Merge branch 'drm-intel-next' of git://anongit.freedesktop.org/drm-intel into drm-next

drm-intel-next-2016-08-22:
- bugfixes and cleanups for rcu-protected requests (Chris)
- atomic modeset fixes for gpu reset on pre-g4x (Maarten&Ville)
- guc submission improvements (Dave Gordon)
- panel power sequence cleanup (Imre)
- better use of stolen and unmappable ggtt (Chris), plus prep work to make that
  happen
- rework of framebuffer offsets, prep for multi-plane framebuffers (Ville)
- fully partial ggtt vmaps, including fenced ones (Chris)
- move lots more of the gem tracking from the object to the vma (Chris)
- tune the command parser (Chris)
- allow fbc without fences on recent platforms (Chris)
- fbc frontbuffer tracking fixes (Chris)
- fast prefaulting using io-mappping.h pgprot caching (Chris)

* 'drm-intel-next' of git://anongit.freedesktop.org/drm-intel: (141 commits)
  io-mapping: Fixup for different names of writecombine
  io-mapping.h: s/PAGE_KERNEL_IO/PAGE_KERNEL/
  drm/i915: Update DRIVER_DATE to 20160822
  drm/i915: Use remap_io_mapping() to prefault all PTE in a single pass
  drm/i915: Embed the io-mapping struct inside drm_i915_private
  io-mapping: Always create a struct to hold metadata about the io-mapping
  drm/i915/fbc: Allow on unfenced surfaces, for recent gen
  drm/i915/fbc: Don't set an illegal fence if unfenced
  drm/i915: Flush delayed fence releases after reset
  drm/i915: Reattach comment, complete type specification
  drm/i915/cmdparser: Accelerate copies from WC memory
  drm/i915/cmdparser: Use binary search for faster register lookup
  drm/i915/cmdparser: Check for SKIP descriptors first
  drm/i915/cmdparser: Compare against the previous command descriptor
  drm/i915/cmdparser: Improve hash function
  drm/i915/cmdparser: Only cache the dst vmap
  drm/i915/cmdparser: Use cached vmappings
  drm/i915/cmdparser: Add the TIMESTAMP register for the other engines
  drm/i915/cmdparser: Make initialisation failure non-fatal
  drm/i915: Stop discarding GTT cache-domain on unbind vma
  ...

arch/x86/kernel/early-quirks.c

@@ -317,16 +317,11 @@ static phys_addr_t __init i85x_stolen_base(int num, int slot, int func,
 static phys_addr_t __init i865_stolen_base(int num, int slot, int func,
 					   size_t stolen_size)
 {
-	u16 toud;
+	u16 toud = 0;
 
-	/*
-	 * FIXME is the graphics stolen memory region
-	 * always at TOUD? Ie. is it always the last
-	 * one to be allocated by the BIOS?
-	 */
 	toud = read_pci_config_16(0, 0, 0, I865_TOUD);
 
-	return (phys_addr_t)toud << 16;
+	return (phys_addr_t)(toud << 16) + i845_tseg_size();
 }
 
 static phys_addr_t __init gen3_stolen_base(int num, int slot, int func,

drivers/char/agp/intel-gtt.c

@@ -845,6 +845,8 @@ void intel_gtt_insert_page(dma_addr_t addr,
 			   unsigned int flags)
 {
 	intel_private.driver->write_entry(addr, pg, flags);
+	if (intel_private.driver->chipset_flush)
+		intel_private.driver->chipset_flush();
 }
 EXPORT_SYMBOL(intel_gtt_insert_page);
 

drivers/gpu/drm/Makefile

@@ -47,7 +47,7 @@ obj-$(CONFIG_DRM_RADEON)+= radeon/
 obj-$(CONFIG_DRM_AMDGPU)+= amd/amdgpu/
 obj-$(CONFIG_DRM_MGA)	+= mga/
 obj-$(CONFIG_DRM_I810)	+= i810/
-obj-$(CONFIG_DRM_I915)  += i915/
+obj-$(CONFIG_DRM_I915)	+= i915/
 obj-$(CONFIG_DRM_MGAG200) += mgag200/
 obj-$(CONFIG_DRM_VC4)  += vc4/
 obj-$(CONFIG_DRM_CIRRUS_QEMU) += cirrus/

drivers/gpu/drm/i915/Makefile

@@ -3,12 +3,16 @@
 # Direct Rendering Infrastructure (DRI) in XFree86 4.1.0 and higher.
 
 subdir-ccflags-$(CONFIG_DRM_I915_WERROR) := -Werror
+subdir-ccflags-y += \
+	$(call as-instr,movntdqa (%eax)$(comma)%xmm0,-DCONFIG_AS_MOVNTDQA)
 
 # Please keep these build lists sorted!
 
 # core driver code
 i915-y := i915_drv.o \
 	  i915_irq.o \
+	  i915_memcpy.o \
+	  i915_mm.o \
 	  i915_params.o \
 	  i915_pci.o \
           i915_suspend.o \
@@ -110,6 +114,6 @@ i915-y += intel_gvt.o
 include $(src)/gvt/Makefile
 endif
 
-obj-$(CONFIG_DRM_I915)  += i915.o
+obj-$(CONFIG_DRM_I915) += i915.o
 
 CFLAGS_i915_trace_points.o := -I$(src)

drivers/gpu/drm/i915/i915_cmd_parser.c

@@ -86,24 +86,25 @@
  * general bitmasking mechanism.
  */
 
-#define STD_MI_OPCODE_MASK  0xFF800000
-#define STD_3D_OPCODE_MASK  0xFFFF0000
-#define STD_2D_OPCODE_MASK  0xFFC00000
-#define STD_MFX_OPCODE_MASK 0xFFFF0000
+#define STD_MI_OPCODE_SHIFT  (32 - 9)
+#define STD_3D_OPCODE_SHIFT  (32 - 16)
+#define STD_2D_OPCODE_SHIFT  (32 - 10)
+#define STD_MFX_OPCODE_SHIFT (32 - 16)
+#define MIN_OPCODE_SHIFT 16
 
 #define CMD(op, opm, f, lm, fl, ...)				\
 	{							\
 		.flags = (fl) | ((f) ? CMD_DESC_FIXED : 0),	\
-		.cmd = { (op), (opm) },				\
+		.cmd = { (op), ~0u << (opm) },			\
 		.length = { (lm) },				\
 		__VA_ARGS__					\
 	}
 
 /* Convenience macros to compress the tables */
-#define SMI STD_MI_OPCODE_MASK
-#define S3D STD_3D_OPCODE_MASK
-#define S2D STD_2D_OPCODE_MASK
-#define SMFX STD_MFX_OPCODE_MASK
+#define SMI STD_MI_OPCODE_SHIFT
+#define S3D STD_3D_OPCODE_SHIFT
+#define S2D STD_2D_OPCODE_SHIFT
+#define SMFX STD_MFX_OPCODE_SHIFT
 #define F true
 #define S CMD_DESC_SKIP
 #define R CMD_DESC_REJECT
@@ -350,6 +351,9 @@ static const struct drm_i915_cmd_descriptor hsw_blt_cmds[] = {
 	CMD(  MI_LOAD_SCAN_LINES_EXCL,          SMI,   !F,  0x3F,   R  ),
 };
 
