[Intel-gfx] [PATCH 36/36] drm/i915: Support per-context user requests for GPU frequency control
Joonas Lahtinen
joonas.lahtinen at linux.intel.com
Fri Nov 16 11:14:59 UTC 2018
Quoting Lionel Landwerlin (2018-11-09 19:51:17)
> I think we have some interest in reviving this for the performance query
> use case.
How are performance queries related?
Regards, Joonas
> Is that on anybody's todo list?
>
> Thanks,
>
> -
> Lionel
>
> On 14/03/2018 09:37, Chris Wilson wrote:
> > Often, we find ourselves facing a workload where the user knows in
> > advance what GPU frequency they require for it to complete in a timely
> > manner, and using past experience they can outperform the HW assisted
> > RPS autotuning. An example might be kodi (HTPC) where they know that
> > video decoding and compositing require a minimum frequency to avoid ever
> > dropping a frame, or conversely know when they are in a powersaving mode
> > and would rather have slower updates than ramp up the GPU frequency and
> > power consumption. Other workloads may defeat the autotuning entirely
> > and need manual control to meet their performance goals, e.g. bursty
> > applications which require low latency.
> >
> > To accommodate the varying needs of different applications, that may be
> > running concurrently, we want a more flexible system than a global limit
> > supplied by sysfs. To this end, we offer the application the option to
> > set their desired frequency bounds on the context itself, and apply those
> > bounds when we execute commands from the application, switching between
> > bounds just as easily as we switch between the clients themselves.
> >
> > The clients can query the range supported by the HW, or at least the
> > range they are restricted to, and then freely select frequencies within
> > that range that they want to run at. (They can select just a single
> > frequency if they so choose.) As this is subject to the global limit
> > supplied by the user in sysfs, and a client can only reduce the range of
> > frequencies they allow the HW to run at, we allow all clients to adjust
> > their request (and not restrict raising the minimum to privileged
> > CAP_SYS_NICE clients).
> >
> > Testcase: igt/gem_ctx_freq
> > Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> > Cc: Joonas Lahtinen <joonas.lahtinen at linux.intel.com>
> > Cc: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
> > Cc: Praveen Paneri <praveen.paneri at intel.com>
> > Cc: Sagar A Kamble <sagar.a.kamble at intel.com>
> > ---
> > drivers/gpu/drm/i915/i915_debugfs.c | 16 ++-
> > drivers/gpu/drm/i915/i915_drv.h | 5 +
> > drivers/gpu/drm/i915/i915_gem_context.c | 54 +++++++++
> > drivers/gpu/drm/i915/i915_gem_context.h | 3 +
> > drivers/gpu/drm/i915/intel_gt_pm.c | 121 ++++++++++++++++---
> > drivers/gpu/drm/i915/intel_gt_pm.h | 4 +
> > drivers/gpu/drm/i915/intel_guc_submission.c | 16 ++-
> > drivers/gpu/drm/i915/intel_lrc.c | 15 +++
> > .../gpu/drm/i915/selftests/i915_mock_selftests.h | 1 +
> > drivers/gpu/drm/i915/selftests/intel_gt_pm.c | 130 +++++++++++++++++++++
> > include/uapi/drm/i915_drm.h | 20 ++++
> > 11 files changed, 368 insertions(+), 17 deletions(-)
