On Sun, 16 Jul 2017 23:37:44 +0900 Masami Hiramatsu <[email protected]> wrote:
> So, the story what the stack said, > > - optimized_callback() calls get_kprobe_ctlblk(), and > this_cpu_ptr(&kprobe_ctlblk) caused a page fault (in apic_timer_interrupt, > does it cause any problem?) > - and following call-chain occured > async_page_fault -> error_entry -> trace_hardirqs_off_thunk -> > trace_hardirqs_off_caller > - "mov %gs:0xc400,%rdx" caused async_page_fault() again. > > Since trace_hardirqs_off_thunk() stores general registers on > the stack, there are some noises. > > [ 114.429637] FS: 00000000021e7880(0000) GS:ffff88001fd40000(0000) > knlGS:0000000000000000 > > So, the problem seems that cpu can not access to per-cpu pages. OK, I got the root cause of this issue. Since at the irqentry code, segment registers are not prepared for kernel yet (e.g. interrupted in user-mode), so we must not optimize it. I found we had already checked that by checking __entry_text_start/end, but that is not enough, we need irqentry_text check too. Here I made another patch, please try it. ----- kprobes/x86: Do not jump-optimize kprobes on irq entry code From: Masami Hiramatsu <[email protected]> Since the segment registers are not prepared for kernel in the irq-entry code, if a kprobe on such code is jump-optimized, accessing per-cpu variables may cause kernel panic. However, if the kprobe is not optimized, it kicks int3 exception and set segment registers correctly. This checks probe-address and if it is in irq-entry code, it prohibits optimizing such kprobes. This means we can continuously probing such interrupt handlers by kprobes but it is not optimized anymore. Signed-off-by: Masami Hiramatsu <[email protected]> Reported-by: Francis Deslauriers <[email protected]> --- arch/x86/entry/entry_64.S | 2 +- arch/x86/include/asm/unwind.h | 1 + arch/x86/kernel/kprobes/opt.c | 4 ++-- arch/x86/kernel/unwind_frame.c | 4 ++-- 4 files changed, 6 insertions(+), 5 deletions(-) diff --git a/arch/x86/entry/entry_64.S b/arch/x86/entry/entry_64.S index a9a8027..95bca8b 100644 --- a/arch/x86/entry/entry_64.S +++ b/arch/x86/entry/entry_64.S @@ -675,7 +675,7 @@ apicinterrupt3 \num trace(\sym) smp_trace(\sym) #endif /* Make sure APIC interrupt handlers end up in the irqentry section: */ -#if defined(CONFIG_FUNCTION_GRAPH_TRACER) || defined(CONFIG_KASAN) +#if defined(CONFIG_FUNCTION_GRAPH_TRACER) || defined(CONFIG_KASAN) || defined(CONFIG_KPROBES) # define PUSH_SECTION_IRQENTRY .pushsection .irqentry.text, "ax" # define POP_SECTION_IRQENTRY .popsection #else diff --git a/arch/x86/include/asm/unwind.h b/arch/x86/include/asm/unwind.h index e667649..a9896fb9 100644 --- a/arch/x86/include/asm/unwind.h +++ b/arch/x86/include/asm/unwind.h @@ -28,6 +28,7 @@ void __unwind_start(struct unwind_state *state, struct task_struct *task, bool unwind_next_frame(struct unwind_state *state); unsigned long unwind_get_return_address(struct unwind_state *state); +bool in_entry_code(unsigned long ip); static inline bool unwind_done(struct unwind_state *state) { diff --git a/arch/x86/kernel/kprobes/opt.c b/arch/x86/kernel/kprobes/opt.c index 69ea0bc..a51c144 100644 --- a/arch/x86/kernel/kprobes/opt.c +++ b/arch/x86/kernel/kprobes/opt.c @@ -39,6 +39,7 @@ #include <asm/insn.h> #include <asm/debugreg.h> #include <asm/set_memory.h> +#include <asm/unwind.h> #include "common.h" @@ -253,8 +254,7 @@ static int can_optimize(unsigned long paddr) * Do not optimize in the entry code due to the unstable * stack handling. */ - if ((paddr >= (unsigned long)__entry_text_start) && - (paddr < (unsigned long)__entry_text_end)) + if (in_entry_code(paddr)) return 0; /* Check there is enough space for a relative jump. */ diff --git a/arch/x86/kernel/unwind_frame.c b/arch/x86/kernel/unwind_frame.c index b9389d7..95123ce 100644 --- a/arch/x86/kernel/unwind_frame.c +++ b/arch/x86/kernel/unwind_frame.c @@ -84,14 +84,14 @@ static size_t regs_size(struct pt_regs *regs) return sizeof(*regs); } -static bool in_entry_code(unsigned long ip) +bool in_entry_code(unsigned long ip) { char *addr = (char *)ip; if (addr >= __entry_text_start && addr < __entry_text_end) return true; -#if defined(CONFIG_FUNCTION_GRAPH_TRACER) || defined(CONFIG_KASAN) +#if defined(CONFIG_FUNCTION_GRAPH_TRACER) || defined(CONFIG_KASAN) || defined(CONFIG_KPROBES) if (addr >= __irqentry_text_start && addr < __irqentry_text_end) return true; #endif -- Masami Hiramatsu <[email protected]>
