Merge tag 'gvt-next-2018-04-23' of https://github.com/intel/gvt-linux into drm-intel-next-queued

- Minor condition check improvment (Gustavo A. R. Silva)
- Non-priviliged batch buffer scan (Yan Zhao)
- Scheduling optimizations (Zhipeng Gong)

Signed-off-by: Jani Nikula <jani.nikula@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/5dafba29-b2bd-6b94-630e-db5c009da7e3@intel.com

drivers/gpu/drm/i915/gvt/cmd_parser.c

@@ -1603,7 +1603,8 @@ static int batch_buffer_needs_scan(struct parser_exec_state *s)
 	if (IS_BROADWELL(gvt->dev_priv) || IS_SKYLAKE(gvt->dev_priv)
 		|| IS_KABYLAKE(gvt->dev_priv)) {
 		/* BDW decides privilege based on address space */
-		if (cmd_val(s, 0) & (1 << 8))
+		if (cmd_val(s, 0) & (1 << 8) &&
+			!(s->vgpu->scan_nonprivbb & (1 << s->ring_id)))
 			return 0;
 	}
 	return 1;
@@ -1617,6 +1618,8 @@ static int find_bb_size(struct parser_exec_state *s, unsigned long *bb_size)
 	bool bb_end = false;
 	struct intel_vgpu *vgpu = s->vgpu;
 	u32 cmd;
+	struct intel_vgpu_mm *mm = (s->buf_addr_type == GTT_BUFFER) ?
+		s->vgpu->gtt.ggtt_mm : s->workload->shadow_mm;
 
 	*bb_size = 0;
 
@@ -1628,18 +1631,22 @@ static int find_bb_size(struct parser_exec_state *s, unsigned long *bb_size)
 	cmd = cmd_val(s, 0);
 	info = get_cmd_info(s->vgpu->gvt, cmd, s->ring_id);
 	if (info == NULL) {
-		gvt_vgpu_err("unknown cmd 0x%x, opcode=0x%x\n",
-				cmd, get_opcode(cmd, s->ring_id));
+		gvt_vgpu_err("unknown cmd 0x%x, opcode=0x%x, addr_type=%s, ring %d, workload=%p\n",
+				cmd, get_opcode(cmd, s->ring_id),
+				(s->buf_addr_type == PPGTT_BUFFER) ?
+				"ppgtt" : "ggtt", s->ring_id, s->workload);
 		return -EBADRQC;
 	}
 	do {
-		if (copy_gma_to_hva(s->vgpu, s->vgpu->gtt.ggtt_mm,
+		if (copy_gma_to_hva(s->vgpu, mm,
 				gma, gma + 4, &cmd) < 0)
 			return -EFAULT;
 		info = get_cmd_info(s->vgpu->gvt, cmd, s->ring_id);
 		if (info == NULL) {
-			gvt_vgpu_err("unknown cmd 0x%x, opcode=0x%x\n",
-				cmd, get_opcode(cmd, s->ring_id));
+			gvt_vgpu_err("unknown cmd 0x%x, opcode=0x%x, addr_type=%s, ring %d, workload=%p\n",
+				cmd, get_opcode(cmd, s->ring_id),
+				(s->buf_addr_type == PPGTT_BUFFER) ?
+				"ppgtt" : "ggtt", s->ring_id, s->workload);
 			return -EBADRQC;
 		}
 
@@ -1665,6 +1672,9 @@ static int perform_bb_shadow(struct parser_exec_state *s)
 	unsigned long gma = 0;
 	unsigned long bb_size;
 	int ret = 0;
+	struct intel_vgpu_mm *mm = (s->buf_addr_type == GTT_BUFFER) ?
+		s->vgpu->gtt.ggtt_mm : s->workload->shadow_mm;
+	unsigned long gma_start_offset = 0;
 
 	/* get the start gm address of the batch buffer */
 	gma = get_gma_bb_from_cmd(s, 1);
@@ -1679,8 +1689,24 @@ static int perform_bb_shadow(struct parser_exec_state *s)
 	if (!bb)
 		return -ENOMEM;
 