> > create mode 100644 drivers/gpu/drm/i915/selftests/intel_gt_pm.c
> >
> > diff --git a/drivers/gpu/drm/i915/i915_debugfs.c b/drivers/gpu/drm/i915/i915_debugfs.c
> > index 7c7afdac8c8c..a21b9164ade8 100644
> > --- a/drivers/gpu/drm/i915/i915_debugfs.c
> > +++ b/drivers/gpu/drm/i915/i915_debugfs.c
> > @@ -2191,6 +2191,7 @@ static int i915_rps_boost_info(struct seq_file *m, void *data)
> > struct drm_device *dev = &dev_priv->drm;
> > struct intel_rps *rps = &dev_priv->gt_pm.rps;
> > struct drm_file *file;
> > + int n;
> >
> > seq_printf(m, "GPU busy? %s [%d requests]\n",
> > yesno(dev_priv->gt.awake), dev_priv->gt.active_requests);
> > @@ -2198,17 +2199,30 @@ static int i915_rps_boost_info(struct seq_file *m, void *data)
> > seq_printf(m, "CPU waiting? %d\n", count_irq_waiters(dev_priv));
> > seq_printf(m, "Boosts outstanding? %d\n",
> > atomic_read(&rps->num_waiters));
> > + seq_printf(m, "Worker pending? %s\n", yesno(work_busy(&rps->work)));
> > seq_printf(m, "Frequency requested %d [%d, %d]\n",
> > intel_gpu_freq(dev_priv, rps->freq),
> > intel_gpu_freq(dev_priv, rps->min),
> > intel_gpu_freq(dev_priv, rps->max));
> > - seq_printf(m, " min hard:%d, soft:%d user:%d; max user:%d, soft: %d hard:%d\n",
> > + seq_printf(m, " min hard:%d, soft:%d, ctx:%d, user:%d; max user:%d, ctx:%d, soft:%d, hard:%d\n",
> > intel_gpu_freq(dev_priv, rps->min_freq_hw),
> > intel_gpu_freq(dev_priv, rps->min_freq_soft),
> > + intel_gpu_freq(dev_priv, rps->min_freq_context),
> > intel_gpu_freq(dev_priv, rps->min_freq_user),
> > intel_gpu_freq(dev_priv, rps->max_freq_user),
> > + intel_gpu_freq(dev_priv, rps->max_freq_context),
> > intel_gpu_freq(dev_priv, rps->max_freq_soft),
> > intel_gpu_freq(dev_priv, rps->max_freq_hw));
> > + seq_printf(m, " engines min: [");
> > + for (n = 0; n < ARRAY_SIZE(rps->min_freq_engine); n++)
> > + seq_printf(m, "%s%d", n ? ", " : "",
> > + intel_gpu_freq(dev_priv, rps->min_freq_engine[n]));
> > + seq_printf(m, "]\n engines max: [");
> > + for (n = 0; n < ARRAY_SIZE(rps->max_freq_engine); n++)
> > + seq_printf(m, "%s%d", n ? ", " : "",
> > + intel_gpu_freq(dev_priv, rps->max_freq_engine[n]));
> > + seq_printf(m, "]\n");
> > +
> > seq_printf(m, " idle:%d, efficient:%d, boost:%d\n",
> > intel_gpu_freq(dev_priv, rps->idle_freq),
> > intel_gpu_freq(dev_priv, rps->efficient_freq),
> > diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h
> > index 82e9a58bd65f..d754d44cfbc2 100644
> > --- a/drivers/gpu/drm/i915/i915_drv.h
> > +++ b/drivers/gpu/drm/i915/i915_drv.h
> > @@ -731,6 +731,7 @@ struct intel_rps_ei {
> >
> > struct intel_rps {
> > struct mutex lock;
> > + spinlock_t engine_lock; /* protects updates to min/max_freq_context */
> > struct work_struct work;
> >
> > bool active;
> > @@ -763,6 +764,10 @@ struct intel_rps {
> > u8 max_freq_user; /* Max frequency permitted by the driver */
> > u8 min_freq_soft;
> > u8 max_freq_soft;
> > + u8 min_freq_context; /* Min frequency permitted by the context */
> > + u8 max_freq_context; /* Max frequency permitted by the context */
> > + u8 min_freq_engine[I915_NUM_ENGINES];
> > + u8 max_freq_engine[I915_NUM_ENGINES];
> >
> > u8 idle_freq; /* Frequency to request when we are idle */
