| Message ID | 7a4866b928c0905ec1074dce1bfc7984c968a58a.1479788404.git.panand@redhat.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
Hi Pratyush, (CC: Mark, mismatched memory attributes in paragraph 3?) On 22/11/16 04:32, Pratyush Anand wrote: > This patch adds support to enable/disable d-cache, which can be used for > faster purgatory sha256 verification. (I'm not clear why we want the sha256, but that is being discussed elsewhere on the thread) > We are supporting only 4K and 64K page sizes. This code will not work if a > hardware is not supporting at least one of these page sizes. Therefore, > D-cache is disabled by default and enabled only when "enable-dcache" is > passed to the kexec(). I don't think the maybe-4K/maybe-64K/maybe-neither logic is needed. It would be a lot simpler to only support one page size, which should be 4K as that is what UEFI requires. (If there are CPUs that only support one size, I bet its 4K!) I would go as far as to generate the page tables at 'kexec -l' time, and only if '/sys/firmware/efi' exists to indicate we booted via UEFI. (and therefore must support 4K pages). This would keep the purgatory code as simple as possible. I don't think the performance difference between 4K and 64K page sizes will be measurable, is purgatory really performance sensitive code? > Since this is an identity mapped system, so VA_BITS will be same as max PA > bits supported. If VA_BITS <= 42 for 64K and <= 39 for 4K then only one > level of page table will be there with block descriptor entries. > Otherwise, For 4K mapping, TTBR points to level 0 lookups, which will have > only table entries pointing to a level 1 lookup. Level 1 will have only > block entries which will map 1GB block. For 64K mapping, TTBR points to > level 1 lookups, which will have only table entries pointing to a level 2 > lookup. Level 2 will have only block entries which will map 512MB block. If This is more complexity to pick a VA size. Why not always use the maximum 48bit VA? The cost is negligible compared to having simpler (easier to review!) purgatory code. By always using 1GB blocks you may be creating aliases with mismatched attributes: * If kdump only reserves 128MB, your 1GB mapping will alias whatever else was in the same 1GB of address space. This could be a reserved region with some other memory attributes. * With kdump, we may have failed to park the other CPUs if they are executing with interrupts masked and haven't yet handled the smp_send_stop() IPI. * One of these other CPUs could be reading/writing in this area as it doesn't belong to the kdump reserved area, just happens to be in the same 1GB. I need to dig through the ARM-ARM to find out what happens next, but I'm pretty sure this is well into the "don't do that" territory. It would be much better to force the memory areas to be a multiple of 2MB and 2MB aligned, which will allow you to use 2M section mappings for memory, (but not the uart). This way we only map regions we had reserved and know are memory. > UART base address and RAM addresses are not at least 1GB and 512MB apart > for 4K and 64K respectively, then mapping result could be unpredictable. In > that case we need to support one more level of granularity, but until > someone needs that keep it like this only. > > We can not allocate dynamic memory in purgatory. Therefore we keep page > table allocation size fixed as (3 * MAX_PAGE_SIZE). (page_table) points to > first level (having only table entries) and (page_table + MAX_PAGE_SIZE) > points to table at next level (having block entries). If index for RAM > area and UART area in first table is not same, then we will need another > next level table which will be located at (page_table + 2 * MAX_PAGE_SIZE). > diff --git a/purgatory/arch/arm64/cache-asm.S b/purgatory/arch/arm64/cache-asm.S > new file mode 100644 > index 000000000000..bef97ef48888 > --- /dev/null > +++ b/purgatory/arch/arm64/cache-asm.S > @@ -0,0 +1,186 @@ > +/* > + * Some of the routines have been copied from Linux Kernel, therefore > + * copying the license as well. > + * > + * Copyright (C) 2001 Deep Blue Solutions Ltd. > + * Copyright (C) 2012 ARM Ltd. > + * Copyright (C) 2015 Pratyush Anand <panand@redhat.com> > + * > + * This program is free software; you can redistribute it and/or modify > + * it under the terms of the GNU General Public License version 2 as > + * published by the Free Software Foundation. > + * > + * This program is distributed in the hope that it will be useful, > + * but WITHOUT ANY WARRANTY; without even the implied warranty of > + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the > + * GNU General Public License for more details. > + * > + * You should have received a copy of the GNU General Public License > + * along with this program. If not, see <http://www.gnu.org/licenses/>. > + */ > + > +#include "cache.h" > + > +/* > + * dcache_line_size - get the minimum D-cache line size from the CTR register. > + */ > + .macro dcache_line_size, reg, tmp > + mrs \tmp, ctr_el0 // read CTR > + ubfm \tmp, \tmp, #16, #19 // cache line size encoding > + mov \reg, #4 // bytes per word > + lsl \reg, \reg, \tmp // actual cache line size > + .endm > + > +/* > + * inval_cache_range(start, end) > + * - x0 - start - start address of region > + * - x1 - end - end address of region > + */ > +.globl inval_cache_range > +inval_cache_range: > + dcache_line_size x2, x3 > + sub x3, x2, #1 > + tst x1, x3 // end cache line aligned? > + bic x1, x1, x3 > + b.eq 1f > + dc civac, x1 // clean & invalidate D / U line > +1: tst x0, x3 // start cache line aligned? > + bic x0, x0, x3 > + b.eq 2f > + dc civac, x0 // clean & invalidate D / U line > + b 3f > +2: dc ivac, x0 // invalidate D / U line > +3: add x0, x0, x2 > + cmp x0, x1 > + b.lo 2b > + dsb sy > + ret > +/* > + * flush_dcache_range(start, end) > + * - x0 - start - start address of region > + * - x1 - end - end address of region > + * > + */ > +.globl flush_dcache_range > +flush_dcache_range: > + dcache_line_size x2, x3 > + sub x3, x2, #1 > + bic x0, x0, x3 > +1: dc civac, x0 // clean & invalidate D line / unified line > + add x0, x0, x2 > + cmp x0, x1 > + b.lo 1b > + dsb sy > + ret > + > +/* > + * invalidate_tlbs_el1() > + */ > +.globl invalidate_tlbs_el1 > +invalidate_tlbs_el1: > + dsb nshst > + tlbi vmalle1 > + dsb nsh > + isb > + ret > + > +/* > + * invalidate_tlbs_el2() > + */ > +.globl invalidate_tlbs_el2 > +invalidate_tlbs_el2: > + dsb nshst > + tlbi alle2 > + dsb nsh > + isb > + ret > + > +/* > + * get_mm_feature_reg0_val - Get information about supported MM > + * features > + */ > +.globl get_mm_feature_reg0_val > +get_mm_feature_reg0_val: > + mrs x0, ID_AA64MMFR0_EL1 > + ret > + > +/* > + * get_current_el - Get information about current exception level > + */ > +.globl get_current_el > +get_current_el: > + mrs x0, CurrentEL > + lsr x0, x0, #2 > + ret > + > +/* > + * invalidate_icache - Invalidate I-cache > + */ > +.globl invalidate_icache > +invalidate_icache: > + ic iallu > + dsb nsh > + isb > + ret > + > +/* > + * set_mair_tcr_ttbr_sctlr_el1(page_table, tcr_flags) - sets MAIR, TCR , TTBR and SCTLR registers > + * x0 - page_table - Page Table Base > + * x1 - tcr_flags - TCR Flags to be set > + */ > +.globl set_mair_tcr_ttbr_sctlr_el1 > +set_mair_tcr_ttbr_sctlr_el1: > + ldr x2, =MEMORY_ATTRIBUTES > + msr mair_el1, x2 > + msr tcr_el1, x1 > + msr ttbr0_el1, x0 > + isb > + mrs x0, sctlr_el1 > + ldr x3, =SCTLR_ELx_FLAGS > + orr x0, x0, x3 > + msr sctlr_el1, x0 > + isb > + ret > + > +/* > + * set_mair_tcr_ttbr_sctlr_el2(page_table, tcr_flags) - sets MAIR, TCR , TTBR and SCTLR registers > + * x0 - page_table - Page Table Base > + * x1 - tcr_flags - TCR Flags to be set > + */ > +.globl set_mair_tcr_ttbr_sctlr_el2 > +set_mair_tcr_ttbr_sctlr_el2: > + ldr x2, =MEMORY_ATTRIBUTES > + msr mair_el2, x2 > + msr tcr_el2, x1 > + msr ttbr0_el2, x0 > + isb > + mrs x0, sctlr_el2 > + ldr x3, =SCTLR_ELx_FLAGS > + orr x0, x0, x3 > + msr sctlr_el2, x0 > + isb > + ret > + > +/* > + * reset_sctlr_el1 - disables cache and mmu > + */ > +.globl reset_sctlr_el1 > +reset_sctlr_el1: > + mrs x0, sctlr_el1 > + bic x0, x0, #SCTLR_ELx_C > + bic x0, x0, #SCTLR_ELx_M > + msr sctlr_el1, x0 > + isb > + ret > + > +/* > + * reset_sctlr_el2 - disables cache and mmu > + */ > +.globl reset_sctlr_el2 > +reset_sctlr_el2: > + mrs x0, sctlr_el2 > + bic x0, x0, #SCTLR_ELx_C > + bic x0, x0, #SCTLR_ELx_M > + msr sctlr_el2, x0 > + isb > + ret > diff --git a/purgatory/arch/arm64/cache.c b/purgatory/arch/arm64/cache.c > new file mode 100644 > index 000000000000..3c7e058ccf11 > --- /dev/null > +++ b/purgatory/arch/arm64/cache.c > @@ -0,0 +1,330 @@ > +/* > + * Copyright (C) 2015 Pratyush Anand <panand@redhat.com> > + * > + * This program is free software; you can redistribute it and/or modify > + * it under the terms of the GNU General Public License version 2 as > + * published by the Free Software Foundation. > + * > + * This program is distributed in the hope that it will be useful, > + * but WITHOUT ANY WARRANTY; without even the implied warranty of > + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the > + * GNU General Public License for more details. > + * > + * You should have received a copy of the GNU General Public License > + * along with this program. If not, see <http://www.gnu.org/licenses/>. > + */ > + > +/* We are supporting only 4K and 64K page sizes. This code will not work if > + * a hardware is not supporting at least one of these page sizes. > + * Therefore, D-cache is disabled by default and enabled only when > + * "enable-dcache" is passed to the kexec(). > + * Since this is an identity mapped system, so VA_BITS will be same as max > + * PA bits supported. If VA_BITS <= 42 for 64K and <= 39 for 4K then only > + * one level of page table will be there with block descriptor entries. > + * Otherwise, For 4K mapping, TTBR points to level 0 lookups, which will > + * have only table entries pointing to a level 1 lookup. Level 1 will have > + * only block entries which will map 1GB block.For 64K mapping, TTBR points > + * to level 1 lookups, which will have only table entries pointing to a > + * level 2 lookup. Level 2 will have only block entries which will map > + * 512MB block. If UART base address and RAM addresses are not at least 1GB > + * and 512MB apart for 4K and 64K respectively, then mapping result could > + * be unpredictable. In that case we need to support one more level of > + * granularity, but until someone needs that keep it like this only. > + * We can not allocate dynamic