summaryrefslogtreecommitdiff
path: root/src/soc/intel/apollolake
diff options
context:
space:
mode:
authorAndrey Petrov <andrey.petrov@intel.com>2016-06-27 13:39:34 -0700
committerMartin Roth <martinroth@google.com>2016-08-28 18:38:48 +0200
commit3f4aece4e07b15a5a2d191873da04b88c8e87049 (patch)
tree95656eb37744843a4e9586c625451b55a5e17251 /src/soc/intel/apollolake
parent7f72c9b30ec543fc5d485dca5f15790d2c4b03f3 (diff)
soc/intel/apollolake: Add CQOS CAR implementation
Add new option to set up Cache-As-RAM by using CQOS, Cache Quality of Service. CQOS allows setting ways of cache in no-fill mode, while keeping other ways in regular evicting mode. This effectively allows using CAR and cache simultaneously. BUG=chrome-os-partner:51959 TEST=switch from NEM to CQOS and back, boot Change-Id: Ic7f9899918f94a5788b02a4fbd2f5d5ba9aaf91d Signed-off-by: Andrey Petrov <andrey.petrov@intel.com> Reviewed-on: https://review.coreboot.org/15455 Tested-by: build bot (Jenkins) Reviewed-by: Aaron Durbin <adurbin@chromium.org>
Diffstat (limited to 'src/soc/intel/apollolake')
-rw-r--r--src/soc/intel/apollolake/Kconfig21
-rw-r--r--src/soc/intel/apollolake/bootblock/cache_as_ram.S64
-rw-r--r--src/soc/intel/apollolake/exit_car.S25
-rw-r--r--src/soc/intel/apollolake/include/soc/cpu.h20
4 files changed, 128 insertions, 2 deletions
diff --git a/src/soc/intel/apollolake/Kconfig b/src/soc/intel/apollolake/Kconfig
index d8e33caf7b..9fa37a1ff3 100644
--- a/src/soc/intel/apollolake/Kconfig
+++ b/src/soc/intel/apollolake/Kconfig
@@ -227,6 +227,27 @@ config NHLT_DA7219
default n
help
Include DSP firmware settings for headset codec.
+choice
+ prompt "Cache-as-ram implementation"
+ default CAR_CQOS
+ help
+ This option allows you to select how cache-as-ram (CAR) is set up.
+
+config CAR_NEM
+ bool "Non-evict mode"
+ help
+ Traditionally, CAR is set up by using Non-Evict mode. This method
+ does not allow CAR and cache to co-exist, because cache fills are
+ block in NEM mode.
+
+config CAR_CQOS
+ bool "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 portion of L2 cache for
+ CAR and use remainder for actual caching.
+
+endchoice
config SPI_FLASH_INCLUDE_ALL_DRIVERS
bool
diff --git a/src/soc/intel/apollolake/bootblock/cache_as_ram.S b/src/soc/intel/apollolake/bootblock/cache_as_ram.S
index e8fae28b5b..65dd4c83f9 100644
--- a/src/soc/intel/apollolake/bootblock/cache_as_ram.S
+++ b/src/soc/intel/apollolake/bootblock/cache_as_ram.S
@@ -124,12 +124,61 @@ clear_var_mtrr:
invd
mov %eax, %cr0
+#if IS_ENABLED(CONFIG_CAR_NEM)
/* Disable cache eviction (setup stage) */
mov $MSR_EVICT_CTL, %ecx
rdmsr
or $0x1, %eax
wrmsr
+#else
+ /*
+ * 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
+#endif
+
+#if IS_ENABLED(CONFIG_CAR_CQOS)
+#if (CONFIG_DCACHE_RAM_SIZE == 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 $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 %bl, %al
+ wrmsr
+
+ /* Set CLOS selector to 0 */
+ mov $MSR_IA32_PQR_ASSOC, %ecx
+ rdmsr
+ and $~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 %bl, %al
+ xor $~0, %al /* invert 8 bits */
+ wrmsr
+#endif
post_code(0x26)
/* Clear the cache memory region. This will also fill up the cache */
@@ -140,11 +189,26 @@ clear_var_mtrr:
post_code(0x27)
+#if IS_ENABLED(CONFIG_CAR_NEM)
/* Disable cache eviction (run stage) */
mov $MSR_EVICT_CTL, %ecx
rdmsr
or $0x2, %eax
wrmsr
+#else
+ /* Cache is populated. Use mask 1 that will block evicts */
+ mov $MSR_IA32_PQR_ASSOC, %ecx
+ rdmsr
+ and $~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
+#endif
post_code(0x28)
diff --git a/src/soc/intel/apollolake/exit_car.S b/src/soc/intel/apollolake/exit_car.S
index 339242b319..e21aa294ce 100644
--- a/src/soc/intel/apollolake/exit_car.S
+++ b/src/soc/intel/apollolake/exit_car.S
@@ -27,14 +27,34 @@ chipset_teardown_car:
*/
pop %ebx
- /* invalidate cache contents. */
- invd
/* Disable MTRRs. */
mov $(MTRR_DEF_TYPE_MSR), %ecx
rdmsr
and $(~(MTRR_DEF_TYPE_EN | MTRR_DEF_TYPE_FIX_EN)), %eax
wrmsr
+#if IS_ENABLED(CONFIG_CAR_CQOS)
+ /* 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 $~IA32_PQR_ASSOC_MASK, %edx
+ wrmsr
+#endif
+ /* invalidate cache contents. */
+ invd
+
+#if IS_ENABLED(CONFIG_CAR_NEM)
/* Knock down bit 1 then bit 0 of NEM control not combining steps. */
mov $(MSR_EVICT_CTL), %ecx
rdmsr
@@ -42,6 +62,7 @@ chipset_teardown_car:
wrmsr
and $(~(1 << 0)), %eax
wrmsr
+#endif
/* Return to caller. */
jmp *%ebx
diff --git a/src/soc/intel/apollolake/include/soc/cpu.h b/src/soc/intel/apollolake/include/soc/cpu.h
index 78fc0b0df6..8887c17e81 100644
--- a/src/soc/intel/apollolake/include/soc/cpu.h
+++ b/src/soc/intel/apollolake/include/soc/cpu.h
@@ -34,6 +34,26 @@ void apollolake_init_cpus(struct device *dev);
#define MSR_EVICT_CTL 0x2e0
#define MSR_EMULATE_PM_TMR 0x121
#define EMULATE_PM_TMR_EN (1 << 16)
+#define MSR_PREFETCH_CTL 0x1a4
+#define PREFETCH_L1_DISABLE (1 << 0)
+#define PREFETCH_L2_DISABLE (1 << 2)
+
+
+#define MSR_L2_QOS_MASK(reg) (0xd10 + reg)
+#define MSR_IA32_PQR_ASSOC 0xc8f
+/* MSR bits 33:32 encode slot number 0-3 */
+#define IA32_PQR_ASSOC_MASK (1 << 0 | 1 << 1)
+/* 16 way cache, 8 bits per QOS, 64 byte cache line, 1024 sets */
+#define CACHE_WAYS 16
+#define CACHE_BITS_PER_MASK 8
+#define CACHE_LINE_SIZE 64
+#define CACHE_SETS 1024
+/*
+ * Each bit in QOS mask controls this many bytes. This is calculated as:
+ * (CACHE_WAYS / CACHE_BITS_PER_MASK) * CACHE_LINE_SIZE * CACHE_SETS
+ */
+#define CACHE_QOS_SIZE_PER_BIT (128 * KiB)
+#define L2_CACHE_SIZE 0x100000
#define BASE_CLOCK_MHZ 100