drm/i915/gvt: Optimize MMIO register handling for some large MMIO blocks

Some of traced MMIO registers are a large continuous section. These
stuffed the MMIO lookup hash table and so waste lots of memory and
get much lower lookup performance.

Here we picked out these sections by special handling. These sections
include:
  o Display pipe registers, total 768.
  o The PVINFO page, total 1024.
  o MCHBAR_MIRROR, total 65536.
  o CSR_MMIO, total 3072.

So we removed 70,400 items from the hash table, and speed up guest
boot time by ~500ms.

v2:
  o add a local function find_mmio_block().
  o fix comments.

Signed-off-by: Changbin Du <changbin.du@intel.com>
Signed-off-by: Zhenyu Wang <zhenyuw@linux.intel.com>

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

@@ -92,11 +92,22 @@ static void write_vreg(struct intel_vgpu *vgpu, unsigned int offset,
 	memcpy(&vgpu_vreg(vgpu, offset), p_data, bytes);
 }
 
+static struct intel_gvt_mmio_info *find_mmio_info(struct intel_gvt *gvt,
+						  unsigned int offset)
+{
+	struct intel_gvt_mmio_info *e;
+
+	hash_for_each_possible(gvt->mmio.mmio_info_table, e, node, offset) {
+		if (e->offset == offset)
+			return e;
+	}
+	return NULL;
+}
+
 static int new_mmio_info(struct intel_gvt *gvt,
 		u32 offset, u32 flags, u32 size,
 		u32 addr_mask, u32 ro_mask, u32 device,
-		int (*read)(struct intel_vgpu *, unsigned int, void *, unsigned int),
-		int (*write)(struct intel_vgpu *, unsigned int, void *, unsigned int))
+		gvt_mmio_func read, gvt_mmio_func write)
 {
 	struct intel_gvt_mmio_info *info, *p;
 	u32 start, end, i;
@@ -116,7 +127,7 @@ static int new_mmio_info(struct intel_gvt *gvt,
 			return -ENOMEM;
 
 		info->offset = i;
-		p = intel_gvt_find_mmio_info(gvt, info->offset);
+		p = find_mmio_info(gvt, info->offset);
 		if (p)
 			gvt_err("dup mmio definition offset %x\n",
 				info->offset);
@@ -1794,10 +1805,6 @@ static int init_generic_mmio_info(struct intel_gvt *gvt)
 	MMIO_D(SPRSCALE(PIPE_C), D_ALL);
 	MMIO_D(SPRSURFLIVE(PIPE_C), D_ALL);
 
-	MMIO_F(LGC_PALETTE(PIPE_A, 0), 4 * 256, 0, 0, 0, D_ALL, NULL, NULL);
-	MMIO_F(LGC_PALETTE(PIPE_B, 0), 4 * 256, 0, 0, 0, D_ALL, NULL, NULL);
-	MMIO_F(LGC_PALETTE(PIPE_C, 0), 4 * 256, 0, 0, 0, D_ALL, NULL, NULL);
-
 	MMIO_D(HTOTAL(TRANSCODER_A), D_ALL);
 	MMIO_D(HBLANK(TRANSCODER_A), D_ALL);
 	MMIO_D(HSYNC(TRANSCODER_A), D_ALL);
@@ -2245,11 +2252,8 @@ static int init_generic_mmio_info(struct intel_gvt *gvt)
 
 	MMIO_DH(GEN6_GDRST, D_ALL, NULL, gdrst_mmio_write);
 	MMIO_F(FENCE_REG_GEN6_LO(0), 0x80, 0, 0, 0, D_ALL, fence_mmio_read, fence_mmio_write);
-	MMIO_F(VGT_PVINFO_PAGE, VGT_PVINFO_SIZE, F_UNALIGN, 0, 0, D_ALL, pvinfo_mmio_read, pvinfo_mmio_write);
 	MMIO_DH(CPU_VGACNTRL, D_ALL, NULL, vga_control_mmio_write);
 
-	MMIO_F(MCHBAR_MIRROR_BASE_SNB, 0x40000, 0, 0, 0, D_ALL, NULL, NULL);
-
 	MMIO_D(TILECTL, D_ALL);
 
 	MMIO_D(GEN6_UCGCTL1, D_ALL);
@@ -2778,7 +2782,6 @@ static int init_skl_mmio_info(struct intel_gvt *gvt)
 	MMIO_D(0x72380, D_SKL_PLUS);
 	MMIO_D(0x7039c, D_SKL_PLUS);
 
