Merge tag 'gvt-next-2017-09-08' of https://github.com/01org/gvt-linux into drm-intel-next-queued

gvt-next-2017-09-08

- PCI config sanitize series (Changbin)
- Workload submission error handling series (Fred)
Signed-off-by: Jani Nikula <jani.nikula@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20170908063155.l54lvpivxntjm7hq@zhen-hp.sh.intel.com

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

@@ -101,7 +101,7 @@ int intel_vgpu_emulate_cfg_read(struct intel_vgpu *vgpu, unsigned int offset,
 	if (WARN_ON(bytes > 4))
 		return -EINVAL;
 
-	if (WARN_ON(offset + bytes > INTEL_GVT_MAX_CFG_SPACE_SZ))
+	if (WARN_ON(offset + bytes > vgpu->gvt->device_info.cfg_space_size))
 		return -EINVAL;
 
 	memcpy(p_data, vgpu_cfg_space(vgpu) + offset, bytes);
@@ -110,13 +110,25 @@ int intel_vgpu_emulate_cfg_read(struct intel_vgpu *vgpu, unsigned int offset,
 
 static int map_aperture(struct intel_vgpu *vgpu, bool map)
 {
-	u64 first_gfn, first_mfn;
+	phys_addr_t aperture_pa = vgpu_aperture_pa_base(vgpu);
+	unsigned long aperture_sz = vgpu_aperture_sz(vgpu);
+	u64 first_gfn;
 	u64 val;
 	int ret;
 
 	if (map == vgpu->cfg_space.bar[INTEL_GVT_PCI_BAR_APERTURE].tracked)
 		return 0;
 
+	if (map) {
+		vgpu->gm.aperture_va = memremap(aperture_pa, aperture_sz,
+						MEMREMAP_WC);
+		if (!vgpu->gm.aperture_va)
+			return -ENOMEM;
+	} else {
+		memunmap(vgpu->gm.aperture_va);
+		vgpu->gm.aperture_va = NULL;
+	}
+
 	val = vgpu_cfg_space(vgpu)[PCI_BASE_ADDRESS_2];
 	if (val & PCI_BASE_ADDRESS_MEM_TYPE_64)
 		val = *(u64 *)(vgpu_cfg_space(vgpu) + PCI_BASE_ADDRESS_2);
@@ -124,14 +136,16 @@ static int map_aperture(struct intel_vgpu *vgpu, bool map)
 		val = *(u32 *)(vgpu_cfg_space(vgpu) + PCI_BASE_ADDRESS_2);
 
 	first_gfn = (val + vgpu_aperture_offset(vgpu)) >> PAGE_SHIFT;
-	first_mfn = vgpu_aperture_pa_base(vgpu) >> PAGE_SHIFT;
 
 	ret = intel_gvt_hypervisor_map_gfn_to_mfn(vgpu, first_gfn,
-						  first_mfn,
-						  vgpu_aperture_sz(vgpu) >>
-						  PAGE_SHIFT, map);
-	if (ret)
+						  aperture_pa >> PAGE_SHIFT,
+						  aperture_sz >> PAGE_SHIFT,
+						  map);
+	if (ret) {
+		memunmap(vgpu->gm.aperture_va);
+		vgpu->gm.aperture_va = NULL;
 		return ret;
+	}
 
 	vgpu->cfg_space.bar[INTEL_GVT_PCI_BAR_APERTURE].tracked = map;
 	return 0;
@@ -197,78 +211,65 @@ static int emulate_pci_command_write(struct intel_vgpu *vgpu,
 static int emulate_pci_bar_write(struct intel_vgpu *vgpu, unsigned int offset,
 	void *p_data, unsigned int bytes)
 {
-	unsigned int bar_index =
-		(rounddown(offset, 8) % PCI_BASE_ADDRESS_0) / 8;
 	u32 new = *(u32 *)(p_data);
 	bool lo = IS_ALIGNED(offset, 8);
 	u64 size;
 	int ret = 0;
 	bool mmio_enabled =
 		vgpu_cfg_space(vgpu)[PCI_COMMAND] & PCI_COMMAND_MEMORY;
+	struct intel_vgpu_pci_bar *bars = vgpu->cfg_space.bar;
 
