powerpc/64s: Add workaround for P9 vector CI load issue

POWER9 DD2.1 and earlier has an issue where some cache inhibited
vector load will return bad data. The workaround is two part, one
firmware/microcode part triggers HMI interrupts when hitting such
loads, the other part is this patch which then emulates the
instructions in Linux.

The affected instructions are limited to lxvd2x, lxvw4x, lxvb16x and
lxvh8x.

When an instruction triggers the HMI, all threads in the core will be
sent to the HMI handler, not just the one running the vector load.

In general, these spurious HMIs are detected by the emulation code and
we just return back to the running process. Unfortunately, if a
spurious interrupt occurs on a vector load that's to normal memory we
have no way to detect that it's spurious (unless we walk the page
tables, which is very expensive). In this case we emulate the load but
we need do so using a vector load itself to ensure 128bit atomicity is
preserved.

Some additional debugfs emulated instruction counters are added also.

Signed-off-by: Michael Neuling <[e-mail]>
Signed-off-by: Benjamin Herrenschmidt <[e-mail]>
[mpe: Switch CONFIG_PPC_BOOK3S_64 to CONFIG_VSX to unbreak the build]
Signed-off-by: Michael Ellerman <[e-mail]>

arch/powerpc/include/asm/emulated_ops.h

···
 	struct ppc_emulated_entry mfdscr;
 	struct ppc_emulated_entry mtdscr;
 	struct ppc_emulated_entry lq_stq;
+	struct ppc_emulated_entry lxvw4x;
+	struct ppc_emulated_entry lxvh8x;
+	struct ppc_emulated_entry lxvd2x;
+	struct ppc_emulated_entry lxvb16x;
 #endif
 } ppc_emulated;
 

arch/powerpc/include/asm/paca.h

···
 	 */
 	u16 in_mce;
 	u8 hmi_event_available;		/* HMI event is available */
+	u8 hmi_p9_special_emu;		/* HMI P9 special emulation */
 #endif
 
 	/* Stuff for accurate time accounting */

arch/powerpc/include/asm/uaccess.h

···
 
 extern long __get_user_bad(void);
 
+/*
+ * This does an atomic 128 byte aligned load from userspace.
+ * Upto caller to do enable_kernel_vmx() before calling!
+ */
+#define __get_user_atomic_128_aligned(kaddr, uaddr, err)		\
+	__asm__ __volatile__(				\
+		"1:	lvx  0,0,%1	# get user\n"	\
+		" 	stvx 0,0,%2	# put kernel\n"	\
+		"2:\n"					\
+		".section .fixup,\"ax\"\n"		\
+		"3:	li %0,%3\n"			\
+		"	b 2b\n"				\
+		".previous\n"				\
+		EX_TABLE(1b, 3b)			\
+		: "=r" (err)			\
+		: "b" (uaddr), "b" (kaddr), "i" (-EFAULT), "0" (err))
+
 #define __get_user_asm(x, addr, err, op)		\
 	__asm__ __volatile__(				\
 		"1:	"op" %1,0(%2)	# get_user\n"	\

arch/powerpc/kernel/exceptions-64s.S

···
 	EXCEPTION_PROLOG_COMMON_3(0xe60)
 	addi	r3,r1,STACK_FRAME_OVERHEAD
 	BRANCH_LINK_TO_FAR(hmi_exception_realmode) /* Function call ABI */
+	cmpdi	cr0,r3,0
+
 	/* Windup the stack. */
 	/* Move original HSRR0 and HSRR1 into the respective regs */
 	ld	r9,_MSR(r1)
···
 	REST_8GPRS(2, r1)
 	REST_GPR(10, r1)
 	ld	r11,_CCR(r1)
+	REST_2GPRS(12, r1)
+	bne	1f
 	mtcr	r11
 	REST_GPR(11, r1)
-	REST_2GPRS(12, r1)
-	/* restore original r1. */
+	ld	r1,GPR1(r1)
+	hrfid
+
+1:	mtcr	r11
+	REST_GPR(11, r1)
 	ld	r1,GPR1(r1)
 