+static const struct drm_i915_cmd_descriptor noop_desc =
+	CMD(MI_NOOP, SMI, F, 1, S);
+
 #undef CMD
 #undef SMI
 #undef S3D
@@ -458,6 +462,7 @@ static const struct drm_i915_reg_descriptor gen7_render_regs[] = {
 	REG32(GEN7_GPGPU_DISPATCHDIMX),
 	REG32(GEN7_GPGPU_DISPATCHDIMY),
 	REG32(GEN7_GPGPU_DISPATCHDIMZ),
+	REG64_IDX(RING_TIMESTAMP, BSD_RING_BASE),
 	REG64_IDX(GEN7_SO_NUM_PRIMS_WRITTEN, 0),
 	REG64_IDX(GEN7_SO_NUM_PRIMS_WRITTEN, 1),
 	REG64_IDX(GEN7_SO_NUM_PRIMS_WRITTEN, 2),
@@ -473,6 +478,7 @@ static const struct drm_i915_reg_descriptor gen7_render_regs[] = {
 	REG32(GEN7_L3SQCREG1),
 	REG32(GEN7_L3CNTLREG2),
 	REG32(GEN7_L3CNTLREG3),
+	REG64_IDX(RING_TIMESTAMP, BLT_RING_BASE),
 };
 
 static const struct drm_i915_reg_descriptor hsw_render_regs[] = {
@@ -502,7 +508,10 @@ static const struct drm_i915_reg_descriptor hsw_render_regs[] = {
 };
 
 static const struct drm_i915_reg_descriptor gen7_blt_regs[] = {
+	REG64_IDX(RING_TIMESTAMP, RENDER_RING_BASE),
+	REG64_IDX(RING_TIMESTAMP, BSD_RING_BASE),
 	REG32(BCS_SWCTRL),
+	REG64_IDX(RING_TIMESTAMP, BLT_RING_BASE),
 };
 
 static const struct drm_i915_reg_descriptor ivb_master_regs[] = {
@@ -691,12 +700,26 @@ struct cmd_node {
  * non-opcode bits being set. But if we don't include those bits, some 3D
  * commands may hash to the same bucket due to not including opcode bits that
  * make the command unique. For now, we will risk hashing to the same bucket.
- *
- * If we attempt to generate a perfect hash, we should be able to look at bits
- * 31:29 of a command from a batch buffer and use the full mask for that
- * client. The existing INSTR_CLIENT_MASK/SHIFT defines can be used for this.
  */
-#define CMD_HASH_MASK STD_MI_OPCODE_MASK
+static inline u32 cmd_header_key(u32 x)
+{
+	u32 shift;
+
+	switch (x >> INSTR_CLIENT_SHIFT) {
+	default:
+	case INSTR_MI_CLIENT:
+		shift = STD_MI_OPCODE_SHIFT;
+		break;
+	case INSTR_RC_CLIENT:
+		shift = STD_3D_OPCODE_SHIFT;
+		break;
+	case INSTR_BC_CLIENT:
+		shift = STD_2D_OPCODE_SHIFT;
+		break;
+	}
+
+	return x >> shift;
+}
 
 static int init_hash_table(struct intel_engine_cs *engine,
 			   const struct drm_i915_cmd_table *cmd_tables,
@@ -720,7 +743,7 @@ static int init_hash_table(struct intel_engine_cs *engine,
 
 			desc_node->desc = desc;
 			hash_add(engine->cmd_hash, &desc_node->node,
-				 desc->cmd.value & CMD_HASH_MASK);
+				 cmd_header_key(desc->cmd.value));
 		}
 	}
 
@@ -746,17 +769,15 @@ static void fini_hash_table(struct intel_engine_cs *engine)
  * Optionally initializes fields related to batch buffer command parsing in the
  * struct intel_engine_cs based on whether the platform requires software
  * command parsing.
- *
- * Return: non-zero if initialization fails
  */
-int intel_engine_init_cmd_parser(struct intel_engine_cs *engine)
+void intel_engine_init_cmd_parser(struct intel_engine_cs *engine)
 {
 	const struct drm_i915_cmd_table *cmd_tables;
 	int cmd_table_count;
 	int ret;
 
 	if (!IS_GEN7(engine->i915))
-		return 0;
+		return;
 
 	switch (engine->id) {
 	case RCS:
@@ -811,24 +832,27 @@ int intel_engine_init_cmd_parser(struct intel_engine_cs *engine)
 		break;
 	default:
 		MISSING_CASE(engine->id);
-		BUG();
+		return;
 	}
 
-	BUG_ON(!validate_cmds_sorted(engine, cmd_tables, cmd_table_count));
-	BUG_ON(!validate_regs_sorted(engine));
-
-	WARN_ON(!hash_empty(engine->cmd_hash));
+	if (!validate_cmds_sorted(engine, cmd_tables, cmd_table_count)) {
+		DRM_ERROR("%s: command descriptions are not sorted\n",
+			  engine->name);
+		return;
+	}
+	if (!validate_regs_sorted(engine)) {
+		DRM_ERROR("%s: registers are not sorted\n", engine->name);
+		return;
+	}
 
 	ret = init_hash_table(engine, cmd_tables, cmd_table_count);
 	if (ret) {
-		DRM_ERROR("CMD: cmd_parser_init failed!\n");
+		DRM_ERROR("%s: initialised failed!\n", engine->name);
 		fini_hash_table(engine);
-		return ret;
+		return;
 	}
 
 	engine->needs_cmd_parser = true;
-
-	return 0;
 }
 
 /**
@@ -853,12 +877,9 @@ find_cmd_in_table(struct intel_engine_cs *engine,
 	struct cmd_node *desc_node;
 
 	hash_for_each_possible(engine->cmd_hash, desc_node, node,
-			       cmd_header & CMD_HASH_MASK) {
+			       cmd_header_key(cmd_header)) {
 		const struct drm_i915_cmd_descriptor *desc = desc_node->desc;
-		u32 masked_cmd = desc->cmd.mask & cmd_header;
-		u32 masked_value = desc->cmd.value & desc->cmd.mask;
-
-		if (masked_cmd == masked_value)
+		if (((cmd_header ^ desc->cmd.value) & desc->cmd.mask) == 0)
 			return desc;
 	}
 
@@ -876,11 +897,14 @@ find_cmd_in_table(struct intel_engine_cs *engine,
 static const struct drm_i915_cmd_descriptor*
 find_cmd(struct intel_engine_cs *engine,
 	 u32 cmd_header,
+	 const struct drm_i915_cmd_descriptor *desc,
 	 struct drm_i915_cmd_descriptor *default_desc)
 {
-	const struct drm_i915_cmd_descriptor *desc;
 	u32 mask;
 
+	if (((cmd_header ^ desc->cmd.value) & desc->cmd.mask) == 0)
+		return desc;
+
 	desc = find_cmd_in_table(engine, cmd_header);
 	if (desc)
 		return desc;
@@ -889,140 +913,127 @@ find_cmd(struct intel_engine_cs *engine,
 	if (!mask)
 		return NULL;
 
-	BUG_ON(!default_desc);
-	default_desc->flags = CMD_DESC_SKIP;
+	default_desc->cmd.value = cmd_header;
+	default_desc->cmd.mask = ~0u << MIN_OPCODE_SHIFT;
 	default_desc->length.mask = mask;
-
+	default_desc->flags = CMD_DESC_SKIP;
 	return default_desc;
 }
 
 static const struct drm_i915_reg_descriptor *
-find_reg(const struct drm_i915_reg_descriptor *table,
-	 int count, u32 addr)
+__find_reg(const struct drm_i915_reg_descriptor *table, int count, u32 addr)
 {
-	int i;
-
-	for (i = 0; i < count; i++) {
-		if (i915_mmio_reg_offset(table[i].addr) == addr)
-			return &table[i];
+	int start = 0, end = count;
+	while (start < end) {
+		int mid = start + (end - start) / 2;
+		int ret = addr - i915_mmio_reg_offset(table[mid].addr);
+		if (ret < 0)
+			end = mid;
+		else if (ret > 0)
+			start = mid + 1;
+		else
+			return &table[mid];
 	}
-
 	return NULL;
 }
 
 static const struct drm_i915_reg_descriptor *
-find_reg_in_tables(const struct drm_i915_reg_table *tables,
-		   int count, bool is_master, u32 addr)
+find_reg(const struct intel_engine_cs *engine, bool is_master, u32 addr)
 {
-	int i;
-	const struct drm_i915_reg_table *table;
-	const struct drm_i915_reg_descriptor *reg;
+	const struct drm_i915_reg_table *table = engine->reg_tables;
+	int count = engine->reg_table_count;
 