-	if (WARN_ON(bar_index >= INTEL_GVT_PCI_BAR_MAX))
-		return -EINVAL;
-
+	/*
+	 * Power-up software can determine how much address
+	 * space the device requires by writing a value of
+	 * all 1's to the register and then reading the value
+	 * back. The device will return 0's in all don't-care
+	 * address bits.
+	 */
 	if (new == 0xffffffff) {
-		/*
-		 * Power-up software can determine how much address
-		 * space the device requires by writing a value of
-		 * all 1's to the register and then reading the value
-		 * back. The device will return 0's in all don't-care
-		 * address bits.
-		 */
-		size = vgpu->cfg_space.bar[bar_index].size;
-		if (lo) {
-			new = rounddown(new, size);
-		} else {
-			u32 val = vgpu_cfg_space(vgpu)[rounddown(offset, 8)];
-			/* for 32bit mode bar it returns all-0 in upper 32
-			 * bit, for 64bit mode bar it will calculate the
-			 * size with lower 32bit and return the corresponding
-			 * value
+		switch (offset) {
+		case PCI_BASE_ADDRESS_0:
+		case PCI_BASE_ADDRESS_1:
+			size = ~(bars[INTEL_GVT_PCI_BAR_GTTMMIO].size -1);
+			intel_vgpu_write_pci_bar(vgpu, offset,
+						size >> (lo ? 0 : 32), lo);
+			/*
+			 * Untrap the BAR, since guest hasn't configured a
+			 * valid GPA
 			 */
-			if (val & PCI_BASE_ADDRESS_MEM_TYPE_64)
-				new &= (~(size-1)) >> 32;
-			else
-				new = 0;
-		}
-		/*
-		 * Unmapp & untrap the BAR, since guest hasn't configured a
-		 * valid GPA
-		 */
-		switch (bar_index) {
-		case INTEL_GVT_PCI_BAR_GTTMMIO:
 			ret = trap_gttmmio(vgpu, false);
 			break;
-		case INTEL_GVT_PCI_BAR_APERTURE:
+		case PCI_BASE_ADDRESS_2:
+		case PCI_BASE_ADDRESS_3:
+			size = ~(bars[INTEL_GVT_PCI_BAR_APERTURE].size -1);
+			intel_vgpu_write_pci_bar(vgpu, offset,
+						size >> (lo ? 0 : 32), lo);
 			ret = map_aperture(vgpu, false);
 			break;
+		default:
+			/* Unimplemented BARs */
+			intel_vgpu_write_pci_bar(vgpu, offset, 0x0, false);
 		}
-		intel_vgpu_write_pci_bar(vgpu, offset, new, lo);
 	} else {
-		/*
-		 * Unmapp & untrap the old BAR first, since guest has
-		 * re-configured the BAR
-		 */
-		switch (bar_index) {
-		case INTEL_GVT_PCI_BAR_GTTMMIO:
-			ret = trap_gttmmio(vgpu, false);
+		switch (offset) {
+		case PCI_BASE_ADDRESS_0:
+		case PCI_BASE_ADDRESS_1:
+			/*
+			 * Untrap the old BAR first, since guest has
+			 * re-configured the BAR
+			 */
+			trap_gttmmio(vgpu, false);
+			intel_vgpu_write_pci_bar(vgpu, offset, new, lo);
+			ret = trap_gttmmio(vgpu, mmio_enabled);
 			break;
-		case INTEL_GVT_PCI_BAR_APERTURE:
-			ret = map_aperture(vgpu, false);
+		case PCI_BASE_ADDRESS_2:
+		case PCI_BASE_ADDRESS_3:
+			map_aperture(vgpu, false);
+			intel_vgpu_write_pci_bar(vgpu, offset, new, lo);
+			ret = map_aperture(vgpu, mmio_enabled);
 			break;
-		}
-		intel_vgpu_write_pci_bar(vgpu, offset, new, lo);
-		/* Track the new BAR */
-		if (mmio_enabled) {
-			switch (bar_index) {
-			case INTEL_GVT_PCI_BAR_GTTMMIO:
-				ret = trap_gttmmio(vgpu, true);
-				break;
-			case INTEL_GVT_PCI_BAR_APERTURE:
-				ret = map_aperture(vgpu, true);
-				break;
-			}
+		default:
+			intel_vgpu_write_pci_bar(vgpu, offset, new, lo);
 		}
 	}
 	return ret;
@@ -288,7 +289,7 @@ int intel_vgpu_emulate_cfg_write(struct intel_vgpu *vgpu, unsigned int offset,
 	if (WARN_ON(bytes > 4))
 		return -EINVAL;
 
-	if (WARN_ON(offset + bytes > INTEL_GVT_MAX_CFG_SPACE_SZ))
+	if (WARN_ON(offset + bytes > vgpu->gvt->device_info.cfg_space_size))
 		return -EINVAL;
 
 	/* First check if it's PCI_COMMAND */
@@ -299,10 +300,7 @@ int intel_vgpu_emulate_cfg_write(struct intel_vgpu *vgpu, unsigned int offset,
 	}
 
 	switch (rounddown(offset, 4)) {
-	case PCI_BASE_ADDRESS_0:
-	case PCI_BASE_ADDRESS_1:
-	case PCI_BASE_ADDRESS_2:
-	case PCI_BASE_ADDRESS_3:
+	case PCI_BASE_ADDRESS_0 ... PCI_BASE_ADDRESS_5:
 		if (WARN_ON(!IS_ALIGNED(offset, 4)))
 			return -EINVAL;
 		return emulate_pci_bar_write(vgpu, offset, p_data, bytes);
@@ -344,7 +342,6 @@ void intel_vgpu_init_cfg_space(struct intel_vgpu *vgpu,
 	struct intel_gvt *gvt = vgpu->gvt;
 	const struct intel_gvt_device_info *info = &gvt->device_info;
 	u16 *gmch_ctl;
-	int i;
 
 	memcpy(vgpu_cfg_space(vgpu), gvt->firmware.cfg_space,
 	       info->cfg_space_size);
@@ -371,13 +368,13 @@ void intel_vgpu_init_cfg_space(struct intel_vgpu *vgpu,
 	 */
 	memset(vgpu_cfg_space(vgpu) + PCI_BASE_ADDRESS_1, 0, 4);
 	memset(vgpu_cfg_space(vgpu) + PCI_BASE_ADDRESS_3, 0, 4);
+	memset(vgpu_cfg_space(vgpu) + PCI_BASE_ADDRESS_4, 0, 8);
 	memset(vgpu_cfg_space(vgpu) + INTEL_GVT_PCI_OPREGION, 0, 4);
 
