On 03/05/2019 00.55, Suraj Jitindar Singh wrote:
> On Thu, 2019-05-02 at 10:01 +0200, Laurent Vivier wrote:
>> On 01/05/2019 09:00, Suraj Jitindar Singh wrote:
>>> Currently the handler for a decrementer exception will simply
>>> reload the
>>> maximum value (0x7FFFFFFF), which will take ~4 seconds to expire
>>> again.
>>> This means that if a vcpu cedes, it will be ~4 seconds between
>>> wakeups.
>>>
>>> The h_cede_tm test is testing a known breakage when a guest cedes
>>> while
>>> suspended. To be sure we cede 500 times to check for the bug.
>>> However
>>> since it takes ~4 seconds to be woken up once we've ceded, we only
>>> get
>>> through ~20 iterations before we reach the 90 seconds timeout and
>>> the
>>> test appears to fail.
>>>
>>> Add an option when registering the decrementer handler to specify
>>> the
>>> value which should be reloaded by the handler, allowing the timeout
>>> to be
>>> chosen.
>>>
>>> Modify the spr test to use the max timeout to preserve existing
>>> behaviour.
>>> Modify the h_cede_tm test to use a 10ms timeout to ensure we can
>>> perform
>>> 500 iterations before hitting the 90 second time limit for a test.
>>>
>>> This means the h_cede_tm test now succeeds rather than timing out.
>>>
>>> Signed-off-by: Suraj Jitindar Singh <sjitindarsingh@xxxxxxxxx>
>>> ---
>>> lib/powerpc/handlers.c | 7 ++++---
>>> powerpc/sprs.c | 3 ++-
>>> powerpc/tm.c | 3 ++-
>>> 3 files changed, 8 insertions(+), 5 deletions(-)
>>>
>>> diff --git a/lib/powerpc/handlers.c b/lib/powerpc/handlers.c
>>> index be8226a..c8721e0 100644
>>> --- a/lib/powerpc/handlers.c
>>> +++ b/lib/powerpc/handlers.c
>>> @@ -12,11 +12,12 @@
>>>
>>> /*
>>> * Generic handler for decrementer exceptions (0x900)
>>> - * Just reset the decrementer back to its maximum value
>>> (0x7FFFFFFF)
>>> + * Just reset the decrementer back to the value specified when
>>> registering the
>>> + * handler
>>> */
>>> -void dec_except_handler(struct pt_regs *regs __unused, void *data
>>> __unused)
>>> +void dec_except_handler(struct pt_regs *regs __unused, void *data)
>>> {
>>> - uint32_t dec = 0x7FFFFFFF;
>>> + uint64_t dec = *((uint64_t *) data);
>>>
>>> asm volatile ("mtdec %0" : : "r" (dec));
>>> }
>>> diff --git a/powerpc/sprs.c b/powerpc/sprs.c
>>> index 6744bd8..3bd6ac7 100644
>>> --- a/powerpc/sprs.c
>>> +++ b/powerpc/sprs.c
>>> @@ -253,6 +253,7 @@ int main(int argc, char **argv)
>>> 0x1234567890ABCDEFULL, 0xFEDCBA0987654321ULL,
>>> -1ULL,
>>> };
>>> + uint64_t decr = 0x7FFFFFFF;
Declare the variable as "static", please, otherwise the pointer to it
might get invalid at the end of the function.
>>> for (i = 1; i < argc; i++) {
>>> if (!strcmp(argv[i], "-w")) {
>>> @@ -288,7 +289,7 @@ int main(int argc, char **argv)
>>> (void) getchar();
>>> } else {
>>> puts("Sleeping...\n");
>>> - handle_exception(0x900, &dec_except_handler,
>>> NULL);
>>> + handle_exception(0x900, &dec_except_handler,
>>> &decr);
>>> asm volatile ("mtdec %0" : : "r" (0x3FFFFFFF));
>>> hcall(H_CEDE);
>>> }
>>> diff --git a/powerpc/tm.c b/powerpc/tm.c
>>> index bd56baa..0f3f543 100644
>>> --- a/powerpc/tm.c
>>> +++ b/powerpc/tm.c
>>> @@ -95,11 +95,12 @@ static bool enable_tm(void)
>>> static void test_h_cede_tm(int argc, char **argv)
>>> {
>>> int i;
>>> + uint64_t decr = 0x3FFFFF;
"static" please.
>>> if (argc > 2)
>>> report_abort("Unsupported argument: '%s'",
>>> argv[2]);
>>>
>>> - handle_exception(0x900, &dec_except_handler, NULL);
>>> + handle_exception(0x900, &dec_except_handler, &decr);
>>
>> Maybe you should also need here:
>>
>> asm volatile ("mtdec %0" : : "r" (decr));
>>
>> To set the first one to the same values as the following ones?
>
> I guess we could get a case where the decrementer is really large (if
> large decrementer is enabled for example) and otherwise we just don't
> wake up...
>
> In practise I'm not sure. But I'll add it if you like
I think I agree with Laurent - it's likely better to add the mtdec here.
Also, while you're at it, could/should we maybe remove the "nodefault"
in unittests.cfg again (or rather the whole "groups" line)? I'm really
afraid that this test does not get enough test coverage due to this...
Thomas
