>
> Ananth, Srikar, I'd suggest this test-case:
>
> #include <stdio.h>
> #include <signal.h>
> #include <ucontext.h>
>
> void *fault_insn;
>
> static inline void *uc_ip(struct ucontext *ctxt)
> {
> return (void*)ctxt->uc_mcontext.gregs[16];
> }
>
> void segv(int sig, siginfo_t *info, void *ctxt)
> {
> static int cnt;
>
> printf("SIGSEGV! ip=%p addr=%p\n", uc_ip(ctxt), info->si_addr);
>
> if (uc_ip(ctxt) != fault_insn)
> printf("ERR!! wrong ip\n");
> if (info->si_addr != (void*)0x12345678)
> printf("ERR!! wrong addr\n");
>
> if (++cnt == 3)
> signal(SIGSEGV, SIG_DFL);
> }
>
> int main(void)
> {
> struct sigaction sa = {
> .sa_sigaction = segv,
> .sa_flags = SA_SIGINFO,
> };
>
> sigaction(SIGSEGV, &sa, NULL);
>
> fault_insn = &&label;
>
> label:
> asm volatile ("movl $0x0,0x12345678");
>
> return 0;
> }
>
> result:
>
> $ ulimit -c unlimited
>
> $ ./segv
> SIGSEGV! ip=0x4006eb addr=0x12345678
> SIGSEGV! ip=0x4006eb addr=0x12345678
> SIGSEGV! ip=0x4006eb addr=0x12345678
> Segmentation fault (core dumped)
>
> $ gdb -c ./core.1826
> ...
> Program terminated with signal 11, Segmentation fault.
> #0 0x00000000004006eb in ?? ()
>
> Now. If you insert uprobe at asm("movl") insn, result should be the same
> or the patches I sent are wrong. In particular, the addr in the coredump
> should be correct too. And consumer->handler() should be called 3 times
> too. This insn is really executed 3 times.
>
> I have no idea how can I test this.
>
I have tested this on both x86_32 and x86_64 and can confirm that the
behaviour is same with or without uprobes placed at that instruction.
This is on the uprobes code with your changes.
However on x86_32; the output is different from x86_64.
On x86_32 (I have additionally printed the uc_ip and fault_insn.
SIGSEGV! ip=0x10246 addr=0x12345678
ERR!! wrong ip uc_ip(ctxt) = 10246 fault_insn = 804856c
SIGSEGV! ip=0x10246 addr=0x12345678
ERR!! wrong ip uc_ip(ctxt) = 10246 fault_insn = 804856c
SIGSEGV! ip=0x10246 addr=0x12345678
ERR!! wrong ip uc_ip(ctxt) = 10246 fault_insn = 804856c
Segmentation fault
the fault_insn matches the address shown in disassemble of gdb.
I still trying to dig up what uc_ip is and why its different on x86_32.
On x86_64 the result is what you pasted above.
Also I was thinking on your suggestion of making abort_xol a weak
function. In which case we could have architecture independent function
in kernel/uprobes.c which is just a wrapper for set_instruction_pointer.
void __weak abort_xol(struct pt_regs *regs, struct uprobe_task *utask)
{
set_instruction_pointer(regs, utask->vaddr);
}
where it would called from uprobe_notify_resume() as
abort_xol(regs, utask);
If other archs would want to do something else, they could override
abort_xol definition.
--
Thanks and Regards
Srikar
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxx. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>