memory in purgatory. Therefore we keep page > + * table allocation size fixed as (3 * MAX_PAGE_SIZE). (page_table) points > + * to first level (having only table entries) and (page_table + > + * MAX_PAGE_SIZE) points to table at next level (having block entries). If > + * index for RAM area and UART area in first table is not same, then we > + * will need another next level table which will be located at (page_table > + * + 2 * MAX_PAGE_SIZE). > + */ > + > +#include <stdint.h> > +#include <string.h> > +#include <purgatory.h> > +#include "cache.h" > + > +static uint64_t page_shift; > +static uint64_t pgtable_level; > +static uint64_t va_bits; > + > +static uint64_t page_table[PAGE_TABLE_SIZE / sizeof(uint64_t)] __attribute__ ((aligned (MAX_PAGE_SIZE))) = { }; > +static uint64_t page_table_used; > + > +#define PAGE_SIZE (1 << page_shift) > +/* > + * is_4k_page_supported - return true if 4k page is supported else > + * false > + */ > +static int is_4k_page_supported(void) > +{ > + return ((get_mm_feature_reg0_val() & ID_AA64MMFR0_TGRAN4_MASK) == > + ID_AA64MMFR0_TGRAN4_SUPPORTED); > +} > + > +/* > + * is_64k_page_supported - return true if 64k page is supported else > + * false > + */ > +static int is_64k_page_supported(void) > +{ > + return ((get_mm_feature_reg0_val() & ID_AA64MMFR0_TGRAN64_MASK) == > + ID_AA64MMFR0_TGRAN64_SUPPORTED); > +} > + > +/* > + * get_ips_bits - return supported IPS bits > + */ > +static uint64_t get_ips_bits(void) > +{ > + return ((get_mm_feature_reg0_val() & ID_AA64MMFR0_PARANGE_MASK) >> > + ID_AA64MMFR0_PARANGE_SHIFT); > +} > + > +/* > + * get_va_bits - return supported VA bits (For identity mapping VA = PA) > + */ > +static uint64_t get_va_bits(void) > +{ > + uint64_t ips = get_ips_bits(); > + > + switch(ips) { > + case ID_AA64MMFR0_PARANGE_48: > + return 48; > + case ID_AA64MMFR0_PARANGE_44: > + return 44; > + case ID_AA64MMFR0_PARANGE_42: > + return 42; > + case ID_AA64MMFR0_PARANGE_40: > + return 40; > + case ID_AA64MMFR0_PARANGE_36: > + return 36; > + default: > + return 32; > + } > +} > + > +/* > + * get_section_shift - get block shift for supported page size > + */ > +static uint64_t get_section_shift(void) > +{ > + if (page_shift == 16) > + return 29; > + else if(page_shift == 12) > + return 30; > + else > + return 0; > +} > + > +/* > + * get_section_mask - get section mask for supported page size > + */ > +static uint64_t get_section_mask(void) > +{ > + if (page_shift == 16) > + return 0x1FFF; > + else if(page_shift == 12) > + return 0x1FF; > + else > + return 0; > +} > + > +/* > + * get_pgdir_shift - get pgdir shift for supported page size > + */ > +static uint64_t get_pgdir_shift(void) > +{ > + if (page_shift == 16) > + return 42; > + else if(page_shift == 12) > + return 39; > + else > + return 0; > +} > + > +/* > + * init_page_table - Initializes page table locations > + */ > + > +static void init_page_table(void) > +{ > + /* > + * Invalidate the page tables to avoid potential dirty cache lines > + * being evicted. > + */ How do these lines get dirty? arm64_relocate_new_kernel() invalidated these pages to PoC before it copied the data. If they were speculatively fetched (I don't know the rules of when/how that happens) they may be wrong, but will be clean and not written back. If we change them in purgatory, you invalidate again from enable_mmu_dcache(). I don't think this is needed. > + inval_cache_range((uint64_t)page_table, > + (uint64_t)page_table + PAGE_TABLE_SIZE); > + memset(page_table, 0, PAGE_TABLE_SIZE); > +} > +/* > + * create_identity_mapping(start, end, flags) > + * start - start address > + * end - end address > + * flags - MMU Flags for Normal or Device type memory > + */ > +static void create_identity_mapping(uint64_t start, uint64_t end, > + uint64_t flags) > +{ > + uint32_t sec_shift, pgdir_shift, sec_mask; > + uint64_t desc, s1, e1, s2, e2; > + uint64_t *table2; > + > + s1 = start; > + e1 = end - 1; > + > + sec_shift = get_section_shift(); > + if (pgtable_level == 1) { > + s1 >>= sec_shift; > + e1 >>= sec_shift; > + do { > + desc = s1 << sec_shift; > + desc |= flags; > + page_table[s1] = desc; > + s1++; > + } while (s1 <= e1); > + } else { > + pgdir_shift = get_pgdir_shift(); > + sec_mask = get_section_mask(); > + s1 >>= pgdir_shift; > + e1 >>= pgdir_shift; > + do { > + /* > + * If there is no table entry then write a new > + * entry else, use old entry > + */ > + if (!page_table[s1]) { > + table2 = &page_table[(++page_table_used * > + MAX_PAGE_SIZE) / > + sizeof(uint64_t)]; > + desc = (uint64_t)table2 | PMD_TYPE_TABLE; > + page_table[s1] = desc; > + } else { > + table2 = (uint64_t *)(page_table[s1] & > + ~PMD_TYPE_MASK); > + } > + s1++; > + s2 = start >> sec_shift; > + s2 &= sec_mask; > + e2 = (end - 1) >> sec_shift; > + e2 &= sec_mask; > + do { > + desc = s2 << sec_shift; > + desc |= flags; > + table2[s2] = desc; > + s2++; > + } while (s2 <= e2); > + } while (s1 <= e1); > + } > +} (I will need to come back to this ... it looks pretty complicated. If you mimic Linux's p?d/pte macros it will be more familiar and easier to read.) > + > +/* > + * enable_mmu_dcache: Enable mmu and D-cache in sctlr_el1 > + */ > +static void enable_mmu_dcache(void) > +{ > + uint64_t tcr_flags = TCR_FLAGS | TCR_T0SZ(va_bits); > + > + switch(page_shift) { > + case 16: > + tcr_flags |= TCR_TG0_64K; > + break; > + case 12: > + tcr_flags |= TCR_TG0_4K; > + break; > + default: > + printf("page shift not supported\n"); > + return; > + } > + /* > + * Since the page tables have been populated with non-cacheable > + * accesses (MMU disabled), invalidate the page tables to remove > + * any speculatively loaded cache lines. > + */ > + inval_cache_range((uint64_t)page_table, > + (uint64_t)page_table + PAGE_TABLE_SIZE); > + > + switch(get_current_el()) { > + case 2: > + invalidate_tlbs_el2(); > + tcr_flags |= (get_ips_bits() << TCR_PS_EL2_SHIFT); > + set_mair_tcr_ttbr_sctlr_el2((uint64_t)page_table, tcr_flags); > + break; > + case 1: > + invalidate_tlbs_el1(); > + tcr_flags |= (get_ips_bits() << TCR_IPS_EL1_SHIFT); > + set_mair_tcr_ttbr_sctlr_el1((uint64_t)page_table, tcr_flags); > + break; > + default: > + return; > + } > + invalidate_icache(); What is this protecting against? We have executed instructions between here and setting the I+M bits in set_mair_tcr_ttbr_sctlr_el1(). (so it may be too late) arm64_relocate_new_kernel() already did 'ic iallu' before it branched into the purgatory code. No executable code has been changed or moved since then, so I don't think this is necessary. > +} > + > +/* > + * enable_dcache: Enable D-cache and set appropriate attributes > + * ram_start - Start address of RAM > + * ram_end - End address of RAM > + * uart_base - Base address of uart > + */ > +int enable_dcache(uint64_t ram_start, uint64_t ram_end, uint64_t uart_base) > +{ > + va_bits = get_va_bits(); > + > + page_table_used = 0; > + if (is_64k_page_supported()) { > + page_shift = 16; > + if (va_bits <= 42) > + pgtable_level = 1; > + else > + pgtable_level = 2; > + } else if (is_4k_page_supported()) { > + page_shift = 12; > + if (va_bits <= 39) > + pgtable_level = 1; > + else > + pgtable_level = 2; > + } else { > + printf("Valid Page Granule not supported by hardware\n"); > + return -1; > + } > + init_page_table(); > + create_identity_mapping(ram_start, ram_end, MM_MMUFLAGS_NORMAL); > + printf("Normal identity mapping created from %lx to %lx\n", > + ram_start, ram_end); > + if (uart_base) { > + create_identity_mapping((uint64_t)uart_base, > + (uint64_t)uart_base + PAGE_SIZE, > + MM_MMUFLAGS_DEVICE); > + printf("Device identity mapping created from %lx to %lx\n", > + (uint64_t)uart_base, > + (uint64_t)uart_base + PAGE_SIZE); > + } > + enable_mmu_dcache(); > + printf("Cache Enabled\n"); > + > + return 0; > +} > + > +/* > + * disable_dcache: Disable D-cache and flush RAM locations > + * ram_start - Start address of RAM > + * ram_end - End address of RAM > + */ > +void disable_dcache(uint64_t ram_start, uint64_t ram_end) > +{ > + switch(get_current_el()) { > + case 2: > + reset_sctlr_el2(); > + break; > + case 1: > + reset_sctlr_el1(); You have C code running between disabling the MMU and cleaning the cache. The compiler is allowed to move data on and off the stack in here, but after disabling the MMU it will see whatever was on the stack before we turned the MMU on. Any data written at the beginning of this function is left in the caches. I'm afraid this sort of stuff needs to be done in assembly! > + break; > + default: > + return; > + } > + invalidate_icache(); > + flush_dcache_range(ram_start, ram_end); > + printf("Cache Disabled\n"); > +} > diff --git a/purgatory/arch/arm64/cache.h b/purgatory/arch/arm64/cache.h > new file mode 100644 > index 000000000000..c988020566e3 > --- /dev/null > +++ b/purgatory/arch/arm64/cache.h > @@ -0,0 +1,79 @@ > +#ifndef __CACHE_H__ > +#define __CACHE_H__ > + > +#define MT_DEVICE_NGNRNE 0 > +#define MT_DEVICE_NGNRE 1 > +#define MT_DEVICE_GRE 2 > +#define MT_NORMAL_NC 3 > +#define MT_NORMAL 4 You only use two of these. I guess this is so the MAIR value matches the kernel? > + > +#ifndef __ASSEMBLER__ > + > +#define MAX_PAGE_SIZE 0x10000 > +#define PAGE_TABLE_SIZE (3 * MAX_PAGE_SIZE) > +#define ID_AA64MMFR0_TGRAN64_SHIFT 24 > +#define ID_AA64MMFR0_TGRAN4_SHIFT 28 > +#define ID_AA64MMFR0_TGRAN64_MASK (0xFUL << ID_AA64MMFR0_TGRAN64_SHIFT) > +#define ID_AA64MMFR0_TGRAN4_MASK (0xFUL << ID_AA64MMFR0_TGRAN4_SHIFT) > +#define ID_AA64MMFR0_TGRAN64_SUPPORTED 0x0 > +#define ID_AA64MMFR0_TGRAN4_SUPPORTED 0x0 > +#define ID_AA64MMFR0_PARANGE_SHIFT 0 > +#define ID_AA64MMFR0_PARANGE_MASK (0xFUL << ID_AA64MMFR0_PARANGE_SHIFT) > +#define ID_AA64MMFR0_PARANGE_48 0x5 > +#define ID_AA64MMFR0_PARANGE_44 0x4 > +#define ID_AA64MMFR0_PARANGE_42 0x3 > +#define ID_AA64MMFR0_PARANGE_40 0x2 > +#define ID_AA64MMFR0_PARANGE_36 0x1 > +#define ID_AA64MMFR0_PARANGE_32 0x0 > + > +#define TCR_TG0_64K (1UL << 14) > +#define TCR_TG0_4K (0UL << 14) > +#define TCR_SHARED_NONE (0UL << 12) > +#define TCR_ORGN_WBWA (1UL << 10) > +#define TCR_IRGN_WBWA (1UL << 8) > +#define TCR_IPS_EL1_SHIFT 32 > +#define TCR_PS_EL2_SHIFT 16 > +#define TCR_T0SZ(x) ((unsigned long)(64 - (x)) << 0) > +#define TCR_FLAGS (TCR_SHARED_NONE | TCR_ORGN_WBWA | TCR_IRGN_WBWA) > + > +#define PMD_TYPE_SECT (1UL << 0) > +#define PMD_TYPE_TABLE (3UL << 0) > +#define PMD_TYPE_MASK 0x3 > +#define PMD_SECT_AF (1UL << 10) > +#define PMD_ATTRINDX(t) ((unsigned long)(t) << 2) > +#define PMD_FLAGS_NORMAL (PMD_TYPE_SECT | PMD_SECT_AF) > +#define PMD_SECT_PXN (1UL << 53) > +#define PMD_SECT_UXN (1UL << 54) > +#define PMD_FLAGS_DEVICE (PMD_TYPE_SECT | PMD_SECT_AF | PMD_SECT_PXN | PMD_SECT_UXN) > +#define MM_MMUFLAGS_NORMAL PMD_ATTRINDX(MT_NORMAL) | PMD_FLAGS_NORMAL > +#define MM_MMUFLAGS_DEVICE PMD_ATTRINDX(MT_DEVICE_NGNRE) | PMD_FLAGS_DEVICE > + > +void disable_dcache(uint64_t ram_start, uint64_t ram_end); > +int enable_dcache(uint64_t ram_start, uint64_t ram_end, uint64_t uart_base); > +uint64_t get_mm_feature_reg0_val(void); > +void inval_cache_range(uint64_t start, uint64_t end); > +void flush_dcache_range(uint64_t start, uint64_t end); > +uint64_t get_current_el(void); > +void set_mair_tcr_ttbr_sctlr_el1(uint64_t page_table, uint64_t tcr_flags); > +void set_mair_tcr_ttbr_sctlr_el2(uint64_t page_table, uint64_t tcr_flags); > +void invalidate_tlbs_el1(void); > +void invalidate_tlbs_el2(void); > +void invalidate_icache(void); > +void reset_sctlr_el1(void); > +void reset_sctlr_el2(void); > +#else > +#define MEMORY_ATTRIBUTES ((0x00 << (MT_DEVICE_NGNRNE*8)) | \ > + (0x04 << (MT_DEVICE_NGNRE*8)) | \ > + (0x0C << (MT_DEVICE_GRE*8)) | \ > + (0x44 << (MT_NORMAL_NC*8)) | \ > + (0xFF << (MT_NORMAL*8))) Again, you only use two of these. > +/* Common SCTLR_ELx flags. */ > +#define SCTLR_ELx_I (1 << 12) > +#define SCTLR_ELx_C (1 << 2) > +#define SCTLR_ELx_M (1 << 0) > + > +#define SCTLR_ELx_FLAGS (SCTLR_ELx_M | SCTLR_ELx_C | SCTLR_ELx_I) > + > +#endif > +#endif > Thanks! James
