Hi Petri, Any more comments? I suggest we should push this patchset firstly. If we have any new requirements, we can create a new ticket to trace it. What's your thought?
2017-02-20 10:37 GMT+08:00 Kevin Wang <[email protected]>: > We don't know how many threads/locks the application will use. The final > timer resolution is determined by applications. So what we are trying to do > here is to find out a MAX resolution in a single thread on this system. So > no matter in the multi-thread environment or other more workloads system, > the resolution should be obvious less than this MAX value. > > I think overrun means the application is not able to keep up with dealing > each timer interrupt. It will miss some interrupts. > > 2017-02-17 17:46 GMT+08:00 Maxim Uvarov <[email protected]>: > >> >> >> On 17 February 2017 at 11:51, Francois Ozog <[email protected]> >> wrote: >> >>> I think this is quite useful for cloud applications that do not know >>> the underlying hardware specs and need to check what is effectively >>> possible. "hardware" resolution is nothing without all surrounding >>> software layers, and virtualization in particular... >>> >>> On 17 February 2017 at 09:43, Kevin Wang <[email protected]> wrote: >>> > clock_getres() is to find the hardware resolution of specific clock >>> source. >>> > It shows 1ns in most of modern system. >>> > What this API trys to do is to find out the resolution without overrun >>> of >>> > this timer. Applications are able to deal with each timer event with >>> that >>> > timer resolution. >>> > >>> >> >> >> Timer run in some thread. With different load you will have different >> latency on timer action. >> Plus you can be in some atomic operation preempted by the kernel. Number >> of 'in fight' timers >> also make sense. I'm not sure that we can do right calculation of that >> value. >> >> btw, what is bad in overrun? And what is overrun in current context? Is >> it hardware time counter overrun >> or variable presented as time in odp overrun? >> >> Maxim. >> >> >>> > 2017-02-17 3:00 GMT+08:00 Maxim Uvarov <[email protected]>: >>> > >>> >> why not clock_getres() ? >>> >> >>> >> Maxim. >>> >> >>> >> On 02/08/17 05:37, Kevin Wang wrote: >>> >> > Implement a new internal function timer_res_init() to detect the max >>> >> > timer resolution without overrun at the ODP init stage. It will >>> check >>> >> > timer resolution from 1ms to 100us, 10us...1ns until the timer is >>> >> > overrun. >>> >> > >>> >> > Signed-off-by: Kevin Wang <[email protected]> >>> >> > --- >>> >> > platform/linux-generic/odp_timer.c | 95 >>> ++++++++++++++++++++++++++++++ >>> >> ++++++++ >>> >> > 1 file changed, 95 insertions(+) >>> >> > >>> >> > diff --git a/platform/linux-generic/odp_timer.c >>> >> b/platform/linux-generic/odp_timer.c >>> >> > index 53fec08..9b884ea 100644 >>> >> > --- a/platform/linux-generic/odp_timer.c >>> >> > +++ b/platform/linux-generic/odp_timer.c >>> >> > @@ -70,6 +70,9 @@ static _odp_atomic_flag_t locks[NUM_LOCKS]; /* >>> >> Multiple locks per cache line! */ >>> >> > #define IDX2LOCK(idx) (&locks[(idx) % NUM_LOCKS]) >>> >> > #endif >>> >> > >>> >> > +/* Timer resolution in nanoseconds */ >>> >> > +static uint64_t timer_res; >>> >> > + >>> >> > /*********************************************************** >>> >> ******************* >>> >> > * Translation between timeout buffer and timeout header >>> >> > ************************************************************ >>> >> *****************/ >>> >> > @@ -188,6 +191,7 @@ typedef struct odp_timer_pool_s { >>> >> > >>> >> > #define MAX_TIMER_POOLS 255 /* Leave one for ODP_TIMER_INVALID */ >>> >> > #define INDEX_BITS 24 >>> >> > +#define TIMER_RES_TEST_LOOP_COUNT 10 >>> >> > static odp_atomic_u32_t num_timer_pools; >>> >> > static odp_timer_pool *timer_pool[MAX_TIMER_POOLS]; >>> >> > >>> >> > @@ -738,6 +742,81 @@ static void *timer_thread(void *arg) >>> >> > return NULL; >>> >> > } >>> >> > >>> >> > +/* Get the max timer resolution without overrun and fill in >>> timer_res >>> >> variable. >>> >> > + * >>> >> > + * Set timer's interval