x86/smpboot: Remove initial_stack on 64-bit

In order to facilitate parallel startup, start to eliminate some of the
global variables passing information to CPUs in the startup path.

However, start by introducing one more: smpboot_control. For now this
merely holds the CPU# of the CPU which is coming up. Each CPU can then
find its own per-cpu data, and everything else it needs can be found
from there, allowing the other global variables to be removed.

First to be removed is initial_stack. Each CPU can load %rsp from its
current_task->thread.sp instead. That is already set up with the correct
idle thread for APs. Set up the .sp field in INIT_THREAD on x86 so that
the BSP also finds a suitable stack pointer in the static per-cpu data
when coming up on first boot.

On resume from S3, the CPU needs a temporary stack because its idle task
is already active. Instead of setting initial_stack, the sleep code can
simply set its own current->thread.sp to point to the temporary stack.
Nobody else cares about ->thread.sp for a thread which is currently on
a CPU, because the true value is actually in the %rsp register. Which
is restored with the rest of the CPU context in do_suspend_lowlevel().

Signed-off-by: Brian Gerst <[email protected]>
Signed-off-by: David Woodhouse <[email protected]>
Signed-off-by: Usama Arif <[email protected]>
Signed-off-by: Thomas Gleixner <[email protected]>
Tested-by: Usama Arif <[email protected]>
Tested-by: Guilherme G. Piccoli <[email protected]>
Reviewed-by: David Woodhouse <[email protected]>
Link: https://lore.kernel.org/r/[email protected]

Author: Brian Gerst

arch/x86/include/asm/processor.h

@@ -647,7 +647,11 @@ static inline void spin_lock_prefetch(const void *x)
 #define KSTK_ESP(task)		(task_pt_regs(task)->sp)
 
 #else
-#define INIT_THREAD { }
+extern unsigned long __end_init_task[];
+
+#define INIT_THREAD {							    \
+	.sp	= (unsigned long)&__end_init_task - sizeof(struct pt_regs), \
+}
 
 extern unsigned long KSTK_ESP(struct task_struct *task);
 

arch/x86/include/asm/smp.h

@@ -199,5 +199,8 @@ extern void nmi_selftest(void);
 #define nmi_selftest() do { } while (0)
 #endif
 
-#endif /* __ASSEMBLY__ */
+extern unsigned int smpboot_control;
+
+#endif /* !__ASSEMBLY__ */
+
 #endif /* _ASM_X86_SMP_H */

arch/x86/kernel/acpi/sleep.c

@@ -111,13 +111,29 @@ int x86_acpi_suspend_lowlevel(void)
 	saved_magic = 0x12345678;
 #else /* CONFIG_64BIT */
 #ifdef CONFIG_SMP
-	initial_stack = (unsigned long)temp_stack + sizeof(temp_stack);
+	/*
+	 * As each CPU starts up, it will find its own stack pointer
+	 * from its current_task->thread.sp. Typically that will be
+	 * the idle thread for a newly-started AP, or even the boot
+	 * CPU which will find it set to &init_task in the static
+	 * per-cpu data.
+	 *
+	 * Make the resuming CPU use the temporary stack at startup
+	 * by setting current->thread.sp to point to that. The true
+	 * %rsp will be restored with the rest of the CPU context,
+	 * by do_suspend_lowlevel(). And unwinders don't care about
+	 * the abuse of ->thread.sp because it's a dead variable
+	 * while the thread is running on the CPU anyway; the true
+	 * value is in the actual %rsp register.
+	 */
+	current->thread.sp = (unsigned long)temp_stack + sizeof(temp_stack);
 	early_gdt_descr.address =
 			(unsigned long)get_cpu_gdt_rw(smp_processor_id());
 	initial_gs = per_cpu_offset(smp_processor_id());
+	smpboot_control = smp_processor_id();
 #endif
 	initial_code = (unsigned long)wakeup_long64;
-       saved_magic = 0x123456789abcdef0L;
+	saved_magic = 0x123456789abcdef0L;
 #endif /* CONFIG_64BIT */
 
 	/*

arch/x86/kernel/asm-offsets.c

@@ -115,6 +115,7 @@ static void __used common(void)
 	OFFSET(TSS_sp1, tss_struct, x86_tss.sp1);
 	OFFSET(TSS_sp2, tss_struct, x86_tss.sp2);
 	OFFSET(X86_top_of_stack, pcpu_hot, top_of_stack);
+	OFFSET(X86_current_task, pcpu_hot, current_task);
 #ifdef CONFIG_CALL_DEPTH_TRACKING
 	OFFSET(X86_call_depth, pcpu_hot, call_depth);
 #endif

arch/x86/kernel/head_64.S

@@ -61,8 +61,8 @@ SYM_CODE_START_NOALIGN(startup_64)
 	 * tables and then reload them.
 	 */
 
