linux_dsm_epyc7002/arch/x86/kernel/head64.c
Kirill A. Shutemov 39b9552281 x86/mm: Optimize boot-time paging mode switching cost
By this point we have functioning boot-time switching between 4- and
5-level paging mode. But naive approach comes with cost.

Numbers below are for kernel build, allmodconfig, 5 times.

CONFIG_X86_5LEVEL=n:

 Performance counter stats for 'sh -c make -j100 -B -k >/dev/null' (5 runs):

   17308719.892691      task-clock:u (msec)       #   26.772 CPUs utilized            ( +-  0.11% )
                 0      context-switches:u        #    0.000 K/sec
                 0      cpu-migrations:u          #    0.000 K/sec
       331,993,164      page-faults:u             #    0.019 M/sec                    ( +-  0.01% )
43,614,978,867,455      cycles:u                  #    2.520 GHz                      ( +-  0.01% )
39,371,534,575,126      stalled-cycles-frontend:u #   90.27% frontend cycles idle     ( +-  0.09% )
28,363,350,152,428      instructions:u            #    0.65  insn per cycle
                                                  #    1.39  stalled cycles per insn  ( +-  0.00% )
 6,316,784,066,413      branches:u                #  364.948 M/sec                    ( +-  0.00% )
   250,808,144,781      branch-misses:u           #    3.97% of all branches          ( +-  0.01% )

     646.531974142 seconds time elapsed                                          ( +-  1.15% )

CONFIG_X86_5LEVEL=y:

 Performance counter stats for 'sh -c make -j100 -B -k >/dev/null' (5 runs):

   17411536.780625      task-clock:u (msec)       #   26.426 CPUs utilized            ( +-  0.10% )
                 0      context-switches:u        #    0.000 K/sec
                 0      cpu-migrations:u          #    0.000 K/sec
       331,868,663      page-faults:u             #    0.019 M/sec                    ( +-  0.01% )
43,865,909,056,301      cycles:u                  #    2.519 GHz                      ( +-  0.01% )
39,740,130,365,581      stalled-cycles-frontend:u #   90.59% frontend cycles idle     ( +-  0.05% )
28,363,358,997,959      instructions:u            #    0.65  insn per cycle
                                                  #    1.40  stalled cycles per insn  ( +-  0.00% )
 6,316,784,937,460      branches:u                #  362.793 M/sec                    ( +-  0.00% )
   251,531,919,485      branch-misses:u           #    3.98% of all branches          ( +-  0.00% )

     658.886307752 seconds time elapsed                                          ( +-  0.92% )

The patch tries to fix the performance regression by using
cpu_feature_enabled(X86_FEATURE_LA57) instead of pgtable_l5_enabled in
all hot code paths. These will statically patch the target code for
additional performance.

CONFIG_X86_5LEVEL=y + the patch:

 Performance counter stats for 'sh -c make -j100 -B -k >/dev/null' (5 runs):

   17381990.268506      task-clock:u (msec)       #   26.907 CPUs utilized            ( +-  0.19% )
                 0      context-switches:u        #    0.000 K/sec
                 0      cpu-migrations:u          #    0.000 K/sec
       331,862,625      page-faults:u             #    0.019 M/sec                    ( +-  0.01% )
43,697,726,320,051      cycles:u                  #    2.514 GHz                      ( +-  0.03% )
39,480,408,690,401      stalled-cycles-frontend:u #   90.35% frontend cycles idle     ( +-  0.05% )
28,363,394,221,388      instructions:u            #    0.65  insn per cycle
                                                  #    1.39  stalled cycles per insn  ( +-  0.00% )
 6,316,794,985,573      branches:u                #  363.410 M/sec                    ( +-  0.00% )
   251,013,232,547      branch-misses:u           #    3.97% of all branches          ( +-  0.01% )

     645.991174661 seconds time elapsed                                          ( +-  1.19% )

Unfortunately, this approach doesn't help with text size:

  vmlinux.before .text size:	8190319
  vmlinux.after .text size:	8200623

The .text section is increased by about 4k. Not sure if we can do anything
about this.

