[RFC PATCH 06/28] x86/percpu: Get rid of absolute per-CPU variable placement

Ard Biesheuvel posted 28 patches 2 months ago
[RFC PATCH 06/28] x86/percpu: Get rid of absolute per-CPU variable placement
Posted by Ard Biesheuvel 2 months ago
From: Ard Biesheuvel <ardb@kernel.org>

For historic reasons, per-CPU symbols on x86_64 are emitted in an
address space that is disjoint from the ordinary kernel VA space,
starting at address 0x0. This splits a per-CPU symbol reference into a
base plus offset, where the base is programmed into the GS segment
register.

This deviates from the usual approach adopted by other SMP
architectures, where the base is a reference to the variable in the
kernel image's per-CPU template area, and the offset is a per-CPU value
accounting for the displacement of that particular CPU's per-CPU region
with respect to the template area. This gives per-CPU variable
references a range that is identical to ordinary references, and
requires no special handling for the startup code, as the offset will
simply be 0x0 up until the point where per-CPU variables are initialized
properly.

The x86_64 approach was needed to accommodate per-task stack protector
cookies, which used to live at a fixed offset of GS+40, requiring GS to
be treated as a base register. This is no longer the case, though, and
so GS can be repurposed as a true per-CPU offset, adopting the same
strategy as other architectures.

This also removes the need for linker tricks to emit the per-CPU ELF
segment at a different virtual address. It also means RIP-relative
per-CPU variables no longer need to be relocated in the opposite
direction when KASLR is applied, which was necessary because the 0x0
based per-CPU region remains in place even when the kernel is moved
around.

Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
---
 arch/x86/include/asm/desc.h      |  1 -
 arch/x86/include/asm/percpu.h    | 22 --------------
 arch/x86/include/asm/processor.h |  5 ++--
 arch/x86/kernel/head64.c         |  2 +-
 arch/x86/kernel/head_64.S        | 12 ++------
 arch/x86/kernel/irq_64.c         |  1 -
 arch/x86/kernel/setup_percpu.c   |  9 +-----
 arch/x86/kernel/vmlinux.lds.S    | 30 --------------------
 arch/x86/platform/pvh/head.S     |  6 ++--
 arch/x86/tools/relocs.c          |  8 +-----
 arch/x86/xen/xen-head.S          | 10 ++-----
 init/Kconfig                     |  1 -
 12 files changed, 13 insertions(+), 94 deletions(-)

diff --git a/arch/x86/include/asm/desc.h b/arch/x86/include/asm/desc.h
index 62dc9f59ea76..ec95fe44fa3a 100644
--- a/arch/x86/include/asm/desc.h
+++ b/arch/x86/include/asm/desc.h
@@ -46,7 +46,6 @@ struct gdt_page {
 } __attribute__((aligned(PAGE_SIZE)));
 
 DECLARE_PER_CPU_PAGE_ALIGNED(struct gdt_page, gdt_page);
-DECLARE_INIT_PER_CPU(gdt_page);
 
 /* Provide the original GDT */
 static inline struct desc_struct *get_cpu_gdt_rw(unsigned int cpu)
diff --git a/arch/x86/include/asm/percpu.h b/arch/x86/include/asm/percpu.h
index c55a79d5feae..1ded1207528d 100644
--- a/arch/x86/include/asm/percpu.h
+++ b/arch/x86/include/asm/percpu.h
@@ -20,12 +20,6 @@
 
 #define PER_CPU_VAR(var)	__percpu(var)__percpu_rel
 
-#ifdef CONFIG_X86_64_SMP
-# define INIT_PER_CPU_VAR(var)  init_per_cpu__##var
-#else
-# define INIT_PER_CPU_VAR(var)  var
-#endif
-
 #else /* !__ASSEMBLY__: */
 
 #include <linux/build_bug.h>
@@ -97,22 +91,6 @@
 #define __percpu_arg(x)		__percpu_prefix "%" #x
 #define __force_percpu_arg(x)	__force_percpu_prefix "%" #x
 
-/*
- * Initialized pointers to per-CPU variables needed for the boot
- * processor need to use these macros to get the proper address
- * offset from __per_cpu_load on SMP.
- *
- * There also must be an entry in vmlinux_64.lds.S
- */
-#define DECLARE_INIT_PER_CPU(var) \
-       extern typeof(var) init_per_cpu_var(var)
-
-#ifdef CONFIG_X86_64_SMP
-# define init_per_cpu_var(var)  init_per_cpu__##var
-#else
-# define init_per_cpu_var(var)  var
-#endif
-
 /*
  * For arch-specific code, we can use direct single-insn ops (they
  * don't give an lvalue though).
diff --git a/arch/x86/include/asm/processor.h b/arch/x86/include/asm/processor.h
index 56bc36116814..d7219e149f24 100644
--- a/arch/x86/include/asm/processor.h
+++ b/arch/x86/include/asm/processor.h
@@ -409,11 +409,12 @@ struct fixed_percpu_data {
 };
 
