drm/xe: Add engine activity support
authorRiana Tauro <riana.tauro@intel.com>
Mon, 24 Feb 2025 05:38:58 +0000 (11:08 +0530)
committerLucas De Marchi <lucas.demarchi@intel.com>
Mon, 24 Feb 2025 20:32:09 +0000 (12:32 -0800)
GuC provides support to read engine counters to calculate the
engine activity. KMD exposes two counters via the PMU interface to
calculate engine activity

Engine Active Ticks(engine-active-ticks) - active ticks of engine
Engine Total Ticks (engine-total-ticks) - total ticks of engine

Engine activity percentage can be calculated as below
Engine activity % = (engine active ticks/engine total ticks) * 100.

v2: fix cosmetic review comments
    add forcewake for gpm_ts (Umesh)

v3: fix CI hooks error
    change function parameters and unpin bo on error
    of allocate_activity_buffers
    fix kernel-doc (Umesh)
    use engine activity (Umesh, Lucas)
    rename xe_engine_activity to xe_guc_engine_*
    fix commit message to use engine activity (Lucas, Umesh)

v4: add forcewake in PMU layer

v5: fix makefile
    use drmm_kcalloc instead of kmalloc_array
    remove managed bo
    skip init for VF
    fix cosmetic review comments (Michal)

Signed-off-by: Riana Tauro <riana.tauro@intel.com>
Reviewed-by: Umesh Nerlige Ramappa <umesh.nerlige.ramappa@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20250224053903.2253539-2-riana.tauro@intel.com
Signed-off-by: Lucas De Marchi <lucas.demarchi@intel.com>
drivers/gpu/drm/xe/Makefile
drivers/gpu/drm/xe/abi/guc_actions_abi.h
drivers/gpu/drm/xe/regs/xe_gt_regs.h
drivers/gpu/drm/xe/xe_guc_engine_activity.c [new file with mode: 0644]
drivers/gpu/drm/xe/xe_guc_engine_activity.h [new file with mode: 0644]
drivers/gpu/drm/xe/xe_guc_engine_activity_types.h [new file with mode: 0644]
drivers/gpu/drm/xe/xe_guc_fwif.h
drivers/gpu/drm/xe/xe_guc_types.h

index be73362ef334ed131df0f0c5f2db665956451dfc..ffc836fa8e6010230d49891f3a4892f3f887e97c 100644 (file)
@@ -60,6 +60,7 @@ xe-y += xe_bb.o \
        xe_guc_capture.o \
        xe_guc_ct.o \
        xe_guc_db_mgr.o \
+       xe_guc_engine_activity.o \
        xe_guc_hwconfig.o \
        xe_guc_id_mgr.o \
        xe_guc_klv_helpers.o \
index fee385532fb0b367e2c24c214f5984da05ad40f4..ec516e838ee89ffd51f481fca6452376d87aae5f 100644 (file)
@@ -140,6 +140,7 @@ enum xe_guc_action {
        XE_GUC_ACTION_REGISTER_CONTEXT_MULTI_LRC = 0x4601,
        XE_GUC_ACTION_CLIENT_SOFT_RESET = 0x5507,
        XE_GUC_ACTION_SET_ENG_UTIL_BUFF = 0x550A,
+       XE_GUC_ACTION_SET_DEVICE_ENGINE_ACTIVITY_BUFFER = 0x550C,
        XE_GUC_ACTION_NOTIFY_MEMORY_CAT_ERROR = 0x6000,
        XE_GUC_ACTION_REPORT_PAGE_FAULT_REQ_DESC = 0x6002,
        XE_GUC_ACTION_PAGE_FAULT_RES_DESC = 0x6003,
index 096859072396d6697fb93cce2f4c4dc10440d6d5..124cc398798e56d66f436f64c55bada74b6a8cde 100644 (file)
 #define   RENDER_AWAKE_STATUS                  REG_BIT(1)
 #define   MEDIA_SLICE0_AWAKE_STATUS            REG_BIT(0)
 