-	for (i = 0; i < count; i++) {
-		table = &tables[i];
+	do {
 		if (!table->master || is_master) {
-			reg = find_reg(table->regs, table->num_regs,
-				       addr);
+			const struct drm_i915_reg_descriptor *reg;
+
+			reg = __find_reg(table->regs, table->num_regs, addr);
 			if (reg != NULL)
 				return reg;
 		}
-	}
+	} while (table++, --count);
 
 	return NULL;
 }
 
-static u32 *vmap_batch(struct drm_i915_gem_object *obj,
-		       unsigned start, unsigned len)
-{
-	int i;
-	void *addr = NULL;
-	struct sg_page_iter sg_iter;
-	int first_page = start >> PAGE_SHIFT;
-	int last_page = (len + start + 4095) >> PAGE_SHIFT;
-	int npages = last_page - first_page;
-	struct page **pages;
-
-	pages = drm_malloc_ab(npages, sizeof(*pages));
-	if (pages == NULL) {
-		DRM_DEBUG_DRIVER("Failed to get space for pages\n");
-		goto finish;
-	}
-
-	i = 0;
-	for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents, first_page) {
-		pages[i++] = sg_page_iter_page(&sg_iter);
-		if (i == npages)
-			break;
-	}
-
-	addr = vmap(pages, i, 0, PAGE_KERNEL);
-	if (addr == NULL) {
-		DRM_DEBUG_DRIVER("Failed to vmap pages\n");
-		goto finish;
-	}
-
-finish:
-	if (pages)
-		drm_free_large(pages);
-	return (u32*)addr;
-}
-
-/* Returns a vmap'd pointer to dest_obj, which the caller must unmap */
-static u32 *copy_batch(struct drm_i915_gem_object *dest_obj,
+/* Returns a vmap'd pointer to dst_obj, which the caller must unmap */
+static u32 *copy_batch(struct drm_i915_gem_object *dst_obj,
 		       struct drm_i915_gem_object *src_obj,
 		       u32 batch_start_offset,
-		       u32 batch_len)
+		       u32 batch_len,
+		       bool *needs_clflush_after)
 {
-	int needs_clflush = 0;
-	void *src_base, *src;
-	void *dst = NULL;
+	unsigned int src_needs_clflush;
+	unsigned int dst_needs_clflush;
+	void *dst, *src;
 	int ret;
 
-	if (batch_len > dest_obj->base.size ||
-	    batch_len + batch_start_offset > src_obj->base.size)
-		return ERR_PTR(-E2BIG);
-
-	if (WARN_ON(dest_obj->pages_pin_count == 0))
-		return ERR_PTR(-ENODEV);
-
-	ret = i915_gem_obj_prepare_shmem_read(src_obj, &needs_clflush);
-	if (ret) {
-		DRM_DEBUG_DRIVER("CMD: failed to prepare shadow batch\n");
+	ret = i915_gem_obj_prepare_shmem_read(src_obj, &src_needs_clflush);
+	if (ret)
 		return ERR_PTR(ret);
-	}
 
-	src_base = vmap_batch(src_obj, batch_start_offset, batch_len);
-	if (!src_base) {
-		DRM_DEBUG_DRIVER("CMD: Failed to vmap batch\n");
-		ret = -ENOMEM;
+	ret = i915_gem_obj_prepare_shmem_write(dst_obj, &dst_needs_clflush);
+	if (ret) {
+		dst = ERR_PTR(ret);
 		goto unpin_src;
 	}
 
-	ret = i915_gem_object_set_to_cpu_domain(dest_obj, true);
-	if (ret) {
-		DRM_DEBUG_DRIVER("CMD: Failed to set shadow batch to CPU\n");
-		goto unmap_src;
+	dst = i915_gem_object_pin_map(dst_obj, I915_MAP_WB);
+	if (IS_ERR(dst))
+		goto unpin_dst;
+
+	src = ERR_PTR(-ENODEV);
+	if (src_needs_clflush &&
+	    i915_memcpy_from_wc((void *)(uintptr_t)batch_start_offset, 0, 0)) {
+		src = i915_gem_object_pin_map(src_obj, I915_MAP_WC);
+		if (!IS_ERR(src)) {
+			i915_memcpy_from_wc(dst,
+					    src + batch_start_offset,
+					    ALIGN(batch_len, 16));
+			i915_gem_object_unpin_map(src_obj);
+		}
 	}
-
-	dst = vmap_batch(dest_obj, 0, batch_len);
-	if (!dst) {
-		DRM_DEBUG_DRIVER("CMD: Failed to vmap shadow batch\n");
-		ret = -ENOMEM;
-		goto unmap_src;
+	if (IS_ERR(src)) {
+		void *ptr;
+		int offset, n;
+
+		offset = offset_in_page(batch_start_offset);
+
+		/* We can avoid clflushing partial cachelines before the write
+		 * if we only every write full cache-lines. Since we know that
+		 * both the source and destination are in multiples of
+		 * PAGE_SIZE, we can simply round up to the next cacheline.
+		 * We don't care about copying too much here as we only
+		 * validate up to the end of the batch.
+		 */
+		if (dst_needs_clflush & CLFLUSH_BEFORE)
+			batch_len = roundup(batch_len,
+					    boot_cpu_data.x86_clflush_size);
+
+		ptr = dst;
+		for (n = batch_start_offset >> PAGE_SHIFT; batch_len; n++) {
+			int len = min_t(int, batch_len, PAGE_SIZE - offset);
+
+			src = kmap_atomic(i915_gem_object_get_page(src_obj, n));
+			if (src_needs_clflush)
+				drm_clflush_virt_range(src + offset, len);
+			memcpy(ptr, src + offset, len);
+			kunmap_atomic(src);
+
+			ptr += len;
+			batch_len -= len;
+			offset = 0;
+		}
 	}
 
-	src = src_base + offset_in_page(batch_start_offset);
-	if (needs_clflush)
-		drm_clflush_virt_range(src, batch_len);
-
-	memcpy(dst, src, batch_len);
+	/* dst_obj is returned with vmap pinned */
+	*needs_clflush_after = dst_needs_clflush & CLFLUSH_AFTER;
 
-unmap_src:
-	vunmap(src_base);
+unpin_dst:
+	i915_gem_obj_finish_shmem_access(dst_obj);
 unpin_src:
-	i915_gem_object_unpin_pages(src_obj);
-
-	return ret ? ERR_PTR(ret) : dst;
+	i915_gem_obj_finish_shmem_access(src_obj);
+	return dst;
 }
 
 /**
@@ -1052,6 +1063,9 @@ static bool check_cmd(const struct intel_engine_cs *engine,
 		      const bool is_master,
 		      bool *oacontrol_set)
 {
+	if (desc->flags & CMD_DESC_SKIP)
+		return true;
+
 	if (desc->flags & CMD_DESC_REJECT) {
 		DRM_DEBUG_DRIVER("CMD: Rejected command: 0x%08X\n", *cmd);
 		return false;
@@ -1076,10 +1090,7 @@ static bool check_cmd(const struct intel_engine_cs *engine,
 		     offset += step) {
 			const u32 reg_addr = cmd[offset] & desc->reg.mask;
 			const struct drm_i915_reg_descriptor *reg =
-				find_reg_in_tables(engine->reg_tables,
-						   engine->reg_table_count,
-						   is_master,
-						   reg_addr);
+				find_reg(engine, is_master, reg_addr);
 
 			if (!reg) {
 				DRM_DEBUG_DRIVER("CMD: Rejected register 0x%08X in command: 0x%08X (exec_id=%d)\n",
@@ -1200,16 +1211,19 @@ int intel_engine_cmd_parser(struct intel_engine_cs *engine,
 			    u32 batch_len,
 			    bool is_master)
 {
-	u32 *cmd, *batch_base, *batch_end;
-	struct drm_i915_cmd_descriptor default_desc = { 0 };
+	u32 *cmd, *batch_end;
+	struct drm_i915_cmd_descriptor default_desc = noop_desc;
+	const struct drm_i915_cmd_descriptor *desc = &default_desc;
 	bool oacontrol_set = false; /* OACONTROL tracking. See check_cmd() */
+	bool needs_clflush_after = false;
 	int ret = 0;
 
-	batch_base = copy_batch(shadow_batch_obj, batch_obj,
-				batch_start_offset, batch_len);
-	if (IS_ERR(batch_base)) {
+	cmd = copy_batch(shadow_batch_obj, batch_obj,
+			 batch_start_offset, batch_len,
+			 &needs_clflush_after);
+	if (IS_ERR(cmd)) {
 		DRM_DEBUG_DRIVER("CMD: Failed to copy batch\n");
-		return PTR_ERR(batch_base);
+		return PTR_ERR(cmd);
 	}
 
