| Message ID | 20161124161033.11456-8-alex.bennee@linaro.org (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On Thu, Nov 24, 2016 at 04:10:29PM +0000, Alex Bennée wrote: > This adds a fairly brain dead torture test for TLB flushes intended for > stressing the MTTCG QEMU build. It takes the usual -smp option for > multiple CPUs. > > By default it CPU0 will do a TLBIALL flush after each cycle. You can > pass options via -append to control additional aspects of the test: > > - "page" flush each page in turn (one per function) > - "self" do the flush after each computation cycle > - "verbose" report progress on each computation cycle > > Signed-off-by: Alex Bennée <alex.bennee@linaro.org> > CC: Mark Rutland <mark.rutland@arm.com> > > --- > v2 > - rename to tlbflush-test > - made makefile changes cleaner > - added self/other flush mode > - create specific prefix > - whitespace fixes > v3 > - using new SMP framework for test runing > v4 > - merge in the unitests.cfg > v5 > - max out at -smp 4 > - printf fmtfix > v7 > - rename to tlbflush-code > - int -> bool flags > --- > arm/Makefile.common | 2 + > arm/tlbflush-code.c | 212 ++++++++++++++++++++++++++++++++++++++++++++++++++++ > arm/unittests.cfg | 24 ++++++ > 3 files changed, 238 insertions(+) > create mode 100644 arm/tlbflush-code.c > > diff --git a/arm/Makefile.common b/arm/Makefile.common > index cca0d9c..de99a6e 100644 > --- a/arm/Makefile.common > +++ b/arm/Makefile.common > @@ -13,6 +13,7 @@ tests-common = $(TEST_DIR)/selftest.flat > tests-common += $(TEST_DIR)/spinlock-test.flat > tests-common += $(TEST_DIR)/pci-test.flat > tests-common += $(TEST_DIR)/gic.flat > +tests-common += $(TEST_DIR)/tlbflush-code.flat > > all: test_cases > > @@ -81,3 +82,4 @@ generated_files = $(asm-offsets) > test_cases: $(generated_files) $(tests-common) $(tests) > > $(TEST_DIR)/selftest.o $(cstart.o): $(asm-offsets) > +$(TEST_DIR)/tlbflush-code.elf: $(cstart.o) $(TEST_DIR)/tlbflush-code.o This should no longer be necessary. > diff --git a/arm/tlbflush-code.c b/arm/tlbflush-code.c > new file mode 100644 > index 0000000..cb5cdc2 > --- /dev/null > +++ b/arm/tlbflush-code.c > @@ -0,0 +1,212 @@ > +/* > + * TLB Flush Race Tests > + * > + * These tests are designed to test for incorrect TLB flush semantics > + * under emulation. The initial CPU will set all the others working a > + * compuation task and will then trigger TLB flushes across the > + * system. It doesn't actually need to re-map anything but the flushes > + * themselves will trigger QEMU's TCG self-modifying code detection > + * which will invalidate any generated code causing re-translation. > + * Eventually the code buffer will fill and a general tb_lush() will > + * be triggered. > + * > + * Copyright (C) 2016, Linaro, Alex Bennée <alex.bennee@linaro.org> > + * > + * This work is licensed under the terms of the GNU LGPL, version 2. > + */ > + > +#include <libcflat.h> > +#include <asm/smp.h> > +#include <asm/cpumask.h> > +#include <asm/barrier.h> > +#include <asm/mmu.h> > + > +#define SEQ_LENGTH 10 > +#define SEQ_HASH 0x7cd707fe > + > +static cpumask_t smp_test_complete; > +static int flush_count = 1000000; > +static bool flush_self; > +static bool flush_page; > +static bool