x86: Add support for newer CAR schemes

Newer Intel SoCs have different ways of setting up cache-as-ram (CAR).
Add support for these along with suitable configuration options.

To make the code cleaner, adjust a few definitions in processor.h so that
they can be used from assembler.

Signed-off-by: Simon Glass <sjg@chromium.org>
Reviewed-by: Bin Meng <bmeng.cn@gmail.com>

arch/x86/Kconfig

@@ -879,4 +879,20 @@ config HIGH_TABLE_SIZE
 	  Increse it if the default size does not fit the board's needs.
 	  This is most likely due to a large ACPI DSDT table is used.
 
+config INTEL_CAR_CQOS
+	bool "Support Intel Cache Quality of Service"
+	help
+	  Cache Quality of Service allows more fine-grained control of cache
+	  usage. As result, it is possible to set up a portion of L2 cache for
+	  CAR and use the remainder for actual caching.
+
+#
+# Each bit in QOS mask controls this many bytes. This is calculated as:
+# (CACHE_WAYS / CACHE_BITS_PER_MASK) * CACHE_LINE_SIZE * CACHE_SETS
+#
+config CACHE_QOS_SIZE_PER_BIT
+	hex
+	depends on INTEL_CAR_CQOS
+	default 0x20000 # 128 KB
+
 endmenu

arch/x86/cpu/intel_common/Makefile

@@ -8,6 +8,14 @@ obj-$(CONFIG_$(SPL_TPL_)X86_32BIT_INIT) += me_status.o
 obj-$(CONFIG_$(SPL_TPL_)X86_32BIT_INIT) += report_platform.o
 obj-$(CONFIG_$(SPL_TPL_)X86_32BIT_INIT) += mrc.o
 endif
+
+ifdef CONFIG_INTEL_CAR_CQOS
+obj-$(CONFIG_TPL_BUILD) += car2.o
+ifndef CONFIG_SPL_BUILD
+obj-y += car2_uninit.o
+endif
+endif
+
 obj-y += cpu.o
 obj-y += fast_spi.o
 obj-y += lpc.o

