Patch set updated for coreboot: cf962a0 Add cache_as_ram_ht.inc with hyper-threading CPU support
Kyösti Mälkki (kyosti.malkki@gmail.com) just uploaded a new patch set to gerrit, which you can find at http://review.coreboot.org/604
-gerrit
commit cf962a0ff185cb8b1d04f4744a0bbd063d521cf5 Author: Kyösti Mälkki kyosti.malkki@gmail.com Date: Wed Feb 1 19:08:23 2012 +0200
Add cache_as_ram_ht.inc with hyper-threading CPU support
This variant of cache_as_ram.inc starts the sibling CPU processors and clears the cache disable bits (CR0.CD) in case a hyper-threading CPU is detected.
A secondary main (named main_no_xip) built into romstage can be executed with XIP cache disabled. On my test setup (Intel e7505) ECC scrub fails if run with XIP enabled.
The code was developed for model_f25 from model_6ex. Some of the cache enable-disable logic seems spurious to me.
Change-Id: Ieabb86a7c47afb3e178cc75bb89dee3efe0c3d18 Signed-off-by: Kyösti Mälkki kyosti.malkki@gmail.com --- src/cpu/intel/car/cache_as_ram_ht.inc | 192 +++++++++++++++++++++++++-------- 1 files changed, 149 insertions(+), 43 deletions(-)
diff --git a/src/cpu/intel/car/cache_as_ram_ht.inc b/src/cpu/intel/car/cache_as_ram_ht.inc index 18ada29..3961535 100644 --- a/src/cpu/intel/car/cache_as_ram_ht.inc +++ b/src/cpu/intel/car/cache_as_ram_ht.inc @@ -3,6 +3,7 @@ * * Copyright (C) 2000,2007 Ronald G. Minnich rminnich@gmail.com * Copyright (C) 2007-2008 coresystems GmbH + * Copyright (C) 2012 Kyösti Mälkki kyosti.malkki@gmail.com * * This program is free software; you can redistribute it and/or modify * it under the terms of the GNU General Public License as published by @@ -21,22 +22,22 @@ #include <cpu/x86/stack.h> #include <cpu/x86/mtrr.h> #include <cpu/x86/post_code.h> +#include <cpu/x86/lapic_def.h>
#define CACHE_AS_RAM_SIZE CONFIG_DCACHE_RAM_SIZE #define CACHE_AS_RAM_BASE CONFIG_DCACHE_RAM_BASE
- /* Save the BIST result. */ - movl %eax, %ebp +#define lapic(x) $(LAPIC_DEFAULT_BASE | LAPIC_ ## x) +#define START_IPI_VECTOR ((CONFIG_AP_SIPI_VECTOR >> 12) & 0xff)
-cache_as_ram: - post_code(0x20) + .code32
- /* Send INIT IPI to all excluding ourself. */ - movl $0x000C4500, %eax - movl $0xFEE00300, %esi - movl %eax, (%esi) + /* Save the BIST result. */ + movl %eax, %ebp
- /* Zero out all fixed range and variable range MTRRs. */ + /* Zero out all fixed range and variable range MTRRs. + * For hyper-threaded CPU MTRRs are shared so we actually + * clear them more than once, but we don't care. */ movl $mtrr_table, %esi movl $((mtrr_table_end - mtrr_table) / 2), %edi xorl %eax, %eax @@ -55,6 +56,98 @@ clear_mtrrs: andl $(~0x00000cff), %eax wrmsr
+ /* For a hyper-threading processor, cache must not be disabled + * on an AP on the same physical package with the BSP. + */ + movl $01, %eax + cpuid + btl $28, %edx + jnc cache_as_ram + bswapl %ebx + cmpb $01, %bh + jbe cache_as_ram + +hyper_threading_cpu: + /* Enable local apic. */ + movl $LAPIC_BASE_MSR, %ecx + rdmsr + andl $(~0x0F), %edx /* MAXPHYWID = 36 */ + andl $(~LAPIC_BASE_MSR_ADDR_MASK), %eax + orl $(LAPIC_DEFAULT_BASE | LAPIC_BASE_MSR_ENABLE), %eax + wrmsr + andl $LAPIC_BASE_MSR_BOOTSTRAP_PROCESSOR, %eax + jnz bsp_init + +ap_init: + /* Do not disable cache (so BSP can enable it). */ + movl %cr0, %eax + andl $(~((1 << 30) | (1 << 29))), %eax + movl %eax, %cr0 + + /* MTRR registers are shared between HT siblings. */ + movl $0x300, %ecx + rdmsr + inc %eax + wrmsr + +ap_halt: + cli +1: hlt + jnz 1b + + +bsp_init: + /* Send INIT IPI to all excluding ourself. */ + movl lapic(ICR), %edi + movl $(LAPIC_DEST_ALLBUT | LAPIC_INT_ASSERT | LAPIC_DM_INIT), %eax +1: movl %eax, (%edi) + movl $0x30, %ecx +2: pause + dec %ecx + jnz 2b + movl (%edi), %ecx + andl $LAPIC_ICR_BUSY, %ecx + jnz 1b + + /* delay 10 ms */ + movl $10000, %ecx +1: inb $0x80, %al + dec %ecx + jnz 1b + + /* Send Start IPI to all excluding ourself. */ + movl lapic(ICR), %edi + movl $(LAPIC_DEST_ALLBUT | LAPIC_DM_STARTUP | START_IPI_VECTOR), %eax +1: movl %eax, (%edi) + movl $0x30, %ecx +2: pause + dec %ecx + jnz 2b + movl (%edi), %ecx + andl $LAPIC_ICR_BUSY, %ecx + jnz 1b + + /* delay 250 us */ + movl $250, %ecx +1: inb $0x80, %al + dec %ecx + jnz 1b + + /* Wait for sibling CPU to start. */ +1: movl $0x300, %ecx + rdmsr + andl %eax, %eax + jnz sipi_complete + + movl $0x30, %ecx +2: pause + dec %ecx + jnz 2b + jmp 1b + +sipi_complete: + +cache_as_ram: /* Set Cache-as-RAM base address. */ movl $(MTRRphysBase_MSR(0)), %ecx movl $(CACHE_AS_RAM_BASE | MTRR_TYPE_WRBACK), %eax @@ -67,36 +160,34 @@ clear_mtrrs: movl $0x0000000f, %edx wrmsr
- /* Enable MTRR. */ + /* Enable variable MTRRs. */ movl $MTRRdefType_MSR, %ecx rdmsr orl $MTRRdefTypeEn, %eax wrmsr
- /* Enable L2 cache. */ - movl $0x11e, %ecx - rdmsr - orl $(1 << 8), %eax - wrmsr - - /* Enable cache (CR0.CD = 0, CR0.NW = 0). */ + /* Enable cache. */ movl %cr0, %eax andl $(~((1 << 30) | (1 << 29))), %eax invd movl %eax, %cr0 + invd
- /* Clear the cache memory reagion. */ - movl $CACHE_AS_RAM_BASE, %esi - movl %esi, %edi - movl $(CACHE_AS_RAM_SIZE / 4), %ecx - // movl $0x23322332, %eax + /* Clear memory for stack. */ + cld xorl %eax, %eax + movl $(CACHE_AS_RAM_BASE), %edi + movl $(CACHE_AS_RAM_SIZE / 4), %ecx rep stosl
+#if 0 /* Enable Cache-as-RAM mode by disabling cache. */ movl %cr0, %eax orl $(1 << 30), %eax + wbinvd movl %eax, %cr0 + wbinvd +#endif
#if CONFIG_XIP_ROM_SIZE /* Enable cache for our code in Flash because we do XIP here */ @@ -117,32 +208,60 @@ clear_mtrrs: wrmsr #endif /* CONFIG_XIP_ROM_SIZE */
- /* Enable cache. */ - movl %cr0, %eax +#if 0 + /* Enable cache (CR0.CD = 0, CR0.NW = 0). */ + movl %cr0, %eax andl $(~((1 << 30) | (1 << 29))), %eax movl %eax, %cr0 +#endif
- /* Set up the stack pointer. */ #if CONFIG_USBDEBUG /* Leave some space for the struct ehci_debug_info. */ - movl $(CACHE_AS_RAM_BASE + CACHE_AS_RAM_SIZE - 4 - 128), %eax + movl $(CACHE_AS_RAM_BASE + CACHE_AS_RAM_SIZE - 128), %esp #else - movl $(CACHE_AS_RAM_BASE + CACHE_AS_RAM_SIZE - 4), %eax + movl $(CACHE_AS_RAM_BASE + CACHE_AS_RAM_SIZE), %esp #endif - movl %eax, %esp
/* Restore the BIST result. */ movl %ebp, %eax movl %esp, %ebp pushl %eax
- post_code(0x23) + post_code(0x23);
- /* Call romstage.c main function. */ call main + addl $4, %esp
post_code(0x2f)
+#if CONFIG_MAIN_WITHOUT_XIP + /* With Intel e7505 memory controller, hardware ECC scrub + * halts and/or flushes call stack if run with XIP enabled. + */ + + /* Disable cache. */ + movl %cr0, %eax + orl $(1 << 30), %eax + movl %eax, %cr0 + + /* Disable Flash XIP. */ + movl $MTRRphysMask_MSR(1), %ecx + movl $0x0, %edx + movl $0x0, %eax + wrmsr + invd + + /* Enable cache. */ + movl %cr0, %eax + andl $~((1 << 30) | (1 << 29)), %eax + movl %eax, %cr0 + + pushl $0x0 + call main_no_xip + addl $4, %esp + movl %esp, %ebp +#endif + post_code(0x30)
/* Disable cache. */ @@ -159,20 +278,7 @@ clear_mtrrs: wrmsr
post_code(0x31) - invd -#if 0 - xorl %eax, %eax - xorl %edx, %edx - movl $MTRRphysBase_MSR(0), %ecx - wrmsr - movl $MTRRphysMask_MSR(0), %ecx - wrmsr - movl $MTRRphysBase_MSR(1), %ecx - wrmsr - movl $MTRRphysMask_MSR(1), %ecx - wrmsr -#endif
post_code(0x33)
-
Kyösti Mälkki