drm/i915/pmu: Expose a PMU interface for perf queries
authorTvrtko Ursulin <tvrtko.ursulin@intel.com>
Tue, 21 Nov 2017 18:18:45 +0000 (18:18 +0000)
committerTvrtko Ursulin <tvrtko.ursulin@intel.com>
Wed, 22 Nov 2017 11:24:57 +0000 (11:24 +0000)
From: Chris Wilson <chris@chris-wilson.co.uk>
From: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
From: Dmitry Rogozhkin <dmitry.v.rogozhkin@intel.com>

The first goal is to be able to measure GPU (and invidual ring) busyness
without having to poll registers from userspace. (Which not only incurs
holding the forcewake lock indefinitely, perturbing the system, but also
runs the risk of hanging the machine.) As an alternative we can use the
perf event counter interface to sample the ring registers periodically
and send those results to userspace.

Functionality we are exporting to userspace is via the existing perf PMU
API and can be exercised via the existing tools. For example:

  perf stat -a -e i915/rcs0-busy/ -I 1000

Will print the render engine busynnes once per second. All the performance
counters can be enumerated (perf list) and have their unit of measure
correctly reported in sysfs.

v1-v2 (Chris Wilson):

v2: Use a common timer for the ring sampling.

v3: (Tvrtko Ursulin)
 * Decouple uAPI from i915 engine ids.
 * Complete uAPI defines.
 * Refactor some code to helpers for clarity.
 * Skip sampling disabled engines.
 * Expose counters in sysfs.
 * Pass in fake regs to avoid null ptr deref in perf core.
 * Convert to class/instance uAPI.
 * Use shared driver code for rc6 residency, power and frequency.

v4: (Dmitry Rogozhkin)
 * Register PMU with .task_ctx_nr=perf_invalid_context
 * Expose cpumask for the PMU with the single CPU in the mask
 * Properly support pmu->stop(): it should call pmu->read()
 * Properly support pmu->del(): it should call stop(event, PERF_EF_UPDATE)
 * Introduce refcounting of event subscriptions.
 * Make pmu.busy_stats a refcounter to avoid busy stats going away
   with some deleted event.
 * Expose cpumask for i915 PMU to avoid multiple events creation of
   the same type followed by counter aggregation by perf-stat.
 * Track CPUs getting online/offline to migrate perf context. If (likely)
   cpumask will initially set CPU0, CONFIG_BOOTPARAM_HOTPLUG_CPU0 will be
   needed to see effect of CPU status tracking.
 * End result is that only global events are supported and perf stat
   works correctly.
 * Deny perf driver level sampling - it is prohibited for uncore PMU.

v5: (Tvrtko Ursulin)

 * Don't hardcode number of engine samplers.
 * Rewrite event ref-counting for correctness and simplicity.
 * Store initial counter value when starting already enabled events
   to correctly report values to all listeners.
 * Fix RC6 residency readout.
 * Comments, GPL header.

v6:
 * Add missing entry to v4 changelog.
 * Fix accounting in CPU hotplug case by copying the approach from
   arch/x86/events/intel/cstate.c. (Dmitry Rogozhkin)

v7:
 * Log failure message only on failure.
 * Remove CPU hotplug notification state on unregister.

v8:
 * Fix error unwind on failed registration.
 * Checkpatch cleanup.

v9:
 * Drop the energy metric, it is available via intel_rapl_perf.
   (Ville Syrjälä)
 * Use HAS_RC6(p). (Chris Wilson)
 * Handle unsupported non-engine events. (Dmitry Rogozhkin)
 * Rebase for intel_rc6_residency_ns needing caller managed
   runtime pm.
 * Drop HAS_RC6 checks from the read callback since creating those
   events will be rejected at init time already.
 * Add counter units to sysfs so perf stat output is nicer.
 * Cleanup the attribute tables for brevity and readability.

v10:
 * Fixed queued accounting.

v11:
 * Move intel_engine_lookup_user to intel_engine_cs.c
 * Commit update. (Joonas Lahtinen)

v12:
 * More accurate sampling. (Chris Wilson)
 * Store and report frequency in MHz for better usability from
   perf stat.
 * Removed metrics: queued, interrupts, rc6 counters.
 * Sample engine busyness based on seqno difference only
   for less MMIO (and forcewake) on all platforms. (Chris Wilson)

v13:
 * Comment spelling, use mul_u32_u32 to work around potential GCC
   issue and somne code alignment changes. (Chris Wilson)

v14:
 * Rebase.

v15:
 * Rebase for RPS refactoring.

v16:
 * Use the dynamic slot in the CPU hotplug state machine so that we are
   free to setup our state as multi-instance. Previously we were re-using
   the CPUHP_AP_PERF_X86_UNCORE_ONLINE slot which is neither used as
   multi-instance, nor owned by our driver to start with.
 * Register the CPU hotplug handlers after the PMU, otherwise the callback
   will get called before the PMU is initialized which can end up in
   perf_pmu_migrate_context with an un-initialized base.
 * Added workaround for a probable bug in cpuhp core.

v17:
 * Remove workaround for the cpuhp bug.

v18:
 * Rebase for drm_i915_gem_engine_class getting upstream before us.

v19:
 * Rebase. (trivial)