arch/x86/cpu/intel_common/car2.S

@@ -0,0 +1,448 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+/*
+ * This file was modified from the coreboot version.
+ *
+ * Copyright (C) 2015-2016 Intel Corp.
+ */
+
+#include <config.h>
+#include <asm/msr-index.h>
+#include <asm/mtrr.h>
+#include <asm/post.h>
+#include <asm/processor.h>
+#include <asm/processor-flags.h>
+
+#define KiB 1024
+
+#define IS_POWER_OF_2(x)	(!((x) & ((x) - 1)))
+
+.global car_init
+car_init:
+	post_code(POST_CAR_START)
+
+	/*
+	 * Use the MTRR default type MSR as a proxy for detecting INIT#.
+	 * Reset the system if any known bits are set in that MSR. That is
+	 * an indication of the CPU not being properly reset.
+	 */
+check_for_clean_reset:
+	mov	$MTRR_DEF_TYPE_MSR, %ecx
+	rdmsr
+	and	$(MTRR_DEF_TYPE_EN | MTRR_DEF_TYPE_FIX_EN), %eax
+	cmp	$0, %eax
+	jz	no_reset
+	/* perform warm reset */
+	movw	$IO_PORT_RESET, %dx
+	movb	$(SYS_RST | RST_CPU), %al
+	outb	%al, %dx
+
+no_reset:
+	post_code(POST_CAR_SIPI)
+
+	/* Clear/disable fixed MTRRs */
+	mov	$fixed_mtrr_list_size, %ebx
+	xor	%eax, %eax
+	xor	%edx, %edx
+
+clear_fixed_mtrr:
+	add	$-2, %ebx
+	movzwl	fixed_mtrr_list(%ebx), %ecx
+	wrmsr
+	jnz	clear_fixed_mtrr
+
+	post_code(POST_CAR_MTRR)
+
+	/* Figure put how many MTRRs we have, and clear them out */
+	mov	$MTRR_CAP_MSR, %ecx
+	rdmsr
+	movzb	%al, %ebx		/* Number of variable MTRRs */
+	mov	$MTRR_PHYS_BASE_MSR(0), %ecx
+	xor	%eax, %eax
+	xor	%edx, %edx
+
+clear_var_mtrr:
+	wrmsr
+	inc	%ecx
+	wrmsr
+	inc	%ecx
+	dec	%ebx
+	jnz	clear_var_mtrr
+
+	post_code(POST_CAR_UNCACHEABLE)
+
+	/* Configure default memory type to uncacheable (UC) */
+	mov	$MTRR_DEF_TYPE_MSR, %ecx
+	rdmsr
+	/* Clear enable bits and set default type to UC */
+	and	$~(MTRR_DEF_TYPE_MASK | MTRR_DEF_TYPE_EN | \
+		 MTRR_DEF_TYPE_FIX_EN), %eax
+	wrmsr
+
+	/*
+	 * Configure MTRR_PHYS_MASK_HIGH for proper addressing above 4GB
+	 * based on the physical address size supported for this processor
+	 * This is based on read from CPUID EAX = 080000008h, EAX bits [7:0]
+	 *
+	 * Examples:
+	 *  MTRR_PHYS_MASK_HIGH = 00000000Fh  For 36 bit addressing
+	 *  MTRR_PHYS_MASK_HIGH = 0000000FFh  For 40 bit addressing
+	 */
+
+	movl	$0x80000008, %eax 	/* Address sizes leaf */
+	cpuid
+	sub	$32, %al
+	movzx	%al, %eax
+	xorl	%esi, %esi
+	bts	%eax, %esi
+	dec	%esi			/* esi <- MTRR_PHYS_MASK_HIGH */
+
+	post_code(POST_CAR_BASE_ADDRESS)
+
+#if IS_POWER_OF_2(CONFIG_DCACHE_RAM_SIZE)
+	/* Configure CAR region as write-back (WB) */
+	mov	$MTRR_PHYS_BASE_MSR(0), %ecx
+	mov	$CONFIG_DCACHE_RAM_BASE, %eax
+	or	$MTRR_TYPE_WRBACK, %eax
+	xor	%edx,%edx
+	wrmsr
+
+	/* Configure the MTRR mask for the size region */
