On Friday 16 November 2012 10:28 AM, Al Viro wrote: >> + ; --------- check for signals/restore-sigmask ------------ >> + bbit0 r9, TIF_SIGPENDING, chk_next_work >> + >> + ; save CALLEE Regs. >> + ; (i) If this signal causes coredump - full regfile needed >> + ; (ii) If signal is SIGTRAP/SIGSTOP, task is being traced thus >> + ; tracer might call PEEKUSR for a CALLEE reg >> + ; >> + ; NOTE: SP will grow up by size of CALLEE Reg-File >> + SAVE_CALLEE_SAVED_USER ; clobbers r12 >> + >> + ; save location of saved Callee Regs @ thread_struct->callee >> + GET_CURR_TASK_FIELD_PTR TASK_THREAD, r10 >> + st sp, [r10, THREAD_CALLEE_REG] >> + >> + bl @do_signal >> + >> + ; unwind SP for cheap discard of Callee saved Regs >> + DISCARD_CALLEE_SAVED_USER > Uh-oh... And what if tracer wanted to modify callee-saved regs? So the solution would be to either unconditionally restore all the 13 callee regs - or add additional state (struct thread) where ptrace makes a note that it wrote to a callee reg which is used here to conditional-ize the restore. Former is simpler to do - although it might ill-affect micro-benchmarks such as LMBench lat_sig. Anyhow correctness comes before optimization. > >> + b resume_user_mode_begin ; loop back to start of U mode ret >> + >> + ; --- notify_resume --- >> +chk_next_work: >> + btst r9, TIF_NOTIFY_RESUME >> + blnz @do_notify_resume >> + >> + ;--------- All things done, go back to Userland ------ >> + >> + b restore_regs > No. After NOTIFY_RESUME stuff you need to recheck SIGPENDING. This should > go to resume_user_mode_begin, not restore regs. Thanks for pointing this - fixed in v2 series. > Another problem here is > IRQ handling - you hit do_signal()/do_notify_resume() with IRQs disabled, > which is broken - you need to re-enable it before going into either. Thanks - fixed that as well. I presume it is broken for "interrupt latency reasons" and not because IRQs could remain disabled all the way into usermode - since given the way RTIE insn works - that won't happen. Although independently this was indeed cause of a WARN_ON_ONCE when bh were enabled with IRQs disabled. -- 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