flush_verbose; > + > +/* > + * Work functions > + * > + * These work functions need to be: > + * > + * - page aligned, so we can flush one function at a time > + * - have branches, so QEMU TCG generates multiple basic blocks > + * - call across pages, so we exercise the TCG basic block slow path > + */ > + > +/* Adler32 */ > +__attribute__((aligned(PAGE_SIZE))) uint32_t hash_array(const void *buf, > + size_t buflen) I think I'd prefer __attribute__((aligned(PAGE_SIZE))) uint32_t hash_array(const void *buf, size_t buflen) to handle the long line > +{ > + const uint8_t *data = (uint8_t *) buf; > + uint32_t s1 = 1; > + uint32_t s2 = 0; > + > + for (size_t n = 0; n < buflen; n++) { > + s1 = (s1 + data[n]) % 65521; > + s2 = (s2 + s1) % 65521; > + } > + return (s2 << 16) | s1; > +} > + > +__attribute__((aligned(PAGE_SIZE))) void create_fib_sequence(int length, > + unsigned int *array) > +{ > + int i; > + > + /* first two values */ > + array[0] = 0; > + array[1] = 1; > + for (i=2; i<length; i++) { > + array[i] = array[i-2] + array[i-1]; > + } please don't use {} for one-liners. Try running the kernel's check_patch on your patches. Applies many places below > +} > + > +__attribute__((aligned(PAGE_SIZE))) unsigned long long factorial(unsigned int n) long line > +{ > + unsigned int i; > + unsigned long long fac = 1; > + for (i=1; i<=n; i++) > + { > + fac = fac * i; > + } > + return fac; > +} > + > +__attribute__((aligned(PAGE_SIZE))) void factorial_array > +(unsigned int n, unsigned int *input, unsigned long long *output) > +{ > + unsigned int i; > + for (i=0; i<n; i++) { > + output[i] = factorial(input[i]); > + } > +} > + > +__attribute__((aligned(PAGE_SIZE))) unsigned int do_computation(void) > +{ > + unsigned int fib_array[SEQ_LENGTH]; > + unsigned long long facfib_array[SEQ_LENGTH]; > + uint32_t fib_hash, facfib_hash; > + > + create_fib_sequence(SEQ_LENGTH, &fib_array[0]); > + fib_hash = hash_array(&fib_array[0], sizeof(fib_array)); > + factorial_array(SEQ_LENGTH, &fib_array[0], &facfib_array[0]); > + facfib_hash = hash_array(&facfib_array[0], sizeof(facfib_array)); > + > + return (fib_hash ^ facfib_hash); > +} > + > +/* This provides a table of the work functions so we can flush each > + * page individually > + */ > +static void * pages[] = {&hash_array, &create_fib_sequence, &factorial, > + &factorial_array, &do_computation}; please put the '*' by pages > + > +static void do_flush(int i) > +{ > + if (flush_page) { > + flush_tlb_page((unsigned long)pages[i % ARRAY_SIZE(pages)]); > + } else { > + flush_tlb_all(); > + } > +} > + > + > +static void just_compute(void) > +{ > + int i, errors = 0; > + int cpu = smp_processor_id(); > + > + uint32_t result; > + > + printf("CPU%d online\n", cpu); > + > + for (i=0; i < flush_count; i++) { > + result = do_computation(); > + > + if (result != SEQ_HASH) { > + errors++; > + printf("CPU%d: seq%d 0x%"PRIx32"!