> > u8 efficient_freq; /* AKA RPe. Pre-determined balanced frequency */
> > diff --git a/drivers/gpu/drm/i915/i915_gem_context.c b/drivers/gpu/drm/i915/i915_gem_context.c
> > index 65bf92658d92..1d36e2a02479 100644
> > --- a/drivers/gpu/drm/i915/i915_gem_context.c
> > +++ b/drivers/gpu/drm/i915/i915_gem_context.c
> > @@ -88,8 +88,10 @@
> > #include <linux/log2.h>
> > #include <drm/drmP.h>
> > #include <drm/i915_drm.h>
> > +
> > #include "i915_drv.h"
> > #include "i915_trace.h"
> > +#include "intel_gt_pm.h"
> >
> > #define ALL_L3_SLICES(dev) (1 << NUM_L3_SLICES(dev)) - 1
> >
> > @@ -281,6 +283,8 @@ __create_hw_context(struct drm_i915_private *dev_priv,
> > list_add_tail(&ctx->link, &dev_priv->contexts.list);
> > ctx->i915 = dev_priv;
> > ctx->priority = I915_PRIORITY_NORMAL;
> > + ctx->min_freq = dev_priv->gt_pm.rps.min_freq_hw;
> > + ctx->max_freq = dev_priv->gt_pm.rps.max_freq_hw;
> >
> > INIT_RADIX_TREE(&ctx->handles_vma, GFP_KERNEL);
> > INIT_LIST_HEAD(&ctx->handles_list);
> > @@ -715,6 +719,7 @@ int i915_gem_context_getparam_ioctl(struct drm_device *dev, void *data,
> > {
> > struct drm_i915_file_private *file_priv = file->driver_priv;
> > struct drm_i915_gem_context_param *args = data;
> > + struct drm_i915_private *i915 = to_i915(dev);
> > struct i915_gem_context *ctx;
> > int ret = 0;
> >
> > @@ -747,6 +752,19 @@ int i915_gem_context_getparam_ioctl(struct drm_device *dev, void *data,
> > case I915_CONTEXT_PARAM_PRIORITY:
> > args->value = ctx->priority;
> > break;
> > + case I915_CONTEXT_PARAM_FREQUENCY:
> > + if (!HAS_RPS(i915) || !HAS_EXECLISTS(i915)) {
> > + ret = -ENODEV;
> > + } else if (args->size) {
> > + ret = -EINVAL;
> > + } else {
> > + u32 min = intel_gpu_freq(i915, ctx->min_freq);
> > + u32 max = intel_gpu_freq(i915, ctx->max_freq);
> > +
> > + args->value = I915_CONTEXT_SET_FREQUENCY(min, max);
> > + }
> > + break;
> > +
> > default:
> > ret = -EINVAL;
> > break;
> > @@ -761,6 +779,7 @@ int i915_gem_context_setparam_ioctl(struct drm_device *dev, void *data,
> > {
> > struct drm_i915_file_private *file_priv = file->driver_priv;
> > struct drm_i915_gem_context_param *args = data;
> > + struct drm_i915_private *i915 = to_i915(dev);
> > struct i915_gem_context *ctx;
> > int ret;
> >
> > @@ -821,6 +840,41 @@ int i915_gem_context_setparam_ioctl(struct drm_device *dev, void *data,
> > ctx->priority = priority;
> > }
> > break;
> > + case I915_CONTEXT_PARAM_FREQUENCY:
> > + if (!HAS_RPS(i915) || !HAS_EXECLISTS(i915)) {
> > + ret = -ENODEV;
> > + } else if (args->size) {
> > + ret = -EINVAL;
> > + } else {
> > + struct intel_rps *rps = &i915->gt_pm.rps;
> > + u32 min, max;
> > +
> > + min = I915_CONTEXT_MIN_FREQUENCY(args->value);
> > + min = intel_freq_opcode(i915, min);
> > +
> > + max = I915_CONTEXT_MAX_FREQUENCY(args->value);
> > + max = intel_freq_opcode(i915, max);
> > +
> > + /*
> > + * As we constrain the frequency request from the
> > + * context (application) by the sysadmin imposed limits,
> > + * it is reasonable to allow the application to
> > + * specify its preferred range within those limits.
> > + * That is we do not need to restrict requesting
> > + * a higher frequency to privileged (CAP_SYS_NICE)
> > + * processes.
