s390: add support for virtually mapped kernel stacks

With virtually mapped kernel stacks the kernel stack overflow detection
is now fault based, every stack has a guard page in the vmalloc space.
The panic_stack is renamed to nodat_stack and is used for all function
that need to run without DAT, e.g. memcpy_real or do_start_kdump.

The main effect is a reduction in the kernel image size as with vmap
stacks the old style overflow checking that adds two instructions per
function is not needed anymore. Result from bloat-o-meter:

add/remove: 20/1 grow/shrink: 13/26854 up/down: 2198/-216240 (-214042)

In regard to performance the micro-benchmark for fork has a hit of a
few microseconds, allocating 4 pages in vmalloc space is more expensive
compare to an order-2 page allocation. But with real workload I could
not find a noticeable difference.

Acked-by: Heiko Carstens <[email protected]>
Signed-off-by: Martin Schwidefsky <[email protected]>

arch/s390/Kconfig

@@ -125,6 +125,7 @@ config S390
 	select HAVE_ARCH_SOFT_DIRTY
 	select HAVE_ARCH_TRACEHOOK
 	select HAVE_ARCH_TRANSPARENT_HUGEPAGE
+	select HAVE_ARCH_VMAP_STACK
 	select HAVE_EBPF_JIT if PACK_STACK && HAVE_MARCH_Z196_FEATURES
 	select HAVE_CMPXCHG_DOUBLE
 	select HAVE_CMPXCHG_LOCAL
@@ -649,6 +650,7 @@ config PACK_STACK
 
 config CHECK_STACK
 	def_bool y
+	depends on !VMAP_STACK
 	prompt "Detect kernel stack overflow"
 	help
 	  This option enables the compiler option -mstack-guard and

arch/s390/include/asm/lowcore.h

@@ -102,9 +102,9 @@ struct lowcore {
 	__u64	current_task;			/* 0x0338 */
 	__u64	kernel_stack;			/* 0x0340 */
 
-	/* Interrupt, panic and restart stack. */
+	/* Interrupt, DAT-off and restartstack. */
 	__u64	async_stack;			/* 0x0348 */
-	__u64	panic_stack;			/* 0x0350 */
+	__u64	nodat_stack;			/* 0x0350 */
 	__u64	restart_stack;			/* 0x0358 */
 
 	/* Restart function and parameter. */

arch/s390/include/asm/processor.h

@@ -162,6 +162,14 @@ struct thread_struct {
 
 typedef struct thread_struct thread_struct;
 
+/*
+ * General size of a stack
+ */
+#define STACK_ORDER 2
+#define STACK_SIZE (PAGE_SIZE << STACK_ORDER)
+#define STACK_INIT_OFFSET \
+	(STACK_SIZE - STACK_FRAME_OVERHEAD - sizeof(struct pt_regs))
+
 /*
  * Stack layout of a C stack frame.
  */

arch/s390/include/asm/thread_info.h

@@ -14,10 +14,7 @@
  * Size of kernel stack for each process
  */
 #define THREAD_SIZE_ORDER 2
-#define ASYNC_ORDER  2
-
 #define THREAD_SIZE (PAGE_SIZE << THREAD_SIZE_ORDER)
-#define ASYNC_SIZE  (PAGE_SIZE << ASYNC_ORDER)
 
 #ifndef __ASSEMBLY__
 #include <asm/lowcore.h>

arch/s390/kernel/asm-offsets.c

@@ -159,7 +159,7 @@ int main(void)
 	OFFSET(__LC_CURRENT, lowcore, current_task);
 	OFFSET(__LC_KERNEL_STACK, lowcore, kernel_stack);
 	OFFSET(__LC_ASYNC_STACK, lowcore, async_stack);
-	OFFSET(__LC_PANIC_STACK, lowcore, panic_stack);
+	OFFSET(__LC_NODAT_STACK, lowcore, nodat_stack);
 	OFFSET(__LC_RESTART_STACK, lowcore, restart_stack);
 	OFFSET(__LC_RESTART_FN, lowcore, restart_fn);
 	OFFSET(__LC_RESTART_DATA, lowcore, restart_data);

