Re: [PATCH RFC] rethook: inline arch_rethook_trampoline_callback() in assembly code
On Wed, Apr 24, 2024 at 5:02 PM Andrii Nakryiko wrote: > > At the lowest level, rethook-based kretprobes on x86-64 architecture go > through arch_rethoook_trampoline() function, manually written in > assembly, which calls into a simple arch_rethook_trampoline_callback() > function, written in C, and only doing a few straightforward field > assignments, before calling further into rethook_trampoline_handler(), > which handles kretprobe callbacks generically. > > Looking at simplicity of arch_rethook_trampoline_callback(), it seems > not really worthwhile to spend an extra function call just to do 4 or > 5 assignments. As such, this patch proposes to "inline" > arch_rethook_trampoline_callback() into arch_rethook_trampoline() by > manually implementing it in an assembly code. > > This has two motivations. First, we do get a bit of runtime speed up by > avoiding function calls. Using BPF selftests's bench tool, we see > 0.6%-0.8% throughput improvement for kretprobe/multi-kretprobe > triggering code path: > > BEFORE (latest probes/for-next) > === > kretprobe : 10.455 ± 0.024M/s > kretprobe-multi: 11.150 ± 0.012M/s > > AFTER (probes/for-next + this patch) > > kretprobe : 10.540 ± 0.009M/s (+0.8%) > kretprobe-multi: 11.219 ± 0.042M/s (+0.6%) > > Second, and no less importantly for some specialized use cases, this > avoids unnecessarily "polluting" LBR records with an extra function call > (recorded as a jump by CPU). This is the case for the retsnoop ([0]) > tool, which relies havily on capturing LBR records to provide users with > lots of insight into kernel internals. > > This RFC patch is only inlining this function for x86-64, but it's > possible to do that for 32-bit x86 arch as well and then remove > arch_rethook_trampoline_callback() implementation altogether. Please let > me know if this change is acceptable and whether I should complete it > with 32-bit "inlining" as well. Thanks! > > [0] > https://nakryiko.com/posts/retsnoop-intro/#peering-deep-into-functions-with-lbr > > Signed-off-by: Andrii Nakryiko > --- > arch/x86/kernel/asm-offsets_64.c | 4 > arch/x86/kernel/rethook.c| 37 +++- > 2 files changed, 36 insertions(+), 5 deletions(-) > > diff --git a/arch/x86/kernel/asm-offsets_64.c > b/arch/x86/kernel/asm-offsets_64.c > index bb65371ea9df..5c444abc540c 100644 > --- a/arch/x86/kernel/asm-offsets_64.c > +++ b/arch/x86/kernel/asm-offsets_64.c > @@ -42,6 +42,10 @@ int main(void) > ENTRY(r14); > ENTRY(r15); > ENTRY(flags); > + ENTRY(ip); > + ENTRY(cs); > + ENTRY(ss); > + ENTRY(orig_ax); > BLANK(); > #undef ENTRY > > diff --git a/arch/x86/kernel/rethook.c b/arch/x86/kernel/rethook.c > index 8a1c0111ae79..3e1c01beebd1 100644 > --- a/arch/x86/kernel/rethook.c > +++ b/arch/x86/kernel/rethook.c > @@ -6,6 +6,7 @@ > #include > #include > #include > +#include > > #include "kprobes/common.h" > > @@ -34,10 +35,36 @@ asm( > " pushq %rsp\n" > " pushfq\n" > SAVE_REGS_STRING > - " movq %rsp, %rdi\n" > - " call arch_rethook_trampoline_callback\n" > + " movq %rsp, %rdi\n" /* $rdi points to regs */ > + /* fixup registers */ > + /* regs->cs = __KERNEL_CS; */ > + " movq $" __stringify(__KERNEL_CS) ", " __stringify(pt_regs_cs) > "(%rdi)\n" > + /* regs->ip = (unsigned long)_rethook_trampoline; */ > + " movq $arch_rethook_trampoline, " __stringify(pt_regs_ip) > "(%rdi)\n" > + /* regs->orig_ax = ~0UL; */ > + " movq $0x, " __stringify(pt_regs_orig_ax) > "(%rdi)\n" > + /* regs->sp += 2*sizeof(long); */ > + " addq $16, " __stringify(pt_regs_sp) "(%rdi)\n" > + /* 2nd arg is frame_pointer = (long *)(regs + 1); */ > + " lea " __stringify(PTREGS_SIZE) "(%rdi), %rsi\n" BTW, all this __stringify() ugliness can be avoided if we move this assembly into its own .S file, like lots of other assembly functions in arch/x86/kernel subdir. That has another benefit of generating better line information in DWARF for those assembly instructions. It's lots more work, so before I do this, I'd like to get confirmation that this change is acceptable in principle. > + /* > +* The return address at 'frame_pointer' is recovered by the > +* arch_rethook_fixup_return() which called from this > +* rethook_trampoline_handler(). > +*/ > + " call rethook_trampoline_handler\n" > + /* > +* Copy FLAGS to 'pt_regs::ss' so we can do RET right after POPF. > +* > +* We don't save/restore %rax below, because we ignore > +* rethook_trampoline_handler result. > +* > +* *(unsigned long *)>ss = regs->flags; > +*/ > + " mov " __stringify(pt_regs_flags) "(%rsp), %rax\n" >