Hi James, Thanks a lot for your review. Its helpful. On Saturday 26 November 2016 12:00 AM, James Morse wrote: > Hi Pratyush, > > (CC: Mark, mismatched memory attributes in paragraph 3?) > > On 22/11/16 04:32, Pratyush Anand wrote: >> This patch adds support to enable/disable d-cache, which can be used for >> faster purgatory sha256 verification. > > (I'm not clear why we want the sha256, but that is being discussed elsewhere on > the thread) > > >> We are supporting only 4K and 64K page sizes. This code will not work if a >> hardware is not supporting at least one of these page sizes. Therefore, >> D-cache is disabled by default and enabled only when "enable-dcache" is >> passed to the kexec(). > > I don't think the maybe-4K/maybe-64K/maybe-neither logic is needed. It would be > a lot simpler to only support one page size, which should be 4K as that is what > UEFI requires. (If there are CPUs that only support one size, I bet its 4K!) Ok.. So, I will implement a new version after considering that 4K will always be supported. If 4K is not supported by hw(which is very unlikely) then there would be no d-cache enabling feature. > > I would go as far as to generate the page tables at 'kexec -l' time, and only if Ok..So you mean that I create a new section which will have page table entries mapping physicalmemory represented by remaining section, and then purgatory can just enable mmu with page table from that section, right? Seems doable. can do that. > '/sys/firmware/efi' exists to indicate we booted via UEFI. (and therefore must > support 4K pages). This would keep the purgatory code as simple as possible. What about reading ID_AA64MMFR0_EL1 instead of /sys/firmware/efi? That can also tell us that whether 4K is supported or not? > > I don't think the performance difference between 4K and 64K page sizes will be > measurable, is purgatory really performance sensitive code? I agree, implementing only 4K will make it very simple. > > >> Since this is an identity mapped system, so VA_BITS will be same as max PA >> bits supported. If VA_BITS <= 42 for 64K and <= 39 for 4K then only one >> level of page table will be there with block descriptor entries. >> Otherwise, For 4K mapping, TTBR points to level 0 lookups, which will have >> only table entries pointing to a level 1 lookup. Level 1 will have only >> block entries which will map 1GB block. For 64K mapping, TTBR points to >> level 1 lookups, which will have only table entries pointing to a level 2 >> lookup. Level 2 will have only block entries which will map 512MB block. If > > This is more complexity to pick a VA size. Why not always use the maximum 48bit > VA? The cost is negligible compared to having simpler (easier to review!) > purgatory code. > > By always using 1GB blocks you may be creating aliases with mismatched attributes: > * If kdump only reserves 128MB, your 1GB mapping will alias whatever else was > in the same 1GB of address space. This could be a reserved region with some > other memory attributes. > * With kdump, we may have failed to park the other CPUs if they are executing > with interrupts masked and haven't yet handled the smp_send_stop() IPI. > * One of these other CPUs could be reading/writing in this area as it doesn't > belong to the kdump reserved area, just happens to be in the same 1GB. > > I need to dig through the ARM-ARM to find out what happens next, but I'm pretty > sure this is well into the "don't do that" territory. > > > It would be much better to force the memory areas to be a multiple of 2MB and > 2MB aligned, which will allow you to use 2M section mappings for memory, (but > not the uart). This way we only map regions we had reserved and know are memory. OK. So, 48 bit VA, 4K page size, 3 level page table with entries in 3rd level representing 2M block size. > > >> UART base address and RAM addresses are not at least 1GB and 512MB apart >> for 4K and 64K respectively, then mapping result could be unpredictable. In >> that case we need to support one more level of granularity, but until >> someone needs that keep it like this only. >> >> We can not allocate dynamic memory in purgatory. Therefore we keep page >> table allocation size fixed as (3 * MAX_PAGE_SIZE). (page_table) points to >> first level (having only table entries) and (page_table + MAX_PAGE_SIZE) >> points to table at next level (having block entries). If index for RAM >> area and UART area in first table is not same, then we will need another >> next level table which will be located at (page_table + 2 * MAX_PAGE_SIZE). > > >> diff --git a/purgatory/arch/arm64/cache-asm.S b/purgatory/arch/arm64/cache-asm.S >> new file mode 100644 >> index 000000000000..bef97ef48888 >> --- /dev/null >> +++ b/purgatory/arch/arm64/cache-asm.S >> @@ -0,0 +1,186 @@ >> +/* >> + * Some of the routines have been copied from Linux Kernel, therefore >> + * copying the license as well. >> + * >> + * Copyright (C) 2001 Deep Blue Solutions Ltd. >> + * Copyright (C) 2012 ARM Ltd. >> + * Copyright (C) 2015 Pratyush Anand <panand@redhat.com> >> + * >> + * This program is free software; you can redistribute it and/or modify >> + * it under the terms of the GNU General Public License version 2 as >> + * published by the Free Software Foundation. >> + * >> + * This program is distributed in the hope that it will be useful, >> + * but WITHOUT ANY WARRANTY; without even the implied warranty of >> + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the >> + * GNU General Public License for more details. >> + * >> + * You should have received a copy of the GNU General Public License >> + * along with this program. If not, see <http://www.gnu.org/licenses/>. >> + */ >> + >> +#include "cache.h" >> + >> +/* >> + * dcache_line_size - get the minimum D-cache line size from the CTR register. >> + */ >> + .macro dcache_line_size, reg, tmp >> + mrs \tmp, ctr_el0 // read CTR >> + ubfm \tmp, \tmp, #16, #19 // cache line size encoding >> + mov \reg, #4 // bytes per word >> + lsl \reg, \reg, \tmp // actual cache line size >> + .endm >> + >> +/* >> + * inval_cache_range(start, end) >> + * - x0 - start - start address of region >> + * - x1 - end - end address of region >> + */ >> +.globl inval_cache_range >> +inval_cache_range: >> + dcache_line_size x2, x3 >> + sub x3, x2, #1 >> + tst x1, x3 // end cache line aligned? >> + bic x1, x1, x3 >> + b.eq 1f >> + dc civac, x1 // clean & invalidate D / U line >> +1: tst x0, x3 // start cache line aligned? >> + bic x0, x0, x3 >> + b.eq 2f >> + dc civac, x0 // clean & invalidate D / U line >> + b 3f >> +2: dc ivac, x0 // invalidate D / U line >> +3: add x0, x0, x2 >> + cmp x0, x1 >> + b.lo 2b >> + dsb sy >> + ret >> +/* >> + * flush_dcache_range(start, end) >> + * - x0 - start - start address of region >> + * - x1 - end - end address of region >> + * >> + */ >> +.globl flush_dcache_range >> +flush_dcache_range: >> + dcache_line_size x2, x3 >> + sub x3, x2, #1 >> + bic x0, x0, x3 >> +1: dc civac, x0 // clean & invalidate D line / unified line >> + add x0, x0, x2 >> + cmp x0, x1 >> + b.lo 1b >> + dsb sy >> + ret >> + >> +/* >> + * invalidate_tlbs_el1() >> + */ >> +.globl invalidate_tlbs_el1 >> +invalidate_tlbs_el1: >> + dsb nshst >> + tlbi vmalle1 >> + dsb nsh >> + isb >> + ret >> + >> +/* >> + * invalidate_tlbs_el2() >> + */ >> +.globl invalidate_tlbs_el2 >> +invalidate_tlbs_el2: >> + dsb nshst >> + tlbi alle2 >> + dsb nsh >> + isb >> + ret >> + >> +/* >> + * get_mm_feature_reg0_val - Get information about supported MM >> + * features >> + */ >> +.globl get_mm_feature_reg0_val >> +get_mm_feature_reg0_val: >> + mrs x0, ID_AA64MMFR0_EL1 >> + ret >> + >> +/* >> + * get_current_el - Get information about current exception level >> + */ >> +.globl get_current_el >> +get_current_el: >> + mrs x0, CurrentEL >> + lsr x0, x0, #2 >> + ret >> + >> +/* >> + * invalidate_icache - Invalidate I-cache >> + */ >> +.globl invalidate_icache >> +invalidate_icache: >> + ic iallu >> + dsb nsh >> + isb >> + ret >> + >> +/* >> + * set_mair_tcr_ttbr_sctlr_el1(page_table, tcr_flags) - sets MAIR, TCR , TTBR and SCTLR registers >> + * x0 - page_table - Page Table Base >> + * x1 - tcr_flags - TCR Flags to be set >> + */ >> +.globl set_mair_tcr_ttbr_sctlr_el1 >> +set_mair_tcr_ttbr_sctlr_el1: >> + ldr x2, =MEMORY_ATTRIBUTES >> + msr mair_el1, x2 >> + msr tcr_el1, x1 >> + msr ttbr0_el1, x0 >> + isb >> + mrs x0, sctlr_el1 >> + ldr x3, =SCTLR_ELx_FLAGS >> + orr x0, x0, x3 >> + msr sctlr_el1, x0 >> + isb >> + ret >> + >> +/* >> + * set_mair_tcr_ttbr_sctlr_el2(page_table, tcr_flags) - sets MAIR, TCR , TTBR and SCTLR registers >> + * x0 - page_table - Page Table Base >> + * x1 - tcr_flags - TCR Flags to be set >> + */ >> +.globl set_mair_tcr_ttbr_sctlr_el2 >> +set_mair_tcr_ttbr_sctlr_el2: >> + ldr x2, =MEMORY_ATTRIBUTES >> + msr mair_el2, x2 >> + msr tcr_el2, x1 >> + msr ttbr0_el2, x0 >> + isb >> + mrs x0, sctlr_el2 >> + ldr x3, =SCTLR_ELx_FLAGS >> + orr x0, x0, x3 >> + msr sctlr_el2, x0 >> + isb >> + ret >> + >> +/* >> + * reset_sctlr_el1 - disables cache and mmu >> + */ >> +.globl reset_sctlr_el1 >> +reset_sctlr_el1: >> + mrs x0, sctlr_el1 >> + bic x0, x0, #SCTLR_ELx_C >> + bic x0, x0, #SCTLR_ELx_M >> + msr sctlr_el1, x0 >> + isb >> + ret >> + >> +/* >> + * reset_sctlr_el2 - disables cache and mmu >> + */ >> +.globl reset_sctlr_el2 >> +reset_sctlr_el2: >> + mrs x0, sctlr_el2 >> + bic x0, x0, #SCTLR_ELx_C >> + bic x0, x0, #SCTLR_ELx_M >> + msr sctlr_el2, x0 >> + isb >> + ret >> diff --git a/purgatory/arch/arm64/cache.c b/purgatory/arch/arm64/cache.c >> new file mode 100644 >> index 000000000000..3c7e058ccf11 >> --- /dev/null >> +++ b/purgatory/arch/arm64/cache.c >> @@ -0,0 +1,330 @@ >> +/* >> + * Copyright (C) 2015 Pratyush Anand <panand@redhat.com> >> + * >> + * This program is free software; you can redistribute it and/or modify >> + * it under the terms of the GNU General Public License version 2 as >> + * published by the Free Software Foundation. >> + * >> + * This program is distributed in the hope that it will be useful, >> + * but WITHOUT ANY WARRANTY; without even the implied warranty of >> + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the >> + * GNU General Public License for more details. >> + * >> + * You should have received a copy of the GNU General Public License >> + * along with this program. If not, see <http://www.gnu.org/licenses/>. >> + */ >> + >> +/* We are supporting only 4K and 64K page sizes. This code will not work if >> + * a hardware is not supporting at least one of these page sizes. >> + * Therefore, D-cache is disabled by default and enabled only when >> + * "enable-dcache" is passed to the kexec(). >> + * Since this is an identity mapped system, so VA_BITS will be same as max >> + * PA bits supported. If VA_BITS <= 42 for 64K and <= 39 for 4K then only >> + * one level of page table will be there with block descriptor entries. >> + * Otherwise, For 4K mapping, TTBR points to level 0 lookups, which will >> + * have only table entries pointing to a level 1 lookup. Level 1 will have >> + * only block entries which will map 1GB block.For 64K mapping, TTBR points >> + * to level 1 lookups, which will have only table entries pointing to a >> + * level 2 lookup. Level 2 will have only block entries which will map >> + * 512MB block. If UART base address and RAM addresses are not at least 1GB >> + * and 512MB apart for 4K and 64K respectively, then mapping result could >> + * be unpredictable. In that case we need to support one more level of >> + * granularity, but until someone needs that keep it like this only. >> + * We can not allocate dynamic memory in purgatory. Therefore we keep page >> + * table allocation size fixed as (3 * MAX_PAGE_SIZE). (page_table) points >> + * to first level (having only table entries) and (page_table + >> + * MAX_PAGE_SIZE) points to table at next level (having block entries). If >> + * index for RAM area and UART area in first table is not same, then we >> + * will need another next level table which will be located at (page_table >> + * + 2 * MAX_PAGE_SIZE). >> + */ >> + >> +#include <stdint.h> >> +#include <string.h> >> +#include <purgatory.h> >> +#include "cache.h" >> + >> +static uint64_t page_shift; >> +static uint64_t pgtable_level; >> +static uint64_t va_bits; >> + >> +static uint64_t page_table[PAGE_TABLE_SIZE / sizeof(uint64_t)] __attribute__ ((aligned (MAX_PAGE_SIZE))) = { }; >> +static uint64_t page_table_used; >> + >> +#define PAGE_SIZE (1 << page_shift) >> +/* >> + * is_4k_page_supported - return true if 4k page is supported else >> + * false >> + */ >> +static int is_4k_page_supported(void) >> +{ >> + return ((get_mm_feature_reg0_val() & ID_AA64MMFR0_TGRAN4_MASK) == >> + ID_AA64MMFR0_TGRAN4_SUPPORTED); >> +} >> + >> +/* >> + * is_64k_page_supported - return true if 64k page is supported else >> + * false >> + */ >> +static int is_64k_page_supported(void) >> +{ >> + return ((get_mm_feature_reg0_val() & ID_AA64MMFR0_TGRAN64_MASK) == >> + ID_AA64MMFR0_TGRAN64_SUPPORTED); >> +} >> + >> +/* >> + * get_ips_bits - return supported IPS bits >> + */ >> +static uint64_t get_ips_bits(void) >> +{ >> + return ((get_mm_feature_reg0_val() & ID_AA64MMFR0_PARANGE_MASK) >> >> + ID_AA64MMFR0_PARANGE_SHIFT); >> +} >> + >> +/* >> + * get_va_bits - return supported VA bits (For identity mapping VA = PA) >> + */ >> +static uint64_t get_va_bits(void) >> +{ >> + uint64_t ips = get_ips_bits(); >> + >> + switch(ips) { >> + case ID_AA64MMFR0_PARANGE_48: >> + return 48; >> + case ID_AA64MMFR0_PARANGE_44: >> + return 44; >> + case ID_AA64MMFR0_PARANGE_42: >> + return 42; >> + case ID_AA64MMFR0_PARANGE_40: >> + return 40; >> + case ID_AA64MMFR0_PARANGE_36: >> + return 36; >> + default: >> + return 32; >> + } >> +} >> + >> +/* >> + * get_section_shift - get block shift for supported page size >> + */ >> +static uint64_t get_section_shift(void) >> +{ >> + if (page_shift == 16) >> + return 29; >> + else if(page_shift == 12) >> + return 30; >> + else >> + return 0; >> +} >> + >> +/* >> + * get_section_mask - get section mask for supported page size >> + */ >> +static uint64_t get_section_mask(void) >> +{ >> + if (page_shift == 16) >> + return 0x1FFF; >> + else if(page_shift == 12) >> + return 0x1FF; >> + else >> + return 0; >> +} >> + >> +/* >> + * get_pgdir_shift - get pgdir shift for supported page size >> + */ >> +static uint64_t get_pgdir_shift(void) >> +{ >> + if (page_shift == 16) >> + return 42; >> + else if(page_shift == 12) >> + return 39; >> + else >> + return 0; >> +} >> + >> +/* >> + * init_page_table - Initializes page table locations >> + */ >> + >> +static void init_page_table(void) >> +{ >> + /* >> + * Invalidate the page tables to avoid potential dirty cache lines >> + * being evicted. >> + */ > > How do these lines get dirty? arm64_relocate_new_kernel() invalidated these > pages to PoC before it copied the data. If they were speculatively fetched (I > don't know the rules of when/how that happens) they may be wrong, but will be > clean and not written back. If we change them in purgatory, you invalidate again > from enable_mmu_dcache(). I don't think this is needed. > Had taken it from kernel arch/arm64/kernel/head.S. But anyway, since this part will go to kexec code as you suggested, so we will not need it there for sure. > >> + inval_cache_range((uint64_t)page_table, >> + (uint64_t)page_table + PAGE_TABLE_SIZE); >> + memset(page_table, 0, PAGE_TABLE_SIZE); >> +} >> +/* >> + * create_identity_mapping(start, end, flags) >> + * start - start address >> + * end - end address >> + * flags - MMU Flags for Normal or Device type memory >> + */ >> +static void create_identity_mapping(uint64_t start, uint64_t end, >> + uint64_t flags) >> +{ >> + uint32_t sec_shift, pgdir_shift, sec_mask; >> + uint64_t desc, s1, e1, s2, e2; >> + uint64_t *table2; >> + >> + s1 = start; >> + e1 = end - 1; >> + >> + sec_shift = get_section_shift(); >> + if (pgtable_level == 1) { >> + s1 >>= sec_shift; >> + e1 >>= sec_shift; >> + do { >> + desc = s1 << sec_shift; >> + desc |= flags; >> + page_table[s1] = desc; >> + s1++; >> + } while (s1 <= e1); >> + } else { >> + pgdir_shift = get_pgdir_shift(); >> + sec_mask = get_section_mask(); >> + s1 >>= pgdir_shift; >> + e1 >>= pgdir_shift; >> + do { >> + /* >> + * If there is no table entry then write a new >> + * entry else, use old entry >> + */ >> + if (!page_table[s1]) { >> + table2 = &page_table[(++page_table_used * >> + MAX_PAGE_SIZE) / >> + sizeof(uint64_t)]; >> + desc = (uint64_t)table2 | PMD_TYPE_TABLE; >> + page_table[s1] = desc; >> + } else { >> + table2 = (uint64_t *)(page_table[s1] & >> + ~PMD_TYPE_MASK); >> + } >> + s1++; >> + s2 = start >> sec_shift; >> + s2 &= sec_mask; >> + e2 = (end - 1) >> sec_shift; >> + e2 &= sec_mask; >> + do { >> + desc = s2 << sec_shift; >> + desc |= flags; >> + table2[s2] = desc; >> + s2++; >> + } while (s2 <= e2); >> + } while (s1 <= e1); >> + } >> +} > > (I will need to come back to this ... it looks pretty complicated. If you mimic > Linux's p?d/pte macros it will be more familiar and easier to read.) Ok, will try to take definitions from head.S, as far as possible. > > >> + >> +/* >> + * enable_mmu_dcache: Enable mmu and D-cache in sctlr_el1 >> + */ >> +static void enable_mmu_dcache(void) >> +{ >> + uint64_t tcr_flags = TCR_FLAGS | TCR_T0SZ(va_bits); >> + >> + switch(page_shift) { >> + case 16: >> + tcr_flags |= TCR_TG0_64K; >> + break; >> + case 12: >> + tcr_flags |= TCR_TG0_4K; >> + break; >> + default: >> + printf("page shift not supported\n"); >> + return; >> + } >> + /* >> + * Since the page tables have been populated with non-cacheable >> + * accesses (MMU disabled), invalidate the page tables to remove >> + * any speculatively loaded cache lines. >> + */ >> + inval_cache_range((uint64_t)page_table, >> + (uint64_t)page_table + PAGE_TABLE_SIZE); >> + >> + switch(get_current_el()) { >> + case 2: >> + invalidate_tlbs_el2(); >> + tcr_flags |= (get_ips_bits() << TCR_PS_EL2_SHIFT); >> + set_mair_tcr_ttbr_sctlr_el2((uint64_t)page_table, tcr_flags); >> + break; >> + case 1: >> + invalidate_tlbs_el1(); >> + tcr_flags |= (get_ips_bits() << TCR_IPS_EL1_SHIFT); >> + set_mair_tcr_ttbr_sctlr_el1((uint64_t)page_table, tcr_flags); >> + break; >> + default: >> + return; >> + } > >> + invalidate_icache(); > > What is this protecting against? We have executed instructions between here and > setting the I+M bits in set_mair_tcr_ttbr_sctlr_el1(). (so it may be too late) > > arm64_relocate_new_kernel() already did 'ic iallu' before it branched into the > purgatory code. No executable code has been changed or moved since then, so I > don't think this is necessary. OK. > > >> +} >> + >> +/* >> + * enable_dcache: Enable D-cache and set appropriate attributes >> + * ram_start - Start address of RAM >> + * ram_end - End address of RAM >> + * uart_base - Base address of uart >> + */ >> +int enable_dcache(uint64_t ram_start, uint64_t ram_end, uint64_t uart_base) >> +{ >> + va_bits = get_va_bits(); >> + >> + page_table_used = 0; >> + if (is_64k_page_supported()) { >> + page_shift = 16; >> + if (va_bits <= 42) >> + pgtable_level = 1; >> + else >> + pgtable_level = 2; >> + } else if (is_4k_page_supported()) { >> + page_shift = 12; >> + if (va_bits <= 39) >> + pgtable_level = 1; >> + else >> + pgtable_level = 2; >> + } else { >> + printf("Valid Page Granule not supported by hardware\n"); >> + return -1; >> + } >> + init_page_table(); >> + create_identity_mapping(ram_start, ram_end, MM_MMUFLAGS_NORMAL); >> + printf("Normal identity mapping created from %lx to %lx\n", >> + ram_start, ram_end); >> + if (uart_base) { >> + create_identity_mapping((uint64_t)uart_base, >> + (uint64_t)uart_base + PAGE_SIZE, >> + MM_MMUFLAGS_DEVICE); >> + printf("Device