tracing/x86_32: Remove non DYNAMIC_FTRACE and mcount support
Steven Rostedt
maintainerv2 cover tracing/x86_32: Remove non DYNAMIC_FTRACE and mcount support
There's no need to support either DYNAMIC_FTRACE=n or mcount (non fentry) in x86_32. As the static tracing (DYNAMIC_FTRACE=n) does not support fentry, we can just remove it, as the static tracing is only around to test the static tracing in generic code as other architectures have it but not DYNAMIC_FTRACE.
This will allow also allow us to remove klp_check_compiler_support() in later patches.
Steven Rostedt (VMware) (2):
ftrace/x86_32: Remove support for non DYNAMIC_FTRACE ftrace/x86: Remove mcount support
----
arch/x86/Kconfig | 11 ++++++ arch/x86/include/asm/ftrace.h | 8 ++--- arch/x86/include/asm/livepatch.h | 3 -- arch/x86/kernel/ftrace_32.S | 75 +++------------------------------------- arch/x86/kernel/ftrace_64.S | 28 +-------------- 5 files changed, 20 insertions(+), 105 deletions(-)
Steven Rostedt
maintainerRFC PATCH v2 1/2 ftrace/x86_32: Remove support for non DYNAMIC_FTRACE
From: "Steven Rostedt (VMware)" <[email protected]>
When DYNAMIC_FTRACE is enabled in the kernel, all the functions that can be traced by the function tracer have a "nop" placeholder at the start of the function. When function tracing is enabled, the nop is converted into a call to the tracing infrastructure where the functions get traced. This also allows for specifying specific functions to trace, and a lot of infrastructure is built on top of this.
When DYNAMIC_FTRACE is not enabled, all the functions have a call to the ftrace trampoline. A check is made to see if a function pointer is the ftrace_stub or not, and if it is not, it calls the function pointer to trace the code. This adds over 10% overhead to the kernel even when tracing is disabled.
When an architecture supports DYNAMIC_FTRACE there really is no reason to use the static tracing. I have kept non DYNAMIC_FTRACE available in x86 so that the generic code for non DYNAMIC_FTRACE can be tested. There is no reason to support non DYNAMIC_FTRACE for both x86_64 and x86_32. As the non DYNAMIC_FTRACE for x86_32 does not even support fentry, and we want to remove mcount completely, there's no reason to keep non DYNAMIC_FTRACE around for x86_32.
--- arch/x86/Kconfig | 11 +++++++++++ arch/x86/kernel/ftrace_32.S | 39 ------------------------------------- 2 files changed, 11 insertions(+), 39 deletions(-)
arch/x86/Kconfig +11 -0
@@ -31,6 +31,17 @@ config X86_64 select X86_DEV_DMA_OPS select ARCH_HAS_SYSCALL_WRAPPER +config FORCE_DYNAMIC_FTRACE + def_bool y + depends on X86_32 + depends on FUNCTION_TRACER + select DYNAMIC_FTRACE + help + We keep the static function tracing (!DYNAMIC_FTRACE) around + in order to test the non static function tracing in the + generic code, as other architectures still use it. But we + only need to keep it around for x86_64. No need to keep it + for x86_32. For x86_32, force DYNAMIC_FTRACE. # # Arch settings #
arch/x86/kernel/ftrace_32.S +0 -39
@@ -18,8 +18,6 @@ EXPORT_SYMBOL(__fentry__) EXPORT_SYMBOL(mcount) #endif -#ifdef CONFIG_DYNAMIC_FTRACE - /* mcount uses a frame pointer even if CONFIG_FRAME_POINTER is not set */ #if !defined(CC_USING_FENTRY) || defined(CONFIG_FRAME_POINTER) # define USING_FRAME_POINTER
@@ -170,43 +168,6 @@ GLOBAL(ftrace_regs_call) lea 3*4(%esp), %esp /* Skip orig_ax, ip and cs */ jmp .Lftrace_ret -#else /* ! CONFIG_DYNAMIC_FTRACE */ - -ENTRY(function_hook) - cmpl $__PAGE_OFFSET, %esp - jb ftrace_stub /* Paging not enabled yet? */ - - cmpl $ftrace_stub, ftrace_trace_function - jnz .Ltrace -#ifdef CONFIG_FUNCTION_GRAPH_TRACER - cmpl $ftrace_stub, ftrace_graph_return - jnz ftrace_graph_caller - - cmpl $ftrace_graph_entry_stub, ftrace_graph_entry - jnz ftrace_graph_caller -#endif -.globl ftrace_stub -ftrace_stub: - ret - - /* taken from glibc */ -.Ltrace: - pushl %eax - pushl %ecx - pushl %edx - movl 0xc(%esp), %eax - movl 0x4(%ebp), %edx - subl $MCOUNT_INSN_SIZE, %eax - - movl ftrace_trace_function, %ecx - CALL_NOSPEC %ecx - - popl %edx - popl %ecx - popl %eax - jmp ftrace_stub -END(function_hook) -#endif /* CONFIG_DYNAMIC_FTRACE */ #ifdef CONFIG_FUNCTION_GRAPH_TRACER ENTRY(ftrace_graph_caller)
Signature · 1 line
2.20.1
Steven Rostedt
maintainerRFC PATCH v2 2/2 ftrace/x86: Remove mcount support
In reply to Steven RostedtFrom: "Steven Rostedt (VMware)" <[email protected]>
There's two methods of enabling function tracing in Linux on x86. One is with just "gcc -pg" and the other is "gcc -pg -mfentry". The former will use calls to a special function "mcount" after the frame is set up in all C functions. The latter will add calls to a special function called "fentry" as the very first instruction of all C functions.