-	MMIO_F(0x80000, 0x3000, 0, 0, 0, D_SKL_PLUS, NULL, NULL);
 	MMIO_D(0x8f074, D_SKL | D_KBL);
 	MMIO_D(0x8f004, D_SKL | D_KBL);
 	MMIO_D(0x8f034, D_SKL | D_KBL);
@@ -2852,26 +2855,36 @@ static int init_skl_mmio_info(struct intel_gvt *gvt)
 	return 0;
 }
 
-/**
- * intel_gvt_find_mmio_info - find MMIO information entry by aligned offset
- * @gvt: GVT device
- * @offset: register offset
- *
- * This function is used to find the MMIO information entry from hash table
- *
- * Returns:
- * pointer to MMIO information entry, NULL if not exists
- */
-struct intel_gvt_mmio_info *intel_gvt_find_mmio_info(struct intel_gvt *gvt,
-	unsigned int offset)
-{
-	struct intel_gvt_mmio_info *e;
+/* Special MMIO blocks. */
+static struct gvt_mmio_block {
+	unsigned int device;
+	i915_reg_t   offset;
+	unsigned int size;
+	gvt_mmio_func read;
+	gvt_mmio_func write;
+} gvt_mmio_blocks[] = {
+	{D_SKL_PLUS, _MMIO(CSR_MMIO_START_RANGE), 0x3000, NULL, NULL},
+	{D_ALL, _MMIO(MCHBAR_MIRROR_BASE_SNB), 0x40000, NULL, NULL},
+	{D_ALL, _MMIO(VGT_PVINFO_PAGE), VGT_PVINFO_SIZE,
+		pvinfo_mmio_read, pvinfo_mmio_write},
+	{D_ALL, LGC_PALETTE(PIPE_A, 0), 1024, NULL, NULL},
+	{D_ALL, LGC_PALETTE(PIPE_B, 0), 1024, NULL, NULL},
+	{D_ALL, LGC_PALETTE(PIPE_C, 0), 1024, NULL, NULL},
+};
 
-	WARN_ON(!IS_ALIGNED(offset, 4));
+static struct gvt_mmio_block *find_mmio_block(struct intel_gvt *gvt,
+					      unsigned int offset)
+{
+	unsigned long device = intel_gvt_get_device_type(gvt);
+	struct gvt_mmio_block *block = gvt_mmio_blocks;
+	int i;
 
-	hash_for_each_possible(gvt->mmio.mmio_info_table, e, node, offset) {
-		if (e->offset == offset)
-			return e;
+	for (i = 0; i < ARRAY_SIZE(gvt_mmio_blocks); i++, block++) {
+		if (!(device & block->device))
+			continue;
+		if (offset >= INTEL_GVT_MMIO_OFFSET(block->offset) &&
+		    offset < INTEL_GVT_MMIO_OFFSET(block->offset) + block->size)
+			return block;
 	}
 	return NULL;
 }
@@ -3056,3 +3069,94 @@ bool intel_gvt_in_force_nonpriv_whitelist(struct intel_gvt *gvt,
 {
 	return in_whitelist(offset);
 }
+
+/**
+ * intel_vgpu_mmio_reg_rw - emulate tracked mmio registers
+ * @vgpu: a vGPU
+ * @offset: register offset
+ * @pdata: data buffer
+ * @bytes: data length
+ *
+ * Returns:
+ * Zero on success, negative error code if failed.
+ */
+int intel_vgpu_mmio_reg_rw(struct intel_vgpu *vgpu, unsigned int offset,
+			   void *pdata, unsigned int bytes, bool is_read)
+{
+	struct intel_gvt *gvt = vgpu->gvt;
+	struct intel_gvt_mmio_info *mmio_info;
+	struct gvt_mmio_block *mmio_block;
+	gvt_mmio_func func;
+	int ret;
+
+	if (WARN_ON(bytes > 4))
+		return -EINVAL;
+
+	/*
+	 * Handle special MMIO blocks.
+	 */
+	mmio_block = find_mmio_block(gvt, offset);
+	if (mmio_block) {
+		func = is_read ? mmio_block->read : mmio_block->write;
+		if (func)
+			return func(vgpu, offset, pdata, bytes);
+		goto default_rw;
+	}
+
+	/*
+	 * Normal tracked MMIOs.
+	 */
+	mmio_info = find_mmio_info(gvt, offset);
+	if (!mmio_info) {
+		if (!vgpu->mmio.disable_warn_untrack)
+			gvt_vgpu_err("untracked MMIO %08x len %d\n",
+				     offset, bytes);
+		goto default_rw;
+	}
+
+	if (WARN_ON(bytes > mmio_info->size))
+		return -EINVAL;
+
+	if (is_read)
+		return mmio_info->read(vgpu, offset, pdata, bytes);
+	else {
+		u64 ro_mask = mmio_info->ro_mask;
+		u32 old_vreg = 0, old_sreg = 0;
+		u64 data = 0;
+
+		if (intel_gvt_mmio_has_mode_mask(gvt, mmio_info->offset)) {
+			old_vreg = vgpu_vreg(vgpu, offset);
+			old_sreg = vgpu_sreg(vgpu, offset);
+		}
+
+		if (likely(!ro_mask))
+			ret = mmio_info->write(vgpu, offset, pdata, bytes);
+		else if (!~ro_mask) {
+			gvt_vgpu_err("try to write RO reg %x\n", offset);
+			return 0;
+		} else {
+			/* keep the RO bits in the virtual register */
+			memcpy(&data, pdata, bytes);
+			data &= ~ro_mask;
+			data |= vgpu_vreg(vgpu, offset) & ro_mask;
+			ret = mmio_info->write(vgpu, offset, &data, bytes);
+		}
+
+		/* higher 16bits of mode ctl regs are mask bits for change */
+		if (intel_gvt_mmio_has_mode_mask(gvt, mmio_info->offset)) {
+			u32 mask = vgpu_vreg(vgpu, offset) >> 16;
+
+			vgpu_vreg(vgpu, offset) = (old_vreg & ~mask)
+					| (vgpu_vreg(vgpu, offset) & mask);
+			vgpu_sreg(vgpu, offset) = (old_sreg & ~mask)
+					| (vgpu_sreg(vgpu, offset) & mask);
+		}
+	}
+
+	return ret;
+
+default_rw:
+	return is_read ?
+		intel_vgpu_default_mmio_read(vgpu, offset, pdata, bytes) :
+		intel_vgpu_default_mmio_write(vgpu, offset, pdata, bytes);
+}