arch/s390/kernel/base.S

@@ -18,7 +18,7 @@
 
 ENTRY(s390_base_mcck_handler)
 	basr	%r13,0
-0:	lg	%r15,__LC_PANIC_STACK	# load panic stack
+0:	lg	%r15,__LC_NODAT_STACK	# load panic stack
 	aghi	%r15,-STACK_FRAME_OVERHEAD
 	larl	%r1,s390_base_mcck_handler_fn
 	lg	%r9,0(%r1)

arch/s390/kernel/dumpstack.c

@@ -77,11 +77,11 @@ void dump_trace(dump_trace_func_t func, void *data, struct task_struct *task,
 	frame_size = STACK_FRAME_OVERHEAD + sizeof(struct pt_regs);
 #ifdef CONFIG_CHECK_STACK
 	sp = __dump_trace(func, data, sp,
-			  S390_lowcore.panic_stack + frame_size - PAGE_SIZE,
-			  S390_lowcore.panic_stack + frame_size);
+			  S390_lowcore.nodat_stack + frame_size - STACK_SIZE,
+			  S390_lowcore.nodat_stack + frame_size);
 #endif
 	sp = __dump_trace(func, data, sp,
-			  S390_lowcore.async_stack + frame_size - ASYNC_SIZE,
+			  S390_lowcore.async_stack + frame_size - STACK_SIZE,
 			  S390_lowcore.async_stack + frame_size);
 	task = task ?: current;
 	__dump_trace(func, data, sp,

arch/s390/kernel/entry.S

@@ -85,14 +85,34 @@ _LPP_OFFSET	= __LC_LPP
 #endif
 	.endm
 
-	.macro	CHECK_STACK stacksize,savearea
+	.macro	CHECK_STACK savearea
 #ifdef CONFIG_CHECK_STACK
-	tml	%r15,\stacksize - CONFIG_STACK_GUARD
+	tml	%r15,STACK_SIZE - CONFIG_STACK_GUARD
 	lghi	%r14,\savearea
 	jz	stack_overflow
 #endif
 	.endm
 
+	.macro	CHECK_VMAP_STACK savearea,oklabel
+#ifdef CONFIG_VMAP_STACK
+	lgr	%r14,%r15
+	nill	%r14,0x10000 - STACK_SIZE
+	oill	%r14,STACK_INIT
+	clg	%r14,__LC_KERNEL_STACK
+	je	\oklabel
+	clg	%r14,__LC_ASYNC_STACK
+	je	\oklabel
+	clg	%r14,__LC_NODAT_STACK
+	je	\oklabel
+	clg	%r14,__LC_RESTART_STACK
+	je	\oklabel
+	lghi	%r14,\savearea
+	j	stack_overflow
+#else
+	j	\oklabel
+#endif
+	.endm
+
 	.macro	SWITCH_ASYNC savearea,timer
 	tmhh	%r8,0x0001		# interrupting from user ?
 	jnz	1f
@@ -104,11 +124,11 @@ _LPP_OFFSET	= __LC_LPP
 	brasl	%r14,cleanup_critical
 	tmhh	%r8,0x0001		# retest problem state after cleanup
 	jnz	1f
-0:	lg	%r14,__LC_ASYNC_STACK	# are we already on the async stack?
+0:	lg	%r14,__LC_ASYNC_STACK	# are we already on the target stack?
 	slgr	%r14,%r15
 	srag	%r14,%r14,STACK_SHIFT
 	jnz	2f
-	CHECK_STACK 1<<STACK_SHIFT,\savearea
+	CHECK_STACK \savearea
 	aghi	%r15,-(STACK_FRAME_OVERHEAD + __PT_SIZE)
 	j	3f
 1:	UPDATE_VTIME %r14,%r15,\timer
@@ -600,9 +620,10 @@ ENTRY(pgm_check_handler)
 	jnz	1f			# -> enabled, can't be a double fault
 	tm	__LC_PGM_ILC+3,0x80	# check for per exception
 	jnz	.Lpgm_svcper		# -> single stepped svc
-1:	CHECK_STACK STACK_SIZE,__LC_SAVE_AREA_SYNC
+1:	CHECK_STACK __LC_SAVE_AREA_SYNC
 	aghi	%r15,-(STACK_FRAME_OVERHEAD + __PT_SIZE)
-	j	4f
+	# CHECK_VMAP_STACK branches to stack_overflow or 4f
+	CHECK_VMAP_STACK __LC_SAVE_AREA_SYNC,4f
 2:	UPDATE_VTIME %r14,%r15,__LC_SYNC_ENTER_TIMER
 	BPENTER __TI_flags(%r12),_TIF_ISOLATE_BP
 	lg	%r15,__LC_KERNEL_STACK
@@ -1136,7 +1157,8 @@ ENTRY(mcck_int_handler)
 	jnz	4f
 	TSTMSK	__LC_MCCK_CODE,MCCK_CODE_PSW_IA_VALID
 	jno	.Lmcck_panic
-4:	SWITCH_ASYNC __LC_GPREGS_SAVE_AREA+64,__LC_MCCK_ENTER_TIMER
+4:	ssm	__LC_PGM_NEW_PSW	# turn dat on, keep irqs off
+	SWITCH_ASYNC __LC_GPREGS_SAVE_AREA+64,__LC_MCCK_ENTER_TIMER
 .Lmcck_skip:
 	lghi	%r14,__LC_GPREGS_SAVE_AREA+64
 	stmg	%r0,%r7,__PT_R0(%r11)
@@ -1163,7 +1185,6 @@ ENTRY(mcck_int_handler)
 	xc	__SF_BACKCHAIN(8,%r1),__SF_BACKCHAIN(%r1)
 	la	%r11,STACK_FRAME_OVERHEAD(%r1)
 	lgr	%r15,%r1
-	ssm	__LC_PGM_NEW_PSW	# turn dat on, keep irqs off
 	TSTMSK	__LC_CPU_FLAGS,_CIF_MCCK_PENDING
 	jno	.Lmcck_return
 	TRACE_IRQS_OFF
@@ -1182,7 +1203,7 @@ ENTRY(mcck_int_handler)
 	lpswe	__LC_RETURN_MCCK_PSW
 