> > + */
> > + if (max < min) {
> > + ret = -EINVAL;
> > + } else if (min < rps->min_freq_hw ||
> > + max > rps->max_freq_hw) {
> > + ret = -EINVAL;
> > + } else {
> > + ctx->min_freq = min;
> > + ctx->max_freq = max;
> > + }
> > + }
> > + break;
> >
> > default:
> > ret = -EINVAL;
> > diff --git a/drivers/gpu/drm/i915/i915_gem_context.h b/drivers/gpu/drm/i915/i915_gem_context.h
> > index 7854262ddfd9..98f7b71a787a 100644
> > --- a/drivers/gpu/drm/i915/i915_gem_context.h
> > +++ b/drivers/gpu/drm/i915/i915_gem_context.h
> > @@ -150,6 +150,9 @@ struct i915_gem_context {
> > */
> > int priority;
> >
> > + u32 min_freq;
> > + u32 max_freq;
> > +
> > /** ggtt_offset_bias: placement restriction for context objects */
> > u32 ggtt_offset_bias;
> >
> > diff --git a/drivers/gpu/drm/i915/intel_gt_pm.c b/drivers/gpu/drm/i915/intel_gt_pm.c
> > index 9705205a26b5..4bbfb4080f8f 100644
> > --- a/drivers/gpu/drm/i915/intel_gt_pm.c
> > +++ b/drivers/gpu/drm/i915/intel_gt_pm.c
> > @@ -402,10 +402,10 @@ static int adjust_rps(struct drm_i915_private *dev_priv, int freq, int adj)
> > GEM_BUG_ON(!rps->active);
> >
> > min = clamp_t(int,
> > - rps->min_freq_soft,
> > + max(rps->min_freq_soft, READ_ONCE(rps->min_freq_context)),
> > rps->min_freq_user, rps->max_freq_user);
> > max = clamp_t(int,
> > - rps->max_freq_soft,
> > + min(rps->max_freq_soft, READ_ONCE(rps->max_freq_context)),
> > min, rps->max_freq_user);
> > if (atomic_read(&rps->num_waiters) && max < rps->boost_freq)
> > max = rps->boost_freq;
> > @@ -809,6 +809,75 @@ void intel_rps_boost(struct i915_request *rq, struct intel_rps_client *client)
> > atomic_inc(client ? &client->boosts : &rps->boosts);
> > }
> >
> > +static void __rps_update_engine(struct intel_rps *rps,
> > + enum intel_engine_id idx,
> > + u32 min, u32 max)
> > +{
> > + unsigned long flags;
> > + bool update = false;
> > + u32 old;
> > + int n;
> > +
> > + GEM_BUG_ON(min > max);
> > +
> > + if (rps->min_freq_engine[idx] != min) {
> > + spin_lock_irqsave(&rps->engine_lock, flags);
> > +
> > + rps->min_freq_engine[idx] = min;
> > +
> > + old = rps->min_freq_context;
> > + rps->min_freq_context = rps->min_freq_engine[0];
> > + for (n = 1; n < ARRAY_SIZE(rps->min_freq_engine); n++)
> > + if (rps->min_freq_engine[n] > rps->min_freq_context)
> > + rps->min_freq_context = rps->min_freq_engine[n];
> > + update |= rps->min_freq_context != old;
> > +
> > + spin_unlock_irqrestore(&rps->engine_lock, flags);
> > + }
> > +
> > + if (rps->max_freq_engine[idx] != max) {
> > + spin_lock_irqsave(&rps->engine_lock, flags);
> > +
> > + rps->max_freq_engine[idx] = max;
> > +
> > + old = rps->max_freq_context;
> > + rps->max_freq_context = rps->max_freq_engine[0];
> > + for (n = 1; n < ARRAY_SIZE(rps->max_freq_engine); n++)
> > + if (rps->max_freq_engine[n] < rps->max_freq_context)
> > + rps->max_freq_context = rps->max_freq_engine[n];
> > + update |= rps->max_freq_context != old;
> > +
> > + spin_unlock_irqrestore(&rps->engine_lock, flags);
> > + }
> > +
> > + /* Kick the RPS worker to apply the updated constraints, as needed */
> > + if (update && !atomic_read(&rps->num_waiters)) {
> > + old = READ_ONCE(rps->freq);
> > + if ((old < min || old > max))
> > + schedule_work(&rps->work);
> > + }
> > +}
> > +
> > +void intel_rps_update_engine(const struct intel_engine_cs *engine,
