[Intel-gfx] [PATCH] i915/query: Correlate engine and cpu timestamps with better accuracy

[Chris Wilson]

Quoting Lionel Landwerlin (2021-03-04 09:45:47)
> On 04/03/2021 10:58, Chris Wilson wrote:
> > Quoting Lionel Landwerlin (2021-03-04 08:28:59)
> >> On 04/03/2021 02:09, Chris Wilson wrote:
> >>> Quoting Umesh Nerlige Ramappa (2021-03-03 21:28:00)
> >>>> Perf measurements rely on CPU and engine timestamps to correlate
> >>>> events of interest across these time domains. Current mechanisms get
> >>>> these timestamps separately and the calculated delta between these
> >>>> timestamps lack enough accuracy.
> >>>>
> >>>> To improve the accuracy of these time measurements to within a few us,
> >>>> add a query that returns the engine and cpu timestamps captured as
> >>>> close to each other as possible.
> >>>>
> >>>> v2: (Tvrtko)
> >>>> - document clock reference used
> >>>> - return cpu timestamp always
> >>>> - capture cpu time just before lower dword of cs timestamp
> >>>>
> >>>> v3: (Chris)
> >>>> - use uncore-rpm
> >>>> - use __query_cs_timestamp helper
> >>>>
> >>>> v4: (Lionel)
> >>>> - Kernel perf subsytem allows users to specify the clock id to be used
> >>>>     in perf_event_open. This clock id is used by the perf subsystem to
> >>>>     return the appropriate cpu timestamp in perf events. Similarly, let
> >>>>     the user pass the clockid to this query so that cpu timestamp
> >>>>     corresponds to the clock id requested.
> >>>>
> >>>> v5: (Tvrtko)
> >>>> - Use normal ktime accessors instead of fast versions
> >>>> - Add more uApi documentation
> >>>>
> >>>> v6: (Lionel)
> >>>> - Move switch out of spinlock
> >>>>
> >>>> v7: (Chris)
> >>>> - cs_timestamp is a misnomer, use cs_cycles instead
> >>>> - return the cs cycle frequency as well in the query
> >>>>
> >>>> v8:
> >>>> - Add platform and engine specific checks
> >>>>
> >>>> v9: (Lionel)
> >>>> - Return 2 cpu timestamps in the query - captured before and after the
> >>>>     register read
> >>>>
> >>>> Signed-off-by: Umesh Nerlige Ramappa <umesh.nerlige.ramappa at intel.com>
> >>>> ---
> >>>>    drivers/gpu/drm/i915/i915_query.c | 144 ++++++++++++++++++++++++++++++
> >>>>    include/uapi/drm/i915_drm.h       |  47 ++++++++++
> >>>>    2 files changed, 191 insertions(+)
> >>>>
> >>>> diff --git a/drivers/gpu/drm/i915/i915_query.c b/drivers/gpu/drm/i915/i915_query.c
> >>>> index fed337ad7b68..acca22ee6014 100644
> >>>> --- a/drivers/gpu/drm/i915/i915_query.c
> >>>> +++ b/drivers/gpu/drm/i915/i915_query.c
> >>>> @@ -6,6 +6,8 @@
> >>>>    
> >>>>    #include <linux/nospec.h>
> >>>>    
> >>>> +#include "gt/intel_engine_pm.h"
> >>>> +#include "gt/intel_engine_user.h"
> >>>>    #include "i915_drv.h"
> >>>>    #include "i915_perf.h"
> >>>>    #include "i915_query.h"
> >>>> @@ -90,6 +92,147 @@ static int query_topology_info(struct drm_i915_private *dev_priv,
> >>>>           return total_length;
> >>>>    }
> >>>>    
> >>>> +typedef u64 (*__ktime_func_t)(void);
> >>>> +static __ktime_func_t __clock_id_to_func(clockid_t clk_id)
> >>>> +{
> >>>> +       /*
> >>>> +        * Use logic same as the perf subsystem to allow user to select the
> >>>> +        * reference clock id to be used for timestamps.
> >>>> +        */
> >>>> +       switch (clk_id) {
> >>>> +       case CLOCK_MONOTONIC:
> >>>> +               return &ktime_get_ns;