+	bb->ppgtt = (s->buf_addr_type == GTT_BUFFER) ? false : true;
+
+	/* the gma_start_offset stores the batch buffer's start gma's
+	 * offset relative to page boundary. so for non-privileged batch
+	 * buffer, the shadowed gem object holds exactly the same page
+	 * layout as original gem object. This is for the convience of
+	 * replacing the whole non-privilged batch buffer page to this
+	 * shadowed one in PPGTT at the same gma address. (this replacing
+	 * action is not implemented yet now, but may be necessary in
+	 * future).
+	 * for prileged batch buffer, we just change start gma address to
+	 * that of shadowed page.
+	 */
+	if (bb->ppgtt)
+		gma_start_offset = gma & ~I915_GTT_PAGE_MASK;
+
 	bb->obj = i915_gem_object_create(s->vgpu->gvt->dev_priv,
-					 roundup(bb_size, PAGE_SIZE));
+			 roundup(bb_size + gma_start_offset, PAGE_SIZE));
 	if (IS_ERR(bb->obj)) {
 		ret = PTR_ERR(bb->obj);
 		goto err_free_bb;
@@ -1701,9 +1727,9 @@ static int perform_bb_shadow(struct parser_exec_state *s)
 		bb->clflush &= ~CLFLUSH_BEFORE;
 	}
 
-	ret = copy_gma_to_hva(s->vgpu, s->vgpu->gtt.ggtt_mm,
+	ret = copy_gma_to_hva(s->vgpu, mm,
 			      gma, gma + bb_size,
-			      bb->va);
+			      bb->va + gma_start_offset);
 	if (ret < 0) {
 		gvt_vgpu_err("fail to copy guest ring buffer\n");
 		ret = -EFAULT;
@@ -1729,7 +1755,7 @@ static int perform_bb_shadow(struct parser_exec_state *s)
 	 * buffer's gma in pair. After all, we don't want to pin the shadow
 	 * buffer here (too early).
 	 */
-	s->ip_va = bb->va;
+	s->ip_va = bb->va + gma_start_offset;
 	s->ip_gma = gma;
 	return 0;
 err_unmap:
@@ -2468,15 +2494,18 @@ static int cmd_parser_exec(struct parser_exec_state *s)
 
 	info = get_cmd_info(s->vgpu->gvt, cmd, s->ring_id);
 	if (info == NULL) {
-		gvt_vgpu_err("unknown cmd 0x%x, opcode=0x%x\n",
-				cmd, get_opcode(cmd, s->ring_id));
+		gvt_vgpu_err("unknown cmd 0x%x, opcode=0x%x, addr_type=%s, ring %d, workload=%p\n",
+				cmd, get_opcode(cmd, s->ring_id),
+				(s->buf_addr_type == PPGTT_BUFFER) ?
+				"ppgtt" : "ggtt", s->ring_id, s->workload);
 		return -EBADRQC;
 	}
 
 	s->info = info;
 
 	trace_gvt_command(vgpu->id, s->ring_id, s->ip_gma, s->ip_va,
-			  cmd_length(s), s->buf_type);
+			  cmd_length(s), s->buf_type, s->buf_addr_type,
+			  s->workload, info->name);
 
 	if (info->handler) {
 		ret = info->handler(s);

drivers/gpu/drm/i915/gvt/debugfs.c

@@ -124,6 +124,68 @@ static int vgpu_mmio_diff_show(struct seq_file *s, void *unused)
 }
 DEFINE_SHOW_ATTRIBUTE(vgpu_mmio_diff);
 