 	/*
@@ -1217,17 +1231,14 @@ int intel_engine_cmd_parser(struct intel_engine_cs *engine,
 	 * large or larger and copy_batch() will write MI_NOPs to the extra
 	 * space. Parsing should be faster in some cases this way.
 	 */
-	batch_end = batch_base + (batch_len / sizeof(*batch_end));
-
-	cmd = batch_base;
+	batch_end = cmd + (batch_len / sizeof(*batch_end));
 	while (cmd < batch_end) {
-		const struct drm_i915_cmd_descriptor *desc;
 		u32 length;
 
 		if (*cmd == MI_BATCH_BUFFER_END)
 			break;
 
-		desc = find_cmd(engine, *cmd, &default_desc);
+		desc = find_cmd(engine, *cmd, desc, &default_desc);
 		if (!desc) {
 			DRM_DEBUG_DRIVER("CMD: Unrecognized command: 0x%08X\n",
 					 *cmd);
@@ -1278,7 +1289,9 @@ int intel_engine_cmd_parser(struct intel_engine_cs *engine,
 		ret = -EINVAL;
 	}
 
-	vunmap(batch_base);
+	if (ret == 0 && needs_clflush_after)
+		drm_clflush_virt_range(shadow_batch_obj->mapping, batch_len);
+	i915_gem_object_unpin_map(shadow_batch_obj);
 
 	return ret;
 }

drivers/gpu/drm/i915/i915_drv.c

@@ -827,6 +827,8 @@ static int i915_driver_init_early(struct drm_i915_private *dev_priv,
 	mutex_init(&dev_priv->wm.wm_mutex);
 	mutex_init(&dev_priv->pps_mutex);
 
+	i915_memcpy_init_early(dev_priv);
+
 	ret = i915_workqueues_init(dev_priv);
 	if (ret < 0)
 		return ret;
@@ -1560,6 +1562,7 @@ static int i915_drm_resume(struct drm_device *dev)
 	i915_gem_resume(dev);
 
 	i915_restore_state(dev);
+	intel_pps_unlock_regs_wa(dev_priv);
 	intel_opregion_setup(dev_priv);
 
 	intel_init_pch_refclk(dev);

drivers/gpu/drm/i915/i915_gem_context.c

@@ -155,9 +155,10 @@ void i915_gem_context_free(struct kref *ctx_ref)
 		if (ce->ring)
 			intel_ring_free(ce->ring);
 
-		i915_gem_object_put(ce->state);
+		i915_vma_put(ce->state);
 	}
 
+	put_pid(ctx->pid);
 	list_del(&ctx->link);
 
 	ida_simple_remove(&ctx->i915->context_hw_ida, ctx->hw_id);
@@ -281,13 +282,24 @@ __create_hw_context(struct drm_device *dev,
 	ctx->ggtt_alignment = get_context_alignment(dev_priv);
 
 	if (dev_priv->hw_context_size) {
-		struct drm_i915_gem_object *obj =
-				i915_gem_alloc_context_obj(dev, dev_priv->hw_context_size);
+		struct drm_i915_gem_object *obj;
+		struct i915_vma *vma;
+
+		obj = i915_gem_alloc_context_obj(dev,
+						 dev_priv->hw_context_size);
 		if (IS_ERR(obj)) {
 			ret = PTR_ERR(obj);
 			goto err_out;
 		}
-		ctx->engine[RCS].state = obj;
+
+		vma = i915_vma_create(obj, &dev_priv->ggtt.base, NULL);
+		if (IS_ERR(vma)) {
+			i915_gem_object_put(obj);
+			ret = PTR_ERR(vma);
+			goto err_out;
+		}
+
+		ctx->engine[RCS].state = vma;
 	}
 
 	/* Default context will never have a file_priv */
@@ -300,6 +312,9 @@ __create_hw_context(struct drm_device *dev,
 		ret = DEFAULT_CONTEXT_HANDLE;
 
 	ctx->file_priv = file_priv;
+	if (file_priv)
+		ctx->pid = get_task_pid(current, PIDTYPE_PID);
+
 	ctx->user_handle = ret;
 	/* NB: Mark all slices as needing a remap so that when the context first
 	 * loads it will restore whatever remap state already exists. If there
@@ -399,7 +414,7 @@ static void i915_gem_context_unpin(struct i915_gem_context *ctx,
 		struct intel_context *ce = &ctx->engine[engine->id];
 
 		if (ce->state)
-			i915_gem_object_ggtt_unpin(ce->state);
+			i915_vma_unpin(ce->state);
 
 		i915_gem_context_put(ctx);
 	}
@@ -568,7 +583,7 @@ mi_set_context(struct drm_i915_gem_request *req, u32 hw_flags)
 	const int num_rings =
 		/* Use an extended w/a on ivb+ if signalling from other rings */
 		i915.semaphores ?
-		hweight32(INTEL_INFO(dev_priv)->ring_mask) - 1 :
+		INTEL_INFO(dev_priv)->num_rings - 1 :
 		0;
 	int len, ret;
 
@@ -621,8 +636,7 @@ mi_set_context(struct drm_i915_gem_request *req, u32 hw_flags)
 	intel_ring_emit(ring, MI_NOOP);
 	intel_ring_emit(ring, MI_SET_CONTEXT);
 	intel_ring_emit(ring,
-			i915_gem_obj_ggtt_offset(req->ctx->engine[RCS].state) |
-			flags);
+			i915_ggtt_offset(req->ctx->engine[RCS].state) | flags);
 	/*
 	 * w/a: MI_SET_CONTEXT must always be followed by MI_NOOP
 	 * WaMiSetContext_Hang:snb,ivb,vlv
@@ -651,7 +665,8 @@ mi_set_context(struct drm_i915_gem_request *req, u32 hw_flags)
 					MI_STORE_REGISTER_MEM |
 					MI_SRM_LRM_GLOBAL_GTT);
 			intel_ring_emit_reg(ring, last_reg);
-			intel_ring_emit(ring, engine->scratch.gtt_offset);
+			intel_ring_emit(ring,
+					i915_ggtt_offset(engine->scratch));
 			intel_ring_emit(ring, MI_NOOP);
 		}
 		intel_ring_emit(ring, MI_ARB_ON_OFF | MI_ARB_ENABLE);
@@ -755,6 +770,7 @@ static int do_rcs_switch(struct drm_i915_gem_request *req)
 	struct i915_gem_context *to = req->ctx;
 	struct intel_engine_cs *engine = req->engine;
 	struct i915_hw_ppgtt *ppgtt = to->ppgtt ?: req->i915->mm.aliasing_ppgtt;
+	struct i915_vma *vma = to->engine[RCS].state;
 	struct i915_gem_context *from;
 	u32 hw_flags;
 	int ret, i;
@@ -762,9 +778,15 @@ static int do_rcs_switch(struct drm_i915_gem_request *req)
 	if (skip_rcs_switch(ppgtt, engine, to))
 		return 0;
 
+	/* Clear this page out of any CPU caches for coherent swap-in/out. */
+	if (!(vma->flags & I915_VMA_GLOBAL_BIND)) {
+		ret = i915_gem_object_set_to_gtt_domain(vma->obj, false);
+		if (ret)
+			return ret;
+	}
+
 	/* Trying to pin first makes error handling easier. */
-	ret = i915_gem_object_ggtt_pin(to->engine[RCS].state, NULL, 0,
-				       to->ggtt_alignment, 0);
+	ret = i915_vma_pin(vma, 0, to->ggtt_alignment, PIN_GLOBAL);
 	if (ret)
 		return ret;
 
@@ -777,18 +799,6 @@ static int do_rcs_switch(struct drm_i915_gem_request *req)
 	 */
 	from = engine->last_context;
 