Signed-off-by: Kirill A. Shuemov <kirill.shutemov@linux.intel.com>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Borislav Petkov <bp@alien8.de>
Cc: Borislav Petkov <bp@suse.de>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: David Woodhouse <dwmw2@infradead.org>
Cc: Josh Poimboeuf <jpoimboe@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: linux-mm@kvack.org
Link: http://lkml.kernel.org/r/20180216114948.68868-4-kirill.shutemov@linux.intel.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-02-21 10:19:18 +01:00

445 lines
12 KiB
C

// SPDX-License-Identifier: GPL-2.0
/*
* prepare to run common code
*
* Copyright (C) 2000 Andrea Arcangeli <andrea@suse.de> SuSE
*/
#define DISABLE_BRANCH_PROFILING
#include <linux/init.h>
#include <linux/linkage.h>
#include <linux/types.h>
#include <linux/kernel.h>
#include <linux/string.h>
#include <linux/percpu.h>
#include <linux/start_kernel.h>
#include <linux/io.h>
#include <linux/memblock.h>
#include <linux/mem_encrypt.h>
#include <asm/processor.h>
#include <asm/proto.h>
#include <asm/smp.h>
#include <asm/setup.h>
#include <asm/desc.h>
#include <asm/pgtable.h>
#include <asm/tlbflush.h>
#include <asm/sections.h>
#include <asm/kdebug.h>
#include <asm/e820/api.h>
#include <asm/bios_ebda.h>
#include <asm/bootparam_utils.h>
#include <asm/microcode.h>
#include <asm/kasan.h>
#ifdef CONFIG_X86_5LEVEL
#undef pgtable_l5_enabled
#define pgtable_l5_enabled __pgtable_l5_enabled
#endif
/*
* Manage page tables very early on.
*/
extern pmd_t early_dynamic_pgts[EARLY_DYNAMIC_PAGE_TABLES][PTRS_PER_PMD];
static unsigned int __initdata next_early_pgt;
pmdval_t early_pmd_flags = __PAGE_KERNEL_LARGE & ~(_PAGE_GLOBAL | _PAGE_NX);
#ifdef CONFIG_X86_5LEVEL
unsigned int __pgtable_l5_enabled __ro_after_init;
EXPORT_SYMBOL(__pgtable_l5_enabled);
unsigned int pgdir_shift __ro_after_init = 39;
EXPORT_SYMBOL(pgdir_shift);
unsigned int ptrs_per_p4d __ro_after_init = 1;
EXPORT_SYMBOL(ptrs_per_p4d);
#endif
#ifdef CONFIG_DYNAMIC_MEMORY_LAYOUT
unsigned long page_offset_base __ro_after_init = __PAGE_OFFSET_BASE_L4;
EXPORT_SYMBOL(page_offset_base);
unsigned long vmalloc_base __ro_after_init = __VMALLOC_BASE_L4;
EXPORT_SYMBOL(vmalloc_base);
unsigned long vmemmap_base __ro_after_init = __VMEMMAP_BASE_L4;
EXPORT_SYMBOL(vmemmap_base);
#endif
#define __head __section(.head.text)
static void __head *fixup_pointer(void *ptr, unsigned long physaddr)
{
return ptr - (void *)_text + (void *)physaddr;
}
static unsigned long __head *fixup_long(void *ptr, unsigned long physaddr)
{
return fixup_pointer(ptr, physaddr);
}
#ifdef CONFIG_X86_5LEVEL
static unsigned int __head *fixup_int(void *ptr, unsigned long physaddr)
{
return fixup_pointer(ptr, physaddr);
}
static bool __head check_la57_support(unsigned long physaddr)
{
if (native_cpuid_eax(0) < 7)
return false;
if (!(native_cpuid_ecx(7) & (1 << (X86_FEATURE_LA57 & 31))))
return false;
*fixup_int(&pgtable_l5_enabled, physaddr) = 1;
*fixup_int(&pgdir_shift, physaddr) = 48;
*fixup_int(&ptrs_per_p4d, physaddr) = 512;