> > + const struct i915_gem_context *ctx)
> > +{
> > + struct intel_rps *rps = &engine->i915->gt_pm.rps;
> > + u32 min, max;
> > +
> > + if (!HAS_RPS(engine->i915))
> > + return;
> > +
> > + if (ctx) {
> > + min = ctx->min_freq;
> > + max = ctx->max_freq;
> > + } else {
> > + min = rps->min_freq_hw;
> > + max = rps->max_freq_hw;
> > + }
> > +
> > + __rps_update_engine(rps, engine->id, min, max);
> > +}
> > +
> > static void gen9_disable_rc6(struct drm_i915_private *dev_priv)
> > {
> > I915_WRITE(GEN6_RC_CONTROL, 0);
> > @@ -2379,12 +2448,41 @@ void intel_gt_pm_sanitize(struct drm_i915_private *dev_priv)
> > gen6_reset_rps_interrupts(dev_priv);
> > }
> >
> > +static void intel_rps_init(struct intel_rps *rps)
> > +{
> > + mutex_init(&rps->lock);
> > + INIT_WORK(&rps->work, intel_rps_work);
> > + spin_lock_init(&rps->engine_lock);
> > +}
> > +
> > +static void intel_rps_init__frequencies(struct intel_rps *rps)
> > +{
> > + int n;
> > +
> > + rps->max_freq_soft = rps->max_freq_hw;
> > + rps->min_freq_soft = rps->min_freq_hw;
> > +
> > + rps->max_freq_context = rps->max_freq_hw;
> > + rps->min_freq_context = rps->min_freq_hw;
> > + for (n = 0; n < ARRAY_SIZE(rps->max_freq_engine); n++) {
> > + rps->max_freq_engine[n] = rps->max_freq_hw;
> > + rps->min_freq_engine[n] = rps->min_freq_hw;
> > + }
> > +
> > + /* Finally allow us to boost to max by default */
> > + rps->boost_freq = rps->max_freq_hw;
> > + rps->idle_freq = rps->min_freq_hw;
> > +
> > + rps->freq = rps->idle_freq;
> > + rps->min = rps->min_freq_hw;
> > + rps->max = rps->max_freq_hw;
> > +}
> > +
> > void intel_gt_pm_init(struct drm_i915_private *dev_priv)
> > {
> > struct intel_rps *rps = &dev_priv->gt_pm.rps;
> >
> > - mutex_init(&rps->lock);
> > - INIT_WORK(&rps->work, intel_rps_work);
> > + intel_rps_init(rps);
> >
> > if (HAS_GUC_SCHED(dev_priv))
> > rps->guc_events = GEN9_GUC_TO_HOST_INT_EVENT;
> > @@ -2449,16 +2547,7 @@ void intel_gt_pm_init(struct drm_i915_private *dev_priv)
> > }
> > }
> >
> > - rps->max_freq_soft = rps->max_freq_hw;
> > - rps->min_freq_soft = rps->min_freq_hw;
> > -
> > - /* Finally allow us to boost to max by default */
> > - rps->boost_freq = rps->max_freq_hw;
> > - rps->idle_freq = rps->min_freq_hw;
> > -
> > - rps->freq = rps->idle_freq;
> > - rps->min = rps->min_freq_hw;
> > - rps->max = rps->max_freq_hw;
> > + intel_rps_init__frequencies(rps);
> >
> > if (HAS_LLC(dev_priv))
> > gen6_update_ring_freq(dev_priv);
> > @@ -2703,3 +2792,7 @@ void gen9_disable_guc_interrupts(struct drm_i915_private *dev_priv)
> >
> > gen9_reset_guc_interrupts(dev_priv);
> > }
> > +
> > +#if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
> > +#include "selftests/intel_gt_pm.c"
> > +#endif
> > diff --git a/drivers/gpu/drm/i915/intel_gt_pm.h b/drivers/gpu/drm/i915/intel_gt_pm.h
> > index 314912c15126..ef3f27eca529 100644
> > --- a/drivers/gpu/drm/i915/intel_gt_pm.h
> > +++ b/drivers/gpu/drm/i915/intel_gt_pm.h
> > @@ -25,7 +25,9 @@
> > #define __INTEL_GT_PM_H__
> >
> > struct drm_i915_private;
> > +struct i915_gem_context;
> > struct i915_request;
> > +struct intel_engine_cs;
> > struct intel_rps_client;
> >
> > void intel_gpu_ips_init(struct drm_i915_private *dev_priv);
> > @@ -47,6 +49,8 @@ void intel_gt_pm_idle(struct drm_i915_private *dev_priv);