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Signed-off-by: Dmitry Rogozhkin <dmitry.v.rogozhkin@intel.com>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Dmitry Rogozhkin <dmitry.v.rogozhkin@intel.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20171121181852.16128-2-tvrtko.ursulin@linux.intel.com
drivers/gpu/drm/i915/Makefile
drivers/gpu/drm/i915/i915_drv.c
drivers/gpu/drm/i915/i915_drv.h
drivers/gpu/drm/i915/i915_pmu.c [new file with mode: 0644]
drivers/gpu/drm/i915/i915_pmu.h [new file with mode: 0644]
drivers/gpu/drm/i915/i915_reg.h
drivers/gpu/drm/i915/intel_engine_cs.c
drivers/gpu/drm/i915/intel_ringbuffer.h
include/uapi/drm/i915_drm.h

index c3649ec..42bc8bd 100644 (file)
@@ -45,6 +45,7 @@ i915-y := i915_drv.o \
 
 i915-$(CONFIG_COMPAT)   += i915_ioc32.o
 i915-$(CONFIG_DEBUG_FS) += i915_debugfs.o intel_pipe_crc.o
+i915-$(CONFIG_PERF_EVENTS) += i915_pmu.o
 
 # GEM code
 i915-y += i915_cmd_parser.o \
index 8dbcb03..0793a27 100644 (file)
@@ -48,6 +48,7 @@
 
 #include "i915_drv.h"
 #include "i915_trace.h"
+#include "i915_pmu.h"
 #include "i915_vgpu.h"
 #include "intel_drv.h"
 #include "intel_uc.h"
@@ -1215,6 +1216,7 @@ static void i915_driver_register(struct drm_i915_private *dev_priv)
        struct drm_device *dev = &dev_priv->drm;
 
        i915_gem_shrinker_init(dev_priv);
+       i915_pmu_register(dev_priv);
 
        /*
         * Notify a valid surface after modesetting,
@@ -1269,6 +1271,7 @@ static void i915_driver_unregister(struct drm_i915_private *dev_priv)
        intel_opregion_unregister(dev_priv);
 
        i915_perf_unregister(dev_priv);
+       i915_pmu_unregister(dev_priv);
 
        i915_teardown_sysfs(dev_priv);
        i915_guc_log_unregister(dev_priv);
index 0191171..5bd5ac4 100644 (file)
@@ -40,6 +40,7 @@
 #include <linux/hash.h>
 #include <linux/intel-iommu.h>
 #include <linux/kref.h>
+#include <linux/perf_event.h>
 #include <linux/pm_qos.h>
 #include <linux/reservation.h>
 #include <linux/shmem_fs.h>
@@ -2290,6 +2291,8 @@ struct drm_i915_private {
        struct i915_gem_context *kernel_context;
        /* Context only to be used for injecting preemption commands */
        struct i915_gem_context *preempt_context;
+       struct intel_engine_cs *engine_class[MAX_ENGINE_CLASS + 1]
+                                           [MAX_ENGINE_INSTANCE + 1];
 
        struct drm_dma_handle *status_page_dmah;
        struct resource mch_res;
@@ -2761,6 +2764,8 @@ struct drm_i915_private {
                int     irq;
        } lpe_audio;
 