-	/*
-	 * Clear this page out of any CPU caches for coherent swap-in/out. Note
-	 * that thanks to write = false in this call and us not setting any gpu
-	 * write domains when putting a context object onto the active list
-	 * (when switching away from it), this won't block.
-	 *
-	 * XXX: We need a real interface to do this instead of trickery.
-	 */
-	ret = i915_gem_object_set_to_gtt_domain(to->engine[RCS].state, false);
-	if (ret)
-		goto unpin_out;
-
 	if (needs_pd_load_pre(ppgtt, engine, to)) {
 		/* Older GENs and non render rings still want the load first,
 		 * "PP_DCLV followed by PP_DIR_BASE register through Load
@@ -797,7 +807,7 @@ static int do_rcs_switch(struct drm_i915_gem_request *req)
 		trace_switch_mm(engine, to);
 		ret = ppgtt->switch_mm(ppgtt, req);
 		if (ret)
-			goto unpin_out;
+			goto err;
 	}
 
 	if (!to->engine[RCS].initialised || i915_gem_context_is_default(to))
@@ -814,7 +824,7 @@ static int do_rcs_switch(struct drm_i915_gem_request *req)
 	if (to != from || (hw_flags & MI_FORCE_RESTORE)) {
 		ret = mi_set_context(req, hw_flags);
 		if (ret)
-			goto unpin_out;
+			goto err;
 	}
 
 	/* The backing object for the context is done after switching to the
@@ -824,8 +834,6 @@ static int do_rcs_switch(struct drm_i915_gem_request *req)
 	 * MI_SET_CONTEXT instead of when the next seqno has completed.
 	 */
 	if (from != NULL) {
-		struct drm_i915_gem_object *obj = from->engine[RCS].state;
-
 		/* As long as MI_SET_CONTEXT is serializing, ie. it flushes the
 		 * whole damn pipeline, we don't need to explicitly mark the
 		 * object dirty. The only exception is that the context must be
@@ -833,11 +841,9 @@ static int do_rcs_switch(struct drm_i915_gem_request *req)
 		 * able to defer doing this until we know the object would be
 		 * swapped, but there is no way to do that yet.
 		 */
-		obj->base.read_domains = I915_GEM_DOMAIN_INSTRUCTION;
-		i915_vma_move_to_active(i915_gem_obj_to_ggtt(obj), req, 0);
-
-		/* obj is kept alive until the next request by its active ref */
-		i915_gem_object_ggtt_unpin(obj);
+		i915_vma_move_to_active(from->engine[RCS].state, req, 0);
+		/* state is kept alive until the next request */
+		i915_vma_unpin(from->engine[RCS].state);
 		i915_gem_context_put(from);
 	}
 	engine->last_context = i915_gem_context_get(to);
@@ -882,8 +888,8 @@ static int do_rcs_switch(struct drm_i915_gem_request *req)
 
 	return 0;
 
-unpin_out:
-	i915_gem_object_ggtt_unpin(to->engine[RCS].state);
+err:
+	i915_vma_unpin(vma);
 	return ret;
 }
 

drivers/gpu/drm/i915/i915_gem_dmabuf.c

@@ -119,7 +119,7 @@ static void *i915_gem_dmabuf_vmap(struct dma_buf *dma_buf)
 	if (ret)
 		return ERR_PTR(ret);
 
-	addr = i915_gem_object_pin_map(obj);
+	addr = i915_gem_object_pin_map(obj, I915_MAP_WB);
 	mutex_unlock(&dev->struct_mutex);
 
 	return addr;

drivers/gpu/drm/i915/i915_gem_evict.c

@@ -47,7 +47,7 @@ gpu_is_idle(struct drm_i915_private *dev_priv)
 }
 
 static bool
-mark_free(struct i915_vma *vma, struct list_head *unwind)
+mark_free(struct i915_vma *vma, unsigned int flags, struct list_head *unwind)
 {
 	if (i915_vma_is_pinned(vma))
 		return false;
@@ -55,6 +55,9 @@ mark_free(struct i915_vma *vma, struct list_head *unwind)
 	if (WARN_ON(!list_empty(&vma->exec_list)))
 		return false;
 
+	if (flags & PIN_NONFAULT && vma->obj->fault_mappable)
+		return false;
+
 	list_add(&vma->exec_list, unwind);
 	return drm_mm_scan_add_block(&vma->node);
 }
@@ -129,7 +132,7 @@ i915_gem_evict_something(struct i915_address_space *vm,
 	phase = phases;
 	do {
 		list_for_each_entry(vma, *phase, vm_link)
-			if (mark_free(vma, &eviction_list))
+			if (mark_free(vma, flags, &eviction_list))
 				goto found;
 	} while (*++phase);
 

drivers/gpu/drm/i915/i915_gem_fence.c

@@ -55,87 +55,85 @@
  * CPU ptes into GTT mmaps (not the GTT ptes themselves) as needed.
  */
 
-static void i965_write_fence_reg(struct drm_device *dev, int reg,
-				 struct drm_i915_gem_object *obj)
+#define pipelined 0
+
+static void i965_write_fence_reg(struct drm_i915_fence_reg *fence,
+				 struct i915_vma *vma)
 {
-	struct drm_i915_private *dev_priv = to_i915(dev);
 	i915_reg_t fence_reg_lo, fence_reg_hi;
 	int fence_pitch_shift;
+	u64 val;
 
-	if (INTEL_INFO(dev)->gen >= 6) {
-		fence_reg_lo = FENCE_REG_GEN6_LO(reg);
-		fence_reg_hi = FENCE_REG_GEN6_HI(reg);
+	if (INTEL_INFO(fence->i915)->gen >= 6) {
+		fence_reg_lo = FENCE_REG_GEN6_LO(fence->id);
+		fence_reg_hi = FENCE_REG_GEN6_HI(fence->id);
 		fence_pitch_shift = GEN6_FENCE_PITCH_SHIFT;
+
 	} else {
-		fence_reg_lo = FENCE_REG_965_LO(reg);
-		fence_reg_hi = FENCE_REG_965_HI(reg);
+		fence_reg_lo = FENCE_REG_965_LO(fence->id);
+		fence_reg_hi = FENCE_REG_965_HI(fence->id);
 		fence_pitch_shift = I965_FENCE_PITCH_SHIFT;
 	}
 
-	/* To w/a incoherency with non-atomic 64-bit register updates,
-	 * we split the 64-bit update into two 32-bit writes. In order
-	 * for a partial fence not to be evaluated between writes, we
-	 * precede the update with write to turn off the fence register,
-	 * and only enable the fence as the last step.
-	 *
-	 * For extra levels of paranoia, we make sure each step lands
-	 * before applying the next step.
-	 */
-	I915_WRITE(fence_reg_lo, 0);
-	POSTING_READ(fence_reg_lo);
-
-	if (obj) {
-		u32 size = i915_gem_obj_ggtt_size(obj);
-		unsigned int tiling = i915_gem_object_get_tiling(obj);
-		unsigned int stride = i915_gem_object_get_stride(obj);
-		uint64_t val;
-
-		/* Adjust fence size to match tiled area */
-		if (tiling != I915_TILING_NONE) {
-			uint32_t row_size = stride *
-				(tiling == I915_TILING_Y ? 32 : 8);
-			size = (size / row_size) * row_size;
-		}
-
-		val = (uint64_t)((i915_gem_obj_ggtt_offset(obj) + size - 4096) &
-				 0xfffff000) << 32;
-		val |= i915_gem_obj_ggtt_offset(obj) & 0xfffff000;
-		val |= (uint64_t)((stride / 128) - 1) << fence_pitch_shift;
-		if (tiling == I915_TILING_Y)
-			val |= 1 << I965_FENCE_TILING_Y_SHIFT;
+	val = 0;
+	if (vma) {
+		unsigned int tiling = i915_gem_object_get_tiling(vma->obj);
+		bool is_y_tiled = tiling == I915_TILING_Y;
+		unsigned int stride = i915_gem_object_get_stride(vma->obj);
+		u32 row_size = stride * (is_y_tiled ? 32 : 8);
+		u32 size = rounddown((u32)vma->node.size, row_size);
+
+		val = ((vma->node.start + size - 4096) & 0xfffff000) << 32;
+		val |= vma->node.start & 0xfffff000;
+		val |= (u64)((stride / 128) - 1) << fence_pitch_shift;
+		if (is_y_tiled)
+			val |= BIT(I965_FENCE_TILING_Y_SHIFT);
 		val |= I965_FENCE_REG_VALID;
+	}
 
-		I915_WRITE(fence_reg_hi, val >> 32);
-		POSTING_READ(fence_reg_hi);
+	if (!pipelined) {
+		struct drm_i915_private *dev_priv = fence->i915;
 
