ARM: implement support for vmap'ed stacks

Wire up the generic support for managing task stack allocations via vmalloc,
and implement the entry code that detects whether we faulted because of a
stack overrun (or future stack overrun caused by pushing the pt_regs array)

While this adds a fair amount of tricky entry asm code, it should be
noted that it only adds a TST + branch to the svc_entry path. The code
implementing the non-trivial handling of the overflow stack is emitted
out-of-line into the .text section.

Since on ARM, we rely on do_translation_fault() to keep PMD level page
table entries that cover the vmalloc region up to date, we need to
ensure that we don't hit such a stale PMD entry when accessing the
stack. So we do a dummy read from the new stack while still running from
the old one on the context switch path, and bump the vmalloc_seq counter
when PMD level entries in the vmalloc range are modified, so that the MM
switch fetches the latest version of the entries.

Note that we need to increase the per-mode stack by 1 word, to gain some
space to stash a GPR until we know it is safe to touch the stack.
However, due to the cacheline alignment of the struct, this does not
actually increase the memory footprint of the struct stack array at all.

Signed-off-by: Ard Biesheuvel <ardb@kernel.org>
Tested-by: Keith Packard <keithpac@amazon.com>
Tested-by: Marc Zyngier <maz@kernel.org>
Tested-by: Vladimir Murzin <vladimir.murzin@arm.com> # ARMv7M

arch/arm/Kconfig

@@ -127,6 +127,7 @@ config ARM
 	select RTC_LIB
 	select SYS_SUPPORTS_APM_EMULATION
 	select THREAD_INFO_IN_TASK if CURRENT_POINTER_IN_TPIDRURO
+	select HAVE_ARCH_VMAP_STACK if MMU && THREAD_INFO_IN_TASK && (!LD_IS_LLD || LLD_VERSION >= 140000)
 	select TRACE_IRQFLAGS_SUPPORT if !CPU_V7M
 	# Above selects are sorted alphabetically; please add new ones
 	# according to that.  Thanks.

arch/arm/include/asm/page.h

@@ -149,6 +149,10 @@ extern void copy_page(void *to, const void *from);
 #include <asm/pgtable-2level-types.h>
 #endif
 
+#ifdef CONFIG_VMAP_STACK
+#define ARCH_PAGE_TABLE_SYNC_MASK	PGTBL_PMD_MODIFIED
+#endif
+
 #endif /* CONFIG_MMU */
 
 typedef struct page *pgtable_t;

arch/arm/include/asm/thread_info.h

@@ -25,6 +25,14 @@
 #define THREAD_SIZE		(PAGE_SIZE << THREAD_SIZE_ORDER)
 #define THREAD_START_SP		(THREAD_SIZE - 8)
 
+#ifdef CONFIG_VMAP_STACK
+#define THREAD_ALIGN		(2 * THREAD_SIZE)
+#else
+#define THREAD_ALIGN		THREAD_SIZE
+#endif
+
+#define OVERFLOW_STACK_SIZE	SZ_4K
+
 #ifndef __ASSEMBLY__
 
 struct task_struct;

arch/arm/kernel/entry-armv.S

@@ -57,6 +57,10 @@ UNWIND(	.setfp	fpreg, sp		)
 	@
 	subs	r2, sp, r0		@ SP above bottom of IRQ stack?
 	rsbscs	r2, r2, #THREAD_SIZE	@ ... and below the top?
+#ifdef CONFIG_VMAP_STACK
+	ldr_l	r2, high_memory, cc	@ End of the linear region
+	cmpcc	r2, r0			@ Stack pointer was below it?
+#endif
 	movcs	sp, r0			@ If so, revert to incoming SP
 
 #ifndef CONFIG_UNWINDER_ARM
@@ -188,13 +192,18 @@ ENDPROC(__und_invalid)
 #define SPFIX(code...)
 #endif
 