+       struct i915_pmu pmu;
+
        /*
         * NOTE: This is the dri1/ums dungeon, don't add stuff here. Your patch
         * will be rejected. Instead look for a better place.
diff --git a/drivers/gpu/drm/i915/i915_pmu.c b/drivers/gpu/drm/i915/i915_pmu.c
new file mode 100644 (file)
index 0000000..01b5ee6
--- /dev/null
@@ -0,0 +1,688 @@
+/*
+ * Copyright © 2017 Intel Corporation
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ *
+ */
+
+#include <linux/perf_event.h>
+#include <linux/pm_runtime.h>
+
+#include "i915_drv.h"
+#include "i915_pmu.h"
+#include "intel_ringbuffer.h"
+
+/* Frequency for the sampling timer for events which need it. */
+#define FREQUENCY 200
+#define PERIOD max_t(u64, 10000, NSEC_PER_SEC / FREQUENCY)
+
+#define ENGINE_SAMPLE_MASK \
+       (BIT(I915_SAMPLE_BUSY) | \
+        BIT(I915_SAMPLE_WAIT) | \
+        BIT(I915_SAMPLE_SEMA))
+
+#define ENGINE_SAMPLE_BITS (1 << I915_PMU_SAMPLE_BITS)
+
+static cpumask_t i915_pmu_cpumask = CPU_MASK_NONE;
+
+static u8 engine_config_sample(u64 config)
+{
+       return config & I915_PMU_SAMPLE_MASK;
+}
+
+static u8 engine_event_sample(struct perf_event *event)
+{
+       return engine_config_sample(event->attr.config);
+}
+
+static u8 engine_event_class(struct perf_event *event)
+{
+       return (event->attr.config >> I915_PMU_CLASS_SHIFT) & 0xff;
+}
+
+static u8 engine_event_instance(struct perf_event *event)
+{
+       return (event->attr.config >> I915_PMU_SAMPLE_BITS) & 0xff;
+}
+
+static bool is_engine_config(u64 config)
+{
+       return config < __I915_PMU_OTHER(0);
+}
+
+static unsigned int config_enabled_bit(u64 config)
+{
+       if (is_engine_config(config))
+               return engine_config_sample(config);
+       else
+               return ENGINE_SAMPLE_BITS + (config - __I915_PMU_OTHER(0));
+}
+
+static u64 config_enabled_mask(u64 config)
+{
+       return BIT_ULL(config_enabled_bit(config));
+}
+
+static bool is_engine_event(struct perf_event *event)
+{
+       return is_engine_config(event->attr.config);
+}
+
+static unsigned int event_enabled_bit(struct perf_event *event)
+{
+       return config_enabled_bit(event->attr.config);
+}
+
+static bool grab_forcewake(struct drm_i915_private *i915, bool fw)
+{
+       if (!fw)
+               intel_uncore_forcewake_get(i915, FORCEWAKE_ALL);
+
+       return true;
+}
+
+static void
+update_sample(struct i915_pmu_sample *sample, u32 unit, u32 val)
+{
+       /*
+        * Since we are doing stochastic sampling for these counters,
+        * average the delta with the previous value for better accuracy.
+        */
+       sample->cur += div_u64(mul_u32_u32(sample->prev + val, unit), 2);
+       sample->prev = val;
+}
+
+static void engines_sample(struct drm_i915_private *dev_priv)
+{
+       struct intel_engine_cs *engine;
+       enum intel_engine_id id;
+       bool fw = false;
+
+       if ((dev_priv->pmu.enable & ENGINE_SAMPLE_MASK) == 0)
+               return;
+
+       if (!dev_priv->gt.awake)
+               return;
+
+       if (!intel_runtime_pm_get_if_in_use(dev_priv))
+               return;
+
+       for_each_engine(engine, dev_priv, id) {
+               u32 current_seqno = intel_engine_get_seqno(engine);
+               u32 last_seqno = intel_engine_last_submit(engine);
+               u32 val;
+
+               val = !i915_seqno_passed(current_seqno, last_seqno);
+
+               update_sample(&engine->pmu.sample[I915_SAMPLE_BUSY],
+                             PERIOD, val);
+
+               if (val && (engine->pmu.enable &
+                   (BIT(I915_SAMPLE_WAIT) | BIT(I915_SAMPLE_SEMA)))) {
+                       fw = grab_forcewake(dev_priv, fw);
+
+                       val = I915_READ_FW(RING_CTL(engine->mmio_base));
+               } else {
+                       val = 0;
+               }
+
+               update_sample(&engine->pmu.sample[I915_SAMPLE_WAIT],
+                             PERIOD, !!(val & RING_WAIT));
+
+               update_sample(&engine->pmu.sample[I915_SAMPLE_SEMA],
+                             PERIOD, !!(val & RING_WAIT_SEMAPHORE));
+       }
+
+       if (fw)
+               intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
+
+       intel_runtime_pm_put(dev_priv);
+}
+
+static void frequency_sample(struct drm_i915_private *dev_priv)
+{
+       if (dev_priv->pmu.enable &
+           config_enabled_mask(I915_PMU_ACTUAL_FREQUENCY)) {
+               u32 val;
+
+               val = dev_priv->gt_pm.rps.cur_freq;
+               if (dev_priv->gt.awake &&
+                   intel_runtime_pm_get_if_in_use(dev_priv)) {
+                       val = intel_get_cagf(dev_priv,
+                                            I915_READ_NOTRACE(GEN6_RPSTAT1));
+                       intel_runtime_pm_put(dev_priv);
+               }
+
+               update_sample(&dev_priv->pmu.sample[__I915_SAMPLE_FREQ_ACT],
+                             1, intel_gpu_freq(dev_priv, val));
+       }
+
+       if (dev_priv->pmu.enable &