> >
> > void intel_gt_pm_irq_handler(struct drm_i915_private *dev_priv, u32 pm_iir);
> >
> > +void intel_rps_update_engine(const struct intel_engine_cs *engine,
> > + const struct i915_gem_context *ctx);
> > void intel_rps_boost(struct i915_request *rq, struct intel_rps_client *rps);
> >
> > int intel_gpu_freq(struct drm_i915_private *dev_priv, int val);
> > diff --git a/drivers/gpu/drm/i915/intel_guc_submission.c b/drivers/gpu/drm/i915/intel_guc_submission.c
> > index 8a8ad2fe158d..d8eaae683186 100644
> > --- a/drivers/gpu/drm/i915/intel_guc_submission.c
> > +++ b/drivers/gpu/drm/i915/intel_guc_submission.c
> > @@ -26,9 +26,12 @@
> > #include <trace/events/dma_fence.h>
> >
> > #include "intel_guc_submission.h"
> > -#include "intel_lrc_reg.h"
> > +
> > #include "i915_drv.h"
> >
> > +#include "intel_gt_pm.h"
> > +#include "intel_lrc_reg.h"
> > +
> > #define GUC_PREEMPT_FINISHED 0x1
> > #define GUC_PREEMPT_BREADCRUMB_DWORDS 0x8
> > #define GUC_PREEMPT_BREADCRUMB_BYTES \
> > @@ -650,6 +653,12 @@ static void guc_submit(struct intel_engine_cs *engine)
> > }
> > }
> >
> > +static void update_rps(struct intel_engine_cs *engine)
> > +{
> > + intel_rps_update_engine(engine,
> > + port_request(engine->execlists.port)->ctx);
> > +}
> > +
> > static void port_assign(struct execlist_port *port, struct i915_request *rq)
> > {
> > GEM_BUG_ON(port_isset(port));
> > @@ -728,6 +737,7 @@ static void guc_dequeue(struct intel_engine_cs *engine)
> > execlists->first = rb;
> > if (submit) {
> > port_assign(port, last);
> > + update_rps(engine);
> > execlists_set_active(execlists, EXECLISTS_ACTIVE_USER);
> > guc_submit(engine);
> > }
> > @@ -757,8 +767,10 @@ static void guc_submission_tasklet(unsigned long data)
> >
> > rq = port_request(&port[0]);
> > }
> > - if (!rq)
> > + if (!rq) {
> > execlists_clear_active(execlists, EXECLISTS_ACTIVE_USER);
> > + intel_rps_update_engine(engine, NULL);
> > + }
> >
> > if (execlists_is_active(execlists, EXECLISTS_ACTIVE_PREEMPT) &&
> > intel_read_status_page(engine, I915_GEM_HWS_PREEMPT_INDEX) ==
> > diff --git a/drivers/gpu/drm/i915/intel_lrc.c b/drivers/gpu/drm/i915/intel_lrc.c
> > index 3a69b367e565..518f7b3db857 100644
> > --- a/drivers/gpu/drm/i915/intel_lrc.c
> > +++ b/drivers/gpu/drm/i915/intel_lrc.c
> > @@ -138,6 +138,7 @@
> > #include "i915_drv.h"
> > #include "i915_gem_render_state.h"
> > #include "intel_lrc_reg.h"
> > +#include "intel_gt_pm.h"
> > #include "intel_mocs.h"
> >
> > #define RING_EXECLIST_QFULL (1 << 0x2)
> > @@ -535,6 +536,12 @@ static void inject_preempt_context(struct intel_engine_cs *engine)
> > execlists_set_active(&engine->execlists, EXECLISTS_ACTIVE_PREEMPT);
> > }
> >
> > +static void update_rps(struct intel_engine_cs *engine)
> > +{
> > + intel_rps_update_engine(engine,
> > + port_request(engine->execlists.port)->ctx);
> > +}
> > +
> > static void execlists_dequeue(struct intel_engine_cs *engine)
> > {
> > struct intel_engine_execlists * const execlists = &engine->execlists;
> > @@ -708,6 +715,7 @@ static void execlists_dequeue(struct intel_engine_cs *engine)
> > spin_unlock_irq(&engine->timeline->lock);
> >
> > if (submit) {
> > + update_rps(engine);
> > execlists_set_active(execlists, EXECLISTS_ACTIVE_USER);
> > execlists_submit_ports(engine);
> > }