identity mapping created from %lx to %lx\n", >> + (uint64_t)uart_base, >> + (uint64_t)uart_base + PAGE_SIZE); >> + } >> + enable_mmu_dcache(); >> + printf("Cache Enabled\n"); >> + >> + return 0; >> +} >> + >> +/* >> + * disable_dcache: Disable D-cache and flush RAM locations >> + * ram_start - Start address of RAM >> + * ram_end - End address of RAM >> + */ >> +void disable_dcache(uint64_t ram_start, uint64_t ram_end) >> +{ >> + switch(get_current_el()) { >> + case 2: >> + reset_sctlr_el2(); >> + break; >> + case 1: >> + reset_sctlr_el1(); > > You have C code running between disabling the MMU and cleaning the cache. The > compiler is allowed to move data on and off the stack in here, but after > disabling the MMU it will see whatever was on the stack before we turned the MMU > on. Any data written at the beginning of this function is left in the caches. > > I'm afraid this sort of stuff needs to be done in assembly! All these routines are self coded in assembly even though they are called from C, so should be safe I think. Anyway, I can keep all of them in assembly as well. > > >> + break; >> + default: >> + return; >> + } >> + invalidate_icache(); >> + flush_dcache_range(ram_start, ram_end); >> + printf("Cache Disabled\n"); >> +} >> diff --git a/purgatory/arch/arm64/cache.h b/purgatory/arch/arm64/cache.h >> new file mode 100644 >> index 000000000000..c988020566e3 >> --- /dev/null >> +++ b/purgatory/arch/arm64/cache.h >> @@ -0,0 +1,79 @@ >> +#ifndef __CACHE_H__ >> +#define __CACHE_H__ >> + >> +#define MT_DEVICE_NGNRNE 0 >> +#define MT_DEVICE_NGNRE 1 >> +#define MT_DEVICE_GRE 2 >> +#define MT_NORMAL_NC 3 >> +#define MT_NORMAL 4 > > You only use two of these. I guess this is so the MAIR value matches the kernel? OK, can remove others. Yes, they are matching with kernel. > > >> + >> +#ifndef __ASSEMBLER__ >> + >> +#define MAX_PAGE_SIZE 0x10000 >> +#define PAGE_TABLE_SIZE (3 * MAX_PAGE_SIZE) >> +#define ID_AA64MMFR0_TGRAN64_SHIFT 24 >> +#define ID_AA64MMFR0_TGRAN4_SHIFT 28 >> +#define ID_AA64MMFR0_TGRAN64_MASK (0xFUL << ID_AA64MMFR0_TGRAN64_SHIFT) >> +#define ID_AA64MMFR0_TGRAN4_MASK (0xFUL << ID_AA64MMFR0_TGRAN4_SHIFT) >> +#define ID_AA64MMFR0_TGRAN64_SUPPORTED 0x0 >> +#define ID_AA64MMFR0_TGRAN4_SUPPORTED 0x0 >> +#define ID_AA64MMFR0_PARANGE_SHIFT 0 >> +#define ID_AA64MMFR0_PARANGE_MASK (0xFUL << ID_AA64MMFR0_PARANGE_SHIFT) >> +#define ID_AA64MMFR0_PARANGE_48 0x5 >> +#define ID_AA64MMFR0_PARANGE_44 0x4 >> +#define ID_AA64MMFR0_PARANGE_42 0x3 >> +#define ID_AA64MMFR0_PARANGE_40 0x2 >> +#define ID_AA64MMFR0_PARANGE_36 0x1 >> +#define ID_AA64MMFR0_PARANGE_32 0x0 >> + >> +#define TCR_TG0_64K (1UL << 14) >> +#define TCR_TG0_4K (0UL << 14) >> +#define TCR_SHARED_NONE (0UL << 12) >> +#define TCR_ORGN_WBWA (1UL << 10) >> +#define TCR_IRGN_WBWA (1UL << 8) >> +#define TCR_IPS_EL1_SHIFT 32 >> +#define TCR_PS_EL2_SHIFT 16 >> +#define TCR_T0SZ(x) ((unsigned long)(64 - (x)) << 0) >> +#define TCR_FLAGS (TCR_SHARED_NONE | TCR_ORGN_WBWA | TCR_IRGN_WBWA) >> + >> +#define PMD_TYPE_SECT (1UL << 0) >> +#define PMD_TYPE_TABLE (3UL << 0) >> +#define PMD_TYPE_MASK 0x3 >> +#define PMD_SECT_AF (1UL << 10) >> +#define PMD_ATTRINDX(t) ((unsigned long)(t) << 2) >> +#define PMD_FLAGS_NORMAL (PMD_TYPE_SECT | PMD_SECT_AF) >> +#define PMD_SECT_PXN (1UL << 53) >> +#define PMD_SECT_UXN (1UL << 54) >> +#define PMD_FLAGS_DEVICE (PMD_TYPE_SECT | PMD_SECT_AF | PMD_SECT_PXN | PMD_SECT_UXN) >> +#define MM_MMUFLAGS_NORMAL PMD_ATTRINDX(MT_NORMAL) | PMD_FLAGS_NORMAL >> +#define MM_MMUFLAGS_DEVICE PMD_ATTRINDX(MT_DEVICE_NGNRE) | PMD_FLAGS_DEVICE >> + >> +void disable_dcache(uint64_t ram_start, uint64_t ram_end); >> +int enable_dcache(uint64_t ram_start, uint64_t ram_end, uint64_t uart_base); >> +uint64_t get_mm_feature_reg0_val(void); >> +void inval_cache_range(uint64_t start, uint64_t end); >> +void flush_dcache_range(uint64_t start, uint64_t end); >> +uint64_t get_current_el(void); >> +void set_mair_tcr_ttbr_sctlr_el1(uint64_t page_table, uint64_t tcr_flags); >> +void set_mair_tcr_ttbr_sctlr_el2(uint64_t page_table, uint64_t tcr_flags); >> +void invalidate_tlbs_el1(void); >> +void invalidate_tlbs_el2(void); >> +void invalidate_icache(void); >> +void reset_sctlr_el1(void); >> +void reset_sctlr_el2(void); >> +#else >> +#define MEMORY_ATTRIBUTES ((0x00 << (MT_DEVICE_NGNRNE*8)) | \ >> + (0x04 << (MT_DEVICE_NGNRE*8)) | \ >> + (0x0C << (MT_DEVICE_GRE*8)) | \ >> + (0x44 << (MT_NORMAL_NC*8)) | \ >> + (0xFF << (MT_NORMAL*8))) > > Again, you only use two of these. OK, will remove others. > > >> +/* Common SCTLR_ELx flags. */ >> +#define SCTLR_ELx_I (1 << 12) >> +#define SCTLR_ELx_C (1 << 2) >> +#define SCTLR_ELx_M (1 << 0) >> + >> +#define SCTLR_ELx_FLAGS (SCTLR_ELx_M | SCTLR_ELx_C | SCTLR_ELx_I) >> + >> +#endif >> +#endif >> ~Pratyush
On Wednesday 14 December 2016 03:08 PM, Pratyush Anand wrote: > >> >> I would go as far as to generate the page tables at 'kexec -l' time, >> and only if > > Ok..So you mean that I create a new section which will have page table > entries mapping physicalmemory represented by remaining section, and > then purgatory can just enable mmu with page table from that section, > right? Seems doable. can do that. I see a problem here. If we create page table as a new segment then, how can we verify in purgatory that sha for page table is correct? We need page table before sha verification start,and we can not rely the page table created by first kernel until it's sha is verified. So a chicken-egg problem. I think, creating page table will just take fraction of second and should be good even in purgatory, What do you say? ~Pratyush
Hi Pratyush, On 14/12/16 09:38, Pratyush Anand wrote: > On Saturday 26 November 2016 12:00 AM, James Morse wrote: >> On 22/11/16 04:32, Pratyush Anand wrote: >>> This patch adds support to enable/disable d-cache, which can be used for >>> faster purgatory sha256 verification. >> >> (I'm not clear why we want the sha256, but that is being discussed elsewhere on >> the thread) >> >> >>> We are supporting only 4K and 64K page sizes. This code will not work if a >>> hardware is not supporting at least one of these page sizes. Therefore, >>> D-cache is disabled by default and enabled only when "enable-dcache" is >>> passed to the kexec(). >> >> I don't think the maybe-4K/maybe-64K/maybe-neither logic is needed. It would be >> a lot simpler to only support one page size, which should be 4K as that is what >> UEFI requires. (If there are CPUs that only support one size, I bet its 4K!) > > Ok.. So, I will implement a new version after considering that 4K will always be > supported. If 4K is not supported by hw(which is very unlikely) then there would > be no d-cache enabling feature. Sounds good tom me. I think its important to keep the purgatory code as small and as simple as possible as its very hard to debug. If we do get bug reports they are likely to be 'it didn't nothing', with no further details. If it only fails on some platform we don't have access to its basically impossible. >> I would go as far as to generate the page tables at 'kexec -l' time, and only if > > Ok..So you mean that I create a new section which will have page table entries > mapping physicalmemory represented by remaining section, and then purgatory can > just enable mmu with page table from that section, right? Seems doable. can do > that. > >> '/sys/firmware/efi' exists to indicate we booted via UEFI. (and therefore must >> support 4K pages). This would keep the purgatory code as simple as possible. > > What about reading ID_AA64MMFR0_EL1 instead of /sys/firmware/efi? That can also > tell us that whether 4K is supported or not? If you're doing it at EL1/EL2 in the purgatory code, sure. But if you generate the page tables at 'kexec -l' time you can't read this register from EL0 so you need another way to guess if 4K pages are supported (or just assume they are and test that register once you're in purgatory). I was looking for some way to print a message at 'kexec -l' time that the sha256 would be slow as 4K wasn't supported. (a message printed at any other time won't get seen). >>> +/* >>> + * disable_dcache: Disable D-cache and flush RAM locations >>> + * ram_start - Start address of RAM >>> + * ram_end - End address of RAM >>> + */ >>> +void disable_dcache(uint64_t ram_start, uint64_t ram_end) >>> +{ >>> + switch(get_current_el()) { >>> + case 2: >>> + reset_sctlr_el2(); >>> + break; >>> + case 1: >>> + reset_sctlr_el1(); >> >> You have C code running between disabling the MMU and cleaning the cache. The >> compiler is allowed to move data on and off the stack in here, but after >> disabling the MMU it will see whatever was on the stack before we turned the MMU >> on. Any data written at the beginning of this function is left in the caches. >> >> I'm afraid this sort of stuff needs to be done in assembly! > > All these routines are self coded in assembly even though they are called > from C, so should be safe I think. Anyway, I can keep all of them in > assembly as well. You can't tell the compiler that the stack data is inaccessible until the dcache clean call completes. Some future version may do really crazy things in here. You can decompile what your compiler version produces to check it doesn't load/store to the stack, but that doesn't mean my compiler version does the same. This is the kind of thing that is extremely difficult to debug, its best not to take the risk. Thanks, James
Hi Pratyush, On 14/12/16 10:12, Pratyush Anand wrote: > On Wednesday 14 December 2016 03:08 PM, Pratyush Anand wrote: >>> I would go as far as to generate the page tables at 'kexec -l' time, >>> and only if >> >> Ok..So you mean that I create a new section which will have page table >> entries mapping physicalmemory represented by remaining section, and >> then purgatory can just enable mmu with page table from that section, >> right? Seems doable. can do that. > > I see a problem here. If we create page table as a new segment then, how can we > verify in purgatory that sha for page table is correct? We need page table > before sha verification start,and we can not rely the page table created by > first kernel until it's sha is verified. So a chicken-egg problem. There is more than one of those! What happens if your sha256 calculation code is corrupted? You have to run it before you know. The same goes for all the purgatory code. This is why I think its better to do this in the kernel before we exit to purgatory, but obviously that doesn't work for kdump. > I think, creating page table will just take fraction of second and should be > good even in purgatory, What do you say? If it's for kdump its best-effort. I think its easier/simpler to generate and debug them at 'kexec -l' time, but if you're worried about the increased area that could be corrupted then do it in purgatory. Thanks, James
