x86/doublefault/32: Move #DF stack and TSS to cpu_entry_area
authorAndy Lutomirski <luto@kernel.org>
Tue, 26 Nov 2019 17:27:16 +0000 (18:27 +0100)
committerIngo Molnar <mingo@kernel.org>
Tue, 26 Nov 2019 20:53:34 +0000 (21:53 +0100)
There are three problems with the current layout of the doublefault
stack and TSS.  First, the TSS is only cacheline-aligned, which is
not enough -- if the hardware portion of the TSS (struct x86_hw_tss)
crosses a page boundary, horrible things happen [0].  Second, the
stack and TSS are global, so simultaneous double faults on different
CPUs will cause massive corruption.  Third, the whole mechanism
won't work if user CR3 is loaded, resulting in a triple fault [1].

Let the doublefault stack and TSS share a page (which prevents the
TSS from spanning a page boundary), make it percpu, and move it into
cpu_entry_area.  Teach the stack dump code about the doublefault
stack.

[0] Real hardware will read past the end of the page onto the next
    *physical* page if a task switch happens.  Virtual machines may
    have any number of bugs, and I would consider it reasonable for
    a VM to summarily kill the guest if it tries to task-switch to
    a page-spanning TSS.

[1] Real hardware triple faults.  At least some VMs seem to hang.
    I'm not sure what's going on.

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/include/asm/cpu_entry_area.h
arch/x86/include/asm/doublefault.h [new file with mode: 0644]
arch/x86/include/asm/pgtable_32_types.h
arch/x86/include/asm/processor.h
arch/x86/kernel/cpu/common.c
arch/x86/kernel/doublefault_32.c
arch/x86/kernel/dumpstack_32.c
arch/x86/mm/cpu_entry_area.c

index ea866c7..8047340 100644 (file)
@@ -65,6 +65,13 @@ enum exception_stack_ordering {
 
 #endif
 
+#ifdef CONFIG_X86_32
+struct doublefault_stack {
+       unsigned long stack[(PAGE_SIZE - sizeof(struct x86_hw_tss)) / sizeof(unsigned long)];
+       struct x86_hw_tss tss;
+} __aligned(PAGE_SIZE);
+#endif
+
 /*
  * cpu_entry_area is a percpu region that contains things needed by the CPU
  * and early entry/exit code.  Real types aren't used for all fields here
@@ -86,6 +93,11 @@ struct cpu_entry_area {
 #endif
        struct entry_stack_page entry_stack_page;
 
+#ifdef CONFIG_X86_32
+       char guard_doublefault_stack[PAGE_SIZE];
+       struct doublefault_stack doublefault_stack;
+#endif
+
        /*
         * On x86_64, the TSS is mapped RO.  On x86_32, it's mapped RW because
         * we need task switches to work, and task switches write to the TSS.
diff --git a/arch/x86/include/asm/doublefault.h b/arch/x86/include/asm/doublefault.h
new file mode 100644 (file)
index 0000000..af9a14a
--- /dev/null
@@ -0,0 +1,13 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _ASM_X86_DOUBLEFAULT_H
+#define _ASM_X86_DOUBLEFAULT_H
+
+#if defined(CONFIG_X86_32) && defined(CONFIG_DOUBLEFAULT)
+extern void doublefault_init_cpu_tss(void);
+#else
+static inline void doublefault_init_cpu_tss(void)
+{
+}
+#endif
+
+#endif /* _ASM_X86_DOUBLEFAULT_H */
index 19f5807..0416d42 100644 (file)
@@ -41,10 +41,11 @@ extern bool __vmalloc_start_set; /* set once high_memory is set */
 #endif
 
 /*
- * Define this here and validate with BUILD_BUG_ON() in pgtable_32.c
- * to avoid include recursion hell
+ * This is an upper bound on sizeof(struct cpu_entry_area) / PAGE_SIZE.
+ * Define this here and validate with BUILD_BUG_ON() in cpu_entry_area.c
+ * to avoid include recursion hell.
  */
-#define CPU_ENTRY_AREA_PAGES   (NR_CPUS * 41)
+#define CPU_ENTRY_AREA_PAGES   (NR_CPUS * 43)
 
 /* The +1 is for the readonly IDT page: */
 #define CPU_ENTRY_AREA_BASE    \
index f6c6300..0340aad 100644 (file)
@@ -166,7 +166,6 @@ enum cpuid_regs_idx {
 extern struct cpuinfo_x86      boot_cpu_data;
 extern struct cpuinfo_x86      new_cpu_data;
 
-extern struct x86_hw_tss       doublefault_tss;
 extern __u32                   cpu_caps_cleared[NCAPINTS + NBUGINTS];
 extern __u32                   cpu_caps_set[NCAPINTS + NBUGINTS];
 
index baa2fed..2e4d902 100644 (file)
@@ -24,6 +24,7 @@
 #include <asm/stackprotector.h>
 #include <asm/perf_event.h>
 #include <asm/mmu_context.h>
+#include <asm/doublefault.h>
 #include <asm/archrandom.h>
 #