At compile time, there is a check to see if gcc supports, -mfentry, and if it does, it will use that, because it is more versatile and less error prone for function tracing.
Starting with v4.19, the minimum gcc supported to build the Linux kernel, was raised to version 4.6. That also happens to be the first gcc version to support -mfentry. Since on x86, using gcc versions from 4.6 and beyond will unconditionally enable the -mfentry, it will no longer use mcount as the method for inserting calls into the C functions of the kernel. This means that there is no point in continuing to maintain mcount in x86.
Remove support for using mcount. This makes the code less complex, and will also allow it to be simplified in the future.
--- arch/x86/include/asm/ftrace.h | 8 +++---- arch/x86/include/asm/livepatch.h | 3 --- arch/x86/kernel/ftrace_32.S | 36 +++++--------------------------- arch/x86/kernel/ftrace_64.S | 28 +------------------------ 4 files changed, 9 insertions(+), 66 deletions(-)
arch/x86/include/asm/ftrace.h +3 -5
@@ -3,12 +3,10 @@#define _ASM_X86_FTRACE_H#ifdef CONFIG_FUNCTION_TRACER -#ifdef CC_USING_FENTRY -# define MCOUNT_ADDR ((unsigned long)(__fentry__)) -#else -# define MCOUNT_ADDR ((unsigned long)(mcount)) -# define HAVE_FUNCTION_GRAPH_FP_TEST +#ifndef CC_USING_FENTRY +# error Compiler does not support fentry?#endif +# define MCOUNT_ADDR ((unsigned long)(__fentry__))#define MCOUNT_INSN_SIZE 5/* sizeof mcount call */ #ifdef CONFIG_DYNAMIC_FTRACE
arch/x86/include/asm/livepatch.h +0 -3
@@ -26,9 +26,6 @@static inline int klp_check_compiler_support (void ) { -#ifndef CC_USING_FENTRY -return 1 ; -#endif return 0 ; }
arch/x86/kernel/ftrace_32.S +5 -31
@@ -10,20 +10,10 @@ #include <asm/ftrace.h> #include <asm/nospec-branch.h> -#ifdef CC_USING_FENTRY # define function_hook __fentry__ EXPORT_SYMBOL(__fentry__) -#else -# define function_hook mcount -EXPORT_SYMBOL(mcount) -#endif - -/* mcount uses a frame pointer even if CONFIG_FRAME_POINTER is not set */ -#if !defined(CC_USING_FENTRY) || defined(CONFIG_FRAME_POINTER) -# define USING_FRAME_POINTER -#endif -#ifdef USING_FRAME_POINTER +#ifdef CONFIG_FRAME_POINTER # define MCOUNT_FRAME 1 /* using frame = true */ #else # define MCOUNT_FRAME 0 /* using frame = false */
@@ -35,8 +25,7 @@ END(function_hook) ENTRY(ftrace_caller) -#ifdef USING_FRAME_POINTER -# ifdef CC_USING_FENTRY +#ifdef CONFIG_FRAME_POINTER /* * Frame pointers are of ip followed by bp. * Since fentry is an immediate jump, we are left with
@@ -47,7 +36,7 @@ ENTRY(ftrace_caller) pushl %ebp movl %esp, %ebp pushl 2*4(%esp) /* function ip */ -# endif + /* For mcount, the function ip is directly above */ pushl %ebp movl %esp, %ebp
@@ -57,7 +46,7 @@ ENTRY(ftrace_caller) pushl %edx pushl $0 /* Pass NULL as regs pointer */ -#ifdef USING_FRAME_POINTER +#ifdef CONFIG_FRAME_POINTER /* Load parent ebp into edx */ movl 4*4(%esp), %edx #else
@@ -80,13 +69,11 @@ ftrace_call: popl %edx popl %ecx popl %eax -#ifdef USING_FRAME_POINTER +#ifdef CONFIG_FRAME_POINTER popl %ebp -# ifdef CC_USING_FENTRY addl $4,%esp /* skip function ip */ popl %ebp /* this is the orig bp */ addl $4, %esp /* skip parent ip */ -# endif #endif .Lftrace_ret: #ifdef CONFIG_FUNCTION_GRAPH_TRACER