+           config_enabled_mask(I915_PMU_REQUESTED_FREQUENCY)) {
+               update_sample(&dev_priv->pmu.sample[__I915_SAMPLE_FREQ_REQ], 1,
+                             intel_gpu_freq(dev_priv,
+                                            dev_priv->gt_pm.rps.cur_freq));
+       }
+}
+
+static enum hrtimer_restart i915_sample(struct hrtimer *hrtimer)
+{
+       struct drm_i915_private *i915 =
+               container_of(hrtimer, struct drm_i915_private, pmu.timer);
+
+       if (i915->pmu.enable == 0)
+               return HRTIMER_NORESTART;
+
+       engines_sample(i915);
+       frequency_sample(i915);
+
+       hrtimer_forward_now(hrtimer, ns_to_ktime(PERIOD));
+       return HRTIMER_RESTART;
+}
+
+static void i915_pmu_event_destroy(struct perf_event *event)
+{
+       WARN_ON(event->parent);
+}
+
+static int engine_event_init(struct perf_event *event)
+{
+       struct drm_i915_private *i915 =
+               container_of(event->pmu, typeof(*i915), pmu.base);
+
+       if (!intel_engine_lookup_user(i915, engine_event_class(event),
+                                     engine_event_instance(event)))
+               return -ENODEV;
+
+       switch (engine_event_sample(event)) {
+       case I915_SAMPLE_BUSY:
+       case I915_SAMPLE_WAIT:
+               break;
+       case I915_SAMPLE_SEMA:
+               if (INTEL_GEN(i915) < 6)
+                       return -ENODEV;
+               break;
+       default:
+               return -ENOENT;
+       }
+
+       return 0;
+}
+
+static int i915_pmu_event_init(struct perf_event *event)
+{
+       struct drm_i915_private *i915 =
+               container_of(event->pmu, typeof(*i915), pmu.base);
+       int cpu, ret;
+
+       if (event->attr.type != event->pmu->type)
+               return -ENOENT;
+
+       /* unsupported modes and filters */
+       if (event->attr.sample_period) /* no sampling */
+               return -EINVAL;
+
+       if (has_branch_stack(event))
+               return -EOPNOTSUPP;
+
+       if (event->cpu < 0)
+               return -EINVAL;
+
+       cpu = cpumask_any_and(&i915_pmu_cpumask,
+                             topology_sibling_cpumask(event->cpu));
+       if (cpu >= nr_cpu_ids)
+               return -ENODEV;
+
+       if (is_engine_event(event)) {
+               ret = engine_event_init(event);
+       } else {
+               ret = 0;
+               switch (event->attr.config) {
+               case I915_PMU_ACTUAL_FREQUENCY:
+                       if (IS_VALLEYVIEW(i915) || IS_CHERRYVIEW(i915))
+                                /* Requires a mutex for sampling! */
+                               ret = -ENODEV;
+               case I915_PMU_REQUESTED_FREQUENCY:
+                       if (INTEL_GEN(i915) < 6)
+                               ret = -ENODEV;
+                       break;
+               default:
+                       ret = -ENOENT;
+                       break;
+               }
+       }
+       if (ret)
+               return ret;
+
+       event->cpu = cpu;
+       if (!event->parent)
+               event->destroy = i915_pmu_event_destroy;
+
+       return 0;
+}
+
+static u64 __i915_pmu_event_read(struct perf_event *event)
+{
+       struct drm_i915_private *i915 =
+               container_of(event->pmu, typeof(*i915), pmu.base);
+       u64 val = 0;
+
+       if (is_engine_event(event)) {
+               u8 sample = engine_event_sample(event);
+               struct intel_engine_cs *engine;
+
+               engine = intel_engine_lookup_user(i915,
+                                                 engine_event_class(event),
+                                                 engine_event_instance(event));
+
+               if (WARN_ON_ONCE(!engine)) {
+                       /* Do nothing */
+               } else {
+                       val = engine->pmu.sample[sample].cur;
+               }
+       } else {
+               switch (event->attr.config) {
+               case I915_PMU_ACTUAL_FREQUENCY:
+                       val =
+                          div_u64(i915->pmu.sample[__I915_SAMPLE_FREQ_ACT].cur,
+                                  FREQUENCY);
+                       break;
+               case I915_PMU_REQUESTED_FREQUENCY:
+                       val =
+                          div_u64(i915->pmu.sample[__I915_SAMPLE_FREQ_REQ].cur,
+                                  FREQUENCY);
+                       break;
+               }
+       }
+
+       return val;
+}
+
+static void i915_pmu_event_read(struct perf_event *event)
+{
+       struct hw_perf_event *hwc = &event->hw;
+       u64 prev, new;
+
+again:
+       prev = local64_read(&hwc->prev_count);
+       new = __i915_pmu_event_read(event);
+
+       if (local64_cmpxchg(&hwc->prev_count, prev, new) != prev)
+               goto again;
+
+       local64_add(new - prev, &event->count);
+}
+
+static void i915_pmu_enable(struct perf_event *event)
+{
+       struct drm_i915_private *i915 =
+               container_of(event->pmu, typeof(*i915), pmu.base);