-	/* Set up the stack for verify_cpu(), similar to initial_stack below */
-	leaq	(__end_init_task - FRAME_SIZE)(%rip), %rsp
+	/* Set up the stack for verify_cpu() */
+	leaq	(__end_init_task - PTREGS_SIZE)(%rip), %rsp
 
 	leaq	_text(%rip), %rdi
 
@@ -241,6 +241,24 @@ SYM_INNER_LABEL(secondary_startup_64_no_verify, SYM_L_GLOBAL)
 	UNWIND_HINT_EMPTY
 	ANNOTATE_NOENDBR // above
 
+#ifdef CONFIG_SMP
+	movl	smpboot_control(%rip), %ecx
+
+	/* Get the per cpu offset for the given CPU# which is in ECX */
+	movq	__per_cpu_offset(,%rcx,8), %rdx
+#else
+	xorl	%edx, %edx /* zero-extended to clear all of RDX */
+#endif /* CONFIG_SMP */
+
+	/*
+	 * Setup a boot time stack - Any secondary CPU will have lost its stack
+	 * by now because the cr3-switch above unmaps the real-mode stack.
+	 *
+	 * RDX contains the per-cpu offset
+	 */
+	movq	pcpu_hot + X86_current_task(%rdx), %rax
+	movq	TASK_threadsp(%rax), %rsp
+
 	/*
 	 * We must switch to a new descriptor in kernel space for the GDT
 	 * because soon the kernel won't have access anymore to the userspace
@@ -275,12 +293,6 @@ SYM_INNER_LABEL(secondary_startup_64_no_verify, SYM_L_GLOBAL)
 	movl	initial_gs+4(%rip),%edx
 	wrmsr
 
-	/*
-	 * Setup a boot time stack - Any secondary CPU will have lost its stack
-	 * by now because the cr3-switch above unmaps the real-mode stack
-	 */
-	movq initial_stack(%rip), %rsp
-
 	/* Setup and Load IDT */
 	pushq	%rsi
 	call	early_setup_idt
@@ -372,7 +384,11 @@ SYM_CODE_END(secondary_startup_64)
 SYM_CODE_START(start_cpu0)
 	ANNOTATE_NOENDBR
 	UNWIND_HINT_EMPTY
-	movq	initial_stack(%rip), %rsp
+
+	/* Find the idle task stack */
+	movq	PER_CPU_VAR(pcpu_hot) + X86_current_task, %rcx
+	movq	TASK_threadsp(%rcx), %rsp
+
 	jmp	.Ljump_to_C_code
 SYM_CODE_END(start_cpu0)
 #endif
@@ -420,12 +436,6 @@ SYM_DATA(initial_gs,	.quad INIT_PER_CPU_VAR(fixed_percpu_data))
 #ifdef CONFIG_AMD_MEM_ENCRYPT
 SYM_DATA(initial_vc_handler,	.quad handle_vc_boot_ghcb)
 #endif
-
-/*
- * The FRAME_SIZE gap is a convention which helps the in-kernel unwinder
- * reliably detect the end of the stack.
- */
-SYM_DATA(initial_stack, .quad init_thread_union + THREAD_SIZE - FRAME_SIZE)
 	__FINITDATA
 
 	__INIT
@@ -660,6 +670,9 @@ SYM_DATA_END(level1_fixmap_pgt)
 SYM_DATA(early_gdt_descr,		.word GDT_ENTRIES*8-1)
 SYM_DATA_LOCAL(early_gdt_descr_base,	.quad INIT_PER_CPU_VAR(gdt_page))
 
+	.align 16
+SYM_DATA(smpboot_control,		.long 0)
+
 	.align 16
 /* This must match the first entry in level2_kernel_pgt */
 SYM_DATA(phys_base, .quad 0x0)

arch/x86/kernel/smpboot.c

@@ -1088,7 +1088,12 @@ static int do_boot_cpu(int apicid, int cpu, struct task_struct *idle,
 	idle->thread.sp = (unsigned long)task_pt_regs(idle);
 	early_gdt_descr.address = (unsigned long)get_cpu_gdt_rw(cpu);
 	initial_code = (unsigned long)start_secondary;
-	initial_stack  = idle->thread.sp;
+
+	if (IS_ENABLED(CONFIG_X86_32)) {
+		initial_stack  = idle->thread.sp;
+	} else {
+		smpboot_control = cpu;
+	}
 
 	/* Enable the espfix hack for this CPU */
 	init_espfix_ap(cpu);

arch/x86/xen/xen-head.S

@@ -49,7 +49,7 @@ SYM_CODE_START(startup_xen)
 	ANNOTATE_NOENDBR
 	cld
 
-	mov initial_stack(%rip), %rsp
+	leaq	(__end_init_task - PTREGS_SIZE)(%rip), %rsp
 
 	/* Set up %gs.
 	 *