 .Lmcck_panic:
-	lg	%r15,__LC_PANIC_STACK
+	lg	%r15,__LC_NODAT_STACK
 	la	%r11,STACK_FRAME_OVERHEAD(%r15)
 	j	.Lmcck_skip
 
@@ -1193,12 +1214,10 @@ ENTRY(restart_int_handler)
 	ALTERNATIVE "", ".insn s,0xb2800000,_LPP_OFFSET", 40
 	stg	%r15,__LC_SAVE_AREA_RESTART
 	lg	%r15,__LC_RESTART_STACK
-	aghi	%r15,-__PT_SIZE			# create pt_regs on stack
-	xc	0(__PT_SIZE,%r15),0(%r15)
-	stmg	%r0,%r14,__PT_R0(%r15)
-	mvc	__PT_R15(8,%r15),__LC_SAVE_AREA_RESTART
-	mvc	__PT_PSW(16,%r15),__LC_RST_OLD_PSW # store restart old psw
-	aghi	%r15,-STACK_FRAME_OVERHEAD	# create stack frame on stack
+	xc	STACK_FRAME_OVERHEAD(__PT_SIZE,%r15),STACK_FRAME_OVERHEAD(%r15)
+	stmg	%r0,%r14,STACK_FRAME_OVERHEAD+__PT_R0(%r15)
+	mvc	STACK_FRAME_OVERHEAD+__PT_R15(8,%r15),__LC_SAVE_AREA_RESTART
+	mvc	STACK_FRAME_OVERHEAD+__PT_PSW(16,%r15),__LC_RST_OLD_PSW
 	xc	0(STACK_FRAME_OVERHEAD,%r15),0(%r15)
 	lg	%r1,__LC_RESTART_FN		# load fn, parm & source cpu
 	lg	%r2,__LC_RESTART_DATA
@@ -1216,14 +1235,14 @@ ENTRY(restart_int_handler)
 
 	.section .kprobes.text, "ax"
 