+       unsigned int bit = event_enabled_bit(event);
+       unsigned long flags;
+
+       spin_lock_irqsave(&i915->pmu.lock, flags);
+
+       /*
+        * Start the sampling timer when enabling the first event.
+        */
+       if (i915->pmu.enable == 0)
+               hrtimer_start_range_ns(&i915->pmu.timer,
+                                      ns_to_ktime(PERIOD), 0,
+                                      HRTIMER_MODE_REL_PINNED);
+
+       /*
+        * Update the bitmask of enabled events and increment
+        * the event reference counter.
+        */
+       GEM_BUG_ON(bit >= I915_PMU_MASK_BITS);
+       GEM_BUG_ON(i915->pmu.enable_count[bit] == ~0);
+       i915->pmu.enable |= BIT_ULL(bit);
+       i915->pmu.enable_count[bit]++;
+
+       /*
+        * For per-engine events the bitmask and reference counting
+        * is stored per engine.
+        */
+       if (is_engine_event(event)) {
+               u8 sample = engine_event_sample(event);
+               struct intel_engine_cs *engine;
+
+               engine = intel_engine_lookup_user(i915,
+                                                 engine_event_class(event),
+                                                 engine_event_instance(event));
+               GEM_BUG_ON(!engine);
+               engine->pmu.enable |= BIT(sample);
+
+               GEM_BUG_ON(sample >= I915_PMU_SAMPLE_BITS);
+               GEM_BUG_ON(engine->pmu.enable_count[sample] == ~0);
+               engine->pmu.enable_count[sample]++;
+       }
+
+       /*
+        * Store the current counter value so we can report the correct delta
+        * for all listeners. Even when the event was already enabled and has
+        * an existing non-zero value.
+        */
+       local64_set(&event->hw.prev_count, __i915_pmu_event_read(event));
+
+       spin_unlock_irqrestore(&i915->pmu.lock, flags);
+}
+
+static void i915_pmu_disable(struct perf_event *event)
+{
+       struct drm_i915_private *i915 =
+               container_of(event->pmu, typeof(*i915), pmu.base);
+       unsigned int bit = event_enabled_bit(event);
+       unsigned long flags;
+
+       spin_lock_irqsave(&i915->pmu.lock, flags);
+
+       if (is_engine_event(event)) {
+               u8 sample = engine_event_sample(event);
+               struct intel_engine_cs *engine;
+
+               engine = intel_engine_lookup_user(i915,
+                                                 engine_event_class(event),
+                                                 engine_event_instance(event));
+               GEM_BUG_ON(!engine);
+               GEM_BUG_ON(sample >= I915_PMU_SAMPLE_BITS);
+               GEM_BUG_ON(engine->pmu.enable_count[sample] == 0);
+               /*
+                * Decrement the reference count and clear the enabled
+                * bitmask when the last listener on an event goes away.
+                */
+               if (--engine->pmu.enable_count[sample] == 0)
+                       engine->pmu.enable &= ~BIT(sample);
+       }
+
+       GEM_BUG_ON(bit >= I915_PMU_MASK_BITS);
+       GEM_BUG_ON(i915->pmu.enable_count[bit] == 0);
+       /*
+        * Decrement the reference count and clear the enabled
+        * bitmask when the last listener on an event goes away.
+        */
+       if (--i915->pmu.enable_count[bit] == 0)
+               i915->pmu.enable &= ~BIT_ULL(bit);
+
+       spin_unlock_irqrestore(&i915->pmu.lock, flags);
+}
+
+static void i915_pmu_event_start(struct perf_event *event, int flags)
+{
+       i915_pmu_enable(event);
+       event->hw.state = 0;
+}
+
+static void i915_pmu_event_stop(struct perf_event *event, int flags)
+{
+       if (flags & PERF_EF_UPDATE)
+               i915_pmu_event_read(event);
+       i915_pmu_disable(event);
+       event->hw.state = PERF_HES_STOPPED;
+}
+
+static int i915_pmu_event_add(struct perf_event *event, int flags)
+{
+       if (flags & PERF_EF_START)
+               i915_pmu_event_start(event, flags);
+
+       return 0;
+}
+
+static void i915_pmu_event_del(struct perf_event *event, int flags)
+{
+       i915_pmu_event_stop(event, PERF_EF_UPDATE);
+}
+
+static int i915_pmu_event_event_idx(struct perf_event *event)
+{
+       return 0;
+}
+
+static ssize_t i915_pmu_format_show(struct device *dev,
+                                   struct device_attribute *attr, char *buf)
+{
+       struct dev_ext_attribute *eattr;
+
+       eattr = container_of(attr, struct dev_ext_attribute, attr);
+       return sprintf(buf, "%s\n", (char *)eattr->var);
+}
+
+#define I915_PMU_FORMAT_ATTR(_name, _config) \
+       (&((struct dev_ext_attribute[]) { \
+               { .attr = __ATTR(_name, 0444, i915_pmu_format_show, NULL), \
+                 .var = (void *)_config, } \
+       })[0].attr.attr)
+
+static struct attribute *i915_pmu_format_attrs[] = {
+       I915_PMU_FORMAT_ATTR(i915_eventid, "config:0-20"),
+       NULL,
+};
+
+static const struct attribute_group i915_pmu_format_attr_group = {
+       .name = "format",
