3 +54,8 @@ file = selftest.flat
smp = $MAX_SMP
extra_params = -append 'smp'
groups = selftest
+
+# Test PMU support
+[pmu]
+file = pmu.flat
+groups = pmu
--
1.8.3.1
On Fri, Nov 18, 2016 at 10:15:41PM -0600, Wei Huang wrote:
> From: Christopher Covington <[email protected]>
>
> Ensure that reads of the PMCCNTR_EL0 are monotonically increasing,
> even for the smallest delta of two subsequent reads.
>
> Signed-off-by: Christopher Covington <[email protected]>
> Signed-off-by: Wei Huang <[email protected]>
> ---
> arm/pmu.c | 156 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
> 1 file changed, 156 insertions(+)
>
> diff --git a/arm/pmu.c b/arm/pmu.c
> index 9d9c53b..fa87de4 100644
> --- a/arm/pmu.c
> +++ b/arm/pmu.c
> @@ -15,6 +15,9 @@
> #include "libcflat.h"
> #include "asm/barrier.h"
>
> +#define PMU_PMCR_E (1 << 0)
> +#define PMU_PMCR_C (1 << 2)
> +#define PMU_PMCR_LC (1 << 6)
> #define PMU_PMCR_N_SHIFT 11
> #define PMU_PMCR_N_MASK 0x1f
> #define PMU_PMCR_ID_SHIFT 16
> @@ -22,6 +25,14 @@
> #define PMU_PMCR_IMP_SHIFT 24
> #define PMU_PMCR_IMP_MASK 0xff
>
> +#define ID_DFR0_PERFMON_SHIFT 24
> +#define ID_DFR0_PERFMON_MASK 0xf
> +
> +#define PMU_CYCLE_IDX 31
> +
> +#define NR_SAMPLES 10
> +
> +static unsigned int pmu_version;
> #if defined(__arm__)
> static inline uint32_t pmcr_read(void)
> {
> @@ -30,6 +41,69 @@ static inline uint32_t pmcr_read(void)
> asm volatile("mrc p15, 0, %0, c9, c12, 0" : "=r" (ret));
> return ret;
> }
> +
> +static inline void pmcr_write(uint32_t value)
> +{
> + asm volatile("mcr p15, 0, %0, c9, c12, 0" : : "r" (value));
> + isb();
> +}
> +
> +static inline void pmselr_write(uint32_t value)
> +{
> + asm volatile("mcr p15, 0, %0, c9, c12, 5" : : "r" (value));
> + isb();
> +}
> +
> +static inline void pmxevtyper_write(uint32_t value)
> +{
> + asm volatile("mcr p15, 0, %0, c9, c13, 1" : : "r" (value));
> +}
> +
> +static inline uint64_t pmccntr_read(void)
> +{
> + uint32_t lo, hi = 0;
> +
> + if (pmu_version == 0x3)
> + asm volatile("mrrc p15, 0, %0, %1, c9" : "=r" (lo), "=r" (hi));
> + else
> + asm volatile("mrc p15, 0, %0, c9, c13, 0" : "=r" (lo));
This is fine for this single test, but I was actually thinking we'd
implement both the 32-bit and 64-bit accessors separately, allowing
us to add an additional test that tries both on AArch32 (PMUv3). I.e.
report_prefix_push("pmccntr");
report("32-bit read", pmccntr_read() != 0); /* test mrc access */
if (pmu_version == 0x3) {
u64 cntr64 = pmccntr_read64(); /* test mrrc access */
report("64-bit read", (cntr64 >> 32) == 0 && (u32)cntr64 != 0);
} else {
report_skip("64-bit read, PMUv3 required");
}
We can refactor when we add that test later though.