+	mov	$MTRR_PHYS_MASK(0), %ecx
+	mov	$CONFIG_DCACHE_RAM_SIZE, %eax	/* size mask */
+	dec	%eax
+	not	%eax
+	or	$MTRR_PHYS_MASK_VALID, %eax
+	movl	%esi, %edx	/* edx <- MTRR_PHYS_MASK_HIGH */
+	wrmsr
+#elif (CONFIG_DCACHE_RAM_SIZE == 768 * KiB) /* 768 KiB */
+	/* Configure CAR region as write-back (WB) */
+	mov	$MTRR_PHYS_BASE_MSR(0), %ecx
+	mov	$CONFIG_DCACHE_RAM_BASE, %eax
+	or	$MTRR_TYPE_WRBACK, %eax
+	xor	%edx,%edx
+	wrmsr
+
+	mov	$MTRR_PHYS_MASK_MSR(0), %ecx
+	mov	$(512 * KiB), %eax	/* size mask */
+	dec	%eax
+	not	%eax
+	or	$MTRR_PHYS_MASK_VALID, %eax
+	movl	%esi, %edx	/* edx <- MTRR_PHYS_MASK_HIGH */
+	wrmsr
+
+	mov	$MTRR_PHYS_BASE_MSR(1), %ecx
+	mov	$(CONFIG_DCACHE_RAM_BASE + 512 * KiB), %eax
+	or	$MTRR_TYPE_WRBACK, %eax
+	xor	%edx,%edx
+	wrmsr
+
+	mov	$MTRR_PHYS_MASK_MSR(1), %ecx
+	mov	$(256 * KiB), %eax	/* size mask */
+	dec	%eax
+	not	%eax
+	or	$MTRR_PHYS_MASK_VALID, %eax
+	movl	%esi, %edx	/* edx <- MTRR_PHYS_MASK_HIGH */
+	wrmsr
+#else
+#error "DCACHE_RAM_SIZE is not a power of 2 and setup code is missing"
+#endif
+	post_code(POST_CAR_FILL)
+
+	/* Enable variable MTRRs */
+	mov	$MTRR_DEF_TYPE_MSR, %ecx
+	rdmsr
+	or	$MTRR_DEF_TYPE_EN, %eax
+	wrmsr
+
+	/* Enable caching */
+	mov	%cr0, %eax
+	and	$~(X86_CR0_CD | X86_CR0_NW), %eax
+	invd
+	mov	%eax, %cr0
+
+#if IS_ENABLED(CONFIG_INTEL_CAR_NEM)
+	jmp	car_nem
+#elif IS_ENABLED(CONFIG_INTEL_CAR_CQOS)
+	jmp	car_cqos
+#elif IS_ENABLED(CONFIG_INTEL_CAR_NEM_ENHANCED)
+	jmp	car_nem_enhanced
+#else
+#error "No CAR mechanism selected:
+#endif
+	jmp	car_init_ret
+
+fixed_mtrr_list:
+	.word	MTRR_FIX_64K_00000_MSR
+	.word	MTRR_FIX_16K_80000_MSR
+	.word	MTRR_FIX_16K_A0000_MSR
+	.word	MTRR_FIX_4K_C0000_MSR
+	.word	MTRR_FIX_4K_C8000_MSR
+	.word	MTRR_FIX_4K_D0000_MSR
+	.word	MTRR_FIX_4K_D8000_MSR
+	.word	MTRR_FIX_4K_E0000_MSR
+	.word	MTRR_FIX_4K_E8000_MSR
+	.word	MTRR_FIX_4K_F0000_MSR
+	.word	MTRR_FIX_4K_F8000_MSR
+fixed_mtrr_list_size = . - fixed_mtrr_list
+
+#if IS_ENABLED(CONFIG_INTEL_CAR_NEM)
+.global car_nem
+car_nem:
+	/* Disable cache eviction (setup stage) */
+	mov	$MSR_EVICT_CTL, %ecx
+	rdmsr
+	or	$0x1, %eax
+	wrmsr
+
+	post_code(0x26)
+
+	/* Clear the cache memory region. This will also fill up the cache */
+	movl	$CONFIG_DCACHE_RAM_BASE, %edi
+	movl	$CONFIG_DCACHE_RAM_SIZE, %ecx
+	shr	$0x02, %ecx
+	xor	%eax, %eax
+	cld
+	rep	stosl
+
+	post_code(0x27)
+
+	/* Disable cache eviction (run stage) */
+	mov	$MSR_EVICT_CTL, %ecx
+	rdmsr
+	or	$0x2, %eax
+	wrmsr
+
+	post_code(0x28)
+
+	jmp	car_init_ret
+
+#elif IS_ENABLED(CONFIG_INTEL_CAR_CQOS)
+.global car_cqos
+car_cqos:
+	/*
+	 * Create CBM_LEN_MASK based on CBM_LEN
+	 * Get CPUID.(EAX=10H, ECX=2H):EAX.CBM_LEN[bits 4:0]