+       .attrs = i915_pmu_format_attrs,
+};
+
+static ssize_t i915_pmu_event_show(struct device *dev,
+                                  struct device_attribute *attr, char *buf)
+{
+       struct dev_ext_attribute *eattr;
+
+       eattr = container_of(attr, struct dev_ext_attribute, attr);
+       return sprintf(buf, "config=0x%lx\n", (unsigned long)eattr->var);
+}
+
+#define I915_EVENT_ATTR(_name, _config) \
+       (&((struct dev_ext_attribute[]) { \
+               { .attr = __ATTR(_name, 0444, i915_pmu_event_show, NULL), \
+                 .var = (void *)_config, } \
+       })[0].attr.attr)
+
+#define I915_EVENT_STR(_name, _str) \
+       (&((struct perf_pmu_events_attr[]) { \
+               { .attr      = __ATTR(_name, 0444, perf_event_sysfs_show, NULL), \
+                 .id        = 0, \
+                 .event_str = _str, } \
+       })[0].attr.attr)
+
+#define I915_EVENT(_name, _config, _unit) \
+       I915_EVENT_ATTR(_name, _config), \
+       I915_EVENT_STR(_name.unit, _unit)
+
+#define I915_ENGINE_EVENT(_name, _class, _instance, _sample) \
+       I915_EVENT_ATTR(_name, __I915_PMU_ENGINE(_class, _instance, _sample)), \
+       I915_EVENT_STR(_name.unit, "ns")
+
+#define I915_ENGINE_EVENTS(_name, _class, _instance) \
+       I915_ENGINE_EVENT(_name##_instance-busy, _class, _instance, I915_SAMPLE_BUSY), \
+       I915_ENGINE_EVENT(_name##_instance-sema, _class, _instance, I915_SAMPLE_SEMA), \
+       I915_ENGINE_EVENT(_name##_instance-wait, _class, _instance, I915_SAMPLE_WAIT)
+
+static struct attribute *i915_pmu_events_attrs[] = {
+       I915_ENGINE_EVENTS(rcs, I915_ENGINE_CLASS_RENDER, 0),
+       I915_ENGINE_EVENTS(bcs, I915_ENGINE_CLASS_COPY, 0),
+       I915_ENGINE_EVENTS(vcs, I915_ENGINE_CLASS_VIDEO, 0),
+       I915_ENGINE_EVENTS(vcs, I915_ENGINE_CLASS_VIDEO, 1),
+       I915_ENGINE_EVENTS(vecs, I915_ENGINE_CLASS_VIDEO_ENHANCE, 0),
+
+       I915_EVENT(actual-frequency,    I915_PMU_ACTUAL_FREQUENCY,    "MHz"),
+       I915_EVENT(requested-frequency, I915_PMU_REQUESTED_FREQUENCY, "MHz"),
+
+       NULL,
+};
+
+static const struct attribute_group i915_pmu_events_attr_group = {
+       .name = "events",
+       .attrs = i915_pmu_events_attrs,
+};
+
+static ssize_t
+i915_pmu_get_attr_cpumask(struct device *dev,
+                         struct device_attribute *attr,
+                         char *buf)
+{
+       return cpumap_print_to_pagebuf(true, buf, &i915_pmu_cpumask);
+}
+
+static DEVICE_ATTR(cpumask, 0444, i915_pmu_get_attr_cpumask, NULL);
+
+static struct attribute *i915_cpumask_attrs[] = {
+       &dev_attr_cpumask.attr,
+       NULL,
+};
+
+static struct attribute_group i915_pmu_cpumask_attr_group = {
+       .attrs = i915_cpumask_attrs,
+};
+
+static const struct attribute_group *i915_pmu_attr_groups[] = {
+       &i915_pmu_format_attr_group,
+       &i915_pmu_events_attr_group,
+       &i915_pmu_cpumask_attr_group,
+       NULL
+};
+
+#ifdef CONFIG_HOTPLUG_CPU
+static int i915_pmu_cpu_online(unsigned int cpu, struct hlist_node *node)
+{
+       struct i915_pmu *pmu = hlist_entry_safe(node, typeof(*pmu), node);
+       unsigned int target;
+
+       GEM_BUG_ON(!pmu->base.event_init);
+
+       target = cpumask_any_and(&i915_pmu_cpumask, &i915_pmu_cpumask);
+       /* Select the first online CPU as a designated reader. */
+       if (target >= nr_cpu_ids)
+               cpumask_set_cpu(cpu, &i915_pmu_cpumask);
+
+       return 0;
+}
+
+static int i915_pmu_cpu_offline(unsigned int cpu, struct hlist_node *node)
+{
+       struct i915_pmu *pmu = hlist_entry_safe(node, typeof(*pmu), node);
+       unsigned int target;
+
+       GEM_BUG_ON(!pmu->base.event_init);
+
+       if (cpumask_test_and_clear_cpu(cpu, &i915_pmu_cpumask)) {
+               target = cpumask_any_but(topology_sibling_cpumask(cpu), cpu);
+               /* Migrate events if there is a valid target */
+               if (target < nr_cpu_ids) {
+                       cpumask_set_cpu(target, &i915_pmu_cpumask);
+                       perf_pmu_migrate_context(&pmu->base, cpu, target);
+               }
+       }
+
+       return 0;
+}
+
+static enum cpuhp_state cpuhp_slot = CPUHP_INVALID;
+#endif
+
+static int i915_pmu_register_cpuhp_state(struct drm_i915_private *i915)
+{
+#ifdef CONFIG_HOTPLUG_CPU
+       enum cpuhp_state slot;
+       int ret;
+
+       ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN,
+                                     "perf/x86/intel/i915:online",
+                                     i915_pmu_cpu_online,
+                                     i915_pmu_cpu_offline);
+       if (ret < 0)
+               return ret;
+
+       slot = ret;
+       ret = cpuhp_state_add_instance(slot, &i915->pmu.node);
+       if (ret) {
+               cpuhp_remove_multi_state(slot);
+               return ret;
+       }
+
+       cpuhp_slot = slot;
+#endif
+       return 0;
+}
+
+static void i915_pmu_unregister_cpuhp_state(struct drm_i915_private *i915)
+{
+#ifdef CONFIG_HOTPLUG_CPU