=0x%x\n", > + cpu, i, result, SEQ_HASH); > + } > + > + if (flush_verbose && (i % 1000) == 0) { > + printf("CPU%d: seq%d\n", cpu, i); > + } > + > + if (flush_self) { > + do_flush(i); > + } > + } > + > + report("CPU%d: Done - Errors: %d\n", errors == 0, cpu, errors); > + > + cpumask_set_cpu(cpu, &smp_test_complete); > + if (cpu != 0) > + halt(); > +} > + > +static void just_flush(void) > +{ > + int cpu = smp_processor_id(); > + int i = 0; > + > + /* set our CPU as done, keep flushing until everyone else > + finished */ Not our comment style > + cpumask_set_cpu(cpu, &smp_test_complete); > + > + while (!cpumask_full(&smp_test_complete)) { > + do_flush(i++); > + } > + > + report("CPU%d: Done - Triggered %d flushes\n", true, cpu, i); > +} > + > +int main(int argc, char **argv) > +{ > + int cpu, i; > + char prefix[100]; > + > + for (i=0; i<argc; i++) { > + char *arg = argv[i]; > + > + if (strcmp(arg, "page") == 0) { > + flush_page = true; > + } > + > + if (strcmp(arg, "self") == 0) { > + flush_self = true; > + } > + > + if (strcmp(arg, "verbose") == 0) { > + flush_verbose = true; > + } > + } > + > + snprintf(prefix, sizeof(prefix), "tlbflush_%s_%s", > + flush_page?"page":"all", > + flush_self?"self":"other"); > + report_prefix_push(prefix); > + > + for_each_present_cpu(cpu) { > + if (cpu == 0) > + continue; > + smp_boot_secondary(cpu, just_compute); > + } > + > + if (flush_self) > + just_compute(); > + else > + just_flush(); > + > + while (!cpumask_full(&smp_test_complete)) > + cpu_relax(); > + > + return report_summary(); > +} > diff --git a/arm/unittests.cfg b/arm/unittests.cfg > index c7392c7..beaae84 100644 > --- a/arm/unittests.cfg > +++ b/arm/unittests.cfg > @@ -72,3 +72,27 @@ file = gic.flat > smp = $MAX_SMP > extra_params = -machine gic-version=3 -append 'ipi' > groups = gic > + > +# TLB Torture Tests > +[tlbflush-code::all_other] We don't use the '::' style anymore, as it doesn't work well with mkstandalone. > +file = tlbflush-code.flat > +smp = $(($MAX_SMP>4?4:$MAX_SMP)) > +groups = tlbflush > + > +[tlbflush-code::page_other] > +file = tlbflush-code.flat > +smp = $(($MAX_SMP>4?4:$MAX_SMP)) > +extra_params = -append 'page' > +groups = tlbflush > + > +[tlbflush-code::all_self] > +file = tlbflush-code.flat > +smp = $(($MAX_SMP>4?4:$MAX_SMP)) > +extra_params = -append 'self' > +groups = tlbflush > + > +[tlbflush-code::page_self] > +file = tlbflush-code.flat > +smp = $(($MAX_SMP>4?4:$MAX_SMP)) > +extra_params = -append 'page self' > +groups = tlbflush > -- > 2.10.1 > I only did a superficial review, but it looks familiar. I guess I've reviewed some of it before. drew
diff --git a/arm/Makefile.common b/arm/Makefile.common index cca0d9c..de99a6e 100644 --- a/arm/Makefile.common +++ b/arm/Makefile.common @@ -13,6 +13,7 @@ tests-common = $(TEST_DIR)/selftest.flat tests-common += $(TEST_DIR)/spinlock-test.flat tests-common += $(TEST_DIR)/pci-test.flat tests-common += $(TEST_DIR)/gic.flat +tests-common += $(TEST_DIR)/tlbflush-code.flat all: test_cases @@ -81,3 +82,4 @@ generated_files = $(asm-offsets) test_cases: $(generated_files) $(tests-common) $(tests) $(TEST_DIR)/selftest.o $(cstart.o): $(asm-offsets) +$(TEST_DIR)/tlbflush-code.elf: $(cstart.o) $(TEST_DIR)/tlbflush-code.o diff --git a/arm/tlbflush-code.c b/arm/tlbflush-code.c new file mode 100644 index 0000000..cb5cdc2 --- /dev/null +++ b/arm/tlbflush-code.c @@ -0,0 +1,212 @@ +/* + * TLB Flush Race Tests + * + * These tests are designed to test for incorrect TLB flush semantics + * under emulation. The initial CPU will set all the others working a + * compuation task and will then trigger TLB flushes across the + * system. It doesn't actually need to re-map anything but the flushes + * themselves will trigger QEMU's TCG self-modifying code detection + * which will invalidate any generated code causing re-translation. + * Eventually the code buffer will fill and a general tb_lush() will + * be triggered. + * + * Copyright (C) 2016, Linaro, Alex Bennée <alex.bennee@linaro.org> + * + * This work is licensed under the terms of the GNU LGPL, version 2. + */ + +#include <libcflat.h> +#include <asm/smp.h> +#include <asm/cpumask.h> +#include <asm/barrier.h> +#include <asm/mmu.h> + +#define SEQ_LENGTH 10 +#define SEQ_HASH 0x7cd707fe + +static cpumask_t smp_test_complete; +static int flush_count = 1000000; +static bool flush_self; +static bool flush_page; +static bool flush_verbose; + +/* + * Work functions + * + * These work functions need to be: + * + * - page aligned, so we can flush one function at a time + * - have branches, so QEMU TCG generates multiple basic blocks + * - call across pages, so we exercise the TCG basic block slow path + */ + +/* Adler32 */ +__attribute__((aligned(PAGE_SIZE))) uint32_t hash_array(const void *buf, + size_t buflen) +{ + const uint8_t *data = (uint8_t *) buf; + uint32_t s1 = 1; + uint32_t s2 = 0; + + for (size_t n = 0; n < buflen; n++) { + s1 = (s1 + data[n]) % 65521; + s2 = (s2 + s1) % 65521; + } + return (s2 << 16) | s1; +} + +__attribute__((aligned(PAGE_SIZE))) void create_fib_sequence(int length, + unsigned int *array) +{ + int i; + + /* first two values */ + array[0] = 0; + array[1] = 1; + for (i=2; i<length; i++) { + array[i] = array[i-2] + array[i-1]; + } +} + +__attribute__((aligned(PAGE_SIZE))) unsigned long long factorial(unsigned int n) +{ + unsigned int i; + unsigned long long fac = 1; + for (i=1; i<=n; i++) + { + fac = fac * i; + } + return fac; +} + +__attribute__((aligned(PAGE_SIZE))) void factorial_array +(unsigned int n, unsigned int *input, unsigned long long *output) +{ + unsigned int i; + for (i=0; i<n; i++) { + output[i] = factorial(input[i]); + } +} + +__attribute__((aligned(PAGE_SIZE))) unsigned int do_computation(void) +{ + unsigned int fib_array[SEQ_LENGTH]; + unsigned long long facfib_array[SEQ_LENGTH]; + uint32_t fib_hash, facfib_hash; + + create_fib_sequence(SEQ_LENGTH, &fib_array[0]); + fib_hash = hash_array(&fib_array[0], sizeof(fib_array)); + factorial_array(SEQ_LENGTH, &fib_array[0], &facfib_array[0]); + facfib_hash = hash_array(&facfib_array[0], sizeof(facfib_array)); + + return (fib_hash ^ facfib_hash); +} + +/* This provides a table of the work functions so we can flush each + * page individually + */ +static void * pages[] = {&hash_array, &create_fib_sequence, &factorial, + &factorial_array, &do_computation}; + +static void do_flush(int i) +{ + if (flush_page) { + flush_tlb_page((unsigned long)pages[i % ARRAY_SIZE(pages)]); + } else { + flush_tlb_all(); + } +} + + +static void just_compute(void) +{ + int i, errors = 0; + int cpu = smp_processor_id(); + + uint32_t result; + + printf("CPU%d online\n", cpu); + + for (i=0; i < flush_count; i++) { + result = do_computation(); + + if (result != SEQ_HASH) { + errors++; + printf("CPU%d: seq%d 0x%"PRIx32"!