> > @@ -982,6 +990,11 @@ static void execlists_submission_tasklet(unsigned long data)
> > engine->name, port->context_id);
> >
> > execlists_port_complete(execlists, port);
> > +
> > + /* Switch to the next request/context */
> > + rq = port_request(port);
> > + intel_rps_update_engine(engine,
> > + rq ? rq->ctx : NULL);
> > } else {
> > port_set(port, port_pack(rq, count));
> > }
> > @@ -1717,6 +1730,8 @@ static void reset_common_ring(struct intel_engine_cs *engine,
> > __unwind_incomplete_requests(engine);
> > spin_unlock(&engine->timeline->lock);
> >
> > + intel_rps_update_engine(engine, NULL);
> > +
> > /* Mark all CS interrupts as complete */
> > execlists->active = 0;
> >
> > diff --git a/drivers/gpu/drm/i915/selftests/i915_mock_selftests.h b/drivers/gpu/drm/i915/selftests/i915_mock_selftests.h
> > index 9a48aa441743..85b6e6d020b7 100644
> > --- a/drivers/gpu/drm/i915/selftests/i915_mock_selftests.h
> > +++ b/drivers/gpu/drm/i915/selftests/i915_mock_selftests.h
> > @@ -14,6 +14,7 @@ selftest(fence, i915_sw_fence_mock_selftests)
> > selftest(scatterlist, scatterlist_mock_selftests)
> > selftest(syncmap, i915_syncmap_mock_selftests)
> > selftest(uncore, intel_uncore_mock_selftests)
> > +selftest(gt_pm, intel_gt_pm_mock_selftests)
> > selftest(breadcrumbs, intel_breadcrumbs_mock_selftests)
> > selftest(timelines, i915_gem_timeline_mock_selftests)
> > selftest(requests, i915_request_mock_selftests)
> > diff --git a/drivers/gpu/drm/i915/selftests/intel_gt_pm.c b/drivers/gpu/drm/i915/selftests/intel_gt_pm.c
> > new file mode 100644
> > index 000000000000..c3871eb9eabb
> > --- /dev/null
> > +++ b/drivers/gpu/drm/i915/selftests/intel_gt_pm.c
> > @@ -0,0 +1,130 @@
> > +/*
> > + * SPDX-License-Identifier: MIT
> > + *
> > + * Copyright © 2018 Intel Corporation
> > + */
> > +
> > +#include "../i915_selftest.h"
> > +#include "i915_random.h"
> > +
> > +#include "mock_gem_device.h"
> > +
> > +static void mock_rps_init(struct drm_i915_private *i915)
> > +{
> > + struct intel_rps *rps = &i915->gt_pm.rps;
> > +
> > + /* Disable the register writes */
> > + mkwrite_device_info(i915)->gen = 0;
> > + mkwrite_device_info(i915)->has_rps = true;
> > +
> > + intel_rps_init(rps);
> > +
> > + rps->min_freq_hw = 0;
> > + rps->max_freq_hw = 255;
> > +
> > + rps->min_freq_user = rps->min_freq_hw;
> > + rps->max_freq_user = rps->max_freq_hw;
> > +
> > + intel_rps_init__frequencies(rps);
> > +}
> > +
> > +static void mock_rps_fini(struct drm_i915_private *i915)
> > +{
> > + struct intel_rps *rps = &i915->gt_pm.rps;
> > +
> > + cancel_work_sync(&rps->work);
> > +}
> > +
> > +static int igt_rps_engine(void *arg)
> > +{
> > + struct drm_i915_private *i915 = arg;
> > + struct intel_rps *rps = &i915->gt_pm.rps;
> > + I915_RND_STATE(prng);
> > + int err;
> > + int i;
> > +
> > + intel_gt_pm_busy(i915); /* Activate RPS */
> > +
> > + /*
> > + * Minimum unit tests for intel_rps_update_engine().
> > + *
> > + * Whenever we call intel_rps_update_engine, it will
> > + * replace the context min/max frequency request for a particular
> > + * engine and then recompute the global max(min)/min(max) over all
> > + * engines. In this mockup, we are limited to checking those
> > + * max(min)/min(max) calculations and then seeing if the rps
> > + * worker uses those bounds.