+static int
+vgpu_scan_nonprivbb_get(void *data, u64 *val)
+{
+	struct intel_vgpu *vgpu = (struct intel_vgpu *)data;
+	*val = vgpu->scan_nonprivbb;
+	return 0;
+}
+
+/*
+ * set/unset bit engine_id of vgpu->scan_nonprivbb to turn on/off scanning
+ * of non-privileged batch buffer. e.g.
+ * if vgpu->scan_nonprivbb=3, then it will scan non-privileged batch buffer
+ * on engine 0 and 1.
+ */
+static int
+vgpu_scan_nonprivbb_set(void *data, u64 val)
+{
+	struct intel_vgpu *vgpu = (struct intel_vgpu *)data;
+	struct drm_i915_private *dev_priv = vgpu->gvt->dev_priv;
+	enum intel_engine_id id;
+	char buf[128], *s;
+	int len;
+
+	val &= (1 << I915_NUM_ENGINES) - 1;
+
+	if (vgpu->scan_nonprivbb == val)
+		return 0;
+
+	if (!val)
+		goto done;
+
+	len = sprintf(buf,
+		"gvt: vgpu %d turns on non-privileged batch buffers scanning on Engines:",
+		vgpu->id);
+
+	s = buf + len;
+
+	for (id = 0; id < I915_NUM_ENGINES; id++) {
+		struct intel_engine_cs *engine;
+
+		engine = dev_priv->engine[id];
+		if (engine && (val & (1 << id))) {
+			len = snprintf(s, 4, "%d, ", engine->id);
+			s += len;
+		} else
+			val &=  ~(1 << id);
+	}
+
+	if (val)
+		sprintf(s, "low performance expected.");
+
+	pr_warn("%s\n", buf);
+
+done:
+	vgpu->scan_nonprivbb = val;
+	return 0;
+}
+
+DEFINE_SIMPLE_ATTRIBUTE(vgpu_scan_nonprivbb_fops,
+			vgpu_scan_nonprivbb_get, vgpu_scan_nonprivbb_set,
+			"0x%llx\n");
+
 /**
  * intel_gvt_debugfs_add_vgpu - register debugfs entries for a vGPU
  * @vgpu: a vGPU
@@ -151,6 +213,11 @@ int intel_gvt_debugfs_add_vgpu(struct intel_vgpu *vgpu)
 	if (!ent)
 		return -ENOMEM;
 
+	ent = debugfs_create_file("scan_nonprivbb", 0644, vgpu->debugfs,
+				 vgpu, &vgpu_scan_nonprivbb_fops);
+	if (!ent)
+		return -ENOMEM;
+
 	return 0;
 }
 

drivers/gpu/drm/i915/gvt/gvt.h

@@ -226,6 +226,7 @@ struct intel_vgpu {
 
 	struct completion vblank_done;
 
+	u32 scan_nonprivbb;
 };
 
 /* validating GM healthy status*/

drivers/gpu/drm/i915/gvt/handlers.c

@@ -1150,6 +1150,7 @@ static int handle_g2v_notification(struct intel_vgpu *vgpu, int notification)
 	switch (notification) {
 	case VGT_G2V_PPGTT_L3_PAGE_TABLE_CREATE:
 		root_entry_type = GTT_TYPE_PPGTT_ROOT_L3_ENTRY;
+		/* fall through */
 	case VGT_G2V_PPGTT_L4_PAGE_TABLE_CREATE:
 		mm = intel_vgpu_get_ppgtt_mm(vgpu, root_entry_type, pdps);
 		return PTR_ERR_OR_ZERO(mm);

drivers/gpu/drm/i915/gvt/sched_policy.c

@@ -53,7 +53,6 @@ struct vgpu_sched_data {
 	bool active;
 
 	ktime_t sched_in_time;
-	ktime_t sched_out_time;
 	ktime_t sched_time;
 	ktime_t left_ts;
 	ktime_t allocated_ts;
@@ -66,17 +65,22 @@ struct gvt_sched_data {
 	struct hrtimer timer;
 	unsigned long period;
 	struct list_head lru_runq_head;
+	ktime_t expire_time;
 };
 
-static void vgpu_update_timeslice(struct intel_vgpu *pre_vgpu)
+static void vgpu_update_timeslice(struct intel_vgpu *vgpu, ktime_t cur_time)
 {
 	ktime_t delta_ts;
-	struct vgpu_sched_data *vgpu_data = pre_vgpu->sched_data;
+	struct vgpu_sched_data *vgpu_data;
 
-	delta_ts = vgpu_data->sched_out_time - vgpu_data->sched_in_time;
+	if (!vgpu || vgpu == vgpu->gvt->idle_vgpu)
+		return;
 
-	vgpu_data->sched_time += delta_ts;
-	vgpu_data->left_ts -= delta_ts;
+	vgpu_data = vgpu->sched_data;
+	delta_ts = ktime_sub(cur_time, vgpu_data->sched_in_time);
+	vgpu_data->sched_time = ktime_add(vgpu_data->sched_time, delta_ts);
+	vgpu_data->left_ts = ktime_sub(vgpu_data->left_ts, delta_ts);
+	vgpu_data->sched_in_time = cur_time;
 }
 