*fixup_long(&page_offset_base, physaddr) = __PAGE_OFFSET_BASE_L5;
*fixup_long(&vmalloc_base, physaddr) = __VMALLOC_BASE_L5;
*fixup_long(&vmemmap_base, physaddr) = __VMEMMAP_BASE_L5;
return true;
}
#else
static bool __head check_la57_support(unsigned long physaddr)
{
return false;
}
#endif
unsigned long __head __startup_64(unsigned long physaddr,
struct boot_params *bp)
{
unsigned long load_delta, *p;
unsigned long pgtable_flags;
pgdval_t *pgd;
p4dval_t *p4d;
pudval_t *pud;
pmdval_t *pmd, pmd_entry;
bool la57;
int i;
unsigned int *next_pgt_ptr;
la57 = check_la57_support(physaddr);
/* Is the address too large? */
if (physaddr >> MAX_PHYSMEM_BITS)
for (;;);
/*
* Compute the delta between the address I am compiled to run at
* and the address I am actually running at.
*/
load_delta = physaddr - (unsigned long)(_text - __START_KERNEL_map);
/* Is the address not 2M aligned? */
if (load_delta & ~PMD_PAGE_MASK)
for (;;);
/* Activate Secure Memory Encryption (SME) if supported and enabled */
sme_enable(bp);
/* Include the SME encryption mask in the fixup value */
load_delta += sme_get_me_mask();
/* Fixup the physical addresses in the page table */
pgd = fixup_pointer(&early_top_pgt, physaddr);
p = pgd + pgd_index(__START_KERNEL_map);
if (la57)
*p = (unsigned long)level4_kernel_pgt;
else
*p = (unsigned long)level3_kernel_pgt;
*p += _PAGE_TABLE_NOENC - __START_KERNEL_map + load_delta;
if (la57) {
p4d = fixup_pointer(&level4_kernel_pgt, physaddr);
p4d[511] += load_delta;
}
pud = fixup_pointer(&level3_kernel_pgt, physaddr);
pud[510] += load_delta;
pud[511] += load_delta;
pmd = fixup_pointer(level2_fixmap_pgt, physaddr);
pmd[506] += load_delta;
/*
* Set up the identity mapping for the switchover. These
* entries should *NOT* have the global bit set! This also
* creates a bunch of nonsense entries but that is fine --
* it avoids problems around wraparound.
*/
next_pgt_ptr = fixup_pointer(&next_early_pgt, physaddr);
pud = fixup_pointer(early_dynamic_pgts[(*next_pgt_ptr)++], physaddr);
pmd = fixup_pointer(early_dynamic_pgts[(*next_pgt_ptr)++], physaddr);
pgtable_flags = _KERNPG_TABLE_NOENC + sme_get_me_mask();
if (la57) {
p4d = fixup_pointer(early_dynamic_pgts[next_early_pgt++], physaddr);
i = (physaddr >> PGDIR_SHIFT) % PTRS_PER_PGD;
pgd[i + 0] = (pgdval_t)p4d + pgtable_flags;
pgd[i + 1] = (pgdval_t)p4d + pgtable_flags;
i = (physaddr >> P4D_SHIFT) % PTRS_PER_P4D;
p4d[i + 0] = (pgdval_t)pud + pgtable_flags;
p4d[i + 1] = (pgdval_t)pud + pgtable_flags;
} else {
i = (physaddr >> PGDIR_SHIFT) % PTRS_PER_PGD;
pgd[i + 0] = (pgdval_t)pud + pgtable_flags;
pgd[i + 1] = (pgdval_t)pud + pgtable_flags;
}
i = (physaddr >> PUD_SHIFT) % PTRS_PER_PUD;
pud[i + 0] = (pudval_t)pmd + pgtable_flags;
pud[i + 1] = (pudval_t)pmd + pgtable_flags;
pmd_entry = __PAGE_KERNEL_LARGE_EXEC & ~_PAGE_GLOBAL;
pmd_entry += sme_get_me_mask();
pmd_entry += physaddr;
for (i = 0; i < DIV_ROUND_UP(_end - _text, PMD_SIZE); i++) {
int idx = i + (physaddr >> PMD_SHIFT) % PTRS_PER_PMD;
pmd[idx] = pmd_entry + i * PMD_SIZE;
}
/*
* Fixup the kernel text+data virtual addresses. Note that