> > + */
> > +
> > + for (i = 0; i < 256 * 256; i++) {
> > + u8 freq = prandom_u32_state(&prng);
> > +
> > + __rps_update_engine(rps, 0, freq, freq);
> > + if (rps->min_freq_context != freq ||
> > + rps->max_freq_context != freq) {
> > + pr_err("Context min/max frequencies not restricted to %d, found [%d, %d]\n",
> > + freq, rps->min_freq_context, rps->max_freq_context);
> > + err = -EINVAL;
> > + goto out;
> > + }
> > + flush_work(&rps->work);
> > +
> > + if (rps->freq != freq) {
> > + pr_err("Tried to restrict frequency to %d, found %d\n",
> > + freq, rps->freq);
> > + err = -EINVAL;
> > + goto out;
> > + }
> > + }
> > +
> > + __rps_update_engine(rps, 0, rps->min_freq_hw, rps->max_freq_hw);
> > + if (rps->min_freq_context != rps->min_freq_hw ||
> > + rps->max_freq_context != rps->max_freq_hw) {
> > + pr_err("Context frequency not restored to [%d, %d], found [%d, %d]\n",
> > + rps->min_freq_hw, rps->min_freq_hw,
> > + rps->min_freq_context, rps->max_freq_context);
> > + err = -EINVAL;
> > + goto out;
> > + }
> > +
> > + for (i = 0; i < I915_NUM_ENGINES; i++)
> > + __rps_update_engine(rps, i, i, 255 - i);
> > + i--;
> > + if (rps->min_freq_context != i) {
> > + pr_err("Minimum context frequency across all engines not raised to %d, found %d\n", i, rps->min_freq_context);
> > + err = -EINVAL;
> > + goto out;
> > + }
> > + if (rps->max_freq_context != 255 - i) {
> > + pr_err("Maxmimum context frequency across all engines not lowered to %d, found %d\n", 255 - i, rps->max_freq_context);
> > + err = -EINVAL;
> > + goto out;
> > + }
> > +
> > + err = 0;
> > +out:
> > + intel_gt_pm_idle(i915);
> > + return err;
> > +}
> > +
> > +int intel_gt_pm_mock_selftests(void)
> > +{
> > + static const struct i915_subtest tests[] = {
> > + SUBTEST(igt_rps_engine),
> > + };
> > + struct drm_i915_private *i915;
> > + int err;
> > +
> > + i915 = mock_gem_device();
> > + if (!i915)
> > + return -ENOMEM;
> > +
> > + mock_rps_init(i915);
> > +
> > + err = i915_subtests(tests, i915);
> > +
> > + mock_rps_fini(i915);
> > + drm_dev_unref(&i915->drm);
> > +
> > + return err;
> > +}
> > diff --git a/include/uapi/drm/i915_drm.h b/include/uapi/drm/i915_drm.h
> > index 7f5634ce8e88..64c6377df769 100644
> > --- a/include/uapi/drm/i915_drm.h
> > +++ b/include/uapi/drm/i915_drm.h
> > @@ -1456,6 +1456,26 @@ struct drm_i915_gem_context_param {
> > #define I915_CONTEXT_MAX_USER_PRIORITY 1023 /* inclusive */
> > #define I915_CONTEXT_DEFAULT_PRIORITY 0
> > #define I915_CONTEXT_MIN_USER_PRIORITY -1023 /* inclusive */
> > +
> > +/*
> > + * I915_CONTEXT_PARAM_FREQUENCY:
> > + *
> > + * Request that when this context runs, the GPU is restricted to run
> > + * in this frequency range; but still contrained by the global user
> > + * restriction specified via sysfs.
> > + *
> > + * The minimum / maximum frequencies are specified in MHz. Each context
> > + * starts in the default unrestricted state, where the range is taken from
> > + * the hardware, and so may be queried.
> > + *
> > + * Note the frequency is only changed on a context switch; if the
> > + * context's frequency is updated whilst the context is currently executing
> > + * the request will not take effect until the next time the context is run.
> > + */
> > +#define I915_CONTEXT_PARAM_FREQUENCY 0x7
> > +#define I915_CONTEXT_MIN_FREQUENCY(x) ((x) & 0xffffffff)
> > +#define I915_CONTEXT_MAX_FREQUENCY(x) ((x) >> 32)
> > +#define I915_CONTEXT_SET_FREQUENCY(min, max) ((__u64)(max) << 32 | (min))
> > __u64 value;
> > };
> >
>
>
> _______________________________________________
> Intel-gfx mailing list
> Intel-gfx at lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/intel-gfx
More information about the Intel-gfx
mailing list