=0x%x\n", + cpu, i, result, SEQ_HASH); + } + + if (flush_verbose && (i % 1000) == 0) { + printf("CPU%d: seq%d\n", cpu, i); + } + + if (flush_self) { + do_flush(i); + } + } + + report("CPU%d: Done - Errors: %d\n", errors == 0, cpu, errors); + + cpumask_set_cpu(cpu, &smp_test_complete); + if (cpu != 0) + halt(); +} + +static void just_flush(void) +{ + int cpu = smp_processor_id(); + int i = 0; + + /* set our CPU as done, keep flushing until everyone else + finished */ + cpumask_set_cpu(cpu, &smp_test_complete); + + while (!cpumask_full(&smp_test_complete)) { + do_flush(i++); + } + + report("CPU%d: Done - Triggered %d flushes\n", true, cpu, i); +} + +int main(int argc, char **argv) +{ + int cpu, i; + char prefix[100]; + + for (i=0; i<argc; i++) { + char *arg = argv[i]; + + if (strcmp(arg, "page") == 0) { + flush_page = true; + } + + if (strcmp(arg, "self") == 0) { + flush_self = true; + } + + if (strcmp(arg, "verbose") == 0) { + flush_verbose = true; + } + } + + snprintf(prefix, sizeof(prefix), "tlbflush_%s_%s", + flush_page?"page":"all", + flush_self?"self":"other"); + report_prefix_push(prefix); + + for_each_present_cpu(cpu) { + if (cpu == 0) + continue; + smp_boot_secondary(cpu, just_compute); + } + + if (flush_self) + just_compute(); + else + just_flush(); + + while (!cpumask_full(&smp_test_complete)) + cpu_relax(); + + return report_summary(); +} diff --git a/arm/unittests.cfg b/arm/unittests.cfg index c7392c7..beaae84 100644 --- a/arm/unittests.cfg +++ b/arm/unittests.cfg @@ -72,3 +72,27 @@ file = gic.flat smp = $MAX_SMP extra_params = -machine gic-version=3 -append 'ipi' groups = gic + +# TLB Torture Tests +[tlbflush-code::all_other] +file = tlbflush-code.flat +smp = $(($MAX_SMP>4?4:$MAX_SMP)) +groups = tlbflush + +[tlbflush-code::page_other] +file = tlbflush-code.flat +smp = $(($MAX_SMP>4?4:$MAX_SMP)) +extra_params = -append 'page' +groups = tlbflush + +[tlbflush-code::all_self] +file = tlbflush-code.flat +smp = $(($MAX_SMP>4?4:$MAX_SMP)) +extra_params = -append 'self' +groups = tlbflush + +[tlbflush-code::page_self] +file = tlbflush-code.flat +smp = $(($MAX_SMP>4?4:$MAX_SMP)) +extra_params = -append 'page self' +groups = tlbflush
This adds a fairly brain dead torture test for TLB flushes intended for stressing the MTTCG QEMU build. It takes the usual -smp option for multiple CPUs. By default it CPU0 will do a TLBIALL flush after each cycle. You can pass options via -append to control additional aspects of the test: - "page" flush each page in turn (one per function) - "self" do the flush after each computation cycle - "verbose" report progress on each computation cycle Signed-off-by: Alex Bennée <alex.bennee@linaro.org> CC: Mark Rutland <mark.rutland@arm.com> --- v2 - rename to tlbflush-test - made makefile changes cleaner - added self/other flush mode - create specific prefix - whitespace fixes v3 - using new SMP framework for test runing v4 - merge in the unitests.cfg v5 - max out at -smp 4 - printf fmtfix v7 - rename to tlbflush-code - int -> bool flags --- arm/Makefile.common | 2 + arm/tlbflush-code.c | 212 ++++++++++++++++++++++++++++++++++++++++++++++++++++ arm/unittests.cfg | 24 ++++++ 3 files changed, 238 insertions(+) create mode 100644 arm/tlbflush-code.c