On Wed, Dec 14, 2016 at 11:16:07AM +0000, James Morse wrote: > Hi Pratyush, > On 14/12/16 09:38, Pratyush Anand wrote: > > On Saturday 26 November 2016 12:00 AM, James Morse wrote: > >> On 22/11/16 04:32, Pratyush Anand wrote: > >>> +/* > >>> + * disable_dcache: Disable D-cache and flush RAM locations > >>> + * ram_start - Start address of RAM > >>> + * ram_end - End address of RAM > >>> + */ > >>> +void disable_dcache(uint64_t ram_start, uint64_t ram_end) > >>> +{ > >>> + switch(get_current_el()) { > >>> + case 2: > >>> + reset_sctlr_el2(); > >>> + break; > >>> + case 1: > >>> + reset_sctlr_el1(); > >> > >> You have C code running between disabling the MMU and cleaning the cache. The > >> compiler is allowed to move data on and off the stack in here, but after > >> disabling the MMU it will see whatever was on the stack before we turned the MMU > >> on. Any data written at the beginning of this function is left in the caches. > >> > >> I'm afraid this sort of stuff needs to be done in assembly! > > > > All these routines are self coded in assembly even though they are called > > from C, so should be safe I think. Anyway, I can keep all of them in > > assembly as well. > > You can't tell the compiler that the stack data is inaccessible until the dcache > clean call completes. Some future version may do really crazy things in here. > You can decompile what your compiler version produces to check it doesn't > load/store to the stack, but that doesn't mean my compiler version does the > same. This is the kind of thing that is extremely difficult to debug, its best > not to take the risk. FWIW, I completely agree. We've been bitten in the past; see commit 5e051531447259e5 ("arm64: convert part of soft_restart() to assembly") for an example. Thanks, Mark.
On Wed, Dec 14, 2016 at 11:16:17AM +0000, James Morse wrote: > Hi Pratyush, > > On 14/12/16 10:12, Pratyush Anand wrote: > > On Wednesday 14 December 2016 03:08 PM, Pratyush Anand wrote: > >>> I would go as far as to generate the page tables at 'kexec -l' time, > >>> and only if > >> > >> Ok..So you mean that I create a new section which will have page table > >> entries mapping physicalmemory represented by remaining section, and > >> then purgatory can just enable mmu with page table from that section, > >> right? Seems doable. can do that. > > > > I see a problem here. If we create page table as a new segment then, how can we > > verify in purgatory that sha for page table is correct? We need page table > > before sha verification start,and we can not rely the page table created by > > first kernel until it's sha is verified. So a chicken-egg problem. > > There is more than one of those! What happens if your sha256 calculation code is > corrupted? You have to run it before you know. The same goes for all the > purgatory code. > > This is why I think its better to do this in the kernel before we exit to > purgatory, but obviously that doesn't work for kdump. I see in an earlier message that the need for sha256 was being discussed in another thread. Do either of you happen to have a pointer to that. To me, it seems like it doesn't come with much benefit for the kdump case given that's best-effort anyway, and as above the verification code could have been be corrupted. In the non-kdump case it's not strictly necessary and seems like a debugging aid rather than a necessary piece of functionality -- if that's the case, a 20 second delay isn't the end of the world... Thanks, Mark.
Hi Mark, On 14/12/16 11:37, Mark Rutland wrote: > On Wed, Dec 14, 2016 at 11:16:17AM +0000, James Morse wrote: >> On 14/12/16 10:12, Pratyush Anand wrote: >>> On Wednesday 14 December 2016 03:08 PM, Pratyush Anand wrote: >>>>> I would go as far as to generate the page tables at 'kexec -l' time, >>>>> and only if >>>> >>>> Ok..So you mean that I create a new section which will have page table >>>> entries mapping physicalmemory represented by remaining section, and >>>> then purgatory can just enable mmu with page table from that section, >>>> right? Seems doable. can do that. >>> >>> I see a problem here. If we create page table as a new segment then, how can we >>> verify in purgatory that sha for page table is correct? We need page table >>> before sha verification start,and we can not rely the page table created by >>> first kernel until it's sha is verified. So a chicken-egg problem. >> >> There is more than one of those! What happens if your sha256 calculation code is >> corrupted? You have to run it before you know. The same goes for all the >> purgatory code. >> >> This is why I think its better to do this in the kernel before we exit to >> purgatory, but obviously that doesn't work for kdump. > > I see in an earlier message that the need for sha256 was being discussed > in another thread. Do either of you happen to have a pointer to that. https://www.spinics.net/lists/arm-kernel/msg544472.html Thanks, James
Hi James, Thanks for your input !! On Wednesday 14 December 2016 04:46 PM, James Morse wrote: > Hi Pratyush, > > On 14/12/16 10:12, Pratyush Anand wrote: >> > On Wednesday 14 December 2016 03:08 PM, Pratyush Anand wrote: >>>> >>> I would go as far as to generate the page tables at 'kexec -l' time, >>>> >>> and only if >>> >> >>> >> Ok..So you mean that I create a new section which will have page table >>> >> entries mapping physicalmemory represented by remaining section, and >>> >> then purgatory can just enable mmu with page table from that section, >>> >> right? Seems doable. can do that. >> > >> > I see a problem here. If we create page table as a new segment then, how can we >> > verify in purgatory that sha for page table is correct? We need page table >> > before sha verification start,and we can not rely the page table created by >> > first kernel until it's sha is verified. So a chicken-egg problem. > There is more than one of those! What happens if your sha256 calculation code is > corrupted? You have to run it before you know. The same goes for all the > purgatory code. > OK, seems reasonable... will do it in kexec code. > This is why I think its better to do this in the kernel before we exit to > purgatory, but obviously that doesn't work for kdump. > > >> > I think, creating page table will just take fraction of second and should be >> > good even in purgatory, What do you say? > If it's for kdump its best-effort. I think its easier/simpler to generate and > debug them at 'kexec -l' time, but if you're worried about the increased area > that could be corrupted then do it in purgatory. > ~Pratyush
On Wednesday 14 December 2016 05:07 PM, Mark Rutland wrote: > On Wed, Dec 14, 2016 at 11:16:17AM +0000, James Morse wrote: >> Hi Pratyush, >> >> On 14/12/16 10:12, Pratyush Anand wrote: >>> On Wednesday 14 December 2016 03:08 PM, Pratyush Anand wrote: >>>>> I would go as far as to generate the page tables at 'kexec -l' time, >>>>> and only if >>>> >>>> Ok..So you mean that I create a new section which will have page table >>>> entries mapping physicalmemory represented by remaining section, and >>>> then purgatory can just enable mmu with page table from that section, >>>> right? Seems doable. can do that. >>> >>> I see a problem here. If we create page table as a new segment then, how can we >>> verify in purgatory that sha for page table is correct? We need page table >>> before sha verification start,and we can not rely the page table created by >>> first kernel until it's sha is verified. So a chicken-egg problem. >> >> There is more than one of those! What happens if your sha256 calculation code is >> corrupted? You have to run it before you know. The same goes for all the >> purgatory code. >> >> This is why I think its better to do this in the kernel before we exit to >> purgatory, but obviously that doesn't work for kdump. > > I see in an earlier message that the need for sha256 was being discussed > in another thread. Do either of you happen to have a pointer to that. > patch 0/2 of this series. > To me, it seems like it doesn't come with much benefit for the kdump > case given that's best-effort anyway, and as above the verification code > could have been be corrupted. In the non-kdump case it's not strictly > necessary and seems like a debugging aid rather than a necessary piece > of functionality -- if that's the case, a 20 second delay isn't the end > of the world... Even for the non-kdump ie `kexec -l` case we do not have a functionality to bypass sha verification in kexec-tools. --lite option with the kexec-tools was discouraged and not accepted. So,it is 20s for both `kexec -l` and `kexec -p`. Also other arch like x86_64 takes negligible time in sha verification. ~Pratyush
Hi, On Wed, Dec 14, 2016 at 05:51:05PM +0530, Pratyush Anand wrote: > > On Wednesday 14 December 2016 05:07 PM, Mark Rutland wrote: > >I see in an earlier message that the need for sha256 was being discussed > >in another thread. Do either of you happen to have a pointer to that. > > patch 0/2 of this series. AFAICT, that just says the the existing sha256 check is slow, not *why* a sha256 check of some description is necessary. I'm still at a loss as to why it is considered necessary, rather than being a debugging aid or sanity check. > >To me, it seems like it doesn't come with much benefit for the kdump > >case given that's best-effort anyway, and as above the verification code > >could have been be corrupted. In the non-kdump case it's not strictly > >necessary and seems like a debugging aid rather than a necessary piece > >of functionality -- if that's the case, a 20 second delay isn't the end > >of the world... > > Even for the non-kdump ie `kexec -l` case we do not have a > functionality to bypass sha verification in kexec-tools. --lite > option with the kexec-tools was discouraged and not accepted. Ok. Do you have a pointer to the thread regarding that, for context? > So,it is 20s for both `kexec -l` and `kexec -p`. Well, unless we can have a --{no-,}sha-check, and make the default NO for arm64. > Also other arch like x86_64 takes negligible time in sha verification. That's certainly an argument for not changing the other architectures, but given it's slow for arm64, we could have a different default... Thanks, Mark.