-	.macro	svc_entry, stack_hole=0, trace=1, uaccess=1
+	.macro	svc_entry, stack_hole=0, trace=1, uaccess=1, overflow_check=1
  UNWIND(.fnstart		)
- UNWIND(.save {r0 - pc}		)
 	sub	sp, sp, #(SVC_REGS_SIZE + \stack_hole)
+ THUMB(	add	sp, r1		)	@ get SP in a GPR without
+ THUMB(	sub	r1, sp, r1	)	@ using a temp register
+
+	.if	\overflow_check
+ UNWIND(.save	{r0 - pc}	)
+	do_overflow_check (SVC_REGS_SIZE + \stack_hole)
+	.endif
+
 #ifdef CONFIG_THUMB2_KERNEL
-	add	sp, r1			@ get SP in a GPR without
-	sub	r1, sp, r1		@ using a temp register
 	tst	r1, #4			@ test stack pointer alignment
 	sub	r1, sp, r1		@ restore original R1
 	sub	sp, r1			@ restore original SP
@@ -827,12 +836,20 @@ ENTRY(__switch_to)
 	str	r7, [r8]
 #endif
 	mov	r0, r5
-#if !defined(CONFIG_THUMB2_KERNEL)
+#if !defined(CONFIG_THUMB2_KERNEL) && !defined(CONFIG_VMAP_STACK)
 	set_current r7
 	ldmia	r4, {r4 - sl, fp, sp, pc}	@ Load all regs saved previously
 #else
 	mov	r1, r7
 	ldmia	r4, {r4 - sl, fp, ip, lr}	@ Load all regs saved previously
+#ifdef CONFIG_VMAP_STACK
+	@
+	@ Do a dummy read from the new stack while running from the old one so
+	@ that we can rely on do_translation_fault() to fix up any stale PMD
+	@ entries covering the vmalloc region.
+	@
+	ldr	r2, [ip]
+#endif
 
 	@ When CONFIG_THREAD_INFO_IN_TASK=n, the update of SP itself is what
 	@ effectuates the task switch, as that is what causes the observable
@@ -849,6 +866,76 @@ ENTRY(__switch_to)
  UNWIND(.fnend		)
 ENDPROC(__switch_to)
 
+#ifdef CONFIG_VMAP_STACK
+	.text
+	.align	2
+__bad_stack:
+	@
+	@ We've just detected an overflow. We need to load the address of this
+	@ CPU's overflow stack into the stack pointer register. We have only one
+	@ scratch register so let's use a sequence of ADDs including one
+	@ involving the PC, and decorate them with PC-relative group
+	@ relocations. As these are ARM only, switch to ARM mode first.
+	@
+	@ We enter here with IP clobbered and its value stashed on the mode
+	@ stack.
+	@
+THUMB(	bx	pc		)
+THUMB(	nop			)
+THUMB(	.arm			)
+	mrc	p15, 0, ip, c13, c0, 4		@ Get per-CPU offset
+
+	.globl	overflow_stack_ptr
+	.reloc	0f, R_ARM_ALU_PC_G0_NC, overflow_stack_ptr
+	.reloc	1f, R_ARM_ALU_PC_G1_NC, overflow_stack_ptr
+	.reloc	2f, R_ARM_LDR_PC_G2, overflow_stack_ptr
+	add	ip, ip, pc
+0:	add	ip, ip, #-4
+1:	add	ip, ip, #0
+2:	ldr	ip, [ip, #4]
+
+	str	sp, [ip, #-4]!			@ Preserve original SP value
+	mov	sp, ip				@ Switch to overflow stack
+	pop	{ip}				@ Original SP in IP
+
+#if defined(CONFIG_UNWINDER_FRAME_POINTER) && defined(CONFIG_CC_IS_GCC)
+	mov	ip, ip				@ mov expected by unwinder
+	push	{fp, ip, lr, pc}		@ GCC flavor frame record
+#else
+	str	ip, [sp, #-8]!			@ store original SP
+	push	{fpreg, lr}			@ Clang flavor frame record
+#endif
+UNWIND( ldr	ip, [r0, #4]	)		@ load exception LR
+UNWIND( str	ip, [sp, #12]	)		@ store in the frame record
+	ldr	ip, [r0, #12]			@ reload IP
+
+	@ Store the original GPRs to the new stack.
+	svc_entry uaccess=0, overflow_check=0
+
+UNWIND( .save   {sp, pc}	)
+UNWIND( .save   {fpreg, lr}	)
+UNWIND( .setfp  fpreg, sp	)
+
+	ldr	fpreg, [sp, #S_SP]		@ Add our frame record
+						@ to the linked list
+#if defined(CONFIG_UNWINDER_FRAME_POINTER) && defined(CONFIG_CC_IS_GCC)
+	ldr	r1, [fp, #4]			@ reload SP at entry
+	add	fp, fp, #12
+#else
+	ldr	r1, [fpreg, #8]
+#endif
+	str	r1, [sp, #S_SP]			@ store in pt_regs
+
+	@ Stash the regs for handle_bad_stack
+	mov	r0, sp
+
+	@ Time to die
+	bl	handle_bad_stack
+	nop
+UNWIND( .fnend			)
+ENDPROC(__bad_stack)
+#endif
+
 	__INIT
 
