um: make stub_exe _start() pure inline asm

Since __attribute__((naked)) cannot be used with functions
containing C statements, just generate the few instructions
it needs in assembly directly.

While at it, fix the stack usage ("1 + 2*x - 1" is odd) and
document what it must do, and why it must adjust the stack.

Fixes: 8508a5e0e9 ("um: Fix misaligned stack in stub_exe")
Link: https://lore.kernel.org/linux-um/CABVgOSntH-uoOFMP5HwMXjx_f1osMnVdhgKRKm4uz6DFm2Lb8Q@mail.gmail.com/
Reviewed-by: David Gow <davidgow@google.com>
Signed-off-by: Johannes Berg <johannes.berg@intel.com>

arch/um/kernel/skas/stub_exe.c

@@ -81,11 +81,15 @@ noinline static void real_init(void)
 
 __attribute__((naked)) void _start(void)
 {
-	char *alloc;
-
-	/* Make enough space for the stub (including space for alignment) */
-	alloc = __builtin_alloca((1 + 2 * STUB_DATA_PAGES - 1) * UM_KERN_PAGE_SIZE);
-	asm volatile("" : "+r,m"(alloc) : : "memory");
-
-	real_init();
+	/*
+	 * Since the stack after exec() starts at the top-most address,
+	 * but that's exactly where we also want to map the stub data
+	 * and code, this must:
+	 *  - push the stack by 1 code and STUB_DATA_PAGES data pages
+	 *  - call real_init()
+	 * This way, real_init() can use the stack normally, while the
+	 * original stack further down (higher address) will become
+	 * inaccessible after the mmap() calls above.
+	 */
+	stub_start(real_init);
 }

arch/x86/um/shared/sysdep/stub_32.h

@@ -123,4 +123,12 @@ static __always_inline void *get_stub_data(void)
 
 	return (void *)ret;
 }
+
+#define stub_start(fn)							\
+	asm volatile (							\
+		"subl %0,%%esp ;"					\
+		"movl %1, %%eax ; "					\
+		"call *%%eax ;"						\
+		:: "i" ((1 + STUB_DATA_PAGES) * UM_KERN_PAGE_SIZE),	\
+		   "i" (&fn))
 #endif

arch/x86/um/shared/sysdep/stub_64.h

@@ -126,4 +126,12 @@ static __always_inline void *get_stub_data(void)
 
 	return (void *)ret;
 }
+
+#define stub_start(fn)							\
+	asm volatile (							\
+		"subq %0,%%rsp ;"					\
+		"movq %1,%%rax ;"					\
+		"call *%%rax ;"						\
+		:: "i" ((1 + STUB_DATA_PAGES) * UM_KERN_PAGE_SIZE),	\
+		   "i" (&fn))
 #endif