[PATCH] powerpc/tm: Fix more userspace r13 corruption

[Nicholas Piggin]

Commit cf13435b730a ("powerpc/tm: Fix userspace r13 corruption") fixes
a problem in treclaim where a SLB miss can occur on the
thread_struct->ckpt_regs while SCRATCH0 is live with the saved user r13
value, clobbering it with the kernel r13 and ultimately resulting in
kernel r13 being stored in ckpt_regs.

There is an equivalent problem in trechkpt where the user r13 value is
loaded into r13 from chkpt_regs to be recheckpointed, but a SLB miss
could occur on ckpt_regs accesses after that, which will result in r13
being clobbered with a kernel value and that will get recheckpointed and
then restored to user registers.

The same memory page is accessed right before this critical window where
a SLB miss could cause corruption, so hitting the bug requires the SLB
entry be removed within a small window of instructions, which is possible
if a SLB related MCE hits there. PAPR also permits the hypervisor to
discard this SLB entry (because slb_shadow->persistent is only set to
SLB_NUM_BOLTED) although it's not known whether any implementations would
do this (KVM does not). So this is an extremely unlikely bug, only found
by inspection.

Fix this by also storing user r13 in a temporary location on the kernel
stack and don't chane the r13 register from kernel r13 until the RI=0
critical section that does not fault.

[ The SCRATCH0 change is not strictly part of the fix, it's only used in
  the RI=0 section so it does not have the same problem as the previous
  SCRATCH0 bug. ]

Signed-off-by: Nicholas Piggin <npiggin at gmail.com>
---
 arch/powerpc/kernel/tm.S | 25 ++++++++++++++++---------
 1 file changed, 16 insertions(+), 9 deletions(-)

diff --git a/arch/powerpc/kernel/tm.S b/arch/powerpc/kernel/tm.S
index 3beecc32940b..5a0f023a26e9 100644
--- a/arch/powerpc/kernel/tm.S
+++ b/arch/powerpc/kernel/tm.S
@@ -443,7 +443,8 @@ restore_gprs:
 
 	REST_GPR(0, r7)				/* GPR0 */
 	REST_GPRS(2, 4, r7)			/* GPR2-4 */
-	REST_GPRS(8, 31, r7)			/* GPR8-31 */
+	REST_GPRS(8, 12, r7)			/* GPR8-12 */
+	REST_GPRS(14, 31, r7)			/* GPR14-31 */
 
 	/* Load up PPR and DSCR here so we don't run with user values for long */
 	mtspr	SPRN_DSCR, r5
@@ -479,18 +480,24 @@ restore_gprs:
 	REST_GPR(6, r7)
 
 	/*
-	 * Store r1 and r5 on the stack so that we can access them after we
-	 * clear MSR RI.
+	 * Store user r1 and r5 and r13 on the stack (in the unused save
+	 * areas / compiler reserved areas), so that we can access them after
+	 * we clear MSR RI.
 	 */
 
 	REST_GPR(5, r7)
 	std	r5, -8(r1)
-	ld	r5, GPR1(r7)
+	ld	r5, GPR13(r7)
 	std	r5, -16(r1)
+	ld	r5, GPR1(r7)
+	std	r5, -24(r1)
 
 	REST_GPR(7, r7)
 
-	/* Clear MSR RI since we are about to use SCRATCH0. EE is already off */
+	/* Stash the stack pointer away for use after recheckpoint */
+	std	r1, PACAR1(r13)
+
+	/* Clear MSR RI since we are about to clobber r13. EE is already off */
 	li	r5, 0
 	mtmsrd	r5, 1
 
@@ -501,9 +508,9 @@ restore_gprs:
 	 * until we turn MSR RI back on.
 	 */
 
-	SET_SCRATCH0(r1)
 	ld	r5, -8(r1)
-	ld	r1, -16(r1)
+	ld	r13, -16(r1)
+	ld	r1, -24(r1)
 
 	/* Commit register state as checkpointed state: */
 	TRECHKPT
@@ -519,9 +526,9 @@ restore_gprs:
 	 */
 
 	GET_PACA(r13)
-	GET_SCRATCH0(r1)
+	ld	r1, PACAR1(r13)
 
-	/* R1 is restored, so we are recoverable again.  EE is still off */
+	/* R13, R1 is restored, so we are recoverable again.  EE is still off */
 	li	r4, MSR_RI
 	mtmsrd	r4, 1
 
-- 
2.23.0