+	 */
+	mov	$0x10, %eax
+	mov	$0x2,  %ecx
+	cpuid
+	and	$0x1f, %eax
+	add	$1, %al
+
+	mov	$1, %ebx
+	mov	%al, %cl
+	shl	%cl, %ebx
+	sub	$1, %ebx
+
+	/* Store the CBM_LEN_MASK in mm3 for later use */
+	movd	%ebx, %mm3
+
+	/*
+	 * Disable both L1 and L2 prefetcher. For yet-to-understood reason,
+	 * prefetchers slow down filling cache with rep stos in CQOS mode.
+	 */
+	mov	$MSR_PREFETCH_CTL, %ecx
+	rdmsr
+	or	$(PREFETCH_L1_DISABLE | PREFETCH_L2_DISABLE), %eax
+	wrmsr
+
+#if (CONFIG_DCACHE_RAM_SIZE == CONFIG_L2_CACHE_SIZE)
+/*
+ * If CAR size is set to full L2 size, mask is calculated as all-zeros.
+ * This is not supported by the CPU/uCode.
+ */
+#error "CQOS CAR may not use whole L2 cache area"
+#endif
+
+	/* Calculate how many bits to be used for CAR */
+	xor	%edx, %edx
+	mov	$CONFIG_DCACHE_RAM_SIZE, %eax	/* dividend */
+	mov	$CONFIG_CACHE_QOS_SIZE_PER_BIT, %ecx	/* divisor */
+	div	%ecx		/* result is in eax */
+	mov	%eax, %ecx	/* save to ecx */
+	mov	$1, %ebx
+	shl	%cl, %ebx
+	sub	$1, %ebx	/* resulting mask is is in ebx */
+
+	/* Set this mask for initial cache fill */
+	mov	$MSR_L2_QOS_MASK(0), %ecx
+	rdmsr
+	mov	%ebx, %eax
+	wrmsr
+
+	/* Set CLOS selector to 0 */
+	mov	$MSR_IA32_PQR_ASSOC, %ecx
+	rdmsr
+	and	$~MSR_IA32_PQR_ASSOC_MASK, %edx	/* select mask 0 */
+	wrmsr
+
+	/* We will need to block CAR region from evicts */
+	mov	$MSR_L2_QOS_MASK(1), %ecx
+	rdmsr
+	/* Invert bits that are to be used for cache */
+	mov	%ebx, %eax
+	xor	$~0, %eax			/* invert 32 bits */
+
+	/*
+	 * Use CBM_LEN_MASK stored in mm3 to set bits based on Capacity Bit
+	 * Mask Length.
+	 */
+	movd	%mm3, %ebx
+	and	%ebx, %eax
+	wrmsr
+
+	post_code(0x26)
+
+	/* Clear the cache memory region. This will also fill up the cache */
+	movl	$CONFIG_DCACHE_RAM_BASE, %edi
+	movl	$CONFIG_DCACHE_RAM_SIZE, %ecx
+	shr	$0x02, %ecx
+	xor	%eax, %eax
+	cld
+	rep	stosl
+
+	post_code(0x27)
+
+	/* Cache is populated. Use mask 1 that will block evicts */
+	mov	$MSR_IA32_PQR_ASSOC, %ecx
+	rdmsr
+	and	$~MSR_IA32_PQR_ASSOC_MASK, %edx	/* clear index bits first */
+	or	$1, %edx			/* select mask 1 */
+	wrmsr
+
+	/* Enable prefetchers */
+	mov	$MSR_PREFETCH_CTL, %ecx
+	rdmsr
+	and	$~(PREFETCH_L1_DISABLE | PREFETCH_L2_DISABLE), %eax
+	wrmsr
+
+	post_code(0x28)
+
+	jmp	car_init_ret
+
+#elif IS_ENABLED(CONFIG_INTEL_CAR_NEM_ENHANCED)
+.global car_nem_enhanced
+car_nem_enhanced:
+	/* Disable cache eviction (setup stage) */
+	mov	$MSR_EVICT_CTL, %ecx
+	rdmsr
+	or	$0x1, %eax
+	wrmsr
+	post_code(0x26)
+
+	/* Create n-way set associativity of cache */
+	xorl	%edi, %edi
+find_llc_subleaf:
+	movl	%edi, %ecx
+	movl	$0x04, %eax
+	cpuid
+	inc	%edi