+#define MISC_STATUS_0                          XE_REG(0xa500)
+
 #define FORCEWAKE_MEDIA_VDBOX(n)               XE_REG(0xa540 + (n) * 4)
 #define FORCEWAKE_MEDIA_VEBOX(n)               XE_REG(0xa560 + (n) * 4)
 #define FORCEWAKE_GSC                          XE_REG(0xa618)
diff --git a/drivers/gpu/drm/xe/xe_guc_engine_activity.c b/drivers/gpu/drm/xe/xe_guc_engine_activity.c
new file mode 100644 (file)
index 0000000..255e63d
--- /dev/null
@@ -0,0 +1,321 @@
+// SPDX-License-Identifier: MIT
+/*
+ * Copyright © 2025 Intel Corporation
+ */
+
+#include <drm/drm_managed.h>
+
+#include "abi/guc_actions_abi.h"
+#include "regs/xe_gt_regs.h"
+
+#include "xe_bo.h"
+#include "xe_force_wake.h"
+#include "xe_gt_printk.h"
+#include "xe_guc.h"
+#include "xe_guc_engine_activity.h"
+#include "xe_guc_ct.h"
+#include "xe_hw_engine.h"
+#include "xe_map.h"
+#include "xe_mmio.h"
+
+#define TOTAL_QUANTA 0x8000
+
+static struct iosys_map engine_activity_map(struct xe_guc *guc, struct xe_hw_engine *hwe)
+{
+       struct xe_guc_engine_activity *engine_activity = &guc->engine_activity;
+       struct engine_activity_buffer *buffer = &engine_activity->device_buffer;
+       u16 guc_class = xe_engine_class_to_guc_class(hwe->class);
+       size_t offset;
+
+       offset = offsetof(struct guc_engine_activity_data,
+                         engine_activity[guc_class][hwe->logical_instance]);
+
+       return IOSYS_MAP_INIT_OFFSET(&buffer->activity_bo->vmap, offset);
+}
+
+static struct iosys_map engine_metadata_map(struct xe_guc *guc)
+{
+       struct xe_guc_engine_activity *engine_activity = &guc->engine_activity;
+       struct engine_activity_buffer *buffer = &engine_activity->device_buffer;
+
+       return buffer->metadata_bo->vmap;
+}
+
+static int allocate_engine_activity_group(struct xe_guc *guc)
+{
+       struct xe_guc_engine_activity *engine_activity = &guc->engine_activity;
+       struct xe_device *xe = guc_to_xe(guc);
+       u32 num_activity_group = 1; /* Will be modified for VF */
+
+       engine_activity->eag  = drmm_kcalloc(&xe->drm, num_activity_group,
+                                            sizeof(struct engine_activity_group), GFP_KERNEL);
+
+       if (!engine_activity->eag)
+               return -ENOMEM;
+
+       engine_activity->num_activity_group = num_activity_group;
+
+       return 0;
+}
+
+static int allocate_engine_activity_buffers(struct xe_guc *guc,
+                                           struct engine_activity_buffer *buffer)
+{
+       u32 metadata_size = sizeof(struct guc_engine_activity_metadata);
+       u32 size = sizeof(struct guc_engine_activity_data);
+       struct xe_gt *gt = guc_to_gt(guc);
+       struct xe_tile *tile = gt_to_tile(gt);
+       struct xe_bo *bo, *metadata_bo;
+
+       metadata_bo = xe_bo_create_pin_map(gt_to_xe(gt), tile, NULL, PAGE_ALIGN(metadata_size),
+                                          ttm_bo_type_kernel, XE_BO_FLAG_SYSTEM |
+                                          XE_BO_FLAG_GGTT | XE_BO_FLAG_GGTT_INVALIDATE);
+
+       if (IS_ERR(metadata_bo))
+               return PTR_ERR(metadata_bo);
+
+       bo = xe_bo_create_pin_map(gt_to_xe(gt), tile, NULL, PAGE_ALIGN(size),
+                                 ttm_bo_type_kernel, XE_BO_FLAG_VRAM_IF_DGFX(tile) |
+                                 XE_BO_FLAG_GGTT | XE_BO_FLAG_GGTT_INVALIDATE);
+
+       if (IS_ERR(bo)) {
+               xe_bo_unpin_map_no_vm(metadata_bo);
+               return PTR_ERR(bo);
+       }
+
+       buffer->metadata_bo = metadata_bo;
+       buffer->activity_bo = bo;
+       return 0;
+}
+
+static void free_engine_activity_buffers(struct engine_activity_buffer *buffer)
+{
+       xe_bo_unpin_map_no_vm(buffer->metadata_bo);
+       xe_bo_unpin_map_no_vm(buffer->activity_bo);
+}
+
+static struct engine_activity *hw_engine_to_engine_activity(struct xe_hw_engine *hwe)
+{
+       struct xe_guc *guc = &hwe->gt->uc.guc;
