On Thu, Feb 11, 2010 at 15:46, Roland McGrath wrote: >> config BLACKFIN >> def_bool y >> select HAVE_ARCH_KGDB >> + select HAVE_ARCH_TRACEHOOK > > Don't define this until you have all its constituents as listed in the > arch/Kconfig comment. I don't see user_regset support. where is user_regset actually used ? i only see it in fs/binfmt_elf.c and core dumps, neither of which work on nommu systems (or at least on Blackfin systems). >> +static inline void >> +syscall_get_arguments(struct task_struct *task, struct pt_regs *regs, >> + unsigned int i, unsigned int n, unsigned long *args) >> +{ >> + /* wtf is "i" ? */ >> + BUG_ON(i); > > i is the starting number. args[0] gets the i'th argument, > args[n - 1] gets the i+n-1'th argument. i dont see anyone calling syscall_get_arguments() with i!=0, and a few other arches are doing the BUG_ON(i) thing too. but should be easy to implement this with memory walking code ... >> +asmlinkage void syscall_trace_leave(struct pt_regs *regs) >> +{ >> + if (test_thread_flag(TIF_SYSCALL_TRACE)) >> + tracehook_report_syscall_exit(regs, 0); >> } > > Is it in fact true that single-step reports still come normally after a > syscall instruction? this is unchanged from the previous Blackfin behavior, and it's how most arches behaved in 2.6.32. but looking in latest mainline, it seems people are changing to: if (test_thread_flag(TIF_SINGLESTEP) || test_thread_flag(TIF_SYSCALL_TRACE)) tracehook_report_syscall_exit(regs, 0); so changing Blackfin too should be straightforward i guess >> @@ -213,7 +213,7 @@ >> */ >> if (regs->syscfg & TRACE_BITS) { >> regs->syscfg &= ~TRACE_BITS; >> - ptrace_notify(SIGTRAP); >> + tracehook_signal_handler(sig, info, ka, regs, 1); >> } > > This call should be made unconditionally, and it should be made after the > signal mask changes have been made (i.e. at the end of handle_signal). I > think it's wrong to clear the single-step flag here. Instead, pass > (regs->syscfg & TRACE_BITS) as the last argument. > > With ptrace, it makes no difference one way or the other because it will > always either explicitly clear or explicitly set single-step before it > resumes. But in future, it will matter. sounds like this issue is unrelated to tracehook and how we've been doing signal/ptrace stuff has always been a little broken ... i'll move it to how most arches seem to do it -- in do_signal after a successful call to handle_signal and after clearing TIF_RESTORE_SIGMASK. thanks for the review -mike -- To unsubscribe from this list: send the line "unsubscribe linux-arch" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html