On Wednesday 14 December 2016 07:14 PM, Mark Rutland wrote: >> Even for the non-kdump ie `kexec -l` case we do not have a >> > functionality to bypass sha verification in kexec-tools. --lite >> > option with the kexec-tools was discouraged and not accepted. > Ok. Do you have a pointer to the thread regarding that, for context? > https://lists.ozlabs.org/pipermail/petitboot/2015-October/000141.html https://lists.ozlabs.org/pipermail/petitboot/2015-October/000136.html ~Pratyush
diff --git a/purgatory/arch/arm64/Makefile b/purgatory/arch/arm64/Makefile index 636abeab17b2..0f80f8165d90 100644 --- a/purgatory/arch/arm64/Makefile +++ b/purgatory/arch/arm64/Makefile @@ -11,6 +11,8 @@ arm64_PURGATORY_EXTRA_CFLAGS = \ arm64_PURGATORY_SRCS += \ purgatory/arch/arm64/entry.S \ + purgatory/arch/arm64/cache-asm.S \ + purgatory/arch/arm64/cache.c \ purgatory/arch/arm64/purgatory-arm64.c dist += \ diff --git a/purgatory/arch/arm64/cache-asm.S b/purgatory/arch/arm64/cache-asm.S new file mode 100644 index 000000000000..bef97ef48888 --- /dev/null +++ b/purgatory/arch/arm64/cache-asm.S @@ -0,0 +1,186 @@ +/* + * Some of the routines have been copied from Linux Kernel, therefore + * copying the license as well. + * + * Copyright (C) 2001 Deep Blue Solutions Ltd. + * Copyright (C) 2012 ARM Ltd. + * Copyright (C) 2015 Pratyush Anand <panand@redhat.com> + * + * This program is free software; you can redistribute it and/or modify + * it under the terms of the GNU General Public License version 2 as + * published by the Free Software Foundation. + * + * This program is distributed in the hope that it will be useful, + * but WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + * GNU General Public License for more details. + * + * You should have received a copy of the GNU General Public License + * along with this program. If not, see <http://www.gnu.org/licenses/>. + */ + +#include "cache.h" + +/* + * dcache_line_size - get the minimum D-cache line size from the CTR register. + */ + .macro dcache_line_size, reg, tmp + mrs \tmp, ctr_el0 // read CTR + ubfm \tmp, \tmp, #16, #19 // cache line size encoding + mov \reg, #4 // bytes per word + lsl \reg, \reg, \tmp // actual cache line size + .endm + +/* + * inval_cache_range(start, end) + * - x0 - start - start address of region + * - x1 - end - end address of region + */ +.globl inval_cache_range +inval_cache_range: + dcache_line_size x2, x3 + sub x3, x2, #1 + tst x1, x3 // end cache line aligned? + bic x1, x1, x3 + b.eq 1f + dc civac, x1 // clean & invalidate D / U line +1: tst x0, x3 // start cache line aligned? + bic x0, x0, x3 + b.eq 2f + dc civac, x0 // clean & invalidate D / U line + b 3f +2: dc ivac, x0 // invalidate D / U line +3: add x0, x0, x2 + cmp x0, x1 + b.lo 2b + dsb sy + ret +/* + * flush_dcache_range(start, end) + * - x0 - start - start address of region + * - x1 - end - end address of region + * + */ +.globl flush_dcache_range +flush_dcache_range: + dcache_line_size x2, x3 + sub x3, x2, #1 + bic x0, x0, x3 +1: dc civac, x0 // clean & invalidate D line / unified line + add x0, x0, x2 + cmp x0, x1 + b.lo 1b + dsb sy + ret + +/* + * invalidate_tlbs_el1() + */ +.globl invalidate_tlbs_el1 +invalidate_tlbs_el1: + dsb nshst + tlbi vmalle1 + dsb nsh + isb + ret + +/* + * invalidate_tlbs_el2() + */ +.globl invalidate_tlbs_el2 +invalidate_tlbs_el2: + dsb nshst + tlbi alle2 + dsb nsh + isb + ret + +/* + * get_mm_feature_reg0_val - Get information about supported MM + * features + */ +.globl get_mm_feature_reg0_val +get_mm_feature_reg0_val: + mrs x0, ID_AA64MMFR0_EL1 + ret + +/* + * get_current_el - Get information about current exception level + */ +.globl get_current_el +get_current_el: + mrs x0, CurrentEL + lsr x0, x0, #2 + ret + +/* + * invalidate_icache - Invalidate I-cache + */ +.globl invalidate_icache +invalidate_icache: + ic iallu + dsb nsh + isb + ret + +/* + * set_mair_tcr_ttbr_sctlr_el1(page_table, tcr_flags) - sets MAIR, TCR , TTBR and SCTLR registers + * x0 - page_table - Page Table Base + * x1 - tcr_flags - TCR Flags to be set + */ +.globl set_mair_tcr_ttbr_sctlr_el1 +set_mair_tcr_ttbr_sctlr_el1: + ldr x2, =MEMORY_ATTRIBUTES + msr mair_el1, x2 + msr tcr_el1, x1 + msr ttbr0_el1, x0 + isb + mrs x0, sctlr_el1 + ldr x3, =SCTLR_ELx_FLAGS + orr x0, x0, x3 + msr sctlr_el1, x0 + isb + ret + +/* + * set_mair_tcr_ttbr_sctlr_el2(page_table, tcr_flags) - sets MAIR, TCR , TTBR and SCTLR registers + * x0 - page_table - Page Table Base + * x1 - tcr_flags - TCR Flags to be set + */ +.globl set_mair_tcr_ttbr_sctlr_el2 +set_mair_tcr_ttbr_sctlr_el2: + ldr x2, =MEMORY_ATTRIBUTES + msr mair_el2, x2 + msr tcr_el2, x1 + msr ttbr0_el2, x0 + isb + mrs x0, sctlr_el2 + ldr x3, =SCTLR_ELx_FLAGS + orr x0, x0, x3 + msr sctlr_el2, x0 + isb + ret + +/* + * reset_sctlr_el1 - disables cache and mmu + */ +.globl reset_sctlr_el1 +reset_sctlr_el1: + mrs x0, sctlr_el1 + bic x0, x0, #SCTLR_ELx_C + bic x0, x0, #SCTLR_ELx_M + msr sctlr_el1, x0 + isb + ret + +/* + * reset_sctlr_el2 - disables cache and mmu + */ +.globl reset_sctlr_el2 +reset_sctlr_el2: + mrs x0, sctlr_el2 + bic x0, x0, #SCTLR_ELx_C + bic x0, x0, #SCTLR_ELx_M + msr sctlr_el2, x0 + isb + ret diff --git a/purgatory/arch/arm64/cache.c b/purgatory/arch/arm64/cache.c new file mode 100644 index 000000000000..3c7e058ccf11 --- /dev/null +++ b/purgatory/arch/arm64/cache.c @@ -0,0 +1,330 @@ +/* + * Copyright (C) 2015 Pratyush Anand <panand@redhat.com> + * + * This program is free software; you can redistribute it and/or modify + * it under the terms of the GNU General Public License version 2 as + * published by the Free Software Foundation. + * + * This program is distributed in the hope that it will be useful, + * but WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + * GNU General Public License for more details. + * + * You should have received a copy of the GNU General Public License + * along with this program. If not, see <http://www.gnu.org/licenses/>. + */ + +/* We are supporting only 4K and 64K page sizes. This code will not work if + * a hardware is not supporting at least one of these page sizes. + * Therefore, D-cache is disabled by default and enabled only when + * "enable-dcache" is passed to the kexec(). + * Since this is an identity mapped system, so VA_BITS will be same as max + * PA bits supported. If VA_BITS <= 42 for 64K and <= 39 for 4K then only + * one level of page table will be there with block descriptor entries. + * Otherwise, For 4K mapping, TTBR points to level 0 lookups, which will + * have only table entries pointing to a level 1 lookup. Level 1 will have + * only block entries which will map 1GB block.For 64K mapping, TTBR points + * to level 1 lookups, which will have only table entries pointing to a + * level 2 lookup. Level 2 will have only block entries which will map + * 512MB block. If UART base address and RAM addresses are not at least 1GB + * and 512MB apart for 4K and 64K respectively, then mapping result could + * be unpredictable. In that case we need to support one more level of + * granularity, but until someone needs that keep it like this only. + * We can not allocate dynamic memory in purgatory. Therefore we keep page + * table allocation size fixed as (3 * MAX_PAGE_SIZE). (page_table) points + * to first level (having only table entries) and (page_table + + * MAX_PAGE_SIZE) points to table at next level (having block entries). If + * index for RAM area and UART area in first table is not same, then we + * will need another next level table which will be located at (page_table + * + 2 * MAX_PAGE_SIZE). + */ + +#include <stdint.h> +#include <string.h> +#include <purgatory.h> +#include "cache.h" + +static uint64_t page_shift; +static uint64_t pgtable_level; +static uint64_t va_bits; + +static uint64_t page_table[PAGE_TABLE_SIZE / sizeof(uint64_t)] __attribute__ ((aligned (MAX_PAGE_SIZE))) = { }; +static uint64_t page_table_used; + +#define PAGE_SIZE (1 << page_shift) +/* + * is_4k_page_supported - return true if 4k page is supported else + * false + */ +static int is_4k_page_supported(void) +{ + return ((get_mm_feature_reg0_val() & ID_AA64MMFR0_TGRAN4_MASK) == + ID_AA64MMFR0_TGRAN4_SUPPORTED); +} + +/* + * is_64k_page_supported - return true if 64k page is supported else + * false + */ +static int is_64k_page_supported(void) +{ + return ((get_mm_feature_reg0_val() & ID_AA64MMFR0_TGRAN64_MASK) == + ID_AA64MMFR0_TGRAN64_SUPPORTED); +} + +/* + * get_ips_bits - return supported IPS bits + */ +static uint64_t get_ips_bits(void) +{ + return ((get_mm_feature_reg0_val() & ID_AA64MMFR0_PARANGE_MASK) >> + ID_AA64MMFR0_PARANGE_SHIFT); +} + +/* + * get_va_bits - return supported VA bits (For identity mapping VA = PA) + */ +static uint64_t get_va_bits(void) +{ + uint64_t ips = get_ips_bits(); + + switch(ips) { + case ID_AA64MMFR0_PARANGE_48: + return 48; + case ID_AA64MMFR0_PARANGE_44: + return 44; + case ID_AA64MMFR0_PARANGE_42: + return 42; + case ID_AA64MMFR0_PARANGE_40: + return 40; + case ID_AA64MMFR0_PARANGE_36: + return 36; + default: + return 32; + } +} + +/* + * get_section_shift - get block shift for supported page size + */ +static uint64_t get_section_shift(void) +{ + if (page_shift == 16) + return 29; + else if(page_shift == 12) + return 30; + else + return 0; +} + +/* + * get_section_mask - get section mask for supported page size + */ +static uint64_t get_section_mask(void) +{ + if (page_shift == 16) + return 0x1FFF; + else if(page_shift == 12) + return 0x1FF; + else + return 0; +} + +/* + * get_pgdir_shift - get pgdir shift for