 	/*
···
 	EXCEPTION_PROLOG_0(PACA_EXGEN)
 	b	tramp_real_hmi_exception
 
-EXC_COMMON_ASYNC(hmi_exception_common, 0xe60, handle_hmi_exception)
-
+EXC_COMMON_BEGIN(hmi_exception_common)
+EXCEPTION_COMMON(PACA_EXGEN, 0xe60, hmi_exception_common, handle_hmi_exception,
+        ret_from_except, FINISH_NAP;ADD_NVGPRS;ADD_RECONCILE;RUNLATCH_ON)
 
 EXC_REAL_OOL_MASKABLE_HV(h_doorbell, 0xe80, 0x20)
 EXC_VIRT_OOL_MASKABLE_HV(h_doorbell, 0x4e80, 0x20, 0xe80)

arch/powerpc/kernel/mce.c

···
 {
 	__this_cpu_inc(irq_stat.hmi_exceptions);
 
+#ifdef CONFIG_PPC_BOOK3S_64
+	/* Workaround for P9 vector CI loads (see p9_hmi_special_emu) */
+	if (pvr_version_is(PVR_POWER9)) {
+		unsigned long hmer = mfspr(SPRN_HMER);
+
+		/* Do we have the debug bit set */
+		if (hmer & PPC_BIT(17)) {
+			hmer &= ~PPC_BIT(17);
+			mtspr(SPRN_HMER, hmer);
+
+			/*
+			 * Now to avoid problems with soft-disable we
+			 * only do the emulation if we are coming from
+			 * user space
+			 */
+			if (user_mode(regs))
+				local_paca->hmi_p9_special_emu = 1;
+
+			/*
+			 * Don't bother going to OPAL if that's the
+			 * only relevant bit.
+			 */
+			if (!(hmer & mfspr(SPRN_HMEER)))
+				return local_paca->hmi_p9_special_emu;
+		}
+	}
+#endif /* CONFIG_PPC_BOOK3S_64 */
+
 	wait_for_subcore_guest_exit();
 
 	if (ppc_md.hmi_exception_early)
···
 
 	wait_for_tb_resync();
 
-	return 0;
+	return 1;
 }

arch/powerpc/kernel/traps.c

···
 #include <linux/kdebug.h>
 #include <linux/ratelimit.h>
 #include <linux/context_tracking.h>
+#include <linux/smp.h>
 
 #include <asm/emulated_ops.h>
 #include <asm/pgtable.h>
···
 	die("System Management Interrupt", regs, SIGABRT);
 }
 