@@ -131,11 +118,7 @@ ENTRY(ftrace_regs_caller) movl 12*4(%esp), %eax /* Load ip (1st parameter) */ subl $MCOUNT_INSN_SIZE, %eax /* Adjust ip */ -#ifdef CC_USING_FENTRY movl 15*4(%esp), %edx /* Load parent ip (2nd parameter) */ -#else - movl 0x4(%ebp), %edx /* Load parent ip (2nd parameter) */ -#endif movl function_trace_op, %ecx /* Save ftrace_pos in 3rd parameter */ pushl %esp /* Save pt_regs as 4th parameter */
@@ -176,13 +159,8 @@ ENTRY(ftrace_graph_caller) pushl %edx movl 3*4(%esp), %eax /* Even with frame pointers, fentry doesn't have one here */ -#ifdef CC_USING_FENTRY lea 4*4(%esp), %edx movl $0, %ecx -#else - lea 0x4(%ebp), %edx - movl (%ebp), %ecx -#endif subl $MCOUNT_INSN_SIZE, %eax call prepare_ftrace_return popl %edx
@@ -195,11 +173,7 @@ END(ftrace_graph_caller) return_to_handler: pushl %eax pushl %edx -#ifdef CC_USING_FENTRY movl $0, %eax -#else - movl %ebp, %eax -#endif call ftrace_return_to_handler movl %eax, %ecx popl %edx
arch/x86/kernel/ftrace_64.S +1 -27
@@ -13,22 +13,12 @@ .code64 .section .entry.text, "ax" -#ifdef CC_USING_FENTRY # define function_hook __fentry__ EXPORT_SYMBOL(__fentry__) -#else -# define function_hook mcount -EXPORT_SYMBOL(mcount) -#endif #ifdef CONFIG_FRAME_POINTER -# ifdef CC_USING_FENTRY /* Save parent and function stack frames (rip and rbp) */ # define MCOUNT_FRAME_SIZE (8+16*2) -# else -/* Save just function stack frame (rip and rbp) */ -# define MCOUNT_FRAME_SIZE (8+16) -# endif #else /* No need to save a stack frame */ # define MCOUNT_FRAME_SIZE 0
@@ -75,17 +65,13 @@ EXPORT_SYMBOL(mcount) * fentry is called before the stack frame is set up, where as mcount * is called afterward. */ -#ifdef CC_USING_FENTRY + /* Save the parent pointer (skip orig rbp and our return address) */ pushq \added+8*2(%rsp) pushq %rbp movq %rsp, %rbp /* Save the return address (now skip orig rbp, rbp and parent) */ pushq \added+8*3(%rsp) -#else - /* Can't assume that rip is before this (unless added was zero) */ - pushq \added+8(%rsp) -#endif pushq %rbp movq %rsp, %rbp #endif /* CONFIG_FRAME_POINTER */
@@ -113,12 +99,7 @@ EXPORT_SYMBOL(mcount) movq %rdx, RBP(%rsp) /* Copy the parent address into %rsi (second parameter) */ -#ifdef CC_USING_FENTRY movq MCOUNT_REG_SIZE+8+\added(%rsp), %rsi -#else - /* %rdx contains original %rbp */ - movq 8(%rdx), %rsi -#endif /* Move RIP to its proper location */ movq MCOUNT_REG_SIZE+\added(%rsp), %rdi
@@ -303,15 +284,8 @@ ENTRY(ftrace_graph_caller) /* Saves rbp into %rdx and fills first parameter */ save_mcount_regs -#ifdef CC_USING_FENTRY leaq MCOUNT_REG_SIZE+8(%rsp), %rsi movq $0, %rdx /* No framepointers needed */ -#else - /* Save address of the return address of traced function */ - leaq 8(%rdx), %rsi - /* ftrace does sanity checks against frame pointers */ - movq (%rdx), %rdx -#endif call prepare_ftrace_return restore_mcount_regs
Signature · 1 line
2.20.1
Steven Rostedt
maintainerQuoted 1 line · On Fri, 10 May 2019 12:35:19 -0400 Steven Rostedt <[email protected]> wrote:
This will allow also allow us to remove
klp_check_compiler_support()
One writes sentences like the above when someone walks in and asks a question in the middle of writing a sentence.