-#ifdef CONFIG_CHECK_STACK
+#if defined(CONFIG_CHECK_STACK) || defined(CONFIG_VMAP_STACK)
 /*
  * The synchronous or the asynchronous stack overflowed. We are dead.
  * No need to properly save the registers, we are going to panic anyway.
  * Setup a pt_regs so that show_trace can provide a good call trace.
  */
 stack_overflow:
-	lg	%r15,__LC_PANIC_STACK	# change to panic stack
+	lg	%r15,__LC_NODAT_STACK	# change to panic stack
 	la	%r11,STACK_FRAME_OVERHEAD(%r15)
 	stmg	%r0,%r7,__PT_R0(%r11)
 	stmg	%r8,%r9,__PT_PSW(%r11)

arch/s390/kernel/entry.h

@@ -86,4 +86,7 @@ DECLARE_PER_CPU(u64, mt_cycles[8]);
 void gs_load_bc_cb(struct pt_regs *regs);
 void set_fs_fixup(void);
 
+unsigned long stack_alloc(void);
+void stack_free(unsigned long stack);
+
 #endif /* _ENTRY_H */

arch/s390/kernel/head64.S

@@ -36,9 +36,7 @@ ENTRY(startup_continue)
 #
 	larl	%r14,init_task
 	stg	%r14,__LC_CURRENT
-	larl	%r15,init_thread_union+THREAD_SIZE
-	stg	%r15,__LC_KERNEL_STACK	# set end of kernel stack
-	aghi	%r15,-STACK_FRAME_OVERHEAD
+	larl	%r15,init_thread_union+THREAD_SIZE-STACK_FRAME_OVERHEAD
 #
 # Early setup functions that may not rely on an initialized bss section,
 # like moving the initrd. Returns with an initialized bss section.

arch/s390/kernel/irq.c

@@ -171,7 +171,7 @@ void do_softirq_own_stack(void)
 	old = current_stack_pointer();
 	/* Check against async. stack address range. */
 	new = S390_lowcore.async_stack;
-	if (((new - old) >> (PAGE_SHIFT + THREAD_SIZE_ORDER)) != 0) {
+	if (((new - old) >> (PAGE_SHIFT + STACK_ORDER)) != 0) {
 		CALL_ON_STACK(__do_softirq, new, 0);
 	} else {
 		/* We are already on the async stack. */

arch/s390/kernel/machine_kexec.c

@@ -142,18 +142,27 @@ static noinline void __machine_kdump(void *image)
 }
 #endif
 
+static unsigned long do_start_kdump(unsigned long addr)
+{
+	struct kimage *image = (struct kimage *) addr;
+	int (*start_kdump)(int) = (void *)image->start;
+	int rc;
+
+	__arch_local_irq_stnsm(0xfb); /* disable DAT */
+	rc = start_kdump(0);
+	__arch_local_irq_stosm(0x04); /* enable DAT */
+	return rc;
+}
+
 /*
  * Check if kdump checksums are valid: We call purgatory with parameter "0"
  */
 static bool kdump_csum_valid(struct kimage *image)
 {
 #ifdef CONFIG_CRASH_DUMP
-	int (*start_kdump)(int) = (void *)image->start;
 	int rc;
 
-	__arch_local_irq_stnsm(0xfb); /* disable DAT */
-	rc = start_kdump(0);
-	__arch_local_irq_stosm(0x04); /* enable DAT */
+	rc = CALL_ON_STACK(do_start_kdump, S390_lowcore.nodat_stack, 1, image);
 	return rc == 0;
 #else
 	return false;

arch/s390/kernel/setup.c

@@ -49,6 +49,7 @@
 #include <linux/crash_dump.h>
 #include <linux/memory.h>
 #include <linux/compat.h>
+#include <linux/start_kernel.h>
 
 #include <asm/ipl.h>
 #include <asm/facility.h>
@@ -303,6 +304,78 @@ early_param("vmalloc", parse_vmalloc);
 
 void *restart_stack __section(.data);
 