-	for (i = 0; i < INTEL_GVT_MAX_BAR_NUM; i++) {
-		vgpu->cfg_space.bar[i].size = pci_resource_len(
-					      gvt->dev_priv->drm.pdev, i * 2);
-		vgpu->cfg_space.bar[i].tracked = false;
-	}
+	vgpu->cfg_space.bar[INTEL_GVT_PCI_BAR_GTTMMIO].size =
+				pci_resource_len(gvt->dev_priv->drm.pdev, 0);
+	vgpu->cfg_space.bar[INTEL_GVT_PCI_BAR_APERTURE].size =
+				pci_resource_len(gvt->dev_priv->drm.pdev, 2);
 }
 
 /**

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

@@ -1576,11 +1576,11 @@ static int batch_buffer_needs_scan(struct parser_exec_state *s)
 	return 1;
 }
 
-static uint32_t find_bb_size(struct parser_exec_state *s)
+static int find_bb_size(struct parser_exec_state *s)
 {
 	unsigned long gma = 0;
 	struct cmd_info *info;
-	uint32_t bb_size = 0;
+	int bb_size = 0;
 	uint32_t cmd_len = 0;
 	bool met_bb_end = false;
 	struct intel_vgpu *vgpu = s->vgpu;
@@ -1637,6 +1637,8 @@ static int perform_bb_shadow(struct parser_exec_state *s)
 
 	/* get the size of the batch buffer */
 	bb_size = find_bb_size(s);
+	if (bb_size < 0)
+		return -EINVAL;
 
 	/* allocate shadow batch buffer */
 	entry_obj = kmalloc(sizeof(*entry_obj), GFP_KERNEL);
@@ -2603,7 +2605,8 @@ static int shadow_workload_ring_buffer(struct intel_vgpu_workload *workload)
 {
 	struct intel_vgpu *vgpu = workload->vgpu;
 	unsigned long gma_head, gma_tail, gma_top, guest_rb_size;
-	u32 *cs;
+	void *shadow_ring_buffer_va;
+	int ring_id = workload->ring_id;
 	int ret;
 
 	guest_rb_size = _RING_CTL_BUF_SIZE(workload->rb_ctl);
@@ -2616,34 +2619,42 @@ static int shadow_workload_ring_buffer(struct intel_vgpu_workload *workload)
 	gma_tail = workload->rb_start + workload->rb_tail;
 	gma_top = workload->rb_start + guest_rb_size;
 
-	/* allocate shadow ring buffer */
-	cs = intel_ring_begin(workload->req, workload->rb_len / sizeof(u32));
-	if (IS_ERR(cs))
-		return PTR_ERR(cs);
+	if (workload->rb_len > vgpu->reserve_ring_buffer_size[ring_id]) {
+		void *va = vgpu->reserve_ring_buffer_va[ring_id];
+		/* realloc the new ring buffer if needed */
+		vgpu->reserve_ring_buffer_va[ring_id] =
+			krealloc(va, workload->rb_len, GFP_KERNEL);
+		if (!vgpu->reserve_ring_buffer_va[ring_id]) {
+			gvt_vgpu_err("fail to alloc reserve ring buffer\n");
+			return -ENOMEM;
+		}
+		vgpu->reserve_ring_buffer_size[ring_id] = workload->rb_len;
+	}
+
+	shadow_ring_buffer_va = vgpu->reserve_ring_buffer_va[ring_id];
 
 	/* get shadow ring buffer va */
-	workload->shadow_ring_buffer_va = cs;
+	workload->shadow_ring_buffer_va = shadow_ring_buffer_va;
 
 	/* head > tail --> copy head <-> top */
 	if (gma_head > gma_tail) {
 		ret = copy_gma_to_hva(vgpu, vgpu->gtt.ggtt_mm,
-				      gma_head, gma_top, cs);
+				      gma_head, gma_top, shadow_ring_buffer_va);
 		if (ret < 0) {
 			gvt_vgpu_err("fail to copy guest ring buffer\n");
 			return ret;
 		}
-		cs += ret / sizeof(u32);
+		shadow_ring_buffer_va += ret;
 		gma_head = workload->rb_start;
 	}
 
 	/* copy head or start <-> tail */
-	ret = copy_gma_to_hva(vgpu, vgpu->gtt.ggtt_mm, gma_head, gma_tail, cs);
+	ret = copy_gma_to_hva(vgpu, vgpu->gtt.ggtt_mm, gma_head, gma_tail,
+				shadow_ring_buffer_va);
 	if (ret < 0) {
 		gvt_vgpu_err("fail to copy guest ring buffer\n");
 		return ret;
 	}
-	cs += ret / sizeof(u32);
-	intel_ring_advance(workload->req, cs);
 	return 0;
 }
 

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

@@ -368,7 +368,7 @@ static void free_workload(struct intel_vgpu_workload *workload)
 #define get_desc_from_elsp_dwords(ed, i) \
 	((struct execlist_ctx_descriptor_format *)&((ed)->data[i * 2]))
 
-static void prepare_shadow_batch_buffer(struct intel_vgpu_workload *workload)
+static int prepare_shadow_batch_buffer(struct intel_vgpu_workload *workload)
 {
 	const int gmadr_bytes = workload->vgpu->gvt->device_info.gmadr_bytes_in_cmd;
 	struct intel_shadow_bb_entry *entry_obj;
@@ -379,7 +379,7 @@ static void prepare_shadow_batch_buffer(struct intel_vgpu_workload *workload)
 
 		vma = i915_gem_object_ggtt_pin(entry_obj->obj, NULL, 0, 4, 0);
 		if (IS_ERR(vma)) {
-			return;
+			return PTR_ERR(vma);
 		}
 