> +
> + return ((uint64_t)hi << 32) | lo;
> +}
> +
> +static inline void pmccntr_write(uint64_t value)
> +{
> + uint32_t lo, hi;
> +
> + lo = value & 0xffffffff;
> + hi = (value >> 32) & 0xffffffff;
> +
> + if (pmu_version == 0x3)
> + asm volatile("mcrr p15, 0, %0, %1, c9" : : "r" (lo), "r" (hi));
> + else
> + asm volatile("mcr p15, 0, %0, c9, c13, 0" : : "r" (lo));
> +}
> +
> +static inline void pmcntenset_write(uint32_t value)
> +{
> + asm volatile("mcr p15, 0, %0, c9, c12, 1" : : "r" (value));
> +}
> +
> +/* PMCCFILTR is an obsolete name for PMXEVTYPER31 in ARMv7 */
> +static inline void pmccfiltr_write(uint32_t value)
> +{
> + pmselr_write(PMU_CYCLE_IDX);
> + pmxevtyper_write(value);
> + isb();
> +}
> +
> +static inline uint32_t id_dfr0_read(void)
> +{
> + uint32_t val;
> +
> + asm volatile("mrc p15, 0, %0, c0, c1, 2" : "=r" (val));
> + return val;
> +}
> #elif defined(__aarch64__)
> static inline uint32_t pmcr_read(void)
> {
> @@ -38,6 +112,44 @@ static inline uint32_t pmcr_read(void)
> asm volatile("mrs %0, pmcr_el0" : "=r" (ret));
> return ret;
> }
> +
> +static inline void pmcr_write(uint32_t value)
> +{
> + asm volatile("msr pmcr_el0, %0" : : "r" (value));
> + isb();
> +}
> +
> +static inline uint64_t pmccntr_read(void)
> +{
> + uint64_t cycles;
> +
> + asm volatile("mrs %0, pmccntr_el0" : "=r" (cycles));
> + return cycles;
> +}
> +
> +static inline void pmccntr_write(uint64_t value)
> +{
> + asm volatile("msr pmccntr_el0, %0" : : "r" (value));
> +}
> +
> +static inline void pmcntenset_write(uint32_t value)
> +{
> + asm volatile("msr pmcntenset_el0, %0" : : "r" (value));
> +}
> +
> +static inline void pmccfiltr_write(uint32_t value)
> +{
> + asm volatile("msr pmccfiltr_el0, %0" : : "r" (value));
> + isb();
> +}
> +
> +static inline uint32_t id_dfr0_read(void)
> +{
> + uint32_t id;
> +
> + asm volatile("mrs %0, id_dfr0_el1" : "=r" (id));
> + return id;
> +}
> #endif
>
> /*
> @@ -64,11 +176,55 @@ static bool check_pmcr(void)
> return ((pmcr >> PMU_PMCR_IMP_SHIFT) & PMU_PMCR_IMP_MASK) != 0;
> }
>
> +/*
> + * Ensure that the cycle counter progresses between back-to-back reads.
> + */
> +static bool check_cycles_increase(void)
> +{
> + bool success = true;
> +
> + pmccntr_write(0);
> + pmcr_write(pmcr_read() | PMU_PMCR_LC | PMU_PMCR_C | PMU_PMCR_E);
> +
> + for (int i = 0; i < NR_SAMPLES; i++) {
> + uint64_t a, b;
> +
> + a = pmccntr_read();
> + b = pmccntr_read();
> +
> + if (a >= b) {
> + printf("Read %"PRId64" then %"PRId64".\n", a, b);
> + success = false;
> + break;
> + }
> + }
> +
> + pmcr_write(pmcr_read() & ~PMU_PMCR_E);
> +
> + return success;
> +}
> +
> +void pmu_init(void)
> +{
> + uint32_t dfr0;
> +
> + /* probe pmu version */
> + dfr0 = id_dfr0_read();
> + pmu_version = (dfr0 >> ID_DFR0_PERFMON_SHIFT) & ID_DFR0_PERFMON_MASK;
> + printf("PMU version: %d\n", pmu_version);
> +
> + /* init for PMU event access, right now only care about cycle count */
> + pmcntenset_write(1 << PMU_CYCLE_IDX);
> + pmccfiltr_write(0); /* count cycles in EL0, EL1, but not EL2 */
These last two register writes are specific to the cycle count test, so
I don't think they belong in a common pmu_init function. I'd keep them
at the top of check_cycles_increase.