 /*

arch/arm/kernel/entry-header.S

@@ -423,3 +423,40 @@ scno	.req	r7		@ syscall number
 tbl	.req	r8		@ syscall table pointer
 why	.req	r8		@ Linux syscall (!= 0)
 tsk	.req	r9		@ current thread_info
+
+	.macro	do_overflow_check, frame_size:req
+#ifdef CONFIG_VMAP_STACK
+	@
+	@ Test whether the SP has overflowed. Task and IRQ stacks are aligned
+	@ so that SP & BIT(THREAD_SIZE_ORDER + PAGE_SHIFT) should always be
+	@ zero.
+	@
+ARM(	tst	sp, #1 << (THREAD_SIZE_ORDER + PAGE_SHIFT)	)
+THUMB(	tst	r1, #1 << (THREAD_SIZE_ORDER + PAGE_SHIFT)	)
+THUMB(	it	ne						)
+	bne	.Lstack_overflow_check\@
+
+	.pushsection	.text
+.Lstack_overflow_check\@:
+	@
+	@ The stack pointer is not pointing to a valid vmap'ed stack, but it
+	@ may be pointing into the linear map instead, which may happen if we
+	@ are already running from the overflow stack. We cannot detect overflow
+	@ in such cases so just carry on.
+	@
+	str	ip, [r0, #12]			@ Stash IP on the mode stack
+	ldr_l	ip, high_memory			@ Start of VMALLOC space
+ARM(	cmp	sp, ip			)	@ SP in vmalloc space?
+THUMB(	cmp	r1, ip			)
+THUMB(	itt	lo			)
+	ldrlo	ip, [r0, #12]			@ Restore IP
+	blo	.Lout\@				@ Carry on
+
+THUMB(	sub	r1, sp, r1		)	@ Restore original R1
+THUMB(	sub	sp, r1			)	@ Restore original SP
+	add	sp, sp, #\frame_size		@ Undo svc_entry's SP change
+	b	__bad_stack			@ Handle VMAP stack overflow
+	.popsection
+.Lout\@:
+#endif
+	.endm

arch/arm/kernel/irq.c

@@ -56,7 +56,14 @@ static void __init init_irq_stacks(void)
 	int cpu;
 