 		/* FIXME: we are not tracking our pinned VMA leaving it
@@ -392,6 +392,7 @@ static void prepare_shadow_batch_buffer(struct intel_vgpu_workload *workload)
 		if (gmadr_bytes == 8)
 			entry_obj->bb_start_cmd_va[2] = 0;
 	}
+	return 0;
 }
 
 static int update_wa_ctx_2_shadow_ctx(struct intel_shadow_wa_ctx *wa_ctx)
@@ -420,7 +421,7 @@ static int update_wa_ctx_2_shadow_ctx(struct intel_shadow_wa_ctx *wa_ctx)
 	return 0;
 }
 
-static void prepare_shadow_wa_ctx(struct intel_shadow_wa_ctx *wa_ctx)
+static int prepare_shadow_wa_ctx(struct intel_shadow_wa_ctx *wa_ctx)
 {
 	struct i915_vma *vma;
 	unsigned char *per_ctx_va =
@@ -428,12 +429,12 @@ static void prepare_shadow_wa_ctx(struct intel_shadow_wa_ctx *wa_ctx)
 		wa_ctx->indirect_ctx.size;
 
 	if (wa_ctx->indirect_ctx.size == 0)
-		return;
+		return 0;
 
 	vma = i915_gem_object_ggtt_pin(wa_ctx->indirect_ctx.obj, NULL,
 				       0, CACHELINE_BYTES, 0);
 	if (IS_ERR(vma)) {
-		return;
+		return PTR_ERR(vma);
 	}
 
 	/* FIXME: we are not tracking our pinned VMA leaving it
@@ -447,26 +448,7 @@ static void prepare_shadow_wa_ctx(struct intel_shadow_wa_ctx *wa_ctx)
 	memset(per_ctx_va, 0, CACHELINE_BYTES);
 
 	update_wa_ctx_2_shadow_ctx(wa_ctx);
-}
-
-static int prepare_execlist_workload(struct intel_vgpu_workload *workload)
-{
-	struct intel_vgpu *vgpu = workload->vgpu;
-	struct execlist_ctx_descriptor_format ctx[2];
-	int ring_id = workload->ring_id;
-
-	intel_vgpu_pin_mm(workload->shadow_mm);
-	intel_vgpu_sync_oos_pages(workload->vgpu);
-	intel_vgpu_flush_post_shadow(workload->vgpu);
-	prepare_shadow_batch_buffer(workload);
-	prepare_shadow_wa_ctx(&workload->wa_ctx);
-	if (!workload->emulate_schedule_in)
-		return 0;
-
-	ctx[0] = *get_desc_from_elsp_dwords(&workload->elsp_dwords, 1);
-	ctx[1] = *get_desc_from_elsp_dwords(&workload->elsp_dwords, 0);
-
-	return emulate_execlist_schedule_in(&vgpu->execlist[ring_id], ctx);
+	return 0;
 }
 
 static void release_shadow_batch_buffer(struct intel_vgpu_workload *workload)
@@ -489,13 +471,62 @@ static void release_shadow_batch_buffer(struct intel_vgpu_workload *workload)
 	}
 }
 
-static void release_shadow_wa_ctx(struct intel_shadow_wa_ctx *wa_ctx)
+static int prepare_execlist_workload(struct intel_vgpu_workload *workload)
 {
-	if (!wa_ctx->indirect_ctx.obj)
-		return;
+	struct intel_vgpu *vgpu = workload->vgpu;
+	struct execlist_ctx_descriptor_format ctx[2];
+	int ring_id = workload->ring_id;
+	int ret;
+
+	ret = intel_vgpu_pin_mm(workload->shadow_mm);
+	if (ret) {
+		gvt_vgpu_err("fail to vgpu pin mm\n");
+		goto out;
+	}
+
+	ret = intel_vgpu_sync_oos_pages(workload->vgpu);
+	if (ret) {
+		gvt_vgpu_err("fail to vgpu sync oos pages\n");
+		goto err_unpin_mm;
+	}
 
-	i915_gem_object_unpin_map(wa_ctx->indirect_ctx.obj);
-	i915_gem_object_put(wa_ctx->indirect_ctx.obj);
+	ret = intel_vgpu_flush_post_shadow(workload->vgpu);
+	if (ret) {
+		gvt_vgpu_err("fail to flush post shadow\n");
+		goto err_unpin_mm;
+	}
+
+	ret = prepare_shadow_batch_buffer(workload);
+	if (ret) {
+		gvt_vgpu_err("fail to prepare_shadow_batch_buffer\n");
+		goto err_unpin_mm;
+	}
+
+	ret = prepare_shadow_wa_ctx(&workload->wa_ctx);
+	if (ret) {
+		gvt_vgpu_err("fail to prepare_shadow_wa_ctx\n");
+		goto err_shadow_batch;
+	}
+
+	if (!workload->emulate_schedule_in)
+		return 0;
+
+	ctx[0] = *get_desc_from_elsp_dwords(&workload->elsp_dwords, 1);
+	ctx[1] = *get_desc_from_elsp_dwords(&workload->elsp_dwords, 0);
+
+	ret = emulate_execlist_schedule_in(&vgpu->execlist[ring_id], ctx);
+	if (!ret)
+		goto out;
+	else
+		gvt_vgpu_err("fail to emulate execlist schedule in\n");
+
+	release_shadow_wa_ctx(&workload->wa_ctx);
+err_shadow_batch:
+	release_shadow_batch_buffer(workload);
+err_unpin_mm:
+	intel_vgpu_unpin_mm(workload->shadow_mm);
+out:
+	return ret;
 }
 
 static int complete_execlist_workload(struct intel_vgpu_workload *workload)
@@ -511,8 +542,10 @@ static int complete_execlist_workload(struct intel_vgpu_workload *workload)
 	gvt_dbg_el("complete workload %p status %d\n", workload,
 			workload->status);
 