-		I915_WRITE(fence_reg_lo, val);
+		/* To w/a incoherency with non-atomic 64-bit register updates,
+		 * we split the 64-bit update into two 32-bit writes. In order
+		 * for a partial fence not to be evaluated between writes, we
+		 * precede the update with write to turn off the fence register,
+		 * and only enable the fence as the last step.
+		 *
+		 * For extra levels of paranoia, we make sure each step lands
+		 * before applying the next step.
+		 */
+		I915_WRITE(fence_reg_lo, 0);
+		POSTING_READ(fence_reg_lo);
+
+		I915_WRITE(fence_reg_hi, upper_32_bits(val));
+		I915_WRITE(fence_reg_lo, lower_32_bits(val));
 		POSTING_READ(fence_reg_lo);
-	} else {
-		I915_WRITE(fence_reg_hi, 0);
-		POSTING_READ(fence_reg_hi);
 	}
 }
 
-static void i915_write_fence_reg(struct drm_device *dev, int reg,
-				 struct drm_i915_gem_object *obj)
+static void i915_write_fence_reg(struct drm_i915_fence_reg *fence,
+				 struct i915_vma *vma)
 {
-	struct drm_i915_private *dev_priv = to_i915(dev);
 	u32 val;
 
-	if (obj) {
-		u32 size = i915_gem_obj_ggtt_size(obj);
-		unsigned int tiling = i915_gem_object_get_tiling(obj);
-		unsigned int stride = i915_gem_object_get_stride(obj);
+	val = 0;
+	if (vma) {
+		unsigned int tiling = i915_gem_object_get_tiling(vma->obj);
+		bool is_y_tiled = tiling == I915_TILING_Y;
+		unsigned int stride = i915_gem_object_get_stride(vma->obj);
 		int pitch_val;
 		int tile_width;
 
-		WARN((i915_gem_obj_ggtt_offset(obj) & ~I915_FENCE_START_MASK) ||
-		     (size & -size) != size ||
-		     (i915_gem_obj_ggtt_offset(obj) & (size - 1)),
-		     "object 0x%08llx [fenceable? %d] not 1M or pot-size (0x%08x) aligned\n",
-		     i915_gem_obj_ggtt_offset(obj), obj->map_and_fenceable, size);
+		WARN((vma->node.start & ~I915_FENCE_START_MASK) ||
+		     !is_power_of_2(vma->node.size) ||
+		     (vma->node.start & (vma->node.size - 1)),
+		     "object 0x%08llx [fenceable? %d] not 1M or pot-size (0x%08llx) aligned\n",
+		     vma->node.start,
+		     i915_vma_is_map_and_fenceable(vma),
+		     vma->node.size);
 
-		if (tiling == I915_TILING_Y && HAS_128_BYTE_Y_TILING(dev))
+		if (is_y_tiled && HAS_128_BYTE_Y_TILING(fence->i915))
 			tile_width = 128;
 		else
 			tile_width = 512;
@@ -144,139 +142,141 @@ static void i915_write_fence_reg(struct drm_device *dev, int reg,
 		pitch_val = stride / tile_width;
 		pitch_val = ffs(pitch_val) - 1;
 
-		val = i915_gem_obj_ggtt_offset(obj);
-		if (tiling == I915_TILING_Y)
-			val |= 1 << I830_FENCE_TILING_Y_SHIFT;
-		val |= I915_FENCE_SIZE_BITS(size);
+		val = vma->node.start;
+		if (is_y_tiled)
+			val |= BIT(I830_FENCE_TILING_Y_SHIFT);
+		val |= I915_FENCE_SIZE_BITS(vma->node.size);
 		val |= pitch_val << I830_FENCE_PITCH_SHIFT;
 		val |= I830_FENCE_REG_VALID;
-	} else
-		val = 0;
+	}
+
+	if (!pipelined) {
+		struct drm_i915_private *dev_priv = fence->i915;
+		i915_reg_t reg = FENCE_REG(fence->id);
 
-	I915_WRITE(FENCE_REG(reg), val);
-	POSTING_READ(FENCE_REG(reg));
+		I915_WRITE(reg, val);
+		POSTING_READ(reg);
+	}
 }
 
-static void i830_write_fence_reg(struct drm_device *dev, int reg,
-				struct drm_i915_gem_object *obj)
+static void i830_write_fence_reg(struct drm_i915_fence_reg *fence,
+				 struct i915_vma *vma)
 {
-	struct drm_i915_private *dev_priv = to_i915(dev);
-	uint32_t val;
+	u32 val;
 
-	if (obj) {
-		u32 size = i915_gem_obj_ggtt_size(obj);
-		unsigned int tiling = i915_gem_object_get_tiling(obj);
-		unsigned int stride = i915_gem_object_get_stride(obj);
-		uint32_t pitch_val;
+	val = 0;
+	if (vma) {
+		unsigned int tiling = i915_gem_object_get_tiling(vma->obj);
+		bool is_y_tiled = tiling == I915_TILING_Y;
+		unsigned int stride = i915_gem_object_get_stride(vma->obj);
+		u32 pitch_val;
 
-		WARN((i915_gem_obj_ggtt_offset(obj) & ~I830_FENCE_START_MASK) ||
-		     (size & -size) != size ||
-		     (i915_gem_obj_ggtt_offset(obj) & (size - 1)),
-		     "object 0x%08llx not 512K or pot-size 0x%08x aligned\n",
-		     i915_gem_obj_ggtt_offset(obj), size);
+		WARN((vma->node.start & ~I830_FENCE_START_MASK) ||
+		     !is_power_of_2(vma->node.size) ||
+		     (vma->node.start & (vma->node.size - 1)),
+		     "object 0x%08llx not 512K or pot-size 0x%08llx aligned\n",
+		     vma->node.start, vma->node.size);
 
 		pitch_val = stride / 128;
 		pitch_val = ffs(pitch_val) - 1;
 
-		val = i915_gem_obj_ggtt_offset(obj);
-		if (tiling == I915_TILING_Y)
-			val |= 1 << I830_FENCE_TILING_Y_SHIFT;
-		val |= I830_FENCE_SIZE_BITS(size);
+		val = vma->node.start;
+		if (is_y_tiled)
+			val |= BIT(I830_FENCE_TILING_Y_SHIFT);
+		val |= I830_FENCE_SIZE_BITS(vma->node.size);
 		val |= pitch_val << I830_FENCE_PITCH_SHIFT;
 		val |= I830_FENCE_REG_VALID;
-	} else
-		val = 0;
+	}
 
-	I915_WRITE(FENCE_REG(reg), val);
-	POSTING_READ(FENCE_REG(reg));
-}
+	if (!pipelined) {
+		struct drm_i915_private *dev_priv = fence->i915;
+		i915_reg_t reg = FENCE_REG(fence->id);
 
-inline static bool i915_gem_object_needs_mb(struct drm_i915_gem_object *obj)
-{
-	return obj && obj->base.read_domains & I915_GEM_DOMAIN_GTT;
+		I915_WRITE(reg, val);
+		POSTING_READ(reg);
+	}
 }
 
-static void i915_gem_write_fence(struct drm_device *dev, int reg,
-				 struct drm_i915_gem_object *obj)
+static void fence_write(struct drm_i915_fence_reg *fence,
+			struct i915_vma *vma)
 {
-	struct drm_i915_private *dev_priv = to_i915(dev);
-
-	/* Ensure that all CPU reads are completed before installing a fence
-	 * and all writes before removing the fence.
+	/* Previous access through the fence register is marshalled by
+	 * the mb() inside the fault handlers (i915_gem_release_mmaps)
+	 * and explicitly managed for internal users.
 	 */
-	if (i915_gem_object_needs_mb(dev_priv->fence_regs[reg].obj))
-		mb();
-
-	WARN(obj &&
-	     (!i915_gem_object_get_stride(obj) ||
-	      !i915_gem_object_get_tiling(obj)),
-	     "bogus fence setup with stride: 0x%x, tiling mode: %i\n",
-	     i915_gem_object_get_stride(obj),
-	     i915_gem_object_get_tiling(obj));
-
-	if (IS_GEN2(dev))
-		i830_write_fence_reg(dev, reg, obj);
-	else if (IS_GEN3(dev))
-		i915_write_fence_reg(dev, reg, obj);
-	else if (INTEL_INFO(dev)->gen >= 4)
-		i965_write_fence_reg(dev, reg, obj);
-
-	/* And similarly be paranoid that no direct access to this region
-	 * is reordered to before the fence is installed.
+
+	if (IS_GEN2(fence->i915))
+		i830_write_fence_reg(fence, vma);
+	else if (IS_GEN3(fence->i915))
+		i915_write_fence_reg(fence, vma);
+	else
+		i965_write_fence_reg(fence, vma);
+
+	/* Access through the fenced region afterwards is
+	 * ordered by the posting reads whilst writing the registers.
 	 */
-	if (i915_gem_object_needs_mb(obj))
-		mb();
-}
 