+       WARN_ON(cpuhp_slot == CPUHP_INVALID);
+       WARN_ON(cpuhp_state_remove_instance(cpuhp_slot, &i915->pmu.node));
+       cpuhp_remove_multi_state(cpuhp_slot);
+#endif
+}
+
+void i915_pmu_register(struct drm_i915_private *i915)
+{
+       int ret;
+
+       if (INTEL_GEN(i915) <= 2) {
+               DRM_INFO("PMU not supported for this GPU.");
+               return;
+       }
+
+       i915->pmu.base.attr_groups      = i915_pmu_attr_groups;
+       i915->pmu.base.task_ctx_nr      = perf_invalid_context;
+       i915->pmu.base.event_init       = i915_pmu_event_init;
+       i915->pmu.base.add              = i915_pmu_event_add;
+       i915->pmu.base.del              = i915_pmu_event_del;
+       i915->pmu.base.start            = i915_pmu_event_start;
+       i915->pmu.base.stop             = i915_pmu_event_stop;
+       i915->pmu.base.read             = i915_pmu_event_read;
+       i915->pmu.base.event_idx        = i915_pmu_event_event_idx;
+
+       spin_lock_init(&i915->pmu.lock);
+       hrtimer_init(&i915->pmu.timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
+       i915->pmu.timer.function = i915_sample;
+
+       ret = perf_pmu_register(&i915->pmu.base, "i915", -1);
+       if (ret)
+               goto err;
+
+       ret = i915_pmu_register_cpuhp_state(i915);
+       if (ret)
+               goto err_unreg;
+
+       return;
+
+err_unreg:
+       perf_pmu_unregister(&i915->pmu.base);
+err:
+       i915->pmu.base.event_init = NULL;
+       DRM_NOTE("Failed to register PMU! (err=%d)\n", ret);
+}
+
+void i915_pmu_unregister(struct drm_i915_private *i915)
+{
+       if (!i915->pmu.base.event_init)
+               return;
+
+       WARN_ON(i915->pmu.enable);
+
+       hrtimer_cancel(&i915->pmu.timer);
+
+       i915_pmu_unregister_cpuhp_state(i915);
+
+       perf_pmu_unregister(&i915->pmu.base);
+       i915->pmu.base.event_init = NULL;
+}
diff --git a/drivers/gpu/drm/i915/i915_pmu.h b/drivers/gpu/drm/i915/i915_pmu.h
new file mode 100644 (file)
index 0000000..1ac8b2e
--- /dev/null
@@ -0,0 +1,104 @@
+/*
+ * Copyright © 2017 Intel Corporation
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ *
+ */
+#ifndef __I915_PMU_H__
+#define __I915_PMU_H__
+
+enum {
+       __I915_SAMPLE_FREQ_ACT = 0,
+       __I915_SAMPLE_FREQ_REQ,
+       __I915_NUM_PMU_SAMPLERS
+};
+
+/**
+ * How many different events we track in the global PMU mask.
+ *
+ * It is also used to know to needed number of event reference counters.
+ */
+#define I915_PMU_MASK_BITS \
+       ((1 << I915_PMU_SAMPLE_BITS) + \
+        (I915_PMU_LAST + 1 - __I915_PMU_OTHER(0)))
+
+struct i915_pmu_sample {
+       u64 cur;
+       u32 prev;
+};
+
+struct i915_pmu {
+       /**
+        * @node: List node for CPU hotplug handling.
+        */
+       struct hlist_node node;
+       /**
+        * @base: PMU base.
+        */
+       struct pmu base;
+       /**
+        * @lock: Lock protecting enable mask and ref count handling.
+        */
+       spinlock_t lock;
+       /**
+        * @timer: Timer for internal i915 PMU sampling.
+        */
+       struct hrtimer timer;
+       /**
+        * @enable: Bitmask of all currently enabled events.
+        *
+        * Bits are derived from uAPI event numbers in a way that low 16 bits
+        * correspond to engine event _sample_ _type_ (I915_SAMPLE_QUEUED is
+        * bit 0), and higher bits correspond to other events (for instance
+        * I915_PMU_ACTUAL_FREQUENCY is bit 16 etc).
+        *
+        * In other words, low 16 bits are not per engine but per engine
+        * sampler type, while the upper bits are directly mapped to other
+        * event types.
+        */
+       u64 enable;
+       /**
+        * @enable_count: Reference counts for the enabled events.
+        *
+        * Array indices are mapped in the same way as bits in the @enable field
+        * and they are used to control sampling on/off when multiple clients
+        * are using the PMU API.
+        */
+       unsigned int enable_count[I915_PMU_MASK_BITS];
+       /**
+        * @sample: Current and previous (raw) counters for sampling events.
+        *
+        * These counters are updated from the i915 PMU sampling timer.
+        *
+        * Only global counters are held here, while the per-engine ones are in
+        * struct intel_engine_cs.
+        */
+       struct i915_pmu_sample sample[__I915_NUM_PMU_SAMPLERS];
+};
+
+#ifdef CONFIG_PERF_EVENTS
+void i915_pmu_register(struct drm_i915_private *i915);
+void i915_pmu_unregister(struct drm_i915_private *i915);
+#else
+static inline void i915_pmu_register(struct drm_i915_private *i915) {}
+static inline void i915_pmu_unregister(struct drm_i915_private *i915) {}
+#endif
+
+#endif
index 96c80fa..09bf043 100644 (file)
@@ -186,6 +186,9 @@ static inline bool i915_mmio_reg_valid(i915_reg_t reg)
 #define VIDEO_ENHANCEMENT_CLASS        2
 #define COPY_ENGINE_CLASS      3
 #define OTHER_CLASS            4
+#define MAX_ENGINE_CLASS       4
+
+#define MAX_ENGINE_INSTANCE    1
 