-	release_shadow_batch_buffer(workload);
-	release_shadow_wa_ctx(&workload->wa_ctx);
+	if (!workload->status) {
+		release_shadow_batch_buffer(workload);
+		release_shadow_wa_ctx(&workload->wa_ctx);
+	}
 
 	if (workload->status || (vgpu->resetting_eng & ENGINE_MASK(ring_id))) {
 		/* if workload->status is not successful means HW GPU
@@ -820,10 +853,21 @@ static void clean_workloads(struct intel_vgpu *vgpu, unsigned long engine_mask)
 
 void intel_vgpu_clean_execlist(struct intel_vgpu *vgpu)
 {
+	enum intel_engine_id i;
+	struct intel_engine_cs *engine;
+
 	clean_workloads(vgpu, ALL_ENGINES);
 	kmem_cache_destroy(vgpu->workloads);
+
+	for_each_engine(engine, vgpu->gvt->dev_priv, i) {
+		kfree(vgpu->reserve_ring_buffer_va[i]);
+		vgpu->reserve_ring_buffer_va[i] = NULL;
+		vgpu->reserve_ring_buffer_size[i] = 0;
+	}
+
 }
 
+#define RESERVE_RING_BUFFER_SIZE		((1 * PAGE_SIZE)/8)
 int intel_vgpu_init_execlist(struct intel_vgpu *vgpu)
 {
 	enum intel_engine_id i;
@@ -843,7 +887,26 @@ int intel_vgpu_init_execlist(struct intel_vgpu *vgpu)
 	if (!vgpu->workloads)
 		return -ENOMEM;
 
+	/* each ring has a shadow ring buffer until vgpu destroyed */
+	for_each_engine(engine, vgpu->gvt->dev_priv, i) {
+		vgpu->reserve_ring_buffer_va[i] =
+			kmalloc(RESERVE_RING_BUFFER_SIZE, GFP_KERNEL);
+		if (!vgpu->reserve_ring_buffer_va[i]) {
+			gvt_vgpu_err("fail to alloc reserve ring buffer\n");
+			goto out;
+		}
+		vgpu->reserve_ring_buffer_size[i] = RESERVE_RING_BUFFER_SIZE;
+	}
 	return 0;
+out:
+	for_each_engine(engine, vgpu->gvt->dev_priv, i) {
+		if (vgpu->reserve_ring_buffer_size[i]) {
+			kfree(vgpu->reserve_ring_buffer_va[i]);
+			vgpu->reserve_ring_buffer_va[i] = NULL;
+			vgpu->reserve_ring_buffer_size[i] = 0;
+		}
+	}
+	return -ENOMEM;
 }
 
 void intel_vgpu_reset_execlist(struct intel_vgpu *vgpu,

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

@@ -1647,14 +1647,13 @@ int intel_vgpu_pin_mm(struct intel_vgpu_mm *mm)
 	if (WARN_ON(mm->type != INTEL_GVT_MM_PPGTT))
 		return 0;
 
-	atomic_inc(&mm->pincount);
-
 	if (!mm->shadowed) {
 		ret = shadow_mm(mm);
 		if (ret)
 			return ret;
 	}
 
+	atomic_inc(&mm->pincount);
 	list_del_init(&mm->lru_list);
 	list_add_tail(&mm->lru_list, &mm->vgpu->gvt->gtt.mm_lru_list_head);
 	return 0;

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

@@ -111,7 +111,7 @@ static void init_device_info(struct intel_gvt *gvt)
 	if (IS_BROADWELL(gvt->dev_priv) || IS_SKYLAKE(gvt->dev_priv)
 		|| IS_KABYLAKE(gvt->dev_priv)) {
 		info->max_support_vgpus = 8;
-		info->cfg_space_size = 256;
+		info->cfg_space_size = PCI_CFG_SPACE_EXP_SIZE;
 		info->mmio_size = 2 * 1024 * 1024;
 		info->mmio_bar = 0;
 		info->gtt_start_offset = 8 * 1024 * 1024;

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

@@ -80,6 +80,7 @@ struct intel_gvt_device_info {
 struct intel_vgpu_gm {
 	u64 aperture_sz;
 	u64 hidden_sz;
+	void *aperture_va;
 	struct drm_mm_node low_gm_node;
 	struct drm_mm_node high_gm_node;
 };
@@ -99,7 +100,6 @@ struct intel_vgpu_mmio {
 	bool disable_warn_untrack;
 };
 
-#define INTEL_GVT_MAX_CFG_SPACE_SZ 256
 #define INTEL_GVT_MAX_BAR_NUM 4
 
 struct intel_vgpu_pci_bar {
@@ -108,7 +108,7 @@ struct intel_vgpu_pci_bar {
 };
 
 struct intel_vgpu_cfg_space {
-	unsigned char virtual_cfg_space[INTEL_GVT_MAX_CFG_SPACE_SZ];
+	unsigned char virtual_cfg_space[PCI_CFG_SPACE_EXP_SIZE];
 	struct intel_vgpu_pci_bar bar[INTEL_GVT_MAX_BAR_NUM];
 };
 
