[Intel-gfx] [PATCH 1/1] i915/query: Correlate engine and cpu timestamps with better accuracy
Jason Ekstrand
jason at jlekstrand.net
Wed Apr 28 20:45:48 UTC 2021
On Wed, Apr 28, 2021 at 3:14 PM Lionel Landwerlin
<lionel.g.landwerlin at intel.com> wrote:
>
> On 28/04/2021 22:54, Jason Ekstrand wrote:
> > On Wed, Apr 28, 2021 at 2:50 PM Lionel Landwerlin
> > <lionel.g.landwerlin at intel.com> wrote:
> >> On 28/04/2021 22:24, Jason Ekstrand wrote:
> >>
> >> On Wed, Apr 28, 2021 at 3:43 AM Jani Nikula <jani.nikula at linux.intel.com> wrote:
> >>
> >> On Tue, 27 Apr 2021, Umesh Nerlige Ramappa <umesh.nerlige.ramappa at intel.com> wrote:
> >>
> >> 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.
> >>
> >> Cc: dri-devel, Jason and Daniel for review.
> >>
> >> Thanks!
> >>
> >> 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
> >>
> >> v10: (Chris)
> >> - Use local_clock() to measure time taken to read lower dword of
> >> register and return it to user.
> >>
> >> v11: (Jani)
> >> - IS_GEN deprecated. User GRAPHICS_VER instead.
> >>
> >> Signed-off-by: Umesh Nerlige Ramappa <umesh.nerlige.ramappa at intel.com>
> >> ---
> >> drivers/gpu/drm/i915/i915_query.c | 145 ++++++++++++++++++++++++++++++
> >> include/uapi/drm/i915_drm.h | 48 ++++++++++
> >> 2 files changed, 193 insertions(+)
> >>
> >> diff --git a/drivers/gpu/drm/i915/i915_query.c b/drivers/gpu/drm/i915/i915_query.c
> >> index fed337ad7b68..2594b93901ac 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 +92,148 @@ 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[1] = local_clock();
> >> + cpu_ts[0] = cpu_clock();
> >> + lower = intel_uncore_read_fw(uncore, lower_reg);
> >> + cpu_ts[1] = local_clock() - cpu_ts[1];
> >> + old_upper = upper;
> >> + upper = intel_uncore_read_fw(uncore, upper_reg);
> >> + } while (upper != old_upper && loop++ < 2);
> >> +
> >> + *cs_ts = (u64)upper << 32 | lower;
> >> +
> >> + return 0;
> >> +}
> >> +
> >> +static int
> >> +__query_cs_cycles(struct intel_engine_cs *engine,
> >> + u64 *cs_ts, u64 *cpu_ts,
> >> + __ktime_func_t cpu_clock)
> >> +{
> >> + struct intel_uncore *uncore = engine->uncore;
> >> + enum forcewake_domains fw_domains;
> >> + u32 base = engine->mmio_base;
> >> + intel_wakeref_t wakeref;
> >> + int ret;
> >> +
> >> + fw_domains = intel_uncore_forcewake_for_reg(uncore,
> >> + RING_TIMESTAMP(base),
> >> + FW_REG_READ);
> >> +
> >> + with_intel_runtime_pm(uncore->rpm, wakeref) {
> >> + spin_lock_irq(&uncore->lock);
> >> + intel_uncore_forcewake_get__locked(uncore, fw_domains);
> >> +
> >> + ret = __read_timestamps(uncore,
> >> + RING_TIMESTAMP(base),
> >> + RING_TIMESTAMP_UDW(base),
> >> + cs_ts,
> >> + cpu_ts,
> >> + cpu_clock);
> >> +
> >> + intel_uncore_forcewake_put__locked(uncore, fw_domains);
> >> + spin_unlock_irq(&uncore->lock);
> >> + }
> >> +
> >> + return ret;
> >> +}
> >> +
> >> +static int
> >> +query_cs_cycles(struct drm_i915_private *i915,
> >> + struct drm_i915_query_item *query_item)
> >> +{