 	for_each_possible_cpu(cpu) {
-		stack = (u8 *)__get_free_pages(GFP_KERNEL, THREAD_SIZE_ORDER);
+		if (!IS_ENABLED(CONFIG_VMAP_STACK))
+			stack = (u8 *)__get_free_pages(GFP_KERNEL,
+						       THREAD_SIZE_ORDER);
+		else
+			stack = __vmalloc_node(THREAD_SIZE, THREAD_ALIGN,
+					       THREADINFO_GFP, NUMA_NO_NODE,
+					       __builtin_return_address(0));
+
 		if (WARN_ON(!stack))
 			break;
 		per_cpu(irq_stack_ptr, cpu) = &stack[THREAD_SIZE];

arch/arm/kernel/setup.c

@@ -141,10 +141,10 @@ EXPORT_SYMBOL(outer_cache);
 int __cpu_architecture __read_mostly = CPU_ARCH_UNKNOWN;
 
 struct stack {
-	u32 irq[3];
-	u32 abt[3];
-	u32 und[3];
-	u32 fiq[3];
+	u32 irq[4];
+	u32 abt[4];
+	u32 und[4];
+	u32 fiq[4];
 } ____cacheline_aligned;
 
 #ifndef CONFIG_CPU_V7M

arch/arm/kernel/sleep.S

@@ -67,6 +67,14 @@ ENTRY(__cpu_suspend)
 	ldr	r4, =cpu_suspend_size
 #endif
 	mov	r5, sp			@ current virtual SP
+#ifdef CONFIG_VMAP_STACK
+	@ Run the suspend code from the overflow stack so we don't have to rely
+	@ on vmalloc-to-phys conversions anywhere in the arch suspend code.
+	@ The original SP value captured in R5 will be restored on the way out.
+	mov_l	r6, overflow_stack_ptr	@ Base pointer
+	mrc	p15, 0, r7, c13, c0, 4	@ Get per-CPU offset
+	ldr	sp, [r6, r7]		@ Address of this CPU's overflow stack
+#endif
 	add	r4, r4, #12		@ Space for pgd, virt sp, phys resume fn
 	sub	sp, sp, r4		@ allocate CPU state on stack
 	ldr	r3, =sleep_save_sp

arch/arm/kernel/traps.c

@@ -121,7 +121,8 @@ void dump_backtrace_stm(u32 *stack, u32 instruction, const char *loglvl)
 static int verify_stack(unsigned long sp)
 {
 	if (sp < PAGE_OFFSET ||
-	    (sp > (unsigned long)high_memory && high_memory != NULL))
+	    (!IS_ENABLED(CONFIG_VMAP_STACK) &&
+	     sp > (unsigned long)high_memory && high_memory != NULL))
 		return -EFAULT;
 
 	return 0;
@@ -291,7 +292,8 @@ static int __die(const char *str, int err, struct pt_regs *regs)
 