@@ -165,6 +165,9 @@ struct intel_vgpu {
 	struct list_head workload_q_head[I915_NUM_ENGINES];
 	struct kmem_cache *workloads;
 	atomic_t running_workload_num;
+	/* 1/2K for each reserve ring buffer */
+	void *reserve_ring_buffer_va[I915_NUM_ENGINES];
+	int reserve_ring_buffer_size[I915_NUM_ENGINES];
 	DECLARE_BITMAP(tlb_handle_pending, I915_NUM_ENGINES);
 	struct i915_gem_context *shadow_ctx;
 	DECLARE_BITMAP(shadow_ctx_desc_updated, I915_NUM_ENGINES);
@@ -474,6 +477,13 @@ int intel_vgpu_emulate_cfg_read(struct intel_vgpu *vgpu, unsigned int offset,
 int intel_vgpu_emulate_cfg_write(struct intel_vgpu *vgpu, unsigned int offset,
 		void *p_data, unsigned int bytes);
 
+static inline u64 intel_vgpu_get_bar_gpa(struct intel_vgpu *vgpu, int bar)
+{
+	/* We are 64bit bar. */
+	return (*(u64 *)(vgpu->cfg_space.virtual_cfg_space + bar)) &
+			PCI_BASE_ADDRESS_MEM_MASK;
+}
+
 void intel_gvt_clean_opregion(struct intel_gvt *gvt);
 int intel_gvt_init_opregion(struct intel_gvt *gvt);
 

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

@@ -609,21 +609,20 @@ static void intel_vgpu_release_work(struct work_struct *work)
 	__intel_vgpu_release(vgpu);
 }
 
-static uint64_t intel_vgpu_get_bar0_addr(struct intel_vgpu *vgpu)
+static uint64_t intel_vgpu_get_bar_addr(struct intel_vgpu *vgpu, int bar)
 {
 	u32 start_lo, start_hi;
 	u32 mem_type;
-	int pos = PCI_BASE_ADDRESS_0;
 
-	start_lo = (*(u32 *)(vgpu->cfg_space.virtual_cfg_space + pos)) &
+	start_lo = (*(u32 *)(vgpu->cfg_space.virtual_cfg_space + bar)) &
 			PCI_BASE_ADDRESS_MEM_MASK;
-	mem_type = (*(u32 *)(vgpu->cfg_space.virtual_cfg_space + pos)) &
+	mem_type = (*(u32 *)(vgpu->cfg_space.virtual_cfg_space + bar)) &
 			PCI_BASE_ADDRESS_MEM_TYPE_MASK;
 
 	switch (mem_type) {
 	case PCI_BASE_ADDRESS_MEM_TYPE_64:
 		start_hi = (*(u32 *)(vgpu->cfg_space.virtual_cfg_space
-						+ pos + 4));
+						+ bar + 4));
 		break;
 	case PCI_BASE_ADDRESS_MEM_TYPE_32:
 	case PCI_BASE_ADDRESS_MEM_TYPE_1M:
@@ -637,6 +636,21 @@ static uint64_t intel_vgpu_get_bar0_addr(struct intel_vgpu *vgpu)
 	return ((u64)start_hi << 32) | start_lo;
 }
 
+static int intel_vgpu_bar_rw(struct intel_vgpu *vgpu, int bar, uint64_t off,
+			     void *buf, unsigned int count, bool is_write)
+{
+	uint64_t bar_start = intel_vgpu_get_bar_addr(vgpu, bar);
+	int ret;
+
+	if (is_write)
+		ret = intel_gvt_ops->emulate_mmio_write(vgpu,
+					bar_start + off, buf, count);
+	else
+		ret = intel_gvt_ops->emulate_mmio_read(vgpu,
+					bar_start + off, buf, count);
+	return ret;
+}
+
 static ssize_t intel_vgpu_rw(struct mdev_device *mdev, char *buf,
 			size_t count, loff_t *ppos, bool is_write)
 {
@@ -661,20 +675,14 @@ static ssize_t intel_vgpu_rw(struct mdev_device *mdev, char *buf,
 						buf, count);
 		break;
 	case VFIO_PCI_BAR0_REGION_INDEX:
-	case VFIO_PCI_BAR1_REGION_INDEX:
-		if (is_write) {
-			uint64_t bar0_start = intel_vgpu_get_bar0_addr(vgpu);
-
-			ret = intel_gvt_ops->emulate_mmio_write(vgpu,
-						bar0_start + pos, buf, count);
-		} else {
-			uint64_t bar0_start = intel_vgpu_get_bar0_addr(vgpu);
-
-			ret = intel_gvt_ops->emulate_mmio_read(vgpu,
-						bar0_start + pos, buf, count);
-		}
+		ret = intel_vgpu_bar_rw(vgpu, PCI_BASE_ADDRESS_0, pos,
+					buf, count, is_write);
 		break;
 	case VFIO_PCI_BAR2_REGION_INDEX:
+		ret = intel_vgpu_bar_rw(vgpu, PCI_BASE_ADDRESS_2, pos,
+					buf, count, is_write);
+		break;
+	case VFIO_PCI_BAR1_REGION_INDEX:
 	case VFIO_PCI_BAR3_REGION_INDEX:
 	case VFIO_PCI_BAR4_REGION_INDEX:
 	case VFIO_PCI_BAR5_REGION_INDEX:
@@ -970,7 +978,7 @@ static long intel_vgpu_ioctl(struct mdev_device *mdev, unsigned int cmd,
 		switch (info.index) {
 		case VFIO_PCI_CONFIG_REGION_INDEX:
 			info.offset = VFIO_PCI_INDEX_TO_OFFSET(info.index);
-			info.size = INTEL_GVT_MAX_CFG_SPACE_SZ;
+			info.size = vgpu->gvt->device_info.cfg_space_size;
 			info.flags = VFIO_REGION_INFO_FLAG_READ |
 				     VFIO_REGION_INFO_FLAG_WRITE;
 			break;