supported page size + */ +static uint64_t get_pgdir_shift(void) +{ + if (page_shift == 16) + return 42; + else if(page_shift == 12) + return 39; + else + return 0; +} + +/* + * init_page_table - Initializes page table locations + */ + +static void init_page_table(void) +{ + /* + * Invalidate the page tables to avoid potential dirty cache lines + * being evicted. + */ + + inval_cache_range((uint64_t)page_table, + (uint64_t)page_table + PAGE_TABLE_SIZE); + memset(page_table, 0, PAGE_TABLE_SIZE); +} +/* + * create_identity_mapping(start, end, flags) + * start - start address + * end - end address + * flags - MMU Flags for Normal or Device type memory + */ +static void create_identity_mapping(uint64_t start, uint64_t end, + uint64_t flags) +{ + uint32_t sec_shift, pgdir_shift, sec_mask; + uint64_t desc, s1, e1, s2, e2; + uint64_t *table2; + + s1 = start; + e1 = end - 1; + + sec_shift = get_section_shift(); + if (pgtable_level == 1) { + s1 >>= sec_shift; + e1 >>= sec_shift; + do { + desc = s1 << sec_shift; + desc |= flags; + page_table[s1] = desc; + s1++; + } while (s1 <= e1); + } else { + pgdir_shift = get_pgdir_shift(); + sec_mask = get_section_mask(); + s1 >>= pgdir_shift; + e1 >>= pgdir_shift; + do { + /* + * If there is no table entry then write a new + * entry else, use old entry + */ + if (!page_table[s1]) { + table2 = &page_table[(++page_table_used * + MAX_PAGE_SIZE) / + sizeof(uint64_t)]; + desc = (uint64_t)table2 | PMD_TYPE_TABLE; + page_table[s1] = desc; + } else { + table2 = (uint64_t *)(page_table[s1] & + ~PMD_TYPE_MASK); + } + s1++; + s2 = start >> sec_shift; + s2 &= sec_mask; + e2 = (end - 1) >> sec_shift; + e2 &= sec_mask; + do { + desc = s2 << sec_shift; + desc |= flags; + table2[s2] = desc; + s2++; + } while (s2 <= e2); + } while (s1 <= e1); + } +} + +/* + * enable_mmu_dcache: Enable mmu and D-cache in sctlr_el1 + */ +static void enable_mmu_dcache(void) +{ + uint64_t tcr_flags = TCR_FLAGS | TCR_T0SZ(va_bits); + + switch(page_shift) { + case 16: + tcr_flags |= TCR_TG0_64K; + break; + case 12: + tcr_flags |= TCR_TG0_4K; + break; + default: + printf("page shift not supported\n"); + return; + } + /* + * Since the page tables have been populated with non-cacheable + * accesses (MMU disabled), invalidate the page tables to remove + * any speculatively loaded cache lines. + */ + inval_cache_range((uint64_t)page_table, + (uint64_t)page_table + PAGE_TABLE_SIZE); + + switch(get_current_el()) { + case 2: + invalidate_tlbs_el2(); + tcr_flags |= (get_ips_bits() << TCR_PS_EL2_SHIFT); + set_mair_tcr_ttbr_sctlr_el2((uint64_t)page_table, tcr_flags); + break; + case 1: + invalidate_tlbs_el1(); + tcr_flags |= (get_ips_bits() << TCR_IPS_EL1_SHIFT); + set_mair_tcr_ttbr_sctlr_el1((uint64_t)page_table, tcr_flags); + break; + default: + return; + } + invalidate_icache(); +} + +/* + * enable_dcache: Enable D-cache and set appropriate attributes + * ram_start - Start address of RAM + * ram_end - End address of RAM + * uart_base - Base address of uart + */ +int enable_dcache(uint64_t ram_start, uint64_t ram_end, uint64_t uart_base) +{ + va_bits = get_va_bits(); + + page_table_used = 0; + if (is_64k_page_supported()) { + page_shift = 16; + if (va_bits <= 42) + pgtable_level = 1; + else + pgtable_level = 2; + } else if (is_4k_page_supported()) { + page_shift = 12; + if (va_bits <= 39) + pgtable_level = 1; + else + pgtable_level = 2; + } else { + printf("Valid Page Granule not supported by hardware\n"); + return -1; + } + init_page_table(); + create_identity_mapping(ram_start, ram_end, MM_MMUFLAGS_NORMAL); + printf("Normal identity mapping created from %lx to %lx\n", + ram_start, ram_end); + if (uart_base) { + create_identity_mapping((uint64_t)uart_base, + (uint64_t)uart_base + PAGE_SIZE, + MM_MMUFLAGS_DEVICE); + printf("Device identity mapping created from %lx to %lx\n", + (uint64_t)uart_base, + (uint64_t)uart_base + PAGE_SIZE); + } + enable_mmu_dcache(); + printf("Cache Enabled\n"); + + return 0; +} + +/* + * disable_dcache: Disable D-cache and flush RAM locations + * ram_start - Start address of RAM + * ram_end - End address of RAM + */ +void disable_dcache(uint64_t ram_start, uint64_t ram_end) +{ + switch(get_current_el()) { + case 2: + reset_sctlr_el2(); + break; + case 1: + reset_sctlr_el1(); + break; + default: + return; + } + invalidate_icache(); + flush_dcache_range(ram_start, ram_end); + printf("Cache Disabled\n"); +} diff --git a/purgatory/arch/arm64/cache.h b/purgatory/arch/arm64/cache.h new file mode 100644 index 000000000000..c988020566e3 --- /dev/null +++ b/purgatory/arch/arm64/cache.h @@ -0,0 +1,79 @@ +#ifndef __CACHE_H__ +#define __CACHE_H__ + +#define MT_DEVICE_NGNRNE 0 +#define MT_DEVICE_NGNRE 1 +#define MT_DEVICE_GRE 2 +#define MT_NORMAL_NC 3 +#define MT_NORMAL 4 + +#ifndef __ASSEMBLER__ + +#define MAX_PAGE_SIZE 0x10000 +#define PAGE_TABLE_SIZE (3 * MAX_PAGE_SIZE) +#define ID_AA64MMFR0_TGRAN64_SHIFT 24 +#define ID_AA64MMFR0_TGRAN4_SHIFT 28 +#define ID_AA64MMFR0_TGRAN64_MASK (0xFUL << ID_AA64MMFR0_TGRAN64_SHIFT) +#define ID_AA64MMFR0_TGRAN4_MASK (0xFUL << ID_AA64MMFR0_TGRAN4_SHIFT) +#define ID_AA64MMFR0_TGRAN64_SUPPORTED 0x0 +#define ID_AA64MMFR0_TGRAN4_SUPPORTED 0x0 +#define ID_AA64MMFR0_PARANGE_SHIFT 0 +#define ID_AA64MMFR0_PARANGE_MASK (0xFUL << ID_AA64MMFR0_PARANGE_SHIFT) +#define ID_AA64MMFR0_PARANGE_48 0x5 +#define ID_AA64MMFR0_PARANGE_44 0x4 +#define ID_AA64MMFR0_PARANGE_42 0x3 +#define ID_AA64MMFR0_PARANGE_40 0x2 +#define ID_AA64MMFR0_PARANGE_36 0x1 +#define ID_AA64MMFR0_PARANGE_32 0x0 + +#define TCR_TG0_64K (1UL << 14) +#define TCR_TG0_4K (0UL << 14) +#define TCR_SHARED_NONE (0UL << 12) +#define TCR_ORGN_WBWA (1UL << 10) +#define TCR_IRGN_WBWA (1UL << 8) +#define TCR_IPS_EL1_SHIFT 32 +#define TCR_PS_EL2_SHIFT 16 +#define TCR_T0SZ(x) ((unsigned long)(64 - (x)) << 0) +#define TCR_FLAGS (TCR_SHARED_NONE | TCR_ORGN_WBWA | TCR_IRGN_WBWA) + +#define PMD_TYPE_SECT (1UL << 0) +#define PMD_TYPE_TABLE (3UL << 0) +#define PMD_TYPE_MASK 0x3 +#define PMD_SECT_AF (1UL << 10) +#define PMD_ATTRINDX(t) ((unsigned long)(t) << 2) +#define PMD_FLAGS_NORMAL (PMD_TYPE_SECT | PMD_SECT_AF) +#define PMD_SECT_PXN (1UL << 53) +#define PMD_SECT_UXN (1UL << 54) +#define PMD_FLAGS_DEVICE (PMD_TYPE_SECT | PMD_SECT_AF | PMD_SECT_PXN | PMD_SECT_UXN) +#define MM_MMUFLAGS_NORMAL PMD_ATTRINDX(MT_NORMAL) | PMD_FLAGS_NORMAL +#define MM_MMUFLAGS_DEVICE PMD_ATTRINDX(MT_DEVICE_NGNRE) | PMD_FLAGS_DEVICE + +void disable_dcache(uint64_t ram_start, uint64_t ram_end); +int enable_dcache(uint64_t ram_start, uint64_t ram_end, uint64_t uart_base); +uint64_t get_mm_feature_reg0_val(void); +void inval_cache_range(uint64_t start, uint64_t end); +void flush_dcache_range(uint64_t start, uint64_t end); +uint64_t get_current_el(void); +void set_mair_tcr_ttbr_sctlr_el1(uint64_t page_table, uint64_t tcr_flags); +void set_mair_tcr_ttbr_sctlr_el2(uint64_t page_table, uint64_t tcr_flags); +void invalidate_tlbs_el1(void); +void invalidate_tlbs_el2(void); +void invalidate_icache(void); +void reset_sctlr_el1(void); +void reset_sctlr_el2(void); +#else +#define MEMORY_ATTRIBUTES ((0x00 << (MT_DEVICE_NGNRNE*8)) | \ + (0x04 << (MT_DEVICE_NGNRE*8)) | \ + (0x0C << (MT_DEVICE_GRE*8)) | \ + (0x44 << (MT_NORMAL_NC*8)) | \ + (0xFF << (MT_NORMAL*8))) + +/* Common SCTLR_ELx flags. */ +#define SCTLR_ELx_I (1 << 12) +#define SCTLR_ELx_C (1 << 2) +#define SCTLR_ELx_M (1 << 0) + +#define SCTLR_ELx_FLAGS (SCTLR_ELx_M | SCTLR_ELx_C | SCTLR_ELx_I) + +#endif +#endif
This patch adds support to enable/disable d-cache, which can be used for faster purgatory sha256 verification. We are supporting only 4K and 64K page sizes. This code will not work if a hardware is not supporting at least one of these page sizes. Therefore, D-cache is disabled by default and enabled only when "enable-dcache" is passed to the kexec(). Since this is an identity mapped system, so VA_BITS will be same as max PA bits supported. If VA_BITS <= 42 for 64K and <= 39 for 4K then only one level of page table will be there with block descriptor entries. Otherwise, For 4K mapping, TTBR points to level 0 lookups, which will have only table entries pointing to a level 1 lookup. Level 1 will have only block entries which will map 1GB block. For 64K mapping, TTBR points to level 1 lookups, which will have only table entries pointing to a level 2 lookup. Level 2 will have only block entries which will map 512MB block. If UART base address and RAM addresses are not at least 1GB and 512MB apart for 4K and 64K respectively, then mapping result could be unpredictable. In that case we need to support one more level of granularity, but until someone needs that keep it like this only. We can not allocate dynamic memory in purgatory. Therefore we keep page table allocation size fixed as (3 * MAX_PAGE_SIZE). (page_table) points to first level (having only table entries) and (page_table + MAX_PAGE_SIZE) points to table at next level (having block entries). If index for RAM area and UART area in first table is not same, then we will need another next level table which will be located at (page_table + 2 * MAX_PAGE_SIZE). Signed-off-by: Pratyush Anand <panand@redhat.com> --- purgatory/arch/arm64/Makefile | 2 + purgatory/arch/arm64/cache-asm.S | 186 ++++++++++++++++++++++ purgatory/arch/arm64/cache.c | 330 +++++++++++++++++++++++++++++++++++++++ purgatory/arch/arm64/cache.h | 79 ++++++++++ 4 files changed, 597 insertions(+) create mode 100644 purgatory/arch/arm64/cache-asm.S create mode 100644 purgatory/arch/arm64/cache.c create mode 100644 purgatory/arch/arm64/cache.h