+       struct engine_activity_group *eag = &guc->engine_activity.eag[0];
+       u16 guc_class = xe_engine_class_to_guc_class(hwe->class);
+
+       return &eag->engine[guc_class][hwe->logical_instance];
+}
+
+static u64 cpu_ns_to_guc_tsc_tick(ktime_t ns, u32 freq)
+{
+       return mul_u64_u32_div(ns, freq, NSEC_PER_SEC);
+}
+
+#define read_engine_activity_record(xe_, map_, field_) \
+       xe_map_rd_field(xe_, map_, 0, struct guc_engine_activity, field_)
+
+#define read_metadata_record(xe_, map_, field_) \
+       xe_map_rd_field(xe_, map_, 0, struct guc_engine_activity_metadata, field_)
+
+static u64 get_engine_active_ticks(struct xe_guc *guc, struct xe_hw_engine *hwe)
+{
+       struct engine_activity *ea = hw_engine_to_engine_activity(hwe);
+       struct guc_engine_activity *cached_activity = &ea->activity;
+       struct guc_engine_activity_metadata *cached_metadata = &ea->metadata;
+       struct xe_guc_engine_activity *engine_activity = &guc->engine_activity;
+       struct iosys_map activity_map, metadata_map;
+       struct xe_device *xe =  guc_to_xe(guc);
+       struct xe_gt *gt = guc_to_gt(guc);
+       u32 last_update_tick, global_change_num;
+       u64 active_ticks, gpm_ts;
+       u16 change_num;
+
+       activity_map = engine_activity_map(guc, hwe);
+       metadata_map = engine_metadata_map(guc);
+       global_change_num = read_metadata_record(xe, &metadata_map, global_change_num);
+
+       /* GuC has not initialized activity data yet, return 0 */
+       if (!global_change_num)
+               goto update;
+
+       if (global_change_num == cached_metadata->global_change_num)
+               goto update;
+
+       cached_metadata->global_change_num = global_change_num;
+       change_num = read_engine_activity_record(xe, &activity_map, change_num);
+
+       if (!change_num || change_num == cached_activity->change_num)
+               goto update;
+
+       /* read engine activity values */
+       last_update_tick = read_engine_activity_record(xe, &activity_map, last_update_tick);
+       active_ticks = read_engine_activity_record(xe, &activity_map, active_ticks);
+
+       /* activity calculations */
+       ea->running = !!last_update_tick;
+       ea->total += active_ticks - cached_activity->active_ticks;
+       ea->active = 0;
+
+       /* cache the counter */
+       cached_activity->change_num = change_num;
+       cached_activity->last_update_tick = last_update_tick;
+       cached_activity->active_ticks = active_ticks;
+
+update:
+       if (ea->running) {
+               gpm_ts = xe_mmio_read64_2x32(&gt->mmio, MISC_STATUS_0) >>
+                        engine_activity->gpm_timestamp_shift;
+               ea->active = lower_32_bits(gpm_ts) - cached_activity->last_update_tick;
+       }
+
+       return ea->total + ea->active;
+}
+
+static u64 get_engine_total_ticks(struct xe_guc *guc, struct xe_hw_engine *hwe)
+{
+       struct engine_activity *ea = hw_engine_to_engine_activity(hwe);
+       struct guc_engine_activity_metadata *cached_metadata = &ea->metadata;
+       struct guc_engine_activity *cached_activity = &ea->activity;
+       struct iosys_map activity_map, metadata_map;
+       struct xe_device *xe = guc_to_xe(guc);
+       ktime_t now, cpu_delta;
+       u64 numerator;
+       u16 quanta_ratio;
+
+       activity_map = engine_activity_map(guc, hwe);
+       metadata_map = engine_metadata_map(guc);
+
+       if (!cached_metadata->guc_tsc_frequency_hz)
+               cached_metadata->guc_tsc_frequency_hz = read_metadata_record(xe, &metadata_map,
+                                                                            guc_tsc_frequency_hz);
+
+       quanta_ratio = read_engine_activity_record(xe, &activity_map, quanta_ratio);
+       cached_activity->quanta_ratio = quanta_ratio;
+
+       /* Total ticks calculations */
+       now = ktime_get();
+       cpu_delta = now - ea->last_cpu_ts;
+       ea->last_cpu_ts = now;