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

@@ -45,8 +45,7 @@
  */
 int intel_vgpu_gpa_to_mmio_offset(struct intel_vgpu *vgpu, u64 gpa)
 {
-	u64 gttmmio_gpa = *(u64 *)(vgpu_cfg_space(vgpu) + PCI_BASE_ADDRESS_0) &
-			  ~GENMASK(3, 0);
+	u64 gttmmio_gpa = intel_vgpu_get_bar_gpa(vgpu, PCI_BASE_ADDRESS_0);
 	return gpa - gttmmio_gpa;
 }
 
@@ -57,6 +56,38 @@ int intel_vgpu_gpa_to_mmio_offset(struct intel_vgpu *vgpu, u64 gpa)
 	(reg >= gvt->device_info.gtt_start_offset \
 	 && reg < gvt->device_info.gtt_start_offset + gvt_ggtt_sz(gvt))
 
+static bool vgpu_gpa_is_aperture(struct intel_vgpu *vgpu, uint64_t gpa)
+{
+	u64 aperture_gpa = intel_vgpu_get_bar_gpa(vgpu, PCI_BASE_ADDRESS_2);
+	u64 aperture_sz = vgpu_aperture_sz(vgpu);
+
+	return gpa >= aperture_gpa && gpa < aperture_gpa + aperture_sz;
+}
+
+static int vgpu_aperture_rw(struct intel_vgpu *vgpu, uint64_t gpa,
+			    void *pdata, unsigned int size, bool is_read)
+{
+	u64 aperture_gpa = intel_vgpu_get_bar_gpa(vgpu, PCI_BASE_ADDRESS_2);
+	u64 offset = gpa - aperture_gpa;
+
+	if (!vgpu_gpa_is_aperture(vgpu, gpa + size - 1)) {
+		gvt_vgpu_err("Aperture rw out of range, offset %llx, size %d\n",
+			     offset, size);
+		return -EINVAL;
+	}
+
+	if (!vgpu->gm.aperture_va) {
+		gvt_vgpu_err("BAR is not enabled\n");
+		return -ENXIO;
+	}
+
+	if (is_read)
+		memcpy(pdata, vgpu->gm.aperture_va + offset, size);
+	else
+		memcpy(vgpu->gm.aperture_va + offset, pdata, size);
+	return 0;
+}
+
 static void failsafe_emulate_mmio_rw(struct intel_vgpu *vgpu, uint64_t pa,
 		void *p_data, unsigned int bytes, bool read)
 {
@@ -133,6 +164,12 @@ int intel_vgpu_emulate_mmio_read(struct intel_vgpu *vgpu, uint64_t pa,
 	}
 	mutex_lock(&gvt->lock);
 
+	if (vgpu_gpa_is_aperture(vgpu, pa)) {
+		ret = vgpu_aperture_rw(vgpu, pa, p_data, bytes, true);
+		mutex_unlock(&gvt->lock);
+		return ret;
+	}
+
 	if (atomic_read(&vgpu->gtt.n_write_protected_guest_page)) {
 		struct intel_vgpu_guest_page *gp;
 
@@ -224,6 +261,12 @@ int intel_vgpu_emulate_mmio_write(struct intel_vgpu *vgpu, uint64_t pa,
 
 	mutex_lock(&gvt->lock);
 
+	if (vgpu_gpa_is_aperture(vgpu, pa)) {
+		ret = vgpu_aperture_rw(vgpu, pa, p_data, bytes, false);
+		mutex_unlock(&gvt->lock);
+		return ret;
+	}
+
 	if (atomic_read(&vgpu->gtt.n_write_protected_guest_page)) {
 		struct intel_vgpu_guest_page *gp;
 

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

@@ -201,6 +201,43 @@ static void shadow_context_descriptor_update(struct i915_gem_context *ctx,
 	ce->lrc_desc = desc;
 }
 
+static int copy_workload_to_ring_buffer(struct intel_vgpu_workload *workload)
+{
+	struct intel_vgpu *vgpu = workload->vgpu;
+	void *shadow_ring_buffer_va;
+	u32 *cs;
+
+	/* allocate shadow ring buffer */
+	cs = intel_ring_begin(workload->req, workload->rb_len / sizeof(u32));
+	if (IS_ERR(cs)) {
+		gvt_vgpu_err("fail to alloc size =%ld shadow  ring buffer\n",
+			workload->rb_len);
+		return PTR_ERR(cs);
+	}
+
+	shadow_ring_buffer_va = workload->shadow_ring_buffer_va;
+
+	/* get shadow ring buffer va */
+	workload->shadow_ring_buffer_va = cs;
+
+	memcpy(cs, shadow_ring_buffer_va,
+			workload->rb_len);
+
+	cs += workload->rb_len / sizeof(u32);
+	intel_ring_advance(workload->req, cs);
+
+	return 0;
+}
+
+void release_shadow_wa_ctx(struct intel_shadow_wa_ctx *wa_ctx)
+{
+	if (!wa_ctx->indirect_ctx.obj)
+		return;
+
+	i915_gem_object_unpin_map(wa_ctx->indirect_ctx.obj);
+	i915_gem_object_put(wa_ctx->indirect_ctx.obj);
+}
+
 /**
  * intel_gvt_scan_and_shadow_workload - audit the workload by scanning and
  * shadow it as well, include ringbuffer,wa_ctx and ctx.
@@ -214,8 +251,10 @@ int intel_gvt_scan_and_shadow_workload(struct intel_vgpu_workload *workload)
 	int ring_id = workload->ring_id;
 	struct i915_gem_context *shadow_ctx = workload->vgpu->shadow_ctx;
 	struct drm_i915_private *dev_priv = workload->vgpu->gvt->dev_priv;
+	struct intel_engine_cs *engine = dev_priv->engine[ring_id];
 	struct drm_i915_gem_request *rq;
 	struct intel_vgpu *vgpu = workload->vgpu;
+	struct intel_ring *ring;
 	int ret;
 
