On Wed, 2019-01-16 at 14:47 -0200, Breno Leitao wrote:
> There is a kernel crash that happens if rt_sigreturn is called inside a
> transactional block.
>
> This crash happens if the kernel hits an in-kernel page fault when
> accessing userspace memory, usually through copy_ckvsx_to_user(). A major
> page fault calls might_sleep() function, which can cause a task reschedule.
> A task reschedule (switch_to()) reclaim and recheckpoint the TM states,
> but, in the signal return path, the checkpointed memory was already
> reclaimed, thus the exception stack has MSR that points to MSR[TS]=0.
>
> When the code returns from might_sleep() and a task reschedule happened,
> then this task is returned with the memory recheckpointed, and
> CPU MSR[TS] = suspended.
>
> This means that there is a side effect at might_sleep() if it is called
> with CPU MSR[TS] = 0 and the task has regs->msr[TS] != 0.
>
> This side effect can cause a TM bad thing, since at the exception entrance,
> the stack saves MSR[TS]=0, and this is what will be used at RFID, but,
> the processor has MSR[TS] = Suspended, and this transition will be invalid
> and a TM Bad thing will be raised, causing the following crash:
>
> Unexpected TM Bad Thing exception at c00000000000e9ec (msr
> 0x8000000302a03031) tm_scratch=800000010280b033
> cpu 0xc: Vector: 700 (Program Check) at [c00000003ff1fd70]
> pc: c00000000000e9ec: fast_exception_return+0x100/0x1bc
> lr: c000000000032948: handle_rt_signal64+0xb8/0xaf0
> sp: c0000004263ebc40
> msr: 8000000302a03031
> current = 0xc000000415050300
> paca = 0xc00000003ffc4080 irqmask: 0x03 irq_happened: 0x01
> pid = 25006, comm = sigfuz
> Linux version 5.0.0-rc1-00001-g3bd6e94bec12 (breno@debian) (gcc version
> 8.2.0 (Debian 8.2.0-3)) #899 SMP Mon Jan 7 11:30:07 EST 2019
> WARNING: exception is not recoverable, can't continue
> enter ? for help
> [c0000004263ebc40] c000000000032948 handle_rt_signal64+0xb8/0xaf0
> (unreliable)
> [c0000004263ebd30] c000000000022780 do_notify_resume+0x2f0/0x430
> [c0000004263ebe20] c00000000000e844 ret_from_except_lite+0x70/0x74
> --- Exception: c00 (System Call) at 00007fffbaac400c
> SP (7fffeca90f40) is in userspace
>
> The solution for this problem is running the sigreturn code with
> regs->msr[TS] disabled, thus, avoiding hitting the side effect above. This
> does not seem to be a problem since regs->msr will be replaced by the
> ucontext value, so, it is being flushed already. In this case, it is
> flushed earlier.
>
> Signed-off-by: Breno Leitao <lei...@debian.org>
Acked-by: Michael Neuling <mi...@neuling.org>
This still applies on powerpc/next so just acking rather than reposting
> ---
> arch/powerpc/kernel/signal_64.c | 27 ++++++++++++++++++++++++++-
> 1 file changed, 26 insertions(+), 1 deletion(-)
>
> diff --git a/arch/powerpc/kernel/signal_64.c b/arch/powerpc/kernel/signal_64.c
> index 6794466f6420..06c299ef6132 100644
> --- a/arch/powerpc/kernel/signal_64.c
> +++ b/arch/powerpc/kernel/signal_64.c
> @@ -565,7 +565,7 @@ static long restore_tm_sigcontexts(struct task_struct
> *tsk,
> preempt_disable();
>
> /* pull in MSR TS bits from user context */
> - regs->msr = (regs->msr & ~MSR_TS_MASK) | (msr & MSR_TS_MASK);
> + regs->msr |= msr & MSR_TS_MASK;
>
> /*
> * Ensure that TM is enabled in regs->msr before we leave the signal
> @@ -745,6 +745,31 @@ SYSCALL_DEFINE0(rt_sigreturn)
> if (MSR_TM_SUSPENDED(mfmsr()))
> tm_reclaim_current(0);
>
> + /*
> + * Disable MSR[TS] bit also, so, if there is an exception in the
> + * code below (as a page fault in copy_ckvsx_to_user()), it does
> + * not recheckpoint this task if there was a context switch inside
> + * the exception.
> + *
> + * A major page fault can indirectly call schedule(). A reschedule
> + * process in the middle of an exception can have a side effect
> + * (Changing the CPU MSR[TS] state), since schedule() is called
> + * with the CPU MSR[TS] disable and returns with MSR[TS]=Suspended
> + * (switch_to() calls tm_recheckpoint() for the 'new' process). In
> + * this case, the process continues to be the same in the CPU, but
> + * the CPU state just changed.
> + *
> + * This can cause a TM Bad Thing, since the MSR in the stack will
> + * have the MSR[TS]=0, and this is what will be used to RFID.
> + *
> + * Clearing MSR[TS] state here will avoid a recheckpoint if there
> + * is any process reschedule in kernel space. The MSR[TS] state
> + * does not need to be saved also, since it will be replaced with
> + * the MSR[TS] that came from user context later, at
> + * restore_tm_sigcontexts.
> + */
> + regs->msr &= ~MSR_TS_MASK;
> +
> if (__get_user(msr, &uc->uc_mcontext.gp_regs[PT_MSR]))
> goto badframe;
> if (MSR_TM_ACTIVE(msr)) {
