Andrew Jones <drjo...@redhat.com> writes:
> On Fri, Jul 24, 2015 at 02:25:06PM +0100, 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 will do a TLBIALL flush after each cycle. If you pass
>> -append "page" to the kernel it will take it in turns to flush each of
>> the computation functions. At the moment it doesn't do any re-mapping of
>> pages but maybe that is something that could be done in the future.
>>
>> [DEV VERSION FOR COMMENT]
>> Signed-off-by: Alex Bennée <alex.ben...@linaro.org>
>> ---
>> arm/tlbflush.c | 163
>> +++++++++++++++++++++++++++++++++++++++++++
>> config/config-arm-common.mak | 4 +-
>> lib/arm/asm/mmu.h | 11 +++
>> 3 files changed, 177 insertions(+), 1 deletion(-)
>> create mode 100644 arm/tlbflush.c
>>
>> diff --git a/arm/tlbflush.c b/arm/tlbflush.c
>> new file mode 100644
>> index 0000000..6eeff18
>> --- /dev/null
>> +++ b/arm/tlbflush.c
>> @@ -0,0 +1,163 @@
>> +#include <libcflat.h>
>> +#include <asm/smp.h>
>> +#include <asm/cpumask.h>
>> +#include <asm/barrier.h>
>> +#include <asm/mmu.h>
>> +
>> +#define SEQ_LENGTH 10
>> +
>> +static cpumask_t smp_test_complete;
>> +static int flush_count = 100000;
>> +static int flush_self = 1;
>> +static int flush_page = 0;
>> +
>> +__attribute__((aligned(0x1000))) unsigned int hash_array(int length,
>> unsigned int *array)
>
> You should use PAGE_SIZE instead of 0x1000 in these attributes, allowing
> the test to also work for aarch64, as we're using 64k pages on
> aarch64.
Good point.
>
>> +{
>> + int i;
>> + unsigned int sum=0;
>> + for (i=0; i<length; i++)
>> + {
>> + unsigned int val = *array++;
>> + sum ^= val;
>> + sum ^= (val >> (val % 16));
>> + sum ^= (val << (val % 32));
>> + }
>> +
>> + return sum;
>> +}
>> +
>> +__attribute__((aligned(0x1000))) 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(0x1000))) 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;
>> +}
>> +
>> +/* do some computationally expensive stuff, return a checksum of the
>> + * results */
>> +__attribute__((aligned(0x1000))) unsigned int do_computation(void)
>> +{
>> + unsigned int fib_array[SEQ_LENGTH];
>> + unsigned long long facfib_array[SEQ_LENGTH];
>> + unsigned int fib_hash, facfib_hash;
>> + int cpu = smp_processor_id();
>> + int i, j;
>> +
>> + create_fib_sequence(SEQ_LENGTH, &fib_array[0]);
>> + fib_hash = hash_array(SEQ_LENGTH, &fib_array[0]);
>> + for (i=0; i<SEQ_LENGTH; i++) {
>> + for (j=0; j<fib_array[i]; j++) {
>> + facfib_array[i] = factorial(fib_array[i]+j);
>> + }
>> + }
>> + facfib_hash = 0;
>> + for (i=0; i<SEQ_LENGTH; i++) {
>> + for (j=0; j<fib_array[i]; j++) {
>> + facfib_hash ^=
>> hash_array(sizeof(facfib_array)/sizeof(unsigned int), (unsigned int
>> *)&facfib_array[0]);
>> + }
>> + }
>> +
>> +#if 0
>> + printf("CPU:%d FIBSEQ ", cpu);
>> + for (i=0; i<SEQ_LENGTH; i++)
>> + printf("%u,", fib_array[i]);
>> + printf("\n");
>> +
>> + printf("CPU:%d FACFIB ", cpu);
>> + for (i=0; i<SEQ_LENGTH; i++)
>> + printf("%llu,", facfib_array[i]);
>> + printf("\n");
>> +#endif
>> +
>> + return (fib_hash ^ facfib_hash);
>> +}
>> +
>> +static void * pages[] = {&hash_array, &create_fib_sequence, &factorial,
>> &do_computation};
>
> I can't comment on whether or not the complexity of do_computation is
> necessary for your test, but it seems like overkill. Comments explaining
> why it's necessary would be good.
OK. From QEMUs TCG point of view I just want to ensure I have more than two
basic blocks per-page region so I can check the block-chaining in-page
and jump caching intra-page which are both affected on flushes. A
computationally complex routine with a known answer would be nicer
though I guess.
>
>> +
>> +static void test_flush(void)
>> +{
>> + int i, errors = 0;
>> + int cpu = smp_processor_id();
>> +
>> + unsigned int ref;
>> +
>> + printf("CPU%d online\n", cpu);
>> +
>> + ref = do_computation();
>
> What makes you sure that the first time you do the computation
> per cpu is correct? I think computing it externally, and saving
> the result, i.e.
>
> #define EXPECTED_RESULT 0x12345678
>
> would be more reliable.
OK.