> >>>> +       case CLOCK_MONOTONIC_RAW:
> >>>> +               return &ktime_get_raw_ns;
> >>>> +       case CLOCK_REALTIME:
> >>>> +               return &ktime_get_real_ns;
> >>>> +       case CLOCK_BOOTTIME:
> >>>> +               return &ktime_get_boottime_ns;
> >>>> +       case CLOCK_TAI:
> >>>> +               return &ktime_get_clocktai_ns;
> >>>> +       default:
> >>>> +               return NULL;
> >>>> +       }
> >>>> +}
> >>>> +
> >>>> +static inline int
> >>>> +__read_timestamps(struct intel_uncore *uncore,
> >>>> +                 i915_reg_t lower_reg,
> >>>> +                 i915_reg_t upper_reg,
> >>>> +                 u64 *cs_ts,
> >>>> +                 u64 *cpu_ts,
> >>>> +                 __ktime_func_t cpu_clock)
> >>>> +{
> >>>> +       u32 upper, lower, old_upper, loop = 0;
> >>>> +
> >>>> +       upper = intel_uncore_read_fw(uncore, upper_reg);
> >>>> +       do {
> >>>> +               cpu_ts[0] = cpu_clock();
> >>>> +               lower = intel_uncore_read_fw(uncore, lower_reg);
> >>>> +               cpu_ts[1] = cpu_clock();
> >>>> +               old_upper = upper;
> >>>> +               upper = intel_uncore_read_fw(uncore, upper_reg);
> >>> Both register reads comprise the timestamp returned to userspace, so
> >>> presumably you want cpu_ts[] to wrap both.
> >>>
> >>>          do {
> >>>                  old_upper = upper;
> >>>
> >>>                  cpu_ts[0] = cpu_clock();
> >>>                  lower = intel_uncore_read_fw(uncore, lower_reg);
> >>>                  upper = intel_uncore_read_fw(uncore, upper_reg);
> >>>                  cpu_ts[1] = cpu_clock();
> >>>          } while (upper != old_upper && loop++ < 2);
> >> Actually if we want the best accuracy we can just deal with the lower dword.
> > Accuracy of what? The lower dword read perhaps, or the accuracy of the
> > sample point for the combined reads for the timestamp, which is closer
> > to an external observer (cpu_clock() implies reference to an external
> > observer).
> >
> > The two clock samples are not even necessarily closely related due to the
> > nmi adjustments. If you wanted an unadjusted elapsed time for the read
> > you can use local_clock() then return the chosen cpu_clock() before plus
> > the elapsed delta from around the read as the estimated error.
> >
> > cpu_ts[1] = local_clock();
> > cpu_ts[0] = cpu_clock();
> > lower = intel_uncore_read_fw(uncore, lower_reg);
> > cpu_ts[1] = local_clock() - cpu_ts[1];
> > -Chris
> 
> Thanks,
> 
> 
> I meant the accuracy of having 2 samples GPU/CPU as close as possible.
> 
> Avoiding to account another register read in there is nice.
> 
> 
> My testing was also mostly done with CLOCK_MONOTONIC_RAW which doesn't 
> seem to be adjusted like CLOCK_MONOTONIC so maybe that why I didn't see 
> the issue.

_RAW is still adjusted for skews, just not coupled into the ntp feedback.
That is less obvious than the other clocks, and why it's preferred for
comparing against other HW sources. But two reads of _RAW are only
monotonic, not necessarily on the same time base. local_clock() is
tsc/arat, so counting the CPU cycles between the two reads with the
frequency (at least on x86) held constant (and arat should be frequency
invariant).

If we want much better accuracy, we are supposed to use cyclecounter_t
and the system_device_crosststamp.
-Chris