-static inline int fence_number(struct drm_i915_private *dev_priv,
-			       struct drm_i915_fence_reg *fence)
-{
-	return fence - dev_priv->fence_regs;
+	fence->dirty = false;
 }
 
-static void i915_gem_object_update_fence(struct drm_i915_gem_object *obj,
-					 struct drm_i915_fence_reg *fence,
-					 bool enable)
+static int fence_update(struct drm_i915_fence_reg *fence,
+			struct i915_vma *vma)
 {
-	struct drm_i915_private *dev_priv = to_i915(obj->base.dev);
-	int reg = fence_number(dev_priv, fence);
+	int ret;
 
-	i915_gem_write_fence(obj->base.dev, reg, enable ? obj : NULL);
+	if (vma) {
+		if (!i915_vma_is_map_and_fenceable(vma))
+			return -EINVAL;
 
-	if (enable) {
-		obj->fence_reg = reg;
-		fence->obj = obj;
-		list_move_tail(&fence->lru_list, &dev_priv->mm.fence_list);
-	} else {
-		obj->fence_reg = I915_FENCE_REG_NONE;
-		fence->obj = NULL;
-		list_del_init(&fence->lru_list);
+		if (WARN(!i915_gem_object_get_stride(vma->obj) ||
+			 !i915_gem_object_get_tiling(vma->obj),
+			 "bogus fence setup with stride: 0x%x, tiling mode: %i\n",
+			 i915_gem_object_get_stride(vma->obj),
+			 i915_gem_object_get_tiling(vma->obj)))
+			return -EINVAL;
+
+		ret = i915_gem_active_retire(&vma->last_fence,
+					     &vma->obj->base.dev->struct_mutex);
+		if (ret)
+			return ret;
 	}
-	obj->fence_dirty = false;
-}
 
-static inline void i915_gem_object_fence_lost(struct drm_i915_gem_object *obj)
-{
-	if (i915_gem_object_is_tiled(obj))
-		i915_gem_release_mmap(obj);
+	if (fence->vma) {
+		ret = i915_gem_active_retire(&fence->vma->last_fence,
+				      &fence->vma->obj->base.dev->struct_mutex);
+		if (ret)
+			return ret;
+	}
 
-	/* As we do not have an associated fence register, we will force
-	 * a tiling change if we ever need to acquire one.
-	 */
-	obj->fence_dirty = false;
-	obj->fence_reg = I915_FENCE_REG_NONE;
-}
+	if (fence->vma && fence->vma != vma) {
+		/* Ensure that all userspace CPU access is completed before
+		 * stealing the fence.
+		 */
+		i915_gem_release_mmap(fence->vma->obj);
 
-static int
-i915_gem_object_wait_fence(struct drm_i915_gem_object *obj)
-{
-	return i915_gem_active_retire(&obj->last_fence,
-				      &obj->base.dev->struct_mutex);
+		fence->vma->fence = NULL;
+		fence->vma = NULL;
+
+		list_move(&fence->link, &fence->i915->mm.fence_list);
+	}
+
+	fence_write(fence, vma);
+
+	if (vma) {
+		if (fence->vma != vma) {
+			vma->fence = fence;
+			fence->vma = vma;
+		}
+
+		list_move_tail(&fence->link, &fence->i915->mm.fence_list);
+	}
+
+	return 0;
 }
 
 /**
- * i915_gem_object_put_fence - force-remove fence for an object
- * @obj: object to map through a fence reg
+ * i915_vma_put_fence - force-remove fence for a VMA
+ * @vma: vma to map linearly (not through a fence reg)
  *
  * This function force-removes any fence from the given object, which is useful
  * if the kernel wants to do untiled GTT access.
@@ -286,70 +286,40 @@ i915_gem_object_wait_fence(struct drm_i915_gem_object *obj)
  * 0 on success, negative error code on failure.
  */
 int
-i915_gem_object_put_fence(struct drm_i915_gem_object *obj)
+i915_vma_put_fence(struct i915_vma *vma)
 {
-	struct drm_i915_private *dev_priv = to_i915(obj->base.dev);
-	struct drm_i915_fence_reg *fence;
-	int ret;
-
-	ret = i915_gem_object_wait_fence(obj);
-	if (ret)
-		return ret;
+	struct drm_i915_fence_reg *fence = vma->fence;
 
-	if (obj->fence_reg == I915_FENCE_REG_NONE)
+	if (!fence)
 		return 0;
 
-	fence = &dev_priv->fence_regs[obj->fence_reg];
-
-	if (WARN_ON(fence->pin_count))
+	if (fence->pin_count)
 		return -EBUSY;
 
-	i915_gem_object_fence_lost(obj);
-	i915_gem_object_update_fence(obj, fence, false);
-
-	return 0;
+	return fence_update(fence, NULL);
 }
 
-static struct drm_i915_fence_reg *
-i915_find_fence_reg(struct drm_device *dev)
+static struct drm_i915_fence_reg *fence_find(struct drm_i915_private *dev_priv)
 {
-	struct drm_i915_private *dev_priv = to_i915(dev);
-	struct drm_i915_fence_reg *reg, *avail;
-	int i;
-
-	/* First try to find a free reg */
-	avail = NULL;
-	for (i = 0; i < dev_priv->num_fence_regs; i++) {
-		reg = &dev_priv->fence_regs[i];
-		if (!reg->obj)
-			return reg;
-
-		if (!reg->pin_count)
-			avail = reg;
-	}
-
-	if (avail == NULL)
-		goto deadlock;
+	struct drm_i915_fence_reg *fence;
 
-	/* None available, try to steal one or wait for a user to finish */
-	list_for_each_entry(reg, &dev_priv->mm.fence_list, lru_list) {
-		if (reg->pin_count)
+	list_for_each_entry(fence, &dev_priv->mm.fence_list, link) {
+		if (fence->pin_count)
 			continue;
 
-		return reg;
+		return fence;
 	}
 
-deadlock:
 	/* Wait for completion of pending flips which consume fences */
-	if (intel_has_pending_fb_unpin(dev))
+	if (intel_has_pending_fb_unpin(&dev_priv->drm))
 		return ERR_PTR(-EAGAIN);
 
 	return ERR_PTR(-EDEADLK);
 }
 
 /**
- * i915_gem_object_get_fence - set up fencing for an object
- * @obj: object to map through a fence reg
+ * i915_vma_get_fence - set up fencing for a vma
+ * @vma: vma to map through a fence reg
  *
  * When mapping objects through the GTT, userspace wants to be able to write
  * to them without having to worry about swizzling if the object is tiled.
@@ -366,103 +336,27 @@ i915_find_fence_reg(struct drm_device *dev)
  * 0 on success, negative error code on failure.
  */
 int
-i915_gem_object_get_fence(struct drm_i915_gem_object *obj)
+i915_vma_get_fence(struct i915_vma *vma)
 {
-	struct drm_device *dev = obj->base.dev;
-	struct drm_i915_private *dev_priv = to_i915(dev);
-	bool enable = i915_gem_object_is_tiled(obj);
-	struct drm_i915_fence_reg *reg;
-	int ret;
-
-	/* Have we updated the tiling parameters upon the object and so
-	 * will need to serialise the write to the associated fence register?
-	 */
-	if (obj->fence_dirty) {
-		ret = i915_gem_object_wait_fence(obj);
-		if (ret)
-			return ret;
-	}
+	struct drm_i915_fence_reg *fence;
+	struct i915_vma *set = i915_gem_object_is_tiled(vma->obj) ? vma : NULL;
 