+#ifdef CONFIG_VSX
+static void p9_hmi_special_emu(struct pt_regs *regs)
+{
+	unsigned int ra, rb, t, i, sel, instr, rc;
+	const void __user *addr;
+	u8 vbuf[16], *vdst;
+	unsigned long ea, msr, msr_mask;
+	bool swap;
+
+	if (__get_user_inatomic(instr, (unsigned int __user *)regs->nip))
+		return;
+
+	/*
+	 * lxvb16x	opcode: 0x7c0006d8
+	 * lxvd2x	opcode: 0x7c000698
+	 * lxvh8x	opcode: 0x7c000658
+	 * lxvw4x	opcode: 0x7c000618
+	 */
+	if ((instr & 0xfc00073e) != 0x7c000618) {
+		pr_devel("HMI vec emu: not vector CI %i:%s[%d] nip=%016lx"
+			 " instr=%08x\n",
+			 smp_processor_id(), current->comm, current->pid,
+			 regs->nip, instr);
+		return;
+	}
+
+	/* Grab vector registers into the task struct */
+	msr = regs->msr; /* Grab msr before we flush the bits */
+	flush_vsx_to_thread(current);
+	enable_kernel_altivec();
+
+	/*
+	 * Is userspace running with a different endian (this is rare but
+	 * not impossible)
+	 */
+	swap = (msr & MSR_LE) != (MSR_KERNEL & MSR_LE);
+
+	/* Decode the instruction */
+	ra = (instr >> 16) & 0x1f;
+	rb = (instr >> 11) & 0x1f;
+	t = (instr >> 21) & 0x1f;
+	if (instr & 1)
+		vdst = (u8 *)&current->thread.vr_state.vr[t];
+	else
+		vdst = (u8 *)&current->thread.fp_state.fpr[t][0];
+
+	/* Grab the vector address */
+	ea = regs->gpr[rb] + (ra ? regs->gpr[ra] : 0);
+	if (is_32bit_task())
+		ea &= 0xfffffffful;
+	addr = (__force const void __user *)ea;
+
+	/* Check it */
+	if (!access_ok(VERIFY_READ, addr, 16)) {
+		pr_devel("HMI vec emu: bad access %i:%s[%d] nip=%016lx"
+			 " instr=%08x addr=%016lx\n",
+			 smp_processor_id(), current->comm, current->pid,
+			 regs->nip, instr, (unsigned long)addr);
+		return;
+	}
+
+	/* Read the vector */
+	rc = 0;
+	if ((unsigned long)addr & 0xfUL)
+		/* unaligned case */
+		rc = __copy_from_user_inatomic(vbuf, addr, 16);
+	else
+		__get_user_atomic_128_aligned(vbuf, addr, rc);
+	if (rc) {
+		pr_devel("HMI vec emu: page fault %i:%s[%d] nip=%016lx"
+			 " instr=%08x addr=%016lx\n",
+			 smp_processor_id(), current->comm, current->pid,
+			 regs->nip, instr, (unsigned long)addr);
+		return;
+	}
+
+	pr_devel("HMI vec emu: emulated vector CI %i:%s[%d] nip=%016lx"
+		 " instr=%08x addr=%016lx\n",
+		 smp_processor_id(), current->comm, current->pid, regs->nip,
+		 instr, (unsigned long) addr);
+
+	/* Grab instruction "selector" */
+	sel = (instr >> 6) & 3;
+
+	/*
+	 * Check to make sure the facility is actually enabled. This
+	 * could happen if we get a false positive hit.
+	 *
+	 * lxvd2x/lxvw4x always check MSR VSX sel = 0,2
+	 * lxvh8x/lxvb16x check MSR VSX or VEC depending on VSR used sel = 1,3
+	 */
+	msr_mask = MSR_VSX;
+	if ((sel & 1) && (instr & 1)) /* lxvh8x & lxvb16x + VSR >= 32 */
+		msr_mask = MSR_VEC;
+	if (!(msr & msr_mask)) {
+		pr_devel("HMI vec emu: MSR fac clear %i:%s[%d] nip=%016lx"
+			 " instr=%08x msr:%016lx\n",
+			 smp_processor_id(), current->comm, current->pid,
+			 regs->nip, instr, msr);
+		return;
+	}
+
+	/* Do logging here before we modify sel based on endian */
+	switch (sel) {
+	case 0:	/* lxvw4x */
+		PPC_WARN_EMULATED(lxvw4x, regs);
+		break;
+	case 1: /* lxvh8x */
+		PPC_WARN_EMULATED(lxvh8x, regs);
+		break;
+	case 2: /* lxvd2x */
+		PPC_WARN_EMULATED(lxvd2x, regs);
+		break;
+	case 3: /* lxvb16x */
+		PPC_WARN_EMULATED(lxvb16x, regs);
+		break;
+	}
+
+#ifdef __LITTLE_ENDIAN__
+	/*
+	 * An LE kernel stores the vector in the task struct as an LE
+	 * byte array (effectively swapping both the components and
+	 * the content of the components). Those instructions expect
+	 * the components to remain in ascending address order, so we
+	 * swap them back.
+	 *
+	 * If we are running a BE user space, the expectation is that
+	 * of a simple memcpy, so forcing the emulation to look like
+	 * a lxvb16x should do the trick.
+	 */
+	if (swap)
+		sel = 3;
+
+	switch (sel) {
+	case 0:	/* lxvw4x */
+		for (i = 0; i < 4; i++)
+			((u32 *)vdst)[i] = ((u32 *)vbuf)[3-i];
+		break;
+	case 1: /* lxvh8x */
+		for (i = 0; i < 8; i++)
+			((u16 *)vdst)[i] = ((u16 *)vbuf)[7-i];
+		break;
+	case 2: /* lxvd2x */
+		for (i = 0; i < 2; i++)
+			((u64 *)vdst)[i] = ((u64 *)vbuf)[1-i];
+		break;
+	case 3: /* lxvb16x */
+		for (i = 0; i < 16; i++)
+			vdst[i] = vbuf[15-i];
+		break;
+	}
+#else /* __LITTLE_ENDIAN__ */
+	/* On a big endian kernel, a BE userspace only needs a memcpy */
+	if (!swap)
+		sel = 3;
+
+	/* Otherwise, we need to swap the content of the components */
+	switch (sel) {
+	case 0:	/* lxvw4x */
+		for (i = 0; i < 4; i++)
+			((u32 *)vdst)[i] = cpu_to_le32(((u32 *)vbuf)[i]);
+		break;
+	case 1: /* lxvh8x */
+		for (i = 0; i < 8; i++)
+			((u16 *)vdst)[i] = cpu_to_le16(((u16 *)vbuf)[i]);
+		break;
+	case 2: /* lxvd2x */
+		for (i = 0; i < 2; i++)
+			((u64 *)vdst)[i] = cpu_to_le64(((u64 *)vbuf)[i]);
+		break;
+	case 3: /* lxvb16x */
+		memcpy(vdst, vbuf, 16);
+		break;
+	}
+#endif /* !__LITTLE_ENDIAN__ */
+
+	/* Go to next instruction */
+	regs->nip += 4;
+}
+#endif /* CONFIG_VSX */
+
 void handle_hmi_exception(struct pt_regs *regs)
 {
 	struct pt_regs *old_regs;
 