 /* PCI config space */
 
index 22c0950..a5a4942 100644 (file)
@@ -205,6 +205,15 @@ intel_engine_setup(struct drm_i915_private *dev_priv,
        GEM_BUG_ON(info->class >= ARRAY_SIZE(intel_engine_classes));
        class_info = &intel_engine_classes[info->class];
 
+       if (GEM_WARN_ON(info->class > MAX_ENGINE_CLASS))
+               return -EINVAL;
+
+       if (GEM_WARN_ON(info->instance > MAX_ENGINE_INSTANCE))
+               return -EINVAL;
+
+       if (GEM_WARN_ON(dev_priv->engine_class[info->class][info->instance]))
+               return -EINVAL;
+
        GEM_BUG_ON(dev_priv->engine[id]);
        engine = kzalloc(sizeof(*engine), GFP_KERNEL);
        if (!engine)
@@ -234,6 +243,7 @@ intel_engine_setup(struct drm_i915_private *dev_priv,
 
        ATOMIC_INIT_NOTIFIER_HEAD(&engine->context_status_notifier);
 
+       dev_priv->engine_class[info->class][info->instance] = engine;
        dev_priv->engine[id] = engine;
        return 0;
 }
@@ -1816,6 +1826,29 @@ void intel_engine_dump(struct intel_engine_cs *engine, struct drm_printer *m)
        drm_printf(m, "\n");
 }
 
+static u8 user_class_map[] = {
+       [I915_ENGINE_CLASS_RENDER] = RENDER_CLASS,
+       [I915_ENGINE_CLASS_COPY] = COPY_ENGINE_CLASS,
+       [I915_ENGINE_CLASS_VIDEO] = VIDEO_DECODE_CLASS,
+       [I915_ENGINE_CLASS_VIDEO_ENHANCE] = VIDEO_ENHANCEMENT_CLASS,
+};
+
+struct intel_engine_cs *
+intel_engine_lookup_user(struct drm_i915_private *i915, u8 class, u8 instance)
+{
+       if (class >= ARRAY_SIZE(user_class_map))
+               return NULL;
+
+       class = user_class_map[class];
+
+       GEM_BUG_ON(class > MAX_ENGINE_CLASS);
+
+       if (instance > MAX_ENGINE_INSTANCE)
+               return NULL;
+
+       return i915->engine_class[class][instance];
+}
+
 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
 #include "selftests/mock_engine.c"
 #endif
index 80cd781..7ee0f18 100644 (file)
@@ -5,6 +5,7 @@
 #include "i915_gem_batch_pool.h"
 #include "i915_gem_request.h"
 #include "i915_gem_timeline.h"
+#include "i915_pmu.h"
 #include "i915_selftest.h"
 
 struct drm_printer;
@@ -338,6 +339,28 @@ struct intel_engine_cs {
                I915_SELFTEST_DECLARE(bool mock : 1);
        } breadcrumbs;
 
+       struct {
+               /**
+                * @enable: Bitmask of enable sample events on this engine.
+                *
+                * Bits correspond to sample event types, for instance
+                * I915_SAMPLE_QUEUED is bit 0 etc.
+                */
+               u32 enable;
+               /**
+                * @enable_count: Reference count for the enabled samplers.
+                *
+                * Index number corresponds to the bit number from @enable.
+                */
+               unsigned int enable_count[I915_PMU_SAMPLE_BITS];
+               /**
+                * @sample: Counter values for sampling events.
+                *
+                * Our internal timer stores the current counters in this field.
+                */
+               struct i915_pmu_sample sample[I915_ENGINE_SAMPLE_MAX];
+       } pmu;
+
        /*
         * A pool of objects to use as shadow copies of client batch buffers
         * when the command parser is enabled. Prevents the client from
@@ -926,4 +949,7 @@ bool intel_engine_can_store_dword(struct intel_engine_cs *engine);
 
 void intel_engine_dump(struct intel_engine_cs *engine, struct drm_printer *p);
 
+struct intel_engine_cs *
+intel_engine_lookup_user(struct drm_i915_private *i915, u8 class, u8 instance);
+
 #endif /* _INTEL_RINGBUFFER_H_ */
index b579859..40e7b43 100644 (file)
@@ -102,6 +102,45 @@ enum drm_i915_gem_engine_class {
        I915_ENGINE_CLASS_INVALID       = -1
 };
 
+/**
+ * DOC: perf_events exposed by i915 through /sys/bus/event_sources/drivers/i915
+ *
+ */
+
+enum drm_i915_pmu_engine_sample {
+       I915_SAMPLE_BUSY = 0,
+       I915_SAMPLE_WAIT = 1,
+       I915_SAMPLE_SEMA = 2,
+       I915_ENGINE_SAMPLE_MAX /* non-ABI */
+};
+
+#define I915_PMU_SAMPLE_BITS (4)
+#define I915_PMU_SAMPLE_MASK (0xf)
+#define I915_PMU_SAMPLE_INSTANCE_BITS (8)
+#define I915_PMU_CLASS_SHIFT \
+       (I915_PMU_SAMPLE_BITS + I915_PMU_SAMPLE_INSTANCE_BITS)
+
+#define __I915_PMU_ENGINE(class, instance, sample) \
+       ((class) << I915_PMU_CLASS_SHIFT | \
+       (instance) << I915_PMU_SAMPLE_BITS | \
+       (sample))
+
+#define I915_PMU_ENGINE_BUSY(class, instance) \
+       __I915_PMU_ENGINE(class, instance, I915_SAMPLE_BUSY)
+
+#define I915_PMU_ENGINE_WAIT(class, instance) \
+       __I915_PMU_ENGINE(class, instance, I915_SAMPLE_WAIT)
+
+#define I915_PMU_ENGINE_SEMA(class, instance) \
+       __I915_PMU_ENGINE(class, instance, I915_SAMPLE_SEMA)
+
+#define __I915_PMU_OTHER(x) (__I915_PMU_ENGINE(0xff, 0xff, 0xf) + 1 + (x))
+
+#define I915_PMU_ACTUAL_FREQUENCY      __I915_PMU_OTHER(0)
+#define I915_PMU_REQUESTED_FREQUENCY   __I915_PMU_OTHER(1)
+
+#define I915_PMU_LAST I915_PMU_REQUESTED_FREQUENCY
+
 /* Each region is a minimum of 16k, and there are at most 255 of them.
  */
 #define I915_NR_TEX_REGIONS 255        /* table size 2k - maximum due to use