 #define GVT_TS_BALANCE_PERIOD_MS 100
@@ -150,11 +154,7 @@ static void try_to_schedule_next_vgpu(struct intel_gvt *gvt)
 	}
 
 	cur_time = ktime_get();
-	if (scheduler->current_vgpu) {
-		vgpu_data = scheduler->current_vgpu->sched_data;
-		vgpu_data->sched_out_time = cur_time;
-		vgpu_update_timeslice(scheduler->current_vgpu);
-	}
+	vgpu_update_timeslice(scheduler->current_vgpu, cur_time);
 	vgpu_data = scheduler->next_vgpu->sched_data;
 	vgpu_data->sched_in_time = cur_time;
 
@@ -226,17 +226,22 @@ out:
 void intel_gvt_schedule(struct intel_gvt *gvt)
 {
 	struct gvt_sched_data *sched_data = gvt->scheduler.sched_data;
-	static uint64_t timer_check;
+	ktime_t cur_time;
 
 	mutex_lock(&gvt->lock);
+	cur_time = ktime_get();
 
 	if (test_and_clear_bit(INTEL_GVT_REQUEST_SCHED,
 				(void *)&gvt->service_request)) {
-		if (!(timer_check++ % GVT_TS_BALANCE_PERIOD_MS))
+		if (cur_time >= sched_data->expire_time) {
 			gvt_balance_timeslice(sched_data);
+			sched_data->expire_time = ktime_add_ms(
+				cur_time, GVT_TS_BALANCE_PERIOD_MS);
+		}
 	}
 	clear_bit(INTEL_GVT_REQUEST_EVENT_SCHED, (void *)&gvt->service_request);
 
+	vgpu_update_timeslice(gvt->scheduler.current_vgpu, cur_time);
 	tbs_sched_func(sched_data);
 
 	mutex_unlock(&gvt->lock);

drivers/gpu/drm/i915/gvt/scheduler.c

@@ -97,7 +97,7 @@ static void sr_oa_regs(struct intel_vgpu_workload *workload,
 		i915_mmio_reg_offset(EU_PERF_CNTL6),
 	};
 
-	if (!workload || !reg_state || workload->ring_id != RCS)
+	if (workload->ring_id != RCS)
 		return;
 
 	if (save) {
@@ -452,12 +452,6 @@ static int prepare_shadow_batch_buffer(struct intel_vgpu_workload *workload)
 	int ret;
 
 	list_for_each_entry(bb, &workload->shadow_bb, list) {
-		bb->vma = i915_gem_object_ggtt_pin(bb->obj, NULL, 0, 0, 0);
-		if (IS_ERR(bb->vma)) {
-			ret = PTR_ERR(bb->vma);
-			goto err;
-		}
-
 		/* For privilge batch buffer and not wa_ctx, the bb_start_cmd_va
 		 * is only updated into ring_scan_buffer, not real ring address
 		 * allocated in later copy_workload_to_ring_buffer. pls be noted
@@ -469,25 +463,53 @@ static int prepare_shadow_batch_buffer(struct intel_vgpu_workload *workload)
 			bb->bb_start_cmd_va = workload->shadow_ring_buffer_va
 				+ bb->bb_offset;
 
-		/* relocate shadow batch buffer */
-		bb->bb_start_cmd_va[1] = i915_ggtt_offset(bb->vma);
-		if (gmadr_bytes == 8)
-			bb->bb_start_cmd_va[2] = 0;
+		if (bb->ppgtt) {
+			/* for non-priv bb, scan&shadow is only for
+			 * debugging purpose, so the content of shadow bb
+			 * is the same as original bb. Therefore,
+			 * here, rather than switch to shadow bb's gma
+			 * address, we directly use original batch buffer's
+			 * gma address, and send original bb to hardware
+			 * directly
+			 */
+			if (bb->clflush & CLFLUSH_AFTER) {
+				drm_clflush_virt_range(bb->va,
+						bb->obj->base.size);
+				bb->clflush &= ~CLFLUSH_AFTER;
+			}
+			i915_gem_obj_finish_shmem_access(bb->obj);
+			bb->accessing = false;
+
+		} else {
+			bb->vma = i915_gem_object_ggtt_pin(bb->obj,
+					NULL, 0, 0, 0);
+			if (IS_ERR(bb->vma)) {
+				ret = PTR_ERR(bb->vma);
+				goto err;
+			}
 
-		/* No one is going to touch shadow bb from now on. */
-		if (bb->clflush & CLFLUSH_AFTER) {
-			drm_clflush_virt_range(bb->va, bb->obj->base.size);
-			bb->clflush &= ~CLFLUSH_AFTER;
-		}
+			/* relocate shadow batch buffer */
+			bb->bb_start_cmd_va[1] = i915_ggtt_offset(bb->vma);
+			if (gmadr_bytes == 8)
+				bb->bb_start_cmd_va[2] = 0;
 