* we might write invalid pmds, when the kernel is relocated
* cleanup_highmap() fixes this up along with the mappings
* beyond _end.
*/
pmd = fixup_pointer(level2_kernel_pgt, physaddr);
for (i = 0; i < PTRS_PER_PMD; i++) {
if (pmd[i] & _PAGE_PRESENT)
pmd[i] += load_delta;
}
/*
* Fixup phys_base - remove the memory encryption mask to obtain
* the true physical address.
*/
*fixup_long(&phys_base, physaddr) += load_delta - sme_get_me_mask();
/* Encrypt the kernel and related (if SME is active) */
sme_encrypt_kernel(bp);
/*
* Return the SME encryption mask (if SME is active) to be used as a
* modifier for the initial pgdir entry programmed into CR3.
*/
return sme_get_me_mask();
}
unsigned long __startup_secondary_64(void)
{
/*
* Return the SME encryption mask (if SME is active) to be used as a
* modifier for the initial pgdir entry programmed into CR3.
*/
return sme_get_me_mask();
}
/* Wipe all early page tables except for the kernel symbol map */
static void __init reset_early_page_tables(void)
{
memset(early_top_pgt, 0, sizeof(pgd_t)*(PTRS_PER_PGD-1));
next_early_pgt = 0;
write_cr3(__sme_pa_nodebug(early_top_pgt));
}
/* Create a new PMD entry */
int __init __early_make_pgtable(unsigned long address, pmdval_t pmd)
{
unsigned long physaddr = address - __PAGE_OFFSET;
pgdval_t pgd, *pgd_p;
p4dval_t p4d, *p4d_p;
pudval_t pud, *pud_p;
pmdval_t *pmd_p;
/* Invalid address or early pgt is done ? */
if (physaddr >= MAXMEM || read_cr3_pa() != __pa_nodebug(early_top_pgt))
return -1;
again:
pgd_p = &early_top_pgt[pgd_index(address)].pgd;
pgd = *pgd_p;
/*
* The use of __START_KERNEL_map rather than __PAGE_OFFSET here is
* critical -- __PAGE_OFFSET would point us back into the dynamic
* range and we might end up looping forever...
*/
if (!pgtable_l5_enabled)
p4d_p = pgd_p;
else if (pgd)
p4d_p = (p4dval_t *)((pgd & PTE_PFN_MASK) + __START_KERNEL_map - phys_base);
else {
if (next_early_pgt >= EARLY_DYNAMIC_PAGE_TABLES) {
reset_early_page_tables();
goto again;
}
p4d_p = (p4dval_t *)early_dynamic_pgts[next_early_pgt++];
memset(p4d_p, 0, sizeof(*p4d_p) * PTRS_PER_P4D);
*pgd_p = (pgdval_t)p4d_p - __START_KERNEL_map + phys_base + _KERNPG_TABLE;
}
p4d_p += p4d_index(address);
p4d = *p4d_p;
if (p4d)
pud_p = (pudval_t *)((p4d & PTE_PFN_MASK) + __START_KERNEL_map - phys_base);
else {
if (next_early_pgt >= EARLY_DYNAMIC_PAGE_TABLES) {
reset_early_page_tables();
goto again;
}
pud_p = (pudval_t *)early_dynamic_pgts[next_early_pgt++];
memset(pud_p, 0, sizeof(*pud_p) * PTRS_PER_PUD);
*p4d_p = (p4dval_t)pud_p - __START_KERNEL_map + phys_base + _KERNPG_TABLE;
}
pud_p += pud_index(address);
pud = *pud_p;
if (pud)
pmd_p = (pmdval_t *)((pud & PTE_PFN_MASK) + __START_KERNEL_map - phys_base);
else {
if (next_early_pgt >= EARLY_DYNAMIC_PAGE_TABLES) {
reset_early_page_tables();
goto again;
}
pmd_p = (pmdval_t *)early_dynamic_pgts[next_early_pgt++];
memset(pmd_p, 0, sizeof(*pmd_p) * PTRS_PER_PMD);
*pud_p = (pudval_t)pmd_p - __START_KERNEL_map + phys_base + _KERNPG_TABLE;
}
pmd_p[pmd_index(address)] = pmd;