+       numerator = (ea->quanta_remainder_ns + cpu_delta) * cached_activity->quanta_ratio;
+       ea->quanta_ns += numerator / TOTAL_QUANTA;
+       ea->quanta_remainder_ns = numerator % TOTAL_QUANTA;
+       ea->quanta = cpu_ns_to_guc_tsc_tick(ea->quanta_ns, cached_metadata->guc_tsc_frequency_hz);
+
+       return ea->quanta;
+}
+
+static int enable_engine_activity_stats(struct xe_guc *guc)
+{
+       struct xe_guc_engine_activity *engine_activity = &guc->engine_activity;
+       struct engine_activity_buffer *buffer = &engine_activity->device_buffer;
+       u32 action[] = {
+               XE_GUC_ACTION_SET_DEVICE_ENGINE_ACTIVITY_BUFFER,
+               xe_bo_ggtt_addr(buffer->metadata_bo),
+               0,
+               xe_bo_ggtt_addr(buffer->activity_bo),
+               0,
+       };
+
+       /* Blocking here to ensure the buffers are ready before reading them */
+       return xe_guc_ct_send_block(&guc->ct, action, ARRAY_SIZE(action));
+}
+
+static void engine_activity_set_cpu_ts(struct xe_guc *guc)
+{
+       struct xe_guc_engine_activity *engine_activity = &guc->engine_activity;
+       struct engine_activity_group *eag = &engine_activity->eag[0];
+       int i, j;
+
+       for (i = 0; i < GUC_MAX_ENGINE_CLASSES; i++)
+               for (j = 0; j < GUC_MAX_INSTANCES_PER_CLASS; j++)
+                       eag->engine[i][j].last_cpu_ts = ktime_get();
+}
+
+static u32 gpm_timestamp_shift(struct xe_gt *gt)
+{
+       u32 reg;
+
+       reg = xe_mmio_read32(&gt->mmio, RPM_CONFIG0);
+
+       return 3 - REG_FIELD_GET(RPM_CONFIG0_CTC_SHIFT_PARAMETER_MASK, reg);
+}
+
+/**
+ * xe_guc_engine_activity_active_ticks - Get engine active ticks
+ * @guc: The GuC object
+ * @hwe: The hw_engine object
+ *
+ * Return: accumulated ticks @hwe was active since engine activity stats were enabled.
+ */
+u64 xe_guc_engine_activity_active_ticks(struct xe_guc *guc, struct xe_hw_engine *hwe)
+{
+       return get_engine_active_ticks(guc, hwe);
+}
+
+/**
+ * xe_guc_engine_activity_total_ticks - Get engine total ticks
+ * @guc: The GuC object
+ * @hwe: The hw_engine object
+ *
+ * Return: accumulated quanta of ticks allocated for the engine
+ */
+u64 xe_guc_engine_activity_total_ticks(struct xe_guc *guc, struct xe_hw_engine *hwe)
+{
+       return get_engine_total_ticks(guc, hwe);
+}
+
+/**
+ * xe_guc_engine_activity_enable_stats - Enable engine activity stats
+ * @guc: The GuC object
+ *
+ * Enable engine activity stats and set initial timestamps
+ */
+void xe_guc_engine_activity_enable_stats(struct xe_guc *guc)
+{
+       int ret;
+
+       ret = enable_engine_activity_stats(guc);
+       if (ret)
+               xe_gt_err(guc_to_gt(guc), "failed to enable activity stats%d\n", ret);
+       else
+               engine_activity_set_cpu_ts(guc);
+}
+
+static void engine_activity_fini(void *arg)
+{
+       struct xe_guc_engine_activity *engine_activity = arg;
+       struct engine_activity_buffer *buffer = &engine_activity->device_buffer;
+
+       free_engine_activity_buffers(buffer);
+}
+
+/**
+ * xe_guc_engine_activity_init - Initialize the engine activity data
+ * @guc: The GuC object
+ *
+ * Return: 0 on success, negative error code otherwise.
+ */
+int xe_guc_engine_activity_init(struct xe_guc *guc)
+{
+       struct xe_guc_engine_activity *engine_activity = &guc->engine_activity;
+       struct xe_gt *gt = guc_to_gt(guc);
+       struct xe_device *xe = gt_to_xe(gt);
+       int ret;
+
+       if (IS_SRIOV_VF(xe))
+               return 0;
+
+       ret = allocate_engine_activity_group(guc);
+       if (ret) {
+               xe_gt_err(gt, "failed to allocate engine activity group (%pe)\n", ERR_PTR(ret));
+               return ret;
+       }
+
+       ret = allocate_engine_activity_buffers(guc, &engine_activity->device_buffer);
+       if (ret) {
+               xe_gt_err(gt, "failed to allocate engine activity buffers (%pe)\n", ERR_PTR(ret));
+               return ret;
+       }
+