-		ret = i915_gem_object_set_to_gtt_domain(bb->obj, false);
-		if (ret)
-			goto err;
+			/* No one is going to touch shadow bb from now on. */
+			if (bb->clflush & CLFLUSH_AFTER) {
+				drm_clflush_virt_range(bb->va,
+						bb->obj->base.size);
+				bb->clflush &= ~CLFLUSH_AFTER;
+			}
 
-		i915_gem_obj_finish_shmem_access(bb->obj);
-		bb->accessing = false;
+			ret = i915_gem_object_set_to_gtt_domain(bb->obj,
+					false);
+			if (ret)
+				goto err;
 
-		i915_vma_move_to_active(bb->vma, workload->req, 0);
+			i915_gem_obj_finish_shmem_access(bb->obj);
+			bb->accessing = false;
+
+			i915_vma_move_to_active(bb->vma, workload->req, 0);
+		}
 	}
 	return 0;
 err:

drivers/gpu/drm/i915/gvt/scheduler.h

@@ -125,6 +125,7 @@ struct intel_vgpu_shadow_bb {
 	unsigned int clflush;
 	bool accessing;
 	unsigned long bb_offset;
+	bool ppgtt;
 };
 
 #define workload_q_head(vgpu, ring_id) \

drivers/gpu/drm/i915/gvt/trace.h

@@ -224,19 +224,25 @@ TRACE_EVENT(oos_sync,
 	TP_printk("%s", __entry->buf)
 );
 
+#define GVT_CMD_STR_LEN 40
 TRACE_EVENT(gvt_command,
-	TP_PROTO(u8 vgpu_id, u8 ring_id, u32 ip_gma, u32 *cmd_va, u32 cmd_len,
-		 u32 buf_type),
+	TP_PROTO(u8 vgpu_id, u8 ring_id, u32 ip_gma, u32 *cmd_va,
+		u32 cmd_len,  u32 buf_type, u32 buf_addr_type,
+		void *workload, char *cmd_name),
 
-	TP_ARGS(vgpu_id, ring_id, ip_gma, cmd_va, cmd_len, buf_type),
+	TP_ARGS(vgpu_id, ring_id, ip_gma, cmd_va, cmd_len, buf_type,
+		buf_addr_type, workload, cmd_name),
 
 	TP_STRUCT__entry(
 		__field(u8, vgpu_id)
 		__field(u8, ring_id)
 		__field(u32, ip_gma)
 		__field(u32, buf_type)
+		__field(u32, buf_addr_type)
 		__field(u32, cmd_len)
+		__field(void*, workload)
 		__dynamic_array(u32, raw_cmd, cmd_len)
+		__array(char, cmd_name, GVT_CMD_STR_LEN)
 	),
 
 	TP_fast_assign(
@@ -244,17 +250,25 @@ TRACE_EVENT(gvt_command,
 		__entry->ring_id = ring_id;
 		__entry->ip_gma = ip_gma;
 		__entry->buf_type = buf_type;
+		__entry->buf_addr_type = buf_addr_type;
 		__entry->cmd_len = cmd_len;
+		__entry->workload = workload;
+		snprintf(__entry->cmd_name, GVT_CMD_STR_LEN, "%s", cmd_name);
 		memcpy(__get_dynamic_array(raw_cmd), cmd_va, cmd_len * sizeof(*cmd_va));
 	),
 
 
-	TP_printk("vgpu%d ring %d: buf_type %u, ip_gma %08x, raw cmd %s",
+	TP_printk("vgpu%d ring %d: address_type %u, buf_type %u, ip_gma %08x,cmd (name=%s,len=%u,raw cmd=%s), workload=%p\n",
 		__entry->vgpu_id,
 		__entry->ring_id,
+		__entry->buf_addr_type,
 		__entry->buf_type,
 		__entry->ip_gma,
-		__print_array(__get_dynamic_array(raw_cmd), __entry->cmd_len, 4))
+		__entry->cmd_name,
+		__entry->cmd_len,
+		__print_array(__get_dynamic_array(raw_cmd),
+			__entry->cmd_len, 4),
+		__entry->workload)
 );
 
 #define GVT_TEMP_STR_LEN 10