 DECLARE_PER_CPU_FIRST(struct fixed_percpu_data, fixed_percpu_data) __visible;
-DECLARE_INIT_PER_CPU(fixed_percpu_data);
 
 static inline unsigned long cpu_kernelmode_gs_base(int cpu)
 {
-	return (unsigned long)&per_cpu(fixed_percpu_data, cpu);
+	extern unsigned long __per_cpu_offset[];
+
+	return IS_ENABLED(CONFIG_SMP) ? __per_cpu_offset[cpu] : 0;
 }
 
 extern asmlinkage void entry_SYSCALL32_ignore(void);
diff --git a/arch/x86/kernel/head64.c b/arch/x86/kernel/head64.c
index 4b9d4557fc94..d4398261ad81 100644
--- a/arch/x86/kernel/head64.c
+++ b/arch/x86/kernel/head64.c
@@ -559,7 +559,7 @@ void early_setup_idt(void)
  */
 void __head startup_64_setup_gdt_idt(void)
 {
-	struct desc_struct *gdt = (void *)(__force unsigned long)init_per_cpu_var(gdt_page.gdt);
+	struct desc_struct *gdt = (void *)(__force unsigned long)gdt_page.gdt;
 	void *handler = NULL;
 
 	struct desc_ptr startup_gdt_descr = {
diff --git a/arch/x86/kernel/head_64.S b/arch/x86/kernel/head_64.S
index 330922b328bf..ab6ccee81493 100644
--- a/arch/x86/kernel/head_64.S
+++ b/arch/x86/kernel/head_64.S
@@ -68,11 +68,10 @@ SYM_CODE_START_NOALIGN(startup_64)
 	/* Set up the stack for verify_cpu() */
 	leaq	__top_init_kernel_stack(%rip), %rsp
 
-	/* Setup GSBASE to allow stack canary access for C code */
+	/* Clear %gs so early per-CPU references target the per-CPU load area */
 	movl	$MSR_GS_BASE, %ecx
-	leaq	INIT_PER_CPU_VAR(fixed_percpu_data)(%rip), %rdx
-	movl	%edx, %eax
-	shrq	$32,  %rdx
+	xorl	%eax, %eax
+	cdq
 	wrmsr
 
 	call	startup_64_setup_gdt_idt
@@ -361,15 +360,10 @@ SYM_INNER_LABEL(common_startup_64, SYM_L_LOCAL)
 
 	/* Set up %gs.
 	 *
-	 * The base of %gs always points to fixed_percpu_data. If the
-	 * stack protector canary is enabled, it is located at %gs:40.
 	 * Note that, on SMP, the boot cpu uses init data section until
 	 * the per cpu areas are set up.
 	 */
 	movl	$MSR_GS_BASE,%ecx
-#ifndef CONFIG_SMP
-	leaq	INIT_PER_CPU_VAR(fixed_percpu_data)(%rip), %rdx
-#endif
 	movl	%edx, %eax
 	shrq	$32, %rdx
 	wrmsr
diff --git a/arch/x86/kernel/irq_64.c b/arch/x86/kernel/irq_64.c
index ade0043ce56e..56bdeecd8ee0 100644
--- a/arch/x86/kernel/irq_64.c
+++ b/arch/x86/kernel/irq_64.c
@@ -27,7 +27,6 @@
 #include <asm/apic.h>
 
 DEFINE_PER_CPU_PAGE_ALIGNED(struct irq_stack, irq_stack_backing_store) __visible;
-DECLARE_INIT_PER_CPU(irq_stack_backing_store);
 
 #ifdef CONFIG_VMAP_STACK
 /*
diff --git a/arch/x86/kernel/setup_percpu.c b/arch/x86/kernel/setup_percpu.c
index b30d6e180df7..57482420ff42 100644
--- a/arch/x86/kernel/setup_percpu.c
+++ b/arch/x86/kernel/setup_percpu.c
@@ -23,17 +23,10 @@
 #include <asm/cpumask.h>
 #include <asm/cpu.h>
 