>
>> +
>> + for (i=0; i < flush_count; i++) {
>> + unsigned int this_ref = do_computation();
>> +
>> + if (this_ref != ref) {
>> + errors++;
>> + printf("CPU%d: seq%d 0x%x!=0x%x\n",
>> + cpu, i, ref, this_ref);
>> + }
>> +
>> + if ((i % 1000) == 0) {
>> + printf("CPU%d: seq%d\n", cpu, i);
>> + }
>> +
>> + if (flush_self) {
>> + if (flush_page) {
>> + int j = (i % (sizeof(pages)/sizeof(void *)));
> libcflat.h has the ARRAY_SIZE macro
OK
>> + flush_tlb_page((unsigned long)pages[j]);
>> + } else {
>> + flush_tlb_all();
>> + }
>> + }
>> + }
>> +
>> + report("CPU%d: Done - Errors: %d\n", errors == 0, cpu, errors);
>> +
>> + cpumask_set_cpu(cpu, &smp_test_complete);
>> + if (cpu != 0)
>> + halt();
>> +}
>> +
>> +int main(int argc, char **argv)
>> +{
>> + int cpu, i;
>> +
>> + report_prefix_push("tlbflush");
>> +
>> + for (i=0; i<argc; i++) {
>> + char *arg = argv[i];
>> +/* printf("arg:%d:%s\n", i, arg); */
>> +
>> + if (strcmp(arg, "page") == 0) {
>> + report_prefix_push("page");
>> + flush_page = 1;
>> + }
>> + }
>> +
>> + for_each_present_cpu(cpu) {
>> + if (cpu == 0)
>> + continue;
>> + smp_boot_secondary(cpu, test_flush);
>> + }
>> +
>> + test_flush();
>> +
>> + while (!cpumask_full(&smp_test_complete))
>> + cpu_relax();
>> +
>> + return report_summary();
>
> As we use the kernel coding style you should run
>
> $KERNEL_SRC/scripts/checkpatch.pl -f arm/tlbflush.c
>
> Also, please rename to tlbflush-test.c to differentiate it
> from an implementation of tlbflush support, and to make
> the standalone test name (if we commit those patches) more
> descriptive.
I'll have another poke at my editor config. It should have been setting
the coding style automatically, although of course explicit local
variables are better ;-)
>
>> +}
>> diff --git a/config/config-arm-common.mak b/config/config-arm-common.mak
>> index 0674daa..5b14db4 100644
>> --- a/config/config-arm-common.mak
>> +++ b/config/config-arm-common.mak
>> @@ -11,7 +11,8 @@ endif
>>
>> tests-common = \
>> $(TEST_DIR)/selftest.flat \
>> - $(TEST_DIR)/spinlock-test.flat
>> + $(TEST_DIR)/spinlock-test.flat \
>> + $(TEST_DIR)/tlbflush.flat
>
> As we're adding tests faster now it's becoming clear that the '\' list
> isn't so great. To add a new test at the bottom we always have to modify
> the last line too. We should either add the new one at the top (right
> below the 'test-common =' line), or change this to a '+=' sequence like
> some other lists are done.
>
>>
>> all: test_cases
>>
>> @@ -72,3 +73,4 @@ test_cases: $(generated_files) $(tests-common) $(tests)
>>
>> $(TEST_DIR)/selftest.elf: $(cstart.o) $(TEST_DIR)/selftest.o
>> $(TEST_DIR)/spinlock-test.elf: $(cstart.o) $(TEST_DIR)/spinlock-test.o
>> +$(TEST_DIR)/tlbflush.elf: $(cstart.o) $(TEST_DIR)/tlbflush.o
>> diff --git a/lib/arm/asm/mmu.h b/lib/arm/asm/mmu.h
>> index c1bd01c..2bb0cde 100644
>> --- a/lib/arm/asm/mmu.h
>> +++ b/lib/arm/asm/mmu.h
>> @@ -14,8 +14,11 @@
>> #define PTE_AF PTE_EXT_AF
>> #define PTE_WBWA L_PTE_MT_WRITEALLOC
>>
>> +/* See B3.18.7 TLB maintenance operations */
>> +
>> static inline void local_flush_tlb_all(void)
>> {
>> + /* TLBIALL */
>> asm volatile("mcr p15, 0, %0, c8, c7, 0" :: "r" (0));
>> dsb();
>> isb();
>> @@ -27,6 +30,14 @@ static inline void flush_tlb_all(void)
>> local_flush_tlb_all();
>> }
>>
>> +static inline void flush_tlb_page(unsigned long vaddr)
>> +{
>> + /* TLBIMVAA */
>> + asm volatile("mcr p15, 0, %0, c8, c7, 3" :: "r" (vaddr));
>> + dsb();
>> + isb();
>> +}
>> +
>> #include <asm/mmu-api.h>
>>
>> #endif /* __ASMARM_MMU_H_ */
>
> This mmu.h change looks good, but please add the arm64
> flush_tlb_page at the same time. And anyway, I guess you'll
> want your test to work for both arm and aarch64?
Yes I will. Currently the MTTCG is arm32 only but this will be expanded.
>
> Thanks,
> drew
--
Alex Bennée
--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