+       engine_activity->gpm_timestamp_shift = gpm_timestamp_shift(gt);
+
+       return devm_add_action_or_reset(gt_to_xe(gt)->drm.dev, engine_activity_fini,
+                                       engine_activity);
+}
diff --git a/drivers/gpu/drm/xe/xe_guc_engine_activity.h b/drivers/gpu/drm/xe/xe_guc_engine_activity.h
new file mode 100644 (file)
index 0000000..e92d245
--- /dev/null
@@ -0,0 +1,18 @@
+/* SPDX-License-Identifier: MIT */
+/*
+ * Copyright © 2025 Intel Corporation
+ */
+
+#ifndef _XE_GUC_ENGINE_ACTIVITY_H_
+#define _XE_GUC_ENGINE_ACTIVITY_H_
+
+#include <linux/types.h>
+
+struct xe_hw_engine;
+struct xe_guc;
+
+int xe_guc_engine_activity_init(struct xe_guc *guc);
+void xe_guc_engine_activity_enable_stats(struct xe_guc *guc);
+u64 xe_guc_engine_activity_active_ticks(struct xe_guc *guc, struct xe_hw_engine *hwe);
+u64 xe_guc_engine_activity_total_ticks(struct xe_guc *guc, struct xe_hw_engine *hwe);
+#endif
diff --git a/drivers/gpu/drm/xe/xe_guc_engine_activity_types.h b/drivers/gpu/drm/xe/xe_guc_engine_activity_types.h
new file mode 100644 (file)
index 0000000..a2ab327
--- /dev/null
@@ -0,0 +1,89 @@
+/* SPDX-License-Identifier: MIT */
+/*
+ * Copyright © 2025 Intel Corporation
+ */
+
+#ifndef _XE_GUC_ENGINE_ACTIVITY_TYPES_H_
+#define _XE_GUC_ENGINE_ACTIVITY_TYPES_H_
+
+#include <linux/types.h>
+
+#include "xe_guc_fwif.h"
+/**
+ * struct engine_activity - Engine specific activity data
+ *
+ * Contains engine specific activity data and snapshot of the
+ * structures from GuC
+ */
+struct engine_activity {
+       /** @active: current activity */
+       u64 active;
+
+       /** @last_cpu_ts: cpu timestamp in nsec of previous sample */
+       u64 last_cpu_ts;
+
+       /** @quanta: total quanta used on HW */
+       u64 quanta;
+
+       /** @quanta_ns: total quanta_ns used on HW */
+       u64 quanta_ns;
+
+       /**
+        * @quanta_remainder_ns: remainder when the CPU time is scaled as
+        * per the quanta_ratio. This remainder is used in subsequent
+        * quanta calculations.
+        */
+       u64 quanta_remainder_ns;
+
+       /** @total: total engine activity */
+       u64 total;
+
+       /** @running: true if engine is running some work */
+       bool running;
+
+       /** @metadata: snapshot of engine activity metadata */
+       struct guc_engine_activity_metadata metadata;
+
+       /** @activity: snapshot of engine activity counter */
+       struct guc_engine_activity activity;
+};
+
+/**
+ * struct engine_activity_group - Activity data for all engines
+ */
+struct engine_activity_group {
+       /** @engine: engine specific activity data */
+       struct engine_activity engine[GUC_MAX_ENGINE_CLASSES][GUC_MAX_INSTANCES_PER_CLASS];
+};
+
+/**
+ * struct engine_activity_buffer - engine activity buffers
+ *
+ * This contains the buffers allocated for metadata and activity data
+ */
+struct engine_activity_buffer {
+       /** @activity_bo: object allocated to hold activity data */
+       struct xe_bo *activity_bo;
+
+       /** @metadata_bo: object allocated to hold activity metadata */
+       struct xe_bo *metadata_bo;
+};
+
+/**
+ * struct xe_guc_engine_activity - Data used by engine activity implementation
+ */
+struct xe_guc_engine_activity {
+       /** @gpm_timestamp_shift: Right shift value for the gpm timestamp */
+       u32 gpm_timestamp_shift;
+
+       /** @num_activity_group: number of activity groups */
+       u32 num_activity_group;
+
+       /** @eag: holds the device level engine activity data */
+       struct engine_activity_group *eag;
+
+       /** @device_buffer: buffer object for global engine activity */
+       struct engine_activity_buffer device_buffer;
+};
+#endif
+
index 057153f89b30b1b3723837b463c3d6db7916047a..6f57578b07cb055e7695e91ef187ce41c9ea2999 100644 (file)
@@ -208,6 +208,25 @@ struct guc_engine_usage {
        struct guc_engine_usage_record engines[GUC_MAX_ENGINE_CLASSES][GUC_MAX_INSTANCES_PER_CLASS];
 } __packed;
 