> >> + struct drm_i915_query_cs_cycles __user *query_ptr;
> >> + struct drm_i915_query_cs_cycles query;
> >> + struct intel_engine_cs *engine;
> >> + __ktime_func_t cpu_clock;
> >> + int ret;
> >> +
> >> + if (GRAPHICS_VER(i915) < 6)
> >> + return -ENODEV;
> >> +
> >> + query_ptr = u64_to_user_ptr(query_item->data_ptr);
> >> + ret = copy_query_item(&query, sizeof(query), sizeof(query), query_item);
> >> + if (ret != 0)
> >> + return ret;
> >> +
> >> + if (query.flags)
> >> + return -EINVAL;
> >> +
> >> + if (query.rsvd)
> >> + return -EINVAL;
> >> +
> >> + cpu_clock = __clock_id_to_func(query.clockid);
> >> + if (!cpu_clock)
> >> + return -EINVAL;
> >> +
> >> + engine = intel_engine_lookup_user(i915,
> >> + query.engine.engine_class,
> >> + query.engine.engine_instance);
> >> + if (!engine)
> >> + return -EINVAL;
> >> +
> >> + if (GRAPHICS_VER(i915) == 6 &&
> >> + query.engine.engine_class != I915_ENGINE_CLASS_RENDER)
> >> + return -ENODEV;
> >> +
> >> + query.cs_frequency = engine->gt->clock_frequency;
> >> + ret = __query_cs_cycles(engine,
> >> + &query.cs_cycles,
> >> + query.cpu_timestamp,
> >> + cpu_clock);
> >> + if (ret)
> >> + return ret;
> >> +
> >> + if (put_user(query.cs_frequency, &query_ptr->cs_frequency))
> >> + return -EFAULT;
> >> +
> >> + if (put_user(query.cpu_timestamp[0], &query_ptr->cpu_timestamp[0]))
> >> + return -EFAULT;
> >> +
> >> + if (put_user(query.cpu_timestamp[1], &query_ptr->cpu_timestamp[1]))
> >> + return -EFAULT;
> >> +
> >> + if (put_user(query.cs_cycles, &query_ptr->cs_cycles))
> >> + return -EFAULT;
> >> +
> >> + return sizeof(query);
> >> +}
> >> +
> >> static int
> >> query_engine_info(struct drm_i915_private *i915,
> >> struct drm_i915_query_item *query_item)
> >> @@ -424,6 +568,7 @@ static int (* const i915_query_funcs[])(struct drm_i915_private *dev_priv,
> >> query_topology_info,
> >> query_engine_info,
> >> query_perf_config,
> >> + query_cs_cycles,
> >> };
> >>
> >> int i915_query_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
> >> diff --git a/include/uapi/drm/i915_drm.h b/include/uapi/drm/i915_drm.h
> >> index 6a34243a7646..08b00f1709b5 100644
> >> --- a/include/uapi/drm/i915_drm.h
> >> +++ b/include/uapi/drm/i915_drm.h
> >> @@ -2230,6 +2230,10 @@ struct drm_i915_query_item {
> >> #define DRM_I915_QUERY_TOPOLOGY_INFO 1
> >> #define DRM_I915_QUERY_ENGINE_INFO 2
> >> #define DRM_I915_QUERY_PERF_CONFIG 3
> >> + /**
> >> + * Query Command Streamer timestamp register.
> >> + */
> >> +#define DRM_I915_QUERY_CS_CYCLES 4
> >> /* Must be kept compact -- no holes and well documented */
> >>
> >> /**
> >> @@ -2397,6 +2401,50 @@ struct drm_i915_engine_info {
> >> __u64 rsvd1[4];
> >> };
> >>
> >> +/**
> >> + * struct drm_i915_query_cs_cycles
> >> + *
> >> + * The query returns the command streamer cycles and the frequency that can be
> >> + * used to calculate the command streamer timestamp. In addition the query
> >> + * returns a set of cpu timestamps that indicate when the command streamer cycle
> >> + * count was captured.