-- Steve
[RFC][PATCH 3/2] livepatch: remove klp_check_compiler_support()
In reply to Steven RostedtFrom: Jiri Kosina <[email protected]>
The only purpose of klp_check_compiler_support() is to make sure that we are not using ftrace on x86 via mcount (because that's executed only after prologue has already happened, and that's too late for livepatching purposes).
Now that mcount is not supported by ftrace any more, there is no need for klp_check_compiler_support() either.
---
I guess it makes most sense to merge this together with mcount removal in one go.
arch/powerpc/include/asm/livepatch.h | 5 ----- arch/s390/include/asm/livepatch.h | 5 ----- arch/x86/include/asm/livepatch.h | 5 ----- kernel/livepatch/core.c | 8 -------- 4 files changed, 23 deletions(-)
arch/powerpc/include/asm/livepatch.h +0 -5
@@ -24,11 +24,6 @@#include <linux/sched/task_stack.h> #ifdef CONFIG_LIVEPATCH -static inline int klp_check_compiler_support (void ) -{ -return 0 ; -} -static inline void klp_arch_set_pc (struct pt_regs *regs,unsigned long ip) { regs->nip = ip;
arch/s390/include/asm/livepatch.h +0 -5
@@ -13,11 +13,6 @@#include <asm/ptrace.h> -static inline int klp_check_compiler_support (void ) -{ -return 0 ; -} -static inline void klp_arch_set_pc (struct pt_regs *regs,unsigned long ip) { regs->psw.addr = ip;
arch/x86/include/asm/livepatch.h +0 -5
@@ -24,11 +24,6 @@#include <asm/setup.h> #include <linux/ftrace.h> -static inline int klp_check_compiler_support (void ) -{ -return 0 ; -} -static inline void klp_arch_set_pc (struct pt_regs *regs,unsigned long ip) { regs->ip = ip;
kernel/livepatch/core.c +0 -8
@@ -1207,14 +1207,6 @@ void klp_module_going(struct module *mod)static int __init klp_init (void ) { -int ret; - - ret =klp_check_compiler_support (); -if (ret) { -pr_info ("Your compiler is too old; turning off.\n" ); -return -EINVAL ; - } - klp_root_kobj =kobject_create_and_add ("livepatch" , kernel_kobj);if (!klp_root_kobj)return -ENOMEM ;
Signature · 2 lines
Jiri Kosina SUSE Labs
Steven Rostedt
maintainerRe: [RFC][PATCH 3/2] livepatch: remove klp_check_compiler_support()
Quoted 10 lines · On Fri, 10 May 2019 23:47:50 +0200 (CEST) Jiri Kosina <[email protected]> wrote:
From: Jiri Kosina <[email protected]>
The only purpose of
klp_check_compiler_support()is to make sure that we are not using ftrace on x86 via mcount (because that's executed only after prologue has already happened, and that's too late for livepatching purposes).Now that mcount is not supported by ftrace any more, there is no need for
klp_check_compiler_support()either.
Quoted 6 lines
Reported-by Linus Torvalds <[email protected]>Signed-off-by Jiri Kosina <[email protected]>---
I guess it makes most sense to merge this together with mcount removal in one go.
Thanks, I applied it to my queue and will start running it through my tests.
-- Steve
Josh Poimboeuf
Acked-byRe: [RFC][PATCH 3/2] livepatch: remove klp_check_compiler_support()
In reply to Jiri KosinaQuoted 10 lines · Jiri Kosina
From: Jiri Kosina <[email protected]>
The only purpose of
klp_check_compiler_support()is to make sure that we are not using ftrace on x86 via mcount (because that's executed only after prologue has already happened, and that's too late for livepatching purposes).Now that mcount is not supported by ftrace any more, there is no need for
klp_check_compiler_support()either.
Quoted 6 lines
Reported-by Linus Torvalds <[email protected]>Signed-off-by Jiri Kosina <[email protected]>---
I guess it makes most sense to merge this together with mcount removal in one go.
Signature · 1 line
Josh