-#ifdef CONFIG_X86_64
-#define BOOT_PERCPU_OFFSET ((unsigned long)__per_cpu_load)
-#else
-#define BOOT_PERCPU_OFFSET 0
-#endif
-
-DEFINE_PER_CPU_READ_MOSTLY(unsigned long, this_cpu_off) = BOOT_PERCPU_OFFSET;
+DEFINE_PER_CPU_READ_MOSTLY(unsigned long, this_cpu_off) = 0;
 EXPORT_PER_CPU_SYMBOL(this_cpu_off);
 
 unsigned long __per_cpu_offset[NR_CPUS] __ro_after_init = {
-	[0 ... NR_CPUS-1] = BOOT_PERCPU_OFFSET,
 };
 EXPORT_SYMBOL(__per_cpu_offset);
 
diff --git a/arch/x86/kernel/vmlinux.lds.S b/arch/x86/kernel/vmlinux.lds.S
index 7f060d873f75..00f82db7b3e1 100644
--- a/arch/x86/kernel/vmlinux.lds.S
+++ b/arch/x86/kernel/vmlinux.lds.S
@@ -103,9 +103,6 @@ PHDRS {
 	text PT_LOAD FLAGS(5);          /* R_E */
 	data PT_LOAD FLAGS(6);          /* RW_ */
 #ifdef CONFIG_X86_64
-#ifdef CONFIG_SMP
-	percpu PT_LOAD FLAGS(6);        /* RW_ */
-#endif
 	init PT_LOAD FLAGS(7);          /* RWE */
 #endif
 	note PT_NOTE FLAGS(0);          /* ___ */
@@ -225,17 +222,6 @@ SECTIONS
 		__init_begin = .; /* paired with __init_end */
 	}
 
-#if defined(CONFIG_X86_64) && defined(CONFIG_SMP)
-	/*
-	 * percpu offsets are zero-based on SMP.  PERCPU_VADDR() changes the
-	 * output PHDR, so the next output section - .init.text - should
-	 * start another segment - init.
-	 */
-	PERCPU_VADDR(INTERNODE_CACHE_BYTES, 0, :percpu)
-	ASSERT(SIZEOF(.data..percpu) < CONFIG_PHYSICAL_START,
-	       "per-CPU data too large - increase CONFIG_PHYSICAL_START")
-#endif
-
 	INIT_TEXT_SECTION(PAGE_SIZE)
 #ifdef CONFIG_X86_64
 	:init
@@ -356,9 +342,7 @@ SECTIONS
 		EXIT_DATA
 	}
 
-#if !defined(CONFIG_X86_64) || !defined(CONFIG_SMP)
 	PERCPU_SECTION(INTERNODE_CACHE_BYTES)
-#endif
 
 	RUNTIME_CONST(shift, d_hash_shift)
 	RUNTIME_CONST(ptr, dentry_hashtable)
@@ -497,20 +481,6 @@ SECTIONS
 	   "kernel image bigger than KERNEL_IMAGE_SIZE");
 
 #ifdef CONFIG_X86_64
-/*
- * Per-cpu symbols which need to be offset from __per_cpu_load
- * for the boot processor.
- */
-#define INIT_PER_CPU(x) init_per_cpu__##x = ABSOLUTE(x) + __per_cpu_load
-INIT_PER_CPU(gdt_page);
-INIT_PER_CPU(fixed_percpu_data);
-INIT_PER_CPU(irq_stack_backing_store);
-
-#ifdef CONFIG_SMP
-. = ASSERT((fixed_percpu_data == 0),
-           "fixed_percpu_data is not at start of per-cpu area");
-#endif
-
 #ifdef CONFIG_MITIGATION_UNRET_ENTRY
 . = ASSERT((retbleed_return_thunk & 0x3f) == 0, "retbleed_return_thunk not cacheline-aligned");
 #endif
diff --git a/arch/x86/platform/pvh/head.S b/arch/x86/platform/pvh/head.S
index a308b79a887c..11245ecdc08d 100644
--- a/arch/x86/platform/pvh/head.S
+++ b/arch/x86/platform/pvh/head.S
@@ -95,9 +95,9 @@ SYM_CODE_START_LOCAL(pvh_start_xen)
 	/* 64-bit entry point. */
 	.code64
 1:
-	/* Set base address in stack canary descriptor. */
+	/* Clear %gs so early per-CPU references target the per-CPU load area */
 	mov $MSR_GS_BASE,%ecx
-	mov $_pa(canary), %eax
+	xor %eax, %eax
 	xor %edx, %edx
 	wrmsr
 