 	/* Just update our place in the LRU if our fence is getting reused. */
-	if (obj->fence_reg != I915_FENCE_REG_NONE) {
-		reg = &dev_priv->fence_regs[obj->fence_reg];
-		if (!obj->fence_dirty) {
-			list_move_tail(&reg->lru_list,
-				       &dev_priv->mm.fence_list);
+	if (vma->fence) {
+		fence = vma->fence;
+		if (!fence->dirty) {
+			list_move_tail(&fence->link,
+				       &fence->i915->mm.fence_list);
 			return 0;
 		}
-	} else if (enable) {
-		if (WARN_ON(!obj->map_and_fenceable))
-			return -EINVAL;
-
-		reg = i915_find_fence_reg(dev);
-		if (IS_ERR(reg))
-			return PTR_ERR(reg);
-
-		if (reg->obj) {
-			struct drm_i915_gem_object *old = reg->obj;
-
-			ret = i915_gem_object_wait_fence(old);
-			if (ret)
-				return ret;
-
-			i915_gem_object_fence_lost(old);
-		}
+	} else if (set) {
+		fence = fence_find(to_i915(vma->vm->dev));
+		if (IS_ERR(fence))
+			return PTR_ERR(fence);
 	} else
 		return 0;
 
-	i915_gem_object_update_fence(obj, reg, enable);
-
-	return 0;
-}
-
-/**
- * i915_gem_object_pin_fence - pin fencing state
- * @obj: object to pin fencing for
- *
- * This pins the fencing state (whether tiled or untiled) to make sure the
- * object is ready to be used as a scanout target. Fencing status must be
- * synchronize first by calling i915_gem_object_get_fence():
- *
- * The resulting fence pin reference must be released again with
- * i915_gem_object_unpin_fence().
- *
- * Returns:
- *
- * True if the object has a fence, false otherwise.
- */
-bool
-i915_gem_object_pin_fence(struct drm_i915_gem_object *obj)
-{
-	if (obj->fence_reg != I915_FENCE_REG_NONE) {
-		struct drm_i915_private *dev_priv = to_i915(obj->base.dev);
-		struct i915_vma *ggtt_vma = i915_gem_obj_to_ggtt(obj);
-
-		WARN_ON(!ggtt_vma ||
-			dev_priv->fence_regs[obj->fence_reg].pin_count >
-			i915_vma_pin_count(ggtt_vma));
-		dev_priv->fence_regs[obj->fence_reg].pin_count++;
-		return true;
-	} else
-		return false;
-}
-
-/**
- * i915_gem_object_unpin_fence - unpin fencing state
- * @obj: object to unpin fencing for
- *
- * This releases the fence pin reference acquired through
- * i915_gem_object_pin_fence. It will handle both objects with and without an
- * attached fence correctly, callers do not need to distinguish this.
- */
-void
-i915_gem_object_unpin_fence(struct drm_i915_gem_object *obj)
-{
-	if (obj->fence_reg != I915_FENCE_REG_NONE) {
-		struct drm_i915_private *dev_priv = to_i915(obj->base.dev);
-		WARN_ON(dev_priv->fence_regs[obj->fence_reg].pin_count <= 0);
-		dev_priv->fence_regs[obj->fence_reg].pin_count--;
-	}
+	return fence_update(fence, set);
 }
 
 /**
@@ -479,17 +373,16 @@ void i915_gem_restore_fences(struct drm_device *dev)
 
 	for (i = 0; i < dev_priv->num_fence_regs; i++) {
 		struct drm_i915_fence_reg *reg = &dev_priv->fence_regs[i];
+		struct i915_vma *vma = reg->vma;
 
 		/*
 		 * Commit delayed tiling changes if we have an object still
 		 * attached to the fence, otherwise just clear the fence.
 		 */
-		if (reg->obj) {
-			i915_gem_object_update_fence(reg->obj, reg,
-						     i915_gem_object_get_tiling(reg->obj));
-		} else {
-			i915_gem_write_fence(dev, i, NULL);
-		}
+		if (vma && !i915_gem_object_is_tiled(vma->obj))
+			vma = NULL;
+
+		fence_update(reg, vma);
 	}
 }
 

drivers/gpu/drm/i915/i915_gem_render_state.h

@@ -24,7 +24,7 @@
 #ifndef _I915_GEM_RENDER_STATE_H_
 #define _I915_GEM_RENDER_STATE_H_
 
-#include <linux/types.h>
+struct drm_i915_gem_request;
 
 int i915_gem_render_state_init(struct drm_i915_gem_request *req);
 

drivers/gpu/drm/i915/i915_gem_stolen.c

@@ -115,17 +115,28 @@ static unsigned long i915_stolen_to_physical(struct drm_device *dev)
 
 		base = bsm & INTEL_BSM_MASK;
 	} else if (IS_I865G(dev)) {
+		u32 tseg_size = 0;
 		u16 toud = 0;
+		u8 tmp;
+
+		pci_bus_read_config_byte(dev->pdev->bus, PCI_DEVFN(0, 0),
+					 I845_ESMRAMC, &tmp);
+
+		if (tmp & TSEG_ENABLE) {
+			switch (tmp & I845_TSEG_SIZE_MASK) {
+			case I845_TSEG_SIZE_512K:
+				tseg_size = KB(512);
+				break;
+			case I845_TSEG_SIZE_1M:
+				tseg_size = MB(1);
+				break;
+			}
+		}
 
-		/*
-		 * FIXME is the graphics stolen memory region
-		 * always at TOUD? Ie. is it always the last
-		 * one to be allocated by the BIOS?
-		 */
 		pci_bus_read_config_word(dev->pdev->bus, PCI_DEVFN(0, 0),
 					 I865_TOUD, &toud);
 
-		base = toud << 16;
+		base = (toud << 16) + tseg_size;
 	} else if (IS_I85X(dev)) {
 		u32 tseg_size = 0;
 		u32 tom;
@@ -685,7 +696,7 @@ i915_gem_object_create_stolen_for_preallocated(struct drm_device *dev,
 	if (gtt_offset == I915_GTT_OFFSET_NONE)
 		return obj;
 
-	vma = i915_gem_obj_lookup_or_create_vma(obj, &ggtt->base);
+	vma = i915_gem_obj_lookup_or_create_vma(obj, &ggtt->base, NULL);
 	if (IS_ERR(vma)) {
 		ret = PTR_ERR(vma);
 		goto err;
@@ -705,6 +716,7 @@ i915_gem_object_create_stolen_for_preallocated(struct drm_device *dev,
 		goto err;
 	}
 
+	vma->pages = obj->pages;
 	vma->flags |= I915_VMA_GLOBAL_BIND;
 	__i915_vma_set_map_and_fenceable(vma);
 	list_move_tail(&vma->vm_link, &ggtt->base.inactive_list);

drivers/gpu/drm/i915/i915_gem_userptr.c

@@ -542,8 +542,6 @@ __i915_gem_userptr_get_pages_worker(struct work_struct *_work)
 			}
 		}
 		obj->userptr.work = ERR_PTR(ret);
-		if (ret)
-			__i915_gem_userptr_set_active(obj, false);
 	}
 
 	obj->userptr.workers--;
@@ -628,15 +626,14 @@ i915_gem_userptr_get_pages(struct drm_i915_gem_object *obj)
 	 * to the vma (discard or cloning) which should prevent the more
 	 * egregious cases from causing harm.
 	 */
-	if (IS_ERR(obj->userptr.work)) {
-		/* active flag will have been dropped already by the worker */
-		ret = PTR_ERR(obj->userptr.work);
-		obj->userptr.work = NULL;
-		return ret;
-	}
-	if (obj->userptr.work)
+
+	if (obj->userptr.work) {
 		/* active flag should still be held for the pending work */
-		return -EAGAIN;
+		if (IS_ERR(obj->userptr.work))
+			return PTR_ERR(obj->userptr.work);
+		else
+			return -EAGAIN;
+	}
 
 	/* Let the mmu-notifier know that we have begun and need cancellation */
 	ret = __i915_gem_userptr_set_active(obj, true);