[PATCH RFC] rethook: inline arch_rethook_trampoline_callback() in assembly code
At the lowest level, rethook-based kretprobes on x86-64 architecture go through arch_rethoook_trampoline() function, manually written in assembly, which calls into a simple arch_rethook_trampoline_callback() function, written in C, and only doing a few straightforward field assignments, before calling further into rethook_trampoline_handler(), which handles kretprobe callbacks generically. Looking at simplicity of arch_rethook_trampoline_callback(), it seems not really worthwhile to spend an extra function call just to do 4 or 5 assignments. As such, this patch proposes to "inline" arch_rethook_trampoline_callback() into arch_rethook_trampoline() by manually implementing it in an assembly code. This has two motivations. First, we do get a bit of runtime speed up by avoiding function calls. Using BPF selftests's bench tool, we see 0.6%-0.8% throughput improvement for kretprobe/multi-kretprobe triggering code path: BEFORE (latest probes/for-next) === kretprobe : 10.455 ± 0.024M/s kretprobe-multi: 11.150 ± 0.012M/s AFTER (probes/for-next + this patch) kretprobe : 10.540 ± 0.009M/s (+0.8%) kretprobe-multi: 11.219 ± 0.042M/s (+0.6%) Second, and no less importantly for some specialized use cases, this avoids unnecessarily "polluting" LBR records with an extra function call (recorded as a jump by CPU). This is the case for the retsnoop ([0]) tool, which relies havily on capturing LBR records to provide users with lots of insight into kernel internals. This RFC patch is only inlining this function for x86-64, but it's possible to do that for 32-bit x86 arch as well and then remove arch_rethook_trampoline_callback() implementation altogether. Please let me know if this change is acceptable and whether I should complete it with 32-bit "inlining" as well. Thanks! [0] https://nakryiko.com/posts/retsnoop-intro/#peering-deep-into-functions-with-lbr Signed-off-by: Andrii Nakryiko --- arch/x86/kernel/asm-offsets_64.c | 4 arch/x86/kernel/rethook.c| 37 +++- 2 files changed, 36 insertions(+), 5 deletions(-) diff --git a/arch/x86/kernel/asm-offsets_64.c b/arch/x86/kernel/asm-offsets_64.c index bb65371ea9df..5c444abc540c 100644 --- a/arch/x86/kernel/asm-offsets_64.c +++ b/arch/x86/kernel/asm-offsets_64.c @@ -42,6 +42,10 @@ int main(void) ENTRY(r14); ENTRY(r15); ENTRY(flags); + ENTRY(ip); + ENTRY(cs); + ENTRY(ss); + ENTRY(orig_ax); BLANK(); #undef ENTRY diff --git a/arch/x86/kernel/rethook.c b/arch/x86/kernel/rethook.c index 8a1c0111ae79..3e1c01beebd1 100644 --- a/arch/x86/kernel/rethook.c +++ b/arch/x86/kernel/rethook.c @@ -6,6 +6,7 @@ #include #include #include +#include #include "kprobes/common.h" @@ -34,10 +35,36 @@ asm( " pushq %rsp\n" " pushfq\n" SAVE_REGS_STRING - " movq %rsp, %rdi\n" - " call arch_rethook_trampoline_callback\n" + " movq %rsp, %rdi\n" /* $rdi points to regs */ + /* fixup registers */ + /* regs->cs = __KERNEL_CS; */ + " movq $" __stringify(__KERNEL_CS) ", " __stringify(pt_regs_cs) "(%rdi)\n" + /* regs->ip = (unsigned long)_rethook_trampoline; */ + " movq $arch_rethook_trampoline, " __stringify(pt_regs_ip) "(%rdi)\n" + /* regs->orig_ax = ~0UL; */ + " movq $0x, " __stringify(pt_regs_orig_ax) "(%rdi)\n" + /* regs->sp += 2*sizeof(long); */ + " addq $16, " __stringify(pt_regs_sp) "(%rdi)\n" + /* 2nd arg is frame_pointer = (long *)(regs + 1); */ + " lea " __stringify(PTREGS_SIZE) "(%rdi), %rsi\n" + /* +* The return address at 'frame_pointer' is recovered by the +* arch_rethook_fixup_return() which called from this +* rethook_trampoline_handler(). +*/ + " call rethook_trampoline_handler\n" + /* +* Copy FLAGS to 'pt_regs::ss' so we can do RET right after POPF. +* +* We don't save/restore %rax below, because we ignore +* rethook_trampoline_handler result. +* +* *(unsigned long *)>ss = regs->flags; +*/ + " mov " __stringify(pt_regs_flags) "(%rsp), %rax\n" + " mov %rax, " __stringify(pt_regs_ss) "(%rsp)\n" RESTORE_REGS_STRING - /* In the callback function, 'regs->flags' is copied to 'regs->ss'. */ + /* We just copied 'regs->flags' into 'regs->ss'. */ " addq $16, %rsp\n" " popfq\n" #else @@ -61,6 +88,7 @@ asm( ); NOKPROBE_SYMBOL(arch_rethook_trampoline); +#ifdef CONFIG_X86_32 /* * Called from arch_rethook_trampoline */ @@ -70,9 +98,7 @@ __used __visible void arch_rethook_trampoline_callback(struct pt_regs *regs) /* fixup registers */ regs->cs = __KERNEL_CS; -#ifdef CONFIG_X86_32 regs->gs = 0; -#endif