with candidate resolutions to get the max >>> >> resolution >>> >> > + * that the timer would not be overrun. >>> >> > + * The candidate resolution value is from 1ms to 100us, 10us...1ns >>> etc. >>> >> > + */ >>> >> > +static int timer_res_init(void) >>> >> > +{ >>> >> > + struct sigevent sigev; >>> >> > + timer_t timerid; >>> >> > + uint64_t res, sec, nsec; >>> >> > + struct itimerspec ispec; >>> >> > + sigset_t sigset; >>> >> > + siginfo_t si; >>> >> > + int loop_cnt; >>> >> > + struct timespec tmo; >>> >> > + >>> >> > + sigev.sigev_notify = SIGEV_THREAD_ID; >>> >> > + sigev._sigev_un._tid = (pid_t)syscall(SYS_gettid); >>> >> > + sigev.sigev_signo = SIGUSR1; >>> >> > + >>> >> > + /* Create timer */ >>> >> > + if (timer_create(CLOCK_MONOTONIC, &sigev, &timerid)) >>> >> > + ODP_ABORT("timer_create() returned error %s\n", >>> >> > + strerror(errno)); >>> >> > + >>> >> > + /* Timer resolution start from 1ms */ >>> >> > + res = ODP_TIME_MSEC_IN_NS; >>> >> > + /* Set initial value of timer_res */ >>> >> > + timer_res = res; >>> >> > + sigemptyset(&sigset); >>> >> > + /* Add SIGUSR1 to sigset */ >>> >> > + sigaddset(&sigset, SIGUSR1); >>> >> > + sigprocmask(SIG_BLOCK, &sigset, NULL); >>> >> > + >>> >> > + while (res > 0) { >>> >> > + /* Loop for 10 times to test the result */ >>> >> > + loop_cnt = TIMER_RES_TEST_LOOP_COUNT; >>> >> > + sec = res / ODP_TIME_SEC_IN_NS; >>> >> > + nsec = res - sec * ODP_TIME_SEC_IN_NS; >>> >> > + >>> >> > + memset(&ispec, 0, sizeof(ispec)); >>> >> > + ispec.it_interval.tv_sec = (time_t)sec; >>> >> > + ispec.it_interval.tv_nsec = (long)nsec; >>> >> > + ispec.it_value.tv_sec = (time_t)sec; >>> >> > + ispec.it_value.tv_nsec = (long)nsec; >>> >> > + >>> >> > + if (timer_settime(timerid, 0, &ispec, NULL)) >>> >> > + ODP_ABORT("timer_settime() returned error >>> %s\n", >>> >> > + strerror(errno)); >>> >> > + /* Set signal wait timeout to 10*res */ >>> >> > + tmo.tv_sec = 0; >>> >> > + tmo.tv_nsec = res * 10; >>> >> > + while (loop_cnt--) { >>> >> > + if (sigtimedwait(&sigset, &si, &tmo) > 0) { >>> >> > + if (timer_getoverrun(timerid)) >>> >> > + /* overrun at this resolution >>> */ >>> >> > + /* goto the end */ >>> >> > + goto timer_res_init_done; >>> >> > + } >>> >> > + } >>> >> > + /* Set timer_res */ >>> >> > + timer_res = res; >>> >> > + /* Test the next timer resolution candidate */ >>> >> > + res /= 10; >>> >> > + } >>> >> > +timer_res_init_done: >>> >> > + if (timer_delete(timerid) != 0) >>> >> > + ODP_ABORT("timer_delete() returned error %s\n", >>> >> > + strerror(errno)); >>> >> > + sigemptyset(&sigset); >>> >> > + sigprocmask(SIG_BLOCK, &sigset, NULL); >>> >> > + return 0; >>> >> > +} >>> >> > + >>> >> > static void itimer_init(odp_timer_pool *tp) >>> >> > { >>> >> > struct sigevent sigev; >>> >> > @@ -795,6 +874,20 @@ static void itimer_fini(odp_timer_pool *tp) >>> >> > * Some parameter checks and error messages >>> >> > * No modificatios of internal state >>> >> > ************************************************************ >>> >> *****************/ >>> >> > +int odp_timer_capability(odp_timer_clk_src_t clk_src, >>> >> > + odp_timer_capability_t *capa) >>> >> > +{ >>> >> > + int ret = 0; >>> >> > + >>> >> > + if (clk_src == ODP_CLOCK_CPU) { >>> >> > + capa->res_ns = timer_res; >>> >> > + } else { >>> >> > + ODP_ERR("ODP timer system doesn't support external >>> clock >>> >> source currently\n"); >>> >> > + ret = -1; >>> >> > + } >>> >> > + return ret; >>> >> > +} >>> >> > + >>> >> > odp_timer_pool_t >>> >> > odp_timer_pool_create(const char *name, >>> >> > const odp_timer_pool_param_t *param) >>> >> > @@ -1003,6 +1096,8 @@ int odp_timer_init_global(void) >>> >> > #endif >>> >> > odp_atomic_init_u32(&num_timer_pools, 0); >>> >> > >>> >> > + timer_res_init(); >>> >> > + >>> >> > block_sigalarm(); >>> >> > >>> >> > return 0; >>> >> > >>> >> >>> >> >>> > >>> > >>> > -- >>> > Thanks, >>> > Kevin >>> >>> >>> >>> -- >>> François-Frédéric Ozog | Director Linaro Networking Group >>> T: +33.67221.6485 >>> [email protected] | Skype: ffozog >>> >> >> > > > -- > Thanks, > Kevin > -- Thanks, Kevin