> +}
> +
> int main(void)
> {
> report_prefix_push("pmu");
>
> + pmu_init();
> report("Control register", check_pmcr());
> + report("Monotonically increasing cycle count", check_cycles_increase());
>
> return report_summary();
> }
> --
> 1.8.3.1
>
Thanks,
drew
On Fri, Nov 18, 2016 at 10:15:42PM -0600, Wei Huang wrote:
> From: Christopher Covington <[email protected]>
>
> Calculate the numbers of cycles per instruction (CPI) implied by ARM
> PMU cycle counter values. The code includes a strict checking facility
> intended for the -icount option in TCG mode in the configuration file.
>
> Signed-off-by: Christopher Covington <[email protected]>
> Signed-off-by: Wei Huang <[email protected]>
> ---
> arm/pmu.c | 111 +++++++++++++++++++++++++++++++++++++++++++++++++++++-
> arm/unittests.cfg | 14 +++++++
> 2 files changed, 124 insertions(+), 1 deletion(-)
>
> diff --git a/arm/pmu.c b/arm/pmu.c
> index fa87de4..b36c4fb 100644
> --- a/arm/pmu.c
> +++ b/arm/pmu.c
> @@ -104,6 +104,25 @@ static inline uint32_t id_dfr0_read(void)
> asm volatile("mrc p15, 0, %0, c0, c1, 2" : "=r" (val));
> return val;
> }
> +
> +/*
> + * Extra instructions inserted by the compiler would be difficult to compensate
> + * for, so hand assemble everything between, and including, the PMCR accesses
> + * to start and stop counting.
> + */
> +static inline void loop(int i, uint32_t pmcr)
Thought you were going to rename this function.
> +{
> + asm volatile(
> + " mcr p15, 0, %[pmcr], c9, c12, 0\n"
> + " isb\n"
> + "1: subs %[i], %[i], #1\n"
> + " bgt 1b\n"
> + " mcr p15, 0, %[z], c9, c12, 0\n"
> + " isb\n"
> + : [i] "+r" (i)
> + : [pmcr] "r" (pmcr), [z] "r" (0)
> + : "cc");
> +}
> #elif defined(__aarch64__)
> static inline uint32_t pmcr_read(void)
> {
> @@ -150,6 +169,25 @@ static inline uint32_t id_dfr0_read(void)
> asm volatile("mrs %0, id_dfr0_el1" : "=r" (id));
> return id;
> }
> +
> +/*
> + * Extra instructions inserted by the compiler would be difficult to compensate
> + * for, so hand assemble everything between, and including, the PMCR accesses
> + * to start and stop counting.
> + */
> +static inline void loop(int i, uint32_t pmcr)
> +{
> + asm volatile(
> + " msr pmcr_el0, %[pmcr]\n"
> + " isb\n"
> + "1: subs %[i], %[i], #1\n"
> + " b.gt 1b\n"
> + " msr pmcr_el0, xzr\n"
> + " isb\n"
> + : [i] "+r" (i)
> + : [pmcr] "r" (pmcr)
> + : "cc");
> +}
> #endif
>
> /*
> @@ -204,6 +242,71 @@ static bool check_cycles_increase(void)
> return success;
> }
>
> +/*
> + * Execute a known number of guest instructions. Only odd instruction counts
> + * greater than or equal to 3 are supported by the in-line assembly code. The
> + * control register (PMCR_EL0) is initialized with the provided value (allowing
> + * for example for the cycle counter or event counters to be reset). At the end
> + * of the exact instruction loop, zero is written to PMCR_EL0 to disable
> + * counting, allowing the cycle counter or event counters to be read at the
> + * leisure of the calling code.