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

@@ -123,7 +123,6 @@ int intel_vgpu_emulate_mmio_read(struct intel_vgpu *vgpu, uint64_t pa,
 		void *p_data, unsigned int bytes)
 {
 	struct intel_gvt *gvt = vgpu->gvt;
-	struct intel_gvt_mmio_info *mmio;
 	unsigned int offset = 0;
 	int ret = -EINVAL;
 
@@ -187,25 +186,8 @@ int intel_vgpu_emulate_mmio_read(struct intel_vgpu *vgpu, uint64_t pa,
 			goto err;
 	}
 
-	mmio = intel_gvt_find_mmio_info(gvt, rounddown(offset, 4));
-	if (mmio) {
-		if (!intel_gvt_mmio_is_unalign(gvt, mmio->offset)) {
-			if (WARN_ON(offset + bytes > mmio->offset + mmio->size))
-				goto err;
-			if (WARN_ON(mmio->offset != offset))
-				goto err;
-		}
-		ret = mmio->read(vgpu, offset, p_data, bytes);
-	} else {
-		ret = intel_vgpu_default_mmio_read(vgpu, offset, p_data, bytes);
-
-		if (!vgpu->mmio.disable_warn_untrack) {
-			gvt_vgpu_err("read untracked MMIO %x(%dB) val %x\n",
-				offset, bytes, *(u32 *)p_data);
-		}
-	}
-
-	if (ret)
+	ret = intel_vgpu_mmio_reg_rw(vgpu, offset, p_data, bytes, true);
+	if (ret < 0)
 		goto err;
 
 	intel_gvt_mmio_set_accessed(gvt, offset);
@@ -232,9 +214,7 @@ int intel_vgpu_emulate_mmio_write(struct intel_vgpu *vgpu, uint64_t pa,
 		void *p_data, unsigned int bytes)
 {
 	struct intel_gvt *gvt = vgpu->gvt;
-	struct intel_gvt_mmio_info *mmio;
 	unsigned int offset = 0;
-	u32 old_vreg = 0, old_sreg = 0;
 	int ret = -EINVAL;
 
 	if (vgpu->failsafe) {
@@ -289,66 +269,10 @@ int intel_vgpu_emulate_mmio_write(struct intel_vgpu *vgpu, uint64_t pa,
 		return ret;
 	}
 
-	mmio = intel_gvt_find_mmio_info(gvt, rounddown(offset, 4));
-	if (!mmio && !vgpu->mmio.disable_warn_untrack)
-		gvt_dbg_mmio("vgpu%d: write untracked MMIO %x len %d val %x\n",
-				vgpu->id, offset, bytes, *(u32 *)p_data);
-
-	if (!intel_gvt_mmio_is_unalign(gvt, offset)) {
-		if (WARN_ON(!IS_ALIGNED(offset, bytes)))
-			goto err;
-	}
-
-	if (mmio) {
-		u64 ro_mask = mmio->ro_mask;
-
-		if (!intel_gvt_mmio_is_unalign(gvt, mmio->offset)) {
-			if (WARN_ON(offset + bytes > mmio->offset + mmio->size))
-				goto err;
-			if (WARN_ON(mmio->offset != offset))
-				goto err;
-		}
-
-		if (intel_gvt_mmio_has_mode_mask(gvt, mmio->offset)) {
-			old_vreg = vgpu_vreg(vgpu, offset);
-			old_sreg = vgpu_sreg(vgpu, offset);
-		}
-
-		if (!ro_mask) {
-			ret = mmio->write(vgpu, offset, p_data, bytes);
-		} else {
-			/* Protect RO bits like HW */
-			u64 data = 0;
-
-			/* all register bits are RO. */
-			if (ro_mask == ~(u64)0) {
-				gvt_vgpu_err("try to write RO reg %x\n",
-					offset);
-				ret = 0;
-				goto out;
-			}
-			/* keep the RO bits in the virtual register */
-			memcpy(&data, p_data, bytes);
-			data &= ~mmio->ro_mask;
-			data |= vgpu_vreg(vgpu, offset) & mmio->ro_mask;
-			ret = mmio->write(vgpu, offset, &data, bytes);
-		}
-
-		/* higher 16bits of mode ctl regs are mask bits for change */
-		if (intel_gvt_mmio_has_mode_mask(gvt, mmio->offset)) {
-			u32 mask = vgpu_vreg(vgpu, offset) >> 16;
-
-			vgpu_vreg(vgpu, offset) = (old_vreg & ~mask)
-				| (vgpu_vreg(vgpu, offset) & mask);
-			vgpu_sreg(vgpu, offset) = (old_sreg & ~mask)
-				| (vgpu_sreg(vgpu, offset) & mask);
-		}
-	} else
-		ret = intel_vgpu_default_mmio_write(vgpu, offset, p_data,
-				bytes);
-	if (ret)
+	ret = intel_vgpu_mmio_reg_rw(vgpu, offset, p_data, bytes, false);
+	if (ret < 0)
 		goto err;
-out:
+
 	intel_gvt_mmio_set_accessed(gvt, offset);
 	mutex_unlock(&gvt->lock);
 	return 0;

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

@@ -52,6 +52,9 @@ struct intel_vgpu;
 #define D_PRE_SKL	(D_BDW)
 #define D_ALL		(D_BDW | D_SKL | D_KBL)
 
+typedef int (*gvt_mmio_func)(struct intel_vgpu *, unsigned int, void *,
+			     unsigned int);
+
 struct intel_gvt_mmio_info {
 	u32 offset;
 	u32 size;
@@ -59,8 +62,8 @@ struct intel_gvt_mmio_info {
 	u32 addr_mask;
 	u64 ro_mask;
 	u32 device;
-	int (*read)(struct intel_vgpu *, unsigned int, void *, unsigned int);
-	int (*write)(struct intel_vgpu *, unsigned int, void *, unsigned int);
+	gvt_mmio_func read;
+	gvt_mmio_func write;
 	u32 addr_range;
 	struct hlist_node node;
 };
@@ -71,8 +74,6 @@ bool intel_gvt_match_device(struct intel_gvt *gvt, unsigned long device);
 int intel_gvt_setup_mmio_info(struct intel_gvt *gvt);
 void intel_gvt_clean_mmio_info(struct intel_gvt *gvt);
 
-struct intel_gvt_mmio_info *intel_gvt_find_mmio_info(struct intel_gvt *gvt,
-						     unsigned int offset);
 #define INTEL_GVT_MMIO_OFFSET(reg) ({ \
 	typeof(reg) __reg = reg; \
 	u32 *offset = (u32 *)&__reg; \
@@ -103,4 +104,8 @@ int intel_vgpu_default_mmio_write(struct intel_vgpu *vgpu, unsigned int offset,
 
 bool intel_gvt_in_force_nonpriv_whitelist(struct intel_gvt *gvt,
 					  unsigned int offset);
+
+int intel_vgpu_mmio_reg_rw(struct intel_vgpu *vgpu, unsigned int offset,
+			   void *pdata, unsigned int bytes, bool is_read);
+
 #endif