powerpc: Fix kernel crash in emulation of vector loads and stores

Commit 350779a2 ("powerpc: Handle most loads and stores in
instruction emulation code", 2017-08-30) changed the register usage
in get_vr and put_vr with the aim of leaving the register number in
r3 untouched on return.  Unfortunately, r6 was not a good choice, as
the callers as of 350779a2 store a MSR value in r6.  Then, in
commit c22435a5 ("powerpc: Emulate FP/vector/VSX loads/stores
correctly when regs not live", 2017-08-30), the saving and restoring
of the MSR got moved into get_vr and put_vr.  Either way, the effect
is that we put a value in MSR that only has the 0x3f8 bits non-zero,
meaning that we are switching to 32-bit mode.  That leads to a crash
like this:

Unable to handle kernel paging request for instruction fetch
Faulting instruction address: 0x0007bea0
Oops: Kernel access of bad area, sig: 11 [#12]
LE SMP NR_CPUS=2048 NUMA PowerNV
Modules linked in: vmx_crypto binfmt_misc ip_tables x_tables autofs4 crc32c_vpmsum
CPU: 6 PID: 32659 Comm: trashy_testcase Tainted: G      D         4.13.0-rc2-00313-gf3026f57e6ed-dirty #23
task: c000000f1bb9e780 task.stack: c000000f1ba98000
NIP:  000000000007bea0 LR: c00000000007b054 CTR: c00000000007be70
REGS: c000000f1ba9b960 TRAP: 0400   Tainted: G      D          (4.13.0-rc2-00313-gf3026f57e6ed-dirty)
MSR:  10000000400010a1 <HV,ME,IR,LE>  CR: 48000228  XER: 00000000
CFAR: c00000000007be74 SOFTE: 1
GPR00: c00000000007b054 c000000f1ba9bbe0 c000000000e6e000 000000000000001d
GPR04: c000000f1ba9bc00 c00000000007be70 00000000000000e8 9000000002009033
GPR08: 0000000002000000 100000000282f033 000000000b0a0900 0000000000001009
GPR12: 0000000000000000 c00000000fd42100 0706050303020100 a5a5a5a5a5a5a5a5
GPR16: 2e2e2e2e2e2de70c 2e2e2e2e2e2e2e2d 0000000000ff00ff 0606040202020000
GPR20: 000000000000005b ffffffffffffffff 0000000003020100 0000000000000000
GPR24: c000000f1ab90020 c000000f1ba9bc00 0000000000000001 0000000000000001
GPR28: c000000f1ba9bc90 c000000f1ba9bea0 000000000b0a0908 0000000000000001
NIP [000000000007bea0] 0x7bea0
LR [c00000000007b054] emulate_loadstore+0x1044/0x1280
Call Trace:
[c000000f1ba9bbe0] [c000000000076b80] analyse_instr+0x60/0x34f0 (unreliable)
[c000000f1ba9bc70] [c00000000007b7ec] emulate_step+0x23c/0x544
[c000000f1ba9bce0] [c000000000053424] arch_uprobe_skip_sstep+0x24/0x40
[c000000f1ba9bd00] [c00000000024b2f8] uprobe_notify_resume+0x598/0xba0
[c000000f1ba9be00] [c00000000001c284] do_notify_resume+0xd4/0xf0
[c000000f1ba9be30] [c00000000000bd44] ret_from_except_lite+0x70/0x74
Instruction dump:
XXXXXXXX XXXXXXXX XXXXXXXX XXXXXXXX XXXXXXXX XXXXXXXX XXXXXXXX XXXXXXXX
XXXXXXXX XXXXXXXX XXXXXXXX XXXXXXXX XXXXXXXX XXXXXXXX XXXXXXXX XXXXXXXX
---[ end trace a7ae7a7f3e0256b5 ]---

To fix this, we just revert to using r3 as before, since the callers
don't rely on r3 being left unmodified.

Fortunately, this can't be triggered by a misaligned load or store,
because vector loads and stores truncate misaligned addresses rather
than taking an alignment interrupt.  It can be triggered using
uprobes.

Fixes: 350779a2 ("powerpc: Handle most loads and stores in instruction emulation code")
Reported-by: Anton Blanchard <anton@ozlabs.org>
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
Tested-by: Anton Blanchard <anton@samba.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>

arch/powerpc/lib/ldstfp.S

@@ -77,7 +77,7 @@ _GLOBAL(get_vr)
 	oris	r7, r6, MSR_VEC@h
 	MTMSRD(r7)
 	isync
-	rlwinm	r6,r3,3,0xf8
+	rlwinm	r3,r3,3,0xf8
 	bcl	20,31,1f
 reg = 0
 	.rept	32
@@ -86,7 +86,7 @@ reg = 0
 reg = reg + 1
 	.endr
 1:	mflr	r5
-	add	r5,r6,r5
+	add	r5,r3,r5
 	mtctr	r5
 	mtlr	r0
 	bctr
@@ -101,7 +101,7 @@ _GLOBAL(put_vr)
 	oris	r7, r6, MSR_VEC@h
 	MTMSRD(r7)
 	isync
-	rlwinm	r6,r3,3,0xf8
+	rlwinm	r3,r3,3,0xf8
 	bcl	20,31,1f
 reg = 0
 	.rept	32
@@ -110,7 +110,7 @@ reg = 0
 reg = reg + 1
 	.endr
 1:	mflr	r5
-	add	r5,r6,r5
+	add	r5,r3,r5
 	mtctr	r5
 	mtlr	r0
 	bctr