+/* Engine Activity stats */
+struct guc_engine_activity {
+       u16 change_num;
+       u16 quanta_ratio;
+       u32 last_update_tick;
+       u64 active_ticks;
+} __packed;
+
+struct guc_engine_activity_data {
+       struct guc_engine_activity engine_activity[GUC_MAX_ENGINE_CLASSES][GUC_MAX_INSTANCES_PER_CLASS];
+} __packed;
+
+struct guc_engine_activity_metadata {
+       u32 guc_tsc_frequency_hz;
+       u32 lag_latency_usec;
+       u32 global_change_num;
+       u32 reserved;
+} __packed;
+
 /* This action will be programmed in C1BC - SOFT_SCRATCH_15_REG */
 enum xe_guc_recv_message {
        XE_GUC_RECV_MSG_CRASH_DUMP_POSTED = BIT(1),
index 573aa6308380a8563590671a7359154c685aed20..63bac64429a5dc5e86c0f8b1877539869ee4d46c 100644 (file)
@@ -13,6 +13,7 @@
 #include "xe_guc_ads_types.h"
 #include "xe_guc_buf_types.h"
 #include "xe_guc_ct_types.h"
+#include "xe_guc_engine_activity_types.h"
 #include "xe_guc_fwif.h"
 #include "xe_guc_log_types.h"
 #include "xe_guc_pc_types.h"
@@ -103,6 +104,9 @@ struct xe_guc {
        /** @relay: GuC Relay Communication used in SR-IOV */
        struct xe_guc_relay relay;
 
+       /** @engine_activity: Device specific engine activity */
+       struct xe_guc_engine_activity engine_activity;
+
        /**
         * @notify_reg: Register which is written to notify GuC of H2G messages
         */