On 05/04/2015, 02:23 PM, Martin Schwidefsky wrote: > On Mon, 4 May 2015 13:40:20 +0200 > Jiri Slaby <jslaby@xxxxxxx> wrote: > >> This means: >> * add a per-thread flag to indicate whether a task is in the old or in >> the new universe, >> * reset it in _slow_ paths of syscall's entry/exit, >> * add helpers around the flag to sched.h, >> * export the status in /proc/<pid>/kgr_in_progress, > >> @@ -217,6 +226,7 @@ ENTRY(system_call) >> mvc __PT_INT_CODE(4,%r11),__LC_SVC_ILC >> stg %r14,__PT_FLAGS(%r11) >> .Lsysc_do_svc: >> + HANDLE_KGRAFT %r12 >> lg %r10,__TI_sysc_table(%r12) # address of system call table >> llgh %r8,__PT_INT_CODE+2(%r11) >> slag %r8,%r8,2 # shift and test for svc 0 > > This is not the slow path, .Lsysc_do_svc is on the main svc path. It is > "only" two instruction but nevertheless this should be avoided. Hi, the commit log says the reset is in the slow path, not the test. But OK, we can optimize, see below. > One way is to combine it with the _TIF_TRACE mechanics: > > .Lsysc_nr_ok: > xc __SF_BACKCHAIN(8,%r15),__SF_BACKCHAIN(%r15) > stg %r2,__PT_ORIG_GPR2(%r11) > stg %r7,STACK_FRAME_OVERHEAD(%r15) > lgf %r9,0(%r8,%r10) # get system call add. > -> tm __TI_flags+6(%r12),_TIF_TRACE>>8 > -> jnz .Lsysc_tracesys > basr %r14,%r9 # call sys_xxxx > stg %r2,__PT_R2(%r11) # store return value > > Add _TIF_KGR_IN_PROGRESS to _TIF_TRACE and branch to a new label, > e.g. to .Lsysc_trace. Distinguish between _TIF_KGR_IN_PROGRESS and > the other trace reasons and either call s390_handle_kgraft or > do_syscall_trace_enter / do_syscall_trace_exit. > > The same for the exit work, add _TIF_KGR_IN_PROGRESS to _TIF_WORK > and sort out the reason in .Lsysc_work. That avoids another two > instructions on the main system call path. I considered this, but there was no space in the word. _TIF_WORK is: TIF_NOTIFY_RESUME 0 TIF_SIGPENDING 1 TIF_NEED_RESCHED 2 TIF_UPROBE 7 _TIF_TRACE is: TIF_SYSCALL_TRACE 3 TIF_SYSCALL_AUDIT 4 TIF_SECCOMP 5 TIF_SYSCALL_TRACEPOINT 6 ===== What I could do is to split them and make this setup: _TIF_WORK: TIF_NOTIFY_RESUME 0 TIF_SIGPENDING 1 TIF_NEED_RESCHED 2 TIF_KGR_IN_PROGRESS_W 3 TIF_UPROBE 7 _TIF_TRACE: TIF_SYSCALL_TRACE 24 TIF_SYSCALL_AUDIT 25 TIF_SECCOMP 26 TIF_SYSCALL_TRACEPOINT 27 TIF_KGR_IN_PROGRESS_T 28 ===== Then make TIF_KGR_IN_PROGRESS_W fire when "tm"-ing _TIF_WORK in "__TI_flags+7". TIF_KGR_IN_PROGRESS_T will work along with _TIF_TRACE using "tm" on "__TI_flags+4". What do you think? thanks, -- js suse labs -- To unsubscribe from this list: send the line "unsubscribe live-patching" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html