Le 07/07/2017 à 11:05, Laurent Vivier a écrit : > Le 07/07/2017 à 04:20, Richard Henderson a écrit : >> We translate gUSA regions atomically in a parallel context. >> But in a serial context a gUSA region may be interrupted. >> In that case, restart the region as the kernel would. >> >> Signed-off-by: Richard Henderson <r...@twiddle.net> >> --- >> linux-user/signal.c | 23 +++++++++++++++++++++++ >> 1 file changed, 23 insertions(+) >> >> diff --git a/linux-user/signal.c b/linux-user/signal.c >> index 3d18d1b..a537778 100644 >> --- a/linux-user/signal.c >> +++ b/linux-user/signal.c >> @@ -3471,6 +3471,25 @@ static abi_ulong get_sigframe(struct target_sigaction >> *ka, >> return (sp - frame_size) & -8ul; >> } >> >> +/* Notice when we're in the middle of a gUSA region and reset. >> + Note that this will only occur for !parallel_cpus, as we will >> + translate such sequences differently in a parallel context. */ >> +static void unwind_gusa(CPUSH4State *regs) >> +{ >> + /* If the stack pointer is sufficiently negative ... */ >> + if ((regs->gregs[15] & 0xc0000000u) == 0xc0000000u >> + /* ... and we haven't completed the sequence ... */ >> + && regs->pc < regs->gregs[0]) { >> + /* Reset the PC to before the gUSA region, as computed from >> + R0 = region end, SP = -(region size), plus one more insn >> + that actually sets SP to the region size. */ >> + regs->pc = regs->gregs[0] + regs->gregs[15] - 2; >> + >> + /* Reset the SP to the saved version in R1. */ >> + regs->gregs[15] = regs->gregs[1]; >> + } >> +} >> + >> static void setup_sigcontext(struct target_sigcontext *sc, >> CPUSH4State *regs, unsigned long mask) >> { >> @@ -3534,6 +3553,8 @@ static void setup_frame(int sig, struct >> target_sigaction *ka, >> abi_ulong frame_addr; >> int i; >> >> + unwind_gusa(regs); >> + >> frame_addr = get_sigframe(ka, regs->gregs[15], sizeof(*frame)); > > I think unwind_gusa() should be moved after the get_sigfram() (in both > cases), because r15 can be updated and the sigframe base lost. >
No, it's stupid, r15 is negative Laurent