x86/xen: Add xenpv_restore_regs_and_return_to_usermode()

In the native case, PER_CPU_VAR(cpu_tss_rw + TSS_sp0) is the
trampoline stack. But XEN pv doesn't use trampoline stack, so
PER_CPU_VAR(cpu_tss_rw + TSS_sp0) is also the kernel stack.

In that case, source and destination stacks are identical, which means
that reusing swapgs_restore_regs_and_return_to_usermode() in XEN pv
would cause %rsp to move up to the top of the kernel stack and leave the
IRET frame below %rsp.

This is dangerous as it can be corrupted if #NMI / #MC hit as either of
these events occurring in the middle of the stack pushing would clobber
data on the (original) stack.

And, with  XEN pv, swapgs_restore_regs_and_return_to_usermode() pushing
the IRET frame on to the original address is useless and error-prone
when there is any future attempt to modify the code.

 [ bp: Massage commit message. ]

Fixes: 7f2590a110b8 ("x86/entry/64: Use a per-CPU trampoline stack for IDT entries")
Signed-off-by: Lai Jiangshan <laijs@linux.alibaba.com>
Signed-off-by: Borislav Petkov <bp@suse.de>
Reviewed-by: Boris Ostrovsky <boris.ostrovsky@oracle.com>
Link: https://lkml.kernel.org/r/20211126101209.8613-4-jiangshanlai@gmail.com
diff --git a/arch/x86/entry/entry_64.S b/arch/x86/entry/entry_64.S
index f9e1c06..97b1f84 100644
--- a/arch/x86/entry/entry_64.S
+++ b/arch/x86/entry/entry_64.S
@@ -574,6 +574,10 @@
 	ud2
 1:
 #endif
+#ifdef CONFIG_XEN_PV
+	ALTERNATIVE "", "jmp xenpv_restore_regs_and_return_to_usermode", X86_FEATURE_XENPV
+#endif
+
 	POP_REGS pop_rdi=0
 
 	/*
diff --git a/arch/x86/xen/xen-asm.S b/arch/x86/xen/xen-asm.S
index 220dd96..444d824 100644
--- a/arch/x86/xen/xen-asm.S
+++ b/arch/x86/xen/xen-asm.S
@@ -20,6 +20,7 @@
 
 #include <linux/init.h>
 #include <linux/linkage.h>
+#include <../entry/calling.h>
 
 .pushsection .noinstr.text, "ax"
 /*
@@ -193,6 +194,25 @@
 SYM_CODE_END(xen_iret)
 
 /*
+ * XEN pv doesn't use trampoline stack, PER_CPU_VAR(cpu_tss_rw + TSS_sp0) is
+ * also the kernel stack.  Reusing swapgs_restore_regs_and_return_to_usermode()
+ * in XEN pv would cause %rsp to move up to the top of the kernel stack and
+ * leave the IRET frame below %rsp, which is dangerous to be corrupted if #NMI
+ * interrupts. And swapgs_restore_regs_and_return_to_usermode() pushing the IRET
+ * frame at the same address is useless.
+ */
+SYM_CODE_START(xenpv_restore_regs_and_return_to_usermode)
+	UNWIND_HINT_REGS
+	POP_REGS
+
+	/* stackleak_erase() can work safely on the kernel stack. */
+	STACKLEAK_ERASE_NOCLOBBER
+
+	addq	$8, %rsp	/* skip regs->orig_ax */
+	jmp xen_iret
+SYM_CODE_END(xenpv_restore_regs_and_return_to_usermode)
+
+/*
  * Xen handles syscall callbacks much like ordinary exceptions, which
  * means we have:
  * - kernel gs