return 0;
}
int __init early_make_pgtable(unsigned long address)
{
unsigned long physaddr = address - __PAGE_OFFSET;
pmdval_t pmd;
pmd = (physaddr & PMD_MASK) + early_pmd_flags;
return __early_make_pgtable(address, pmd);
}
/* Don't add a printk in there. printk relies on the PDA which is not initialized
yet. */
static void __init clear_bss(void)
{
memset(__bss_start, 0,
(unsigned long) __bss_stop - (unsigned long) __bss_start);
}
static unsigned long get_cmd_line_ptr(void)
{
unsigned long cmd_line_ptr = boot_params.hdr.cmd_line_ptr;
cmd_line_ptr |= (u64)boot_params.ext_cmd_line_ptr << 32;
return cmd_line_ptr;
}
static void __init copy_bootdata(char *real_mode_data)
{
char * command_line;
unsigned long cmd_line_ptr;
/*
* If SME is active, this will create decrypted mappings of the
* boot data in advance of the copy operations.
*/
sme_map_bootdata(real_mode_data);
memcpy(&boot_params, real_mode_data, sizeof boot_params);
sanitize_boot_params(&boot_params);
cmd_line_ptr = get_cmd_line_ptr();
if (cmd_line_ptr) {
command_line = __va(cmd_line_ptr);
memcpy(boot_command_line, command_line, COMMAND_LINE_SIZE);
}
/*
* The old boot data is no longer needed and won't be reserved,
* freeing up that memory for use by the system. If SME is active,
* we need to remove the mappings that were created so that the
* memory doesn't remain mapped as decrypted.
*/
sme_unmap_bootdata(real_mode_data);
}
asmlinkage __visible void __init x86_64_start_kernel(char * real_mode_data)
{
/*
* Build-time sanity checks on the kernel image and module
* area mappings. (these are purely build-time and produce no code)
*/
BUILD_BUG_ON(MODULES_VADDR < __START_KERNEL_map);
BUILD_BUG_ON(MODULES_VADDR - __START_KERNEL_map < KERNEL_IMAGE_SIZE);
BUILD_BUG_ON(MODULES_LEN + KERNEL_IMAGE_SIZE > 2*PUD_SIZE);
BUILD_BUG_ON((__START_KERNEL_map & ~PMD_MASK) != 0);
BUILD_BUG_ON((MODULES_VADDR & ~PMD_MASK) != 0);
BUILD_BUG_ON(!(MODULES_VADDR > __START_KERNEL));
MAYBE_BUILD_BUG_ON(!(((MODULES_END - 1) & PGDIR_MASK) ==
(__START_KERNEL & PGDIR_MASK)));
BUILD_BUG_ON(__fix_to_virt(__end_of_fixed_addresses) <= MODULES_END);
cr4_init_shadow();
/* Kill off the identity-map trampoline */
reset_early_page_tables();
clear_bss();
clear_page(init_top_pgt);
/*
* SME support may update early_pmd_flags to include the memory
* encryption mask, so it needs to be called before anything
* that may generate a page fault.
*/
sme_early_init();
kasan_early_init();
idt_setup_early_handler();
copy_bootdata(__va(real_mode_data));
/*
* Load microcode early on BSP.
*/
load_ucode_bsp();
/* set init_top_pgt kernel high mapping*/
init_top_pgt[511] = early_top_pgt[511];
x86_64_start_reservations(real_mode_data);
}
void __init x86_64_start_reservations(char *real_mode_data)
{
/* version is always not zero if it is copied */
if (!boot_params.hdr.version)
copy_bootdata(__va(real_mode_data));
x86_early_init_platform_quirks();
switch (boot_params.hdr.hardware_subarch) {
case X86_SUBARCH_INTEL_MID:
x86_intel_mid_early_setup();
break;
default:
break;
}
start_kernel();
}