> + */
> +static void measure_instrs(int num, uint32_t pmcr)
> +{
> + int i = (num - 1) / 2;
> +
> + assert(num >= 3 && ((num - 1) % 2 == 0));
> + loop(i, pmcr);
> +}
> +
> +/*
> + * Measure cycle counts for various known instruction counts. Ensure that the
> + * cycle counter progresses (similar to check_cycles_increase() but with more
> + * instructions and using reset and stop controls). If supplied a positive,
> + * nonzero CPI parameter, also strictly check that every measurement matches
> + * it. Strict CPI checking is used to test -icount mode.
> + */
> +static bool check_cpi(int cpi)
> +{
> + uint32_t pmcr = pmcr_read() | PMU_PMCR_LC | PMU_PMCR_C | PMU_PMCR_E;
> +
> + if (cpi > 0)
> + printf("Checking for CPI=%d.\n", cpi);
> + printf("instrs : cycles0 cycles1 ...\n");
> +
> + for (unsigned int i = 3; i < 300; i += 32) {
> + uint64_t avg, sum = 0;
> +
> + printf("%d :", i);
> + for (int j = 0; j < NR_SAMPLES; j++) {
> + uint64_t cycles;
> +
> + pmccntr_write(0);
> + measure_instrs(i, pmcr);
> + cycles = pmccntr_read();
> + printf(" %"PRId64"", cycles);
> +
> + /*
> + * The cycles taken by the loop above should fit in
> + * 32 bits easily. We check the upper 32 bits of the
> + * cycle counter to make sure there is no supprise.
> + */
> + if (!cycles || (cpi > 0 && cycles != i * cpi) ||
> + (cycles & 0xffffffff00000000)) {
(cycles >> 32) != 0 would look better.
> + printf("\n");
We have 3 cases where we return false here. How about doing the tests
separately and adding descriptive print statements for each?
if (!cycles) {
printf("\ncycles not incrementing!\n");
return false;
} else if (cpi > 0 && cycles != i * cpi) {
...
} else if ((cycles >> 32) != 0) {
...
}
> + return false;
> + }
> +
> + sum += cycles;
> + }
> + avg = sum / NR_SAMPLES;
> + printf(" sum=%"PRId64" avg=%"PRId64" avg_ipc=%"PRId64" "
> + "avg_cpi=%"PRId64"\n", sum, avg, i / avg, avg / i);
> + }
> +
> + return true;
> +}
> +
> void pmu_init(void)
> {
> uint32_t dfr0;
> @@ -218,13 +321,19 @@ void pmu_init(void)
> pmccfiltr_write(0); /* count cycles in EL0, EL1, but not EL2 */
> }
>
> -int main(void)
> +int main(int argc, char *argv[])
> {
> + int cpi = 0;
> +
> + if (argc >= 1)
^ this '=' shouldn't be here
> + cpi = atol(argv[0]);
^ sigh, this is still zero...
Looks like you forgot all my comments from the last review round...
> +
> report_prefix_push("pmu");
>
> pmu_init();
> report("Control register", check_pmcr());
> report("Monotonically increasing cycle count", check_cycles_increase());
> + report("Cycle/instruction ratio", check_cpi(cpi));
>
> return report_summary();
> }
> diff --git a/arm/unittests.cfg b/arm/unittests.cfg
> index 7645180..2050dc8 100644
> --- a/arm/unittests.cfg
> +++ b/arm/unittests.cfg
> @@ -59,3 +59,17 @@ groups = selftest
> [pmu]
> file = pmu.flat
> groups = pmu
> +
> +# Test PMU support (TCG) with -icount IPC=1
> +[pmu-tcg-icount-1]
> +file = pmu.flat
> +extra_params = -icount 0 -append '1'
> +groups = pmu
> +accel = tcg
> +
> +# Test PMU support (TCG) with -icount IPC=256
> +[pmu-tcg-icount-256]
> +file = pmu.flat
> +extra_params = -icount 8 -append '256'
> +groups = pmu
> +accel = tcg
> --
> 1.8.3.1
>
drew