powerpc/process: Fix flush_all_to_thread for SPE
authorFelipe Rechia <felipe.rechia@datacom.com.br>
Wed, 24 Oct 2018 13:57:22 +0000 (10:57 -0300)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Sun, 1 Dec 2019 08:17:01 +0000 (09:17 +0100)
[ Upstream commit e901378578c62202594cba0f6c076f3df365ec91 ]

Fix a bug introduced by the creation of flush_all_to_thread() for
processors that have SPE (Signal Processing Engine) and use it to
compute floating-point operations.

>From userspace perspective, the problem was seen in attempts of
computing floating-point operations which should generate exceptions.
For example:

  fork();
  float x = 0.0 / 0.0;
  isnan(x);           // forked process returns False (should be True)

The operation above also should always cause the SPEFSCR FINV bit to
be set. However, the SPE floating-point exceptions were turned off
after a fork().

Kernel versions prior to the bug used flush_spe_to_thread(), which
first saves SPEFSCR register values in tsk->thread and then calls
giveup_spe(tsk).

After commit 579e633e764e, the save_all() function was called first
to giveup_spe(), and then the SPEFSCR register values were saved in
tsk->thread. This would save the SPEFSCR register values after
disabling SPE for that thread, causing the bug described above.

Fixes 579e633e764e ("powerpc: create flush_all_to_thread()")
Signed-off-by: Felipe Rechia <felipe.rechia@datacom.com.br>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Signed-off-by: Sasha Levin <sashal@kernel.org>
arch/powerpc/kernel/process.c

index 909c940..02b69a6 100644 (file)
@@ -575,12 +575,11 @@ void flush_all_to_thread(struct task_struct *tsk)
        if (tsk->thread.regs) {
                preempt_disable();
                BUG_ON(tsk != current);
-               save_all(tsk);
-
 #ifdef CONFIG_SPE
                if (tsk->thread.regs->msr & MSR_SPE)
                        tsk->thread.spefscr = mfspr(SPRN_SPEFSCR);
 #endif
+               save_all(tsk);
 
                preempt_enable();
        }