hi Cyril,
On Wed, Jun 08, 2016 at 02:00:34PM +1000, Cyril Bur wrote:
quoted hunk ↗ jump to hunk
@@ -1108,11 +1084,11 @@ struct task_struct *__switch_to(struct task_struct *prev,
*/
save_sprs(&prev->thread);
- __switch_to_tm(prev);
-
/* Save FPU, Altivec, VSX and SPE state */
giveup_all(prev);
+ __switch_to_tm(prev);
+
There should be a bug.
giveup_all() will clear MSR[FP] bit.
__switch_to_tm() reads that bit to decide whether the FP
register needs to be flushed to thread_struct.
=== tm_reclaim() (invoked by __switch_to_tm)========================
andi. r0, r4, MSR_FP
beq dont_backup_fp
addi r7, r3, THREAD_CKFPSTATE
SAVE_32FPRS_VSRS(0, R6, R7) /* r6 scratch, r7 transact fp
state */
mffs fr0
stfd fr0,FPSTATE_FPSCR(r7)
dont_backup_fp:
=============================
But now the __switch_to_tm() is moved behind giveup_all().
So __switch_to_tm() loses MSR[FP] and cannot decide whether saving ckpt FPU or not.
The same applies to VMX/VSX.
Thanks,
- Simon