+	and	$0xe0, %al	/* EAX[7:5] = Cache Level */
+	cmp	$0x60, %al	/* Check to see if it is LLC */
+	jnz	find_llc_subleaf
+
+	/*
+	 * Set MSR 0xC91 IA32_L3_MASK_! = 0xE/0xFE/0xFFE/0xFFFE
+	 * for 4/8/16 way of LLC
+	*/
+	shr	$22, %ebx
+	inc	%ebx
+	/* Calculate n-way associativity of LLC */
+	mov	%bl, %cl
+
+	/*
+	 * Maximizing RO cacheability while locking in the CAR to a
+	 * single way since that particular way won't be victim candidate
+	 * for evictions.
+	 * This has been done after programing LLC_WAY_MASK_1 MSR
+	 * with desired LLC way as mentioned below.
+	 *
+	 * Hence create Code and Data Size as per request
+	 * Code Size (RO) : Up to 16M
+	 * Data Size (RW) : Up to 256K
+	 */
+	movl	$0x01, %eax
+	/*
+	 * LLC Ways -> LLC_WAY_MASK_1:
+	 *  4: 0x000E
+	 *  8: 0x00FE
+	 * 12: 0x0FFE
+	 * 16: 0xFFFE
+	 *
+	 * These MSRs contain one bit per each way of LLC
+	 * - If this bit is '0' - the way is protected from eviction
+	 * - If this bit is '1' - the way is not protected from eviction
+	 */
+	shl	%cl, %eax
+	subl	$0x02, %eax
+	movl	$MSR_IA32_L3_MASK_1, %ecx
+	xorl	%edx, %edx
+	wrmsr
+	/*
+	 * Set MSR 0xC92 IA32_L3_MASK_2 = 0x1
+	 *
+	 * For SKL SOC, data size remains 256K consistently.
+	 * Hence, creating 1-way associative cache for Data
+	*/
+	mov	$MSR_IA32_L3_MASK_2, %ecx
+	mov	$0x01, %eax
+	xorl	%edx, %edx
+	wrmsr
+	/*
+	 * Set MSR_IA32_PQR_ASSOC = 0x02
+	 *
+	 * Possible values:
+	 * 0: Default value, no way mask should be applied
+	 * 1: Apply way mask 1 to LLC
+	 * 2: Apply way mask 2 to LLC
+	 * 3: Shouldn't be use in NEM Mode
+	 */
+	movl	$MSR_IA32_PQR_ASSOC, %ecx
+	movl	$0x02, %eax
+	xorl	%edx, %edx
+	wrmsr
+
+	movl	$CONFIG_DCACHE_RAM_BASE, %edi
+	movl	$CONFIG_DCACHE_RAM_SIZE, %ecx
+	shr	$0x02, %ecx
+	xor	%eax, %eax
+	cld
+	rep	stosl
+	/*
+	 * Set MSR_IA32_PQR_ASSOC = 0x01
+	 * At this stage we apply LLC_WAY_MASK_1 to the cache.
+	 * i.e. way 0 is protected from eviction.
+	*/
+	movl	$MSR_IA32_PQR_ASSOC, %ecx
+	movl	$0x01, %eax
+	xorl	%edx, %edx
+	wrmsr
+
+	post_code(0x27)
+	/*
+	 * Enable No-Eviction Mode Run State by setting
+	 * NO_EVICT_MODE MSR 2E0h bit [1] = '1'.
+	 */
+
+	movl	$MSR_EVICT_CTL, %ecx
+	rdmsr
+	orl	$0x02, %eax
+	wrmsr
+
+	post_code(0x28)
+
+	jmp	car_init_ret
+#endif
+
+#if CONFIG_IS_ENABLED(X86_16BIT_INIT)
+_dt_ucode_base_size:
+	/* These next two fields are filled in by binman */
+.globl ucode_base
+ucode_base:	/* Declared in microcode.h */
+	.long	0			/* microcode base */
+.globl ucode_size
+ucode_size:	/* Declared in microcode.h */
+	.long	0			/* microcode size */
+	.long	CONFIG_SYS_MONITOR_BASE	/* code region base */
+	.long	CONFIG_SYS_MONITOR_LEN	/* code region size */
+#endif

