x86/doublefault/32: Rewrite the x86_32 #DF handler and unify with 64-bit

The old x86_32 doublefault_fn() was old and crufty, and it did not
even try to recover.  do_double_fault() is much nicer.  Rewrite the
32-bit double fault code to sanitize CPU state and call
do_double_fault().  This is mostly an exercise i386 archaeology.

With this patch applied, 32-bit double faults get a real stack trace,
just like 64-bit double faults.

[ mingo: merged the patch to a later kernel base. ]

Signed-off-by: Andy Lutomirski <luto@kernel.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>

arch/x86/kernel/traps.c

@@ -306,8 +306,23 @@ __visible void __noreturn handle_stack_overflow(const char *message,
 }
 #endif
 
-#ifdef CONFIG_X86_64
-/* Runs on IST stack */
+#if defined(CONFIG_X86_64) || defined(CONFIG_DOUBLEFAULT)
+/*
+ * Runs on an IST stack for x86_64 and on a special task stack for x86_32.
+ *
+ * On x86_64, this is more or less a normal kernel entry.  Notwithstanding the
+ * SDM's warnings about double faults being unrecoverable, returning works as
+ * expected.  Presumably what the SDM actually means is that the CPU may get
+ * the register state wrong on entry, so returning could be a bad idea.
+ *
+ * Various CPU engineers have promised that double faults due to an IRET fault
+ * while the stack is read-only are, in fact, recoverable.
+ *
+ * On x86_32, this is entered through a task gate, and regs are synthesized
+ * from the TSS.  Returning is, in principle, okay, but changes to regs will
+ * be lost.  If, for some reason, we need to return to a context with modified
+ * regs, the shim code could be adjusted to synchronize the registers.
+ */
 dotraplinkage void do_double_fault(struct pt_regs *regs, long error_code, unsigned long cr2)
 {
 	static const char str[] = "double fault";