On Tue, 2017-10-24 at 15:13 -0200, Breno Leitao wrote:
> From: Breno Leitao <breno.lei...@debian.org>
>
> On commit commit f48e91e87e67 ("powerpc/tm: Fix FP and VMX register
> corruption"), we check ckpt_regs.msr to see if a feature (as VEC, VSX
> and FP) is disabled (thus the hot registers might be bogus during the
> reclaim), and then copy the previously saved thread registers, with the
> non-bogus values, into the checkpoint area for a later trecheckpoint.
> This mechanism is used to recheckpoints the proper register values when
> a transaction started using the bogus registers, and these values were
> sent to the memory checkpoint area.
>
> I see a problem on this code that ckpt_regs.msr is not properly set when
> using it, as for example, when there is a vsx_unavailable_tm() in a code
> like the following, the ckpt_regs.msg[FP] is 0;
>
> 1: sleep_until_{fp,vec,vsx} = 0
> 2: fadd
> 3: tbegin.
> 4: beq
> 5: xxmrghd
> 6: tend.
>
> In this case, line 5 will raise an vsx_unavailable_tm() exception, and
> the ckpt_regs.msr[FP] will be zero before memcpy() block, executing the
> memcpy even with the the FP registers hot. That is not correct because
> we executed a float point instruction on line 2, and MSR[FP] was set to
> 1.
>
> Fortunately this does not cause a big problem as I can see, other than
> this extra memcpy() because treclaim() will later overwrite this wrong
> copied value, since it relies on the correct MSR value, which was
> updated by giveup_all->check_if_tm_restore_required. There might be a
> problem when laziness is being turned on, but I was not able to
> reproduce it.
I believe this analysis is correct, I have come to the same conclusion
in the past. I've also done a bunch of testing with variants of this
patch and haven't seen a difference, however, I do believe the code is
more correct with this patch.
Signed-off-by: Cyril Bur <cyril...@gmail.com>
Having said all that, nothing rules out that our tests simply aren't
good enough ;)
>
> The solution I am proposing is updating ckpt_regs.msr before using it.
>
> Signed-off-by: Breno Leitao <lei...@debian.org>
> Signed-off-by: Gustavo Romero <gusbrom...@gmail.com>
> CC: Cyril Bur <cyril...@gmail.com>
> CC: Michael Neuling <mi...@neuling.org>
> ---
> arch/powerpc/kernel/process.c | 5 +++--
> 1 file changed, 3 insertions(+), 2 deletions(-)
>
> diff --git a/arch/powerpc/kernel/process.c b/arch/powerpc/kernel/process.c
> index c051dc2b42ad..773e9c5594e7 100644
> --- a/arch/powerpc/kernel/process.c
> +++ b/arch/powerpc/kernel/process.c
> @@ -860,6 +860,9 @@ static void tm_reclaim_thread(struct thread_struct *thr,
> if (!MSR_TM_SUSPENDED(mfmsr()))
> return;
>
> + /* Give up all the registers and set ckpt_regs.msr */
> + giveup_all(container_of(thr, struct task_struct, thread));
> +
> /*
> * If we are in a transaction and FP is off then we can't have
> * used FP inside that transaction. Hence the checkpointed
> @@ -879,8 +882,6 @@ static void tm_reclaim_thread(struct thread_struct *thr,
> memcpy(&thr->ckvr_state, &thr->vr_state,
> sizeof(struct thread_vr_state));
>
> - giveup_all(container_of(thr, struct task_struct, thread));
> -
> tm_reclaim(thr, thr->ckpt_regs.msr, cause);
> }
>