> >> + */
> >> +struct drm_i915_query_cs_cycles {
> >> + /** Engine for which command streamer cycles is queried. */
> >> + struct i915_engine_class_instance engine;
> >>
> >> Why is this per-engine? Do we actually expect it to change between
> >> engines?
> >>
> >>
> >> Each engine has its own timestamp register.
> >>
> >>
> >> If so, we may have a problem because Vulkan expects a
> >> unified timestamp domain for all command streamer timestamp queries.
> >>
> >>
> >> I don't think it does : "
> >>
> >> Timestamps may only be meaningfully compared if they are written by commands submitted to the same queue.
> > Yes but vkGetCalibratedTimestampsEXT() doesn't take a queue or even a
> > queue family.
>
>
> I know, I brought up the issue recently. See khronos issue 2551.
I guess this is what I get for not attending the Vulkan SI call
anymore. Small price to pay....
So the answer is that we just stop exposing the DEVICE time domain as
soon as we start using anything other than RENDER? Seems a bit rough
but should be doable.
> You might not like the resolution... I did propose to do a rev2 of the
> extension to let the user specify the queue.
>
> We can still do that in the future.
Yeah, I think we'll want to do something if we care about this
extension. One option would be to make it take a queue family.
Another would be to expose it as one domain per queue family.
Anyway... that's a discussion for another forum.
> > Also, VkPhysicalDeviceLimits::timestampPeriod gives a
> > single timestampPeriod for all queues.
>
>
> That is fine for us, we should have the same period on all command
> streamers.
I guess I've got no problem returning the period as part of this
query. ANV should probably assert that it's what it expects, though.
> -Lionel
>
>
> > It's possible that Vulkan
> > messed up real bad there but I thought we did a HW survey at the time
> > and determined that it was ok.
> >
> > --Jason
> >
> >
> >> " [1]
> >>
> >>
> >> [1] : https://www.khronos.org/registry/vulkan/specs/1.2-extensions/man/html/vkCmdWriteTimestamp.html
> >>
> >>
> >> -Lionel
> >>
> >>
> >>
> >> --Jason
> >>
> >>
> >> + /** Must be zero. */
> >> + __u32 flags;
> >> +
> >> + /**
> >> + * Command streamer cycles as read from the command streamer
> >> + * register at 0x358 offset.
> >> + */
> >> + __u64 cs_cycles;
> >> +
> >> + /** Frequency of the cs cycles in Hz. */
> >> + __u64 cs_frequency;
> >> +
> >> + /**
> >> + * CPU timestamps in ns. cpu_timestamp[0] is captured before reading the
> >> + * cs_cycles register using the reference clockid set by the user.
> >> + * cpu_timestamp[1] is the time taken in ns to read the lower dword of
> >> + * the cs_cycles register.
> >> + */
> >> + __u64 cpu_timestamp[2];
I think the API would be more clear if we had separate cpu_timestamp
and cpu_delta fields or something like that. That or make
cpu_timestamp[1] the end time rather than a delta. It's weird to have
an array where the first entry is absolute and the second entry is a
delta.
--Jason
> >> +
> >> + /**
> >> + * Reference clock id for CPU timestamp. For definition, see
> >> + * clock_gettime(2) and perf_event_open(2). Supported clock ids are
> >> + * CLOCK_MONOTONIC, CLOCK_MONOTONIC_RAW, CLOCK_REALTIME, CLOCK_BOOTTIME,
> >> + * CLOCK_TAI.
> >> + */
> >> + __s32 clockid;
> >> +
> >> + /** Must be zero. */
> >> + __u32 rsvd;
> >> +};
> >> +
> >> /**
> >> * struct drm_i915_query_engine_info
> >> *
> >>
> >> --
> >> Jani Nikula, Intel Open Source Graphics Center
> >>
> >>
>
More information about the Intel-gfx
mailing list