@@ -161,8 +161,6 @@ SYM_DATA_START_LOCAL(gdt_start)
 SYM_DATA_END_LABEL(gdt_start, SYM_L_LOCAL, gdt_end)
 
 	.balign 16
-SYM_DATA_LOCAL(canary, .fill 48, 1, 0)
-
 SYM_DATA_START_LOCAL(early_stack)
 	.fill BOOT_STACK_SIZE, 1, 0
 SYM_DATA_END_LABEL(early_stack, SYM_L_LOCAL, early_stack_end)
diff --git a/arch/x86/tools/relocs.c b/arch/x86/tools/relocs.c
index 880f0f2e465e..10add45b99f1 100644
--- a/arch/x86/tools/relocs.c
+++ b/arch/x86/tools/relocs.c
@@ -88,7 +88,6 @@ static const char * const	sym_regex_kernel[S_NSYMTYPES] = {
 	"(jiffies|jiffies_64)|"
 #if ELF_BITS == 64
 	"__per_cpu_load|"
-	"init_per_cpu__.*|"
 	"__end_rodata_hpage_align|"
 #endif
 	"__vvar_page|"
@@ -785,10 +784,6 @@ static void percpu_init(void)
  * The GNU linker incorrectly associates:
  *	__init_begin
  *	__per_cpu_load
- *
- * The "gold" linker incorrectly associates:
- *	init_per_cpu__fixed_percpu_data
- *	init_per_cpu__gdt_page
  */
 static int is_percpu_sym(ElfW(Sym) *sym, const char *symname)
 {
@@ -796,8 +791,7 @@ static int is_percpu_sym(ElfW(Sym) *sym, const char *symname)
 
 	return (shndx == per_cpu_shndx) &&
 		strcmp(symname, "__init_begin") &&
-		strcmp(symname, "__per_cpu_load") &&
-		strncmp(symname, "init_per_cpu_", 13);
+		strcmp(symname, "__per_cpu_load");
 }
 
 
diff --git a/arch/x86/xen/xen-head.S b/arch/x86/xen/xen-head.S
index 758bcd47b72d..faadac7c29e6 100644
--- a/arch/x86/xen/xen-head.S
+++ b/arch/x86/xen/xen-head.S
@@ -51,15 +51,9 @@ SYM_CODE_START(startup_xen)
 
 	leaq	__top_init_kernel_stack(%rip), %rsp
 
-	/* Set up %gs.
-	 *
-	 * The base of %gs always points to fixed_percpu_data.  If the
-	 * stack protector canary is enabled, it is located at %gs:40.
-	 * Note that, on SMP, the boot cpu uses init data section until
-	 * the per cpu areas are set up.
-	 */
+	/* Clear %gs so early per-CPU references target the per-CPU load area */
 	movl	$MSR_GS_BASE,%ecx
-	movq	$INIT_PER_CPU_VAR(fixed_percpu_data),%rax
+	xorl	%eax, %eax
 	cdq
 	wrmsr
 
diff --git a/init/Kconfig b/init/Kconfig
index b05467014041..be8a9a786d3c 100644
--- a/init/Kconfig
+++ b/init/Kconfig
@@ -1838,7 +1838,6 @@ config KALLSYMS_ALL
 config KALLSYMS_ABSOLUTE_PERCPU
 	bool
 	depends on KALLSYMS
-	default X86_64 && SMP
 
 # end of the "standard kernel features (expert users)" menu
 
-- 
2.46.0.792.g87dc391469-goog
Re: [RFC PATCH 06/28] x86/percpu: Get rid of absolute per-CPU variable placement
Posted by Christoph Lameter (Ampere) 2 months ago
On Wed, 25 Sep 2024, Ard Biesheuvel wrote:

> The x86_64 approach was needed to accommodate per-task stack protector
> cookies, which used to live at a fixed offset of GS+40, requiring GS to
> be treated as a base register. This is no longer the case, though, and
> so GS can be repurposed as a true per-CPU offset, adopting the same
> strategy as other architectures.
>
> This also removes the need for linker tricks to emit the per-CPU ELF
> segment at a different virtual address. It also means RIP-relative
> per-CPU variables no longer need to be relocated in the opposite
> direction when KASLR is applied, which was necessary because the 0x0
> based per-CPU region remains in place even when the kernel is moved
> around.

Looks like a good cleanup. Hope it does not break anything that relies on
structures %GS points to.

Reviewed-by: Christoph Lameter <cl@linux.com>