 	if (!user_mode(regs) || in_interrupt()) {
 		dump_mem(KERN_EMERG, "Stack: ", regs->ARM_sp,
-			 ALIGN(regs->ARM_sp, THREAD_SIZE));
+			 ALIGN(regs->ARM_sp - THREAD_SIZE, THREAD_ALIGN)
+			 + THREAD_SIZE);
 		dump_backtrace(regs, tsk, KERN_EMERG);
 		dump_instr(KERN_EMERG, regs);
 	}
@@ -838,3 +840,77 @@ void __init early_trap_init(void *vectors_base)
 	 */
 #endif
 }
+
+#ifdef CONFIG_VMAP_STACK
+
+DECLARE_PER_CPU(u8 *, irq_stack_ptr);
+
+asmlinkage DEFINE_PER_CPU(u8 *, overflow_stack_ptr);
+
+static int __init allocate_overflow_stacks(void)
+{
+	u8 *stack;
+	int cpu;
+
+	for_each_possible_cpu(cpu) {
+		stack = (u8 *)__get_free_page(GFP_KERNEL);
+		if (WARN_ON(!stack))
+			return -ENOMEM;
+		per_cpu(overflow_stack_ptr, cpu) = &stack[OVERFLOW_STACK_SIZE];
+	}
+	return 0;
+}
+early_initcall(allocate_overflow_stacks);
+
+asmlinkage void handle_bad_stack(struct pt_regs *regs)
+{
+	unsigned long tsk_stk = (unsigned long)current->stack;
+	unsigned long irq_stk = (unsigned long)this_cpu_read(irq_stack_ptr);
+	unsigned long ovf_stk = (unsigned long)this_cpu_read(overflow_stack_ptr);
+
+	console_verbose();
+	pr_emerg("Insufficient stack space to handle exception!");
+
+	pr_emerg("Task stack:     [0x%08lx..0x%08lx]\n",
+		 tsk_stk, tsk_stk + THREAD_SIZE);
+	pr_emerg("IRQ stack:      [0x%08lx..0x%08lx]\n",
+		 irq_stk - THREAD_SIZE, irq_stk);
+	pr_emerg("Overflow stack: [0x%08lx..0x%08lx]\n",
+		 ovf_stk - OVERFLOW_STACK_SIZE, ovf_stk);
+
+	die("kernel stack overflow", regs, 0);
+}
+
+/*
+ * Normally, we rely on the logic in do_translation_fault() to update stale PMD
+ * entries covering the vmalloc space in a task's page tables when it first
+ * accesses the region in question. Unfortunately, this is not sufficient when
+ * the task stack resides in the vmalloc region, as do_translation_fault() is a
+ * C function that needs a stack to run.
+ *
+ * So we need to ensure that these PMD entries are up to date *before* the MM
+ * switch. As we already have some logic in the MM switch path that takes care
+ * of this, let's trigger it by bumping the counter every time the core vmalloc
+ * code modifies a PMD entry in the vmalloc region.
+ */
+void arch_sync_kernel_mappings(unsigned long start, unsigned long end)
+{
+	if (start > VMALLOC_END || end < VMALLOC_START)
+		return;
+
+	/*
+	 * This hooks into the core vmalloc code to receive notifications of
+	 * any PMD level changes that have been made to the kernel page tables.
+	 * This means it should only be triggered once for every MiB worth of
+	 * vmalloc space, given that we don't support huge vmalloc/vmap on ARM,
+	 * and that kernel PMD level table entries are rarely (if ever)
+	 * updated.
+	 *
+	 * This means that the counter is going to max out at ~250 for the
+	 * typical case. If it overflows, something entirely unexpected has
+	 * occurred so let's throw a warning if that happens.
+	 */
+	WARN_ON(++init_mm.context.vmalloc_seq == UINT_MAX);
+}
+
+#endif

arch/arm/kernel/unwind.c

@@ -389,7 +389,8 @@ int unwind_frame(struct stackframe *frame)
 
 	/* store the highest address on the stack to avoid crossing it*/
 	ctrl.sp_low = frame->sp;
-	ctrl.sp_high = ALIGN(ctrl.sp_low, THREAD_SIZE);
+	ctrl.sp_high = ALIGN(ctrl.sp_low - THREAD_SIZE, THREAD_ALIGN)
+		       + THREAD_SIZE;
 
 	pr_debug("%s(pc = %08lx lr = %08lx sp = %08lx)\n", __func__,
 		 frame->pc, frame->lr, frame->sp);

arch/arm/kernel/vmlinux.lds.S

@@ -138,12 +138,12 @@ SECTIONS
 #ifdef CONFIG_STRICT_KERNEL_RWX
 	. = ALIGN(1<<SECTION_SHIFT);
 #else
-	. = ALIGN(THREAD_SIZE);
+	. = ALIGN(THREAD_ALIGN);
 #endif
 	__init_end = .;
 
 	_sdata = .;
-	RW_DATA(L1_CACHE_BYTES, PAGE_SIZE, THREAD_SIZE)
+	RW_DATA(L1_CACHE_BYTES, PAGE_SIZE, THREAD_ALIGN)
 	_edata = .;
 
 	BSS_SECTION(0, 0, 0)