 	lockdep_assert_held(&dev_priv->drm.struct_mutex);
@@ -231,35 +270,56 @@ int intel_gvt_scan_and_shadow_workload(struct intel_vgpu_workload *workload)
 		shadow_context_descriptor_update(shadow_ctx,
 					dev_priv->engine[ring_id]);
 
-	rq = i915_gem_request_alloc(dev_priv->engine[ring_id], shadow_ctx);
-	if (IS_ERR(rq)) {
-		gvt_vgpu_err("fail to allocate gem request\n");
-		ret = PTR_ERR(rq);
-		goto out;
-	}
-
-	gvt_dbg_sched("ring id %d get i915 gem request %p\n", ring_id, rq);
-
-	workload->req = i915_gem_request_get(rq);
-
 	ret = intel_gvt_scan_and_shadow_ringbuffer(workload);
 	if (ret)
-		goto out;
+		goto err_scan;
 
 	if ((workload->ring_id == RCS) &&
 	    (workload->wa_ctx.indirect_ctx.size != 0)) {
 		ret = intel_gvt_scan_and_shadow_wa_ctx(&workload->wa_ctx);
 		if (ret)
-			goto out;
+			goto err_scan;
+	}
+
+	/* pin shadow context by gvt even the shadow context will be pinned
+	 * when i915 alloc request. That is because gvt will update the guest
+	 * context from shadow context when workload is completed, and at that
+	 * moment, i915 may already unpined the shadow context to make the
+	 * shadow_ctx pages invalid. So gvt need to pin itself. After update
+	 * the guest context, gvt can unpin the shadow_ctx safely.
+	 */
+	ring = engine->context_pin(engine, shadow_ctx);
+	if (IS_ERR(ring)) {
+		ret = PTR_ERR(ring);
+		gvt_vgpu_err("fail to pin shadow context\n");
+		goto err_shadow;
 	}
 
 	ret = populate_shadow_context(workload);
 	if (ret)
-		goto out;
+		goto err_unpin;
 
+	rq = i915_gem_request_alloc(dev_priv->engine[ring_id], shadow_ctx);
+	if (IS_ERR(rq)) {
+		gvt_vgpu_err("fail to allocate gem request\n");
+		ret = PTR_ERR(rq);
+		goto err_unpin;
+	}
+
+	gvt_dbg_sched("ring id %d get i915 gem request %p\n", ring_id, rq);
+
+	workload->req = i915_gem_request_get(rq);
+	ret = copy_workload_to_ring_buffer(workload);
+	if (ret)
+		goto err_unpin;
 	workload->shadowed = true;
+	return 0;
 
-out:
+err_unpin:
+	engine->context_unpin(engine, shadow_ctx);
+err_shadow:
+	release_shadow_wa_ctx(&workload->wa_ctx);
+err_scan:
 	return ret;
 }
 
@@ -269,8 +329,6 @@ static int dispatch_workload(struct intel_vgpu_workload *workload)
 	struct i915_gem_context *shadow_ctx = workload->vgpu->shadow_ctx;
 	struct drm_i915_private *dev_priv = workload->vgpu->gvt->dev_priv;
 	struct intel_engine_cs *engine = dev_priv->engine[ring_id];
-	struct intel_vgpu *vgpu = workload->vgpu;
-	struct intel_ring *ring;
 	int ret = 0;
 
 	gvt_dbg_sched("ring id %d prepare to dispatch workload %p\n",
@@ -284,22 +342,10 @@ static int dispatch_workload(struct intel_vgpu_workload *workload)
 
 	if (workload->prepare) {
 		ret = workload->prepare(workload);
-		if (ret)
+		if (ret) {
+			engine->context_unpin(engine, shadow_ctx);
 			goto out;
-	}
-
-	/* pin shadow context by gvt even the shadow context will be pinned
-	 * when i915 alloc request. That is because gvt will update the guest
-	 * context from shadow context when workload is completed, and at that
-	 * moment, i915 may already unpined the shadow context to make the
-	 * shadow_ctx pages invalid. So gvt need to pin itself. After update
-	 * the guest context, gvt can unpin the shadow_ctx safely.
-	 */
-	ring = engine->context_pin(engine, shadow_ctx);
-	if (IS_ERR(ring)) {
-		ret = PTR_ERR(ring);
-		gvt_vgpu_err("fail to pin shadow context\n");
-		goto out;
+		}
 	}
 
 out:

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

@@ -140,4 +140,5 @@ int intel_vgpu_init_gvt_context(struct intel_vgpu *vgpu);
 
 void intel_vgpu_clean_gvt_context(struct intel_vgpu *vgpu);
 
+void release_shadow_wa_ctx(struct intel_shadow_wa_ctx *wa_ctx);
 #endif