+unsigned long stack_alloc(void)
+{
+#ifdef CONFIG_VMAP_STACK
+	return (unsigned long)
+		__vmalloc_node_range(STACK_SIZE, STACK_SIZE,
+				     VMALLOC_START, VMALLOC_END,
+				     THREADINFO_GFP,
+				     PAGE_KERNEL, 0, NUMA_NO_NODE,
+				     __builtin_return_address(0));
+#else
+	return __get_free_pages(GFP_KERNEL, STACK_ORDER);
+#endif
+}
+
+void stack_free(unsigned long stack)
+{
+#ifdef CONFIG_VMAP_STACK
+	vfree((void *) stack);
+#else
+	free_pages(stack, STACK_ORDER);
+#endif
+}
+
+int __init arch_early_irq_init(void)
+{
+	unsigned long stack;
+
+	stack = __get_free_pages(GFP_KERNEL, STACK_ORDER);
+	if (!stack)
+		panic("Couldn't allocate async stack");
+	S390_lowcore.async_stack = stack + STACK_INIT_OFFSET;
+	return 0;
+}
+
+static int __init async_stack_realloc(void)
+{
+	unsigned long old, new;
+
+	old = S390_lowcore.async_stack - STACK_INIT_OFFSET;
+	new = stack_alloc();
+	if (!new)
+		panic("Couldn't allocate async stack");
+	S390_lowcore.async_stack = new + STACK_INIT_OFFSET;
+	free_pages(old, STACK_ORDER);
+	return 0;
+}
+early_initcall(async_stack_realloc);
+
+void __init arch_call_rest_init(void)
+{
+	struct stack_frame *frame;
+	unsigned long stack;
+
+	stack = stack_alloc();
+	if (!stack)
+		panic("Couldn't allocate kernel stack");
+	current->stack = (void *) stack;
+#ifdef CONFIG_VMAP_STACK
+	current->stack_vm_area = (void *) stack;
+#endif
+	set_task_stack_end_magic(current);
+	stack += STACK_INIT_OFFSET;
+	S390_lowcore.kernel_stack = stack;
+	frame = (struct stack_frame *) stack;
+	memset(frame, 0, sizeof(*frame));
+	/* Branch to rest_init on the new stack, never returns */
+	asm volatile(
+		"	la	15,0(%[_frame])\n"
+		"	jg	rest_init\n"
+		: : [_frame] "a" (frame));
+}
+
 static void __init setup_lowcore(void)
 {
 	struct lowcore *lc;
@@ -329,14 +402,8 @@ static void __init setup_lowcore(void)
 		PSW_MASK_DAT | PSW_MASK_MCHECK;
 	lc->io_new_psw.addr = (unsigned long) io_int_handler;
 	lc->clock_comparator = clock_comparator_max;
-	lc->kernel_stack = ((unsigned long) &init_thread_union)
+	lc->nodat_stack = ((unsigned long) &init_thread_union)
 		+ THREAD_SIZE - STACK_FRAME_OVERHEAD - sizeof(struct pt_regs);
-	lc->async_stack = (unsigned long)
-		memblock_virt_alloc(ASYNC_SIZE, ASYNC_SIZE)
-		+ ASYNC_SIZE - STACK_FRAME_OVERHEAD - sizeof(struct pt_regs);
-	lc->panic_stack = (unsigned long)
-		memblock_virt_alloc(PAGE_SIZE, PAGE_SIZE)
-		+ PAGE_SIZE - STACK_FRAME_OVERHEAD - sizeof(struct pt_regs);
 	lc->current_task = (unsigned long)&init_task;
 	lc->lpp = LPP_MAGIC;
 	lc->machine_flags = S390_lowcore.machine_flags;
@@ -357,8 +424,12 @@ static void __init setup_lowcore(void)
 	lc->last_update_timer = S390_lowcore.last_update_timer;
 	lc->last_update_clock = S390_lowcore.last_update_clock;
 
-	restart_stack = memblock_virt_alloc(ASYNC_SIZE, ASYNC_SIZE);
-	restart_stack += ASYNC_SIZE;
+	/*
+	 * Allocate the global restart stack which is the same for
+	 * all CPUs in cast *one* of them does a PSW restart.
+	 */
+	restart_stack = memblock_virt_alloc(STACK_SIZE, STACK_SIZE);
+	restart_stack += STACK_INIT_OFFSET;
 
 	/*
 	 * Set up PSW restart to call ipl.c:do_restart(). Copy the relevant