 	old_regs = set_irq_regs(regs);
 	irq_enter();
+
+#ifdef CONFIG_VSX
+	/* Real mode flagged P9 special emu is needed */
+	if (local_paca->hmi_p9_special_emu) {
+		local_paca->hmi_p9_special_emu = 0;
+
+		/*
+		 * We don't want to take page faults while doing the
+		 * emulation, we just replay the instruction if necessary.
+		 */
+		pagefault_disable();
+		p9_hmi_special_emu(regs);
+		pagefault_enable();
+	}
+#endif /* CONFIG_VSX */
 
 	if (ppc_md.handle_hmi_exception)
 		ppc_md.handle_hmi_exception(regs);
···
 	WARN_EMULATED_SETUP(mfdscr),
 	WARN_EMULATED_SETUP(mtdscr),
 	WARN_EMULATED_SETUP(lq_stq),
+	WARN_EMULATED_SETUP(lxvw4x),
+	WARN_EMULATED_SETUP(lxvh8x),
+	WARN_EMULATED_SETUP(lxvd2x),
+	WARN_EMULATED_SETUP(lxvb16x),
 #endif
 };
 

arch/powerpc/platforms/powernv/smp.c

···
 
 static void pnv_smp_setup_cpu(int cpu)
 {
+	/*
+	 * P9 workaround for CI vector load (see traps.c),
+	 * enable the corresponding HMI interrupt
+	 */
+	if (pvr_version_is(PVR_POWER9))
+		mtspr(SPRN_HMEER, mfspr(SPRN_HMEER) | PPC_BIT(17));
+
 	if (xive_enabled())
 		xive_smp_setup_cpu();
 	else if (cpu != boot_cpuid)