arch/x86/cpu/intel_common/car2_uninit.S

@@ -0,0 +1,87 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+/*
+ * Copyright 2017 Intel Corp.
+ * Copyright 2019 Google LLC
+ * Taken from coreboot file exit_car.S
+ */
+
+#include <config.h>
+#include <asm/msr-index.h>
+#include <asm/mtrr.h>
+
+.text
+.global car_uninit
+car_uninit:
+
+	/*
+	 * Retrieve return address from stack as it will get trashed below if
+	 * execution is utilizing the cache-as-ram stack.
+	 */
+	pop	%ebx
+
+	/* Disable MTRRs */
+	mov	$(MTRR_DEF_TYPE_MSR), %ecx
+	rdmsr
+	and	$(~(MTRR_DEF_TYPE_EN | MTRR_DEF_TYPE_FIX_EN)), %eax
+	wrmsr
+
+#ifdef CONFIG_INTEL_CAR_NEM
+.global car_nem_teardown
+car_nem_teardown:
+
+	/* invalidate cache contents */
+	invd
+
+	/* Knock down bit 1 then bit 0 of NEM control not combining steps */
+	mov	$(MSR_EVICT_CTL), %ecx
+	rdmsr
+	and	$(~(1 << 1)), %eax
+	wrmsr
+	and	$(~(1 << 0)), %eax
+	wrmsr
+
+#elif IS_ENABLED(CONFIG_INTEL_CAR_CQOS)
+.global car_cqos_teardown
+car_cqos_teardown:
+
+	/* Go back to all-evicting mode, set both masks to all-1s */
+	mov	$MSR_L2_QOS_MASK(0), %ecx
+	rdmsr
+	mov	$~0, %al
+	wrmsr
+
+	mov	$MSR_L2_QOS_MASK(1), %ecx
+	rdmsr
+	mov	$~0, %al
+	wrmsr
+
+	/* Reset CLOS selector to 0 */
+	mov	$MSR_IA32_PQR_ASSOC, %ecx
+	rdmsr
+	and	$~MSR_IA32_PQR_ASSOC_MASK, %edx
+	wrmsr
+
+#elif IS_ENABLED(CONFIG_INTEL_CAR_NEM_ENHANCED)
+.global car_nem_enhanced_teardown
+car_nem_enhanced_teardown:
+
+	/* invalidate cache contents */
+	invd
+
+	/* Knock down bit 1 then bit 0 of NEM control not combining steps */
+	mov	$(MSR_EVICT_CTL), %ecx
+	rdmsr
+	and	$(~(1 << 1)), %eax
+	wrmsr
+	and	$(~(1 << 0)), %eax
+	wrmsr
+
+	/* Reset CLOS selector to 0 */
+	mov	$IA32_PQR_ASSOC, %ecx
+	rdmsr
+	and	$~IA32_PQR_ASSOC_MASK, %edx
+	wrmsr
+#endif
+
+	/* Return to caller */
+	jmp	*%ebx

arch/x86/include/asm/processor.h

@@ -25,8 +25,6 @@
 /* Length of the public header on Intel microcode blobs */
 #define UCODE_HEADER_LEN	0x30
 
-#ifndef __ASSEMBLY__
-
 /*
  * This register is documented in (for example) the Intel Atom Processor E3800
  * Product Family Datasheet in "PCU - Power Management Controller (PMC)".
@@ -37,11 +35,11 @@
  */
 #define IO_PORT_RESET		0xcf9
 
-enum {
+#define SYS_RST		(1 << 1)	/* 0 for soft reset, 1 for hard reset */
-	SYS_RST		= 1 << 1,	/* 0 for soft reset, 1 for hard reset */
+#define RST_CPU		(1 << 2)	/* initiate reset */
-	RST_CPU		= 1 << 2,	/* initiate reset */
+#define FULL_RST	(1 << 3)	/* full power cycle */
-	FULL_RST	= 1 << 3,	/* full power cycle */
+
-};
+#ifndef __ASSEMBLY__
 
 static inline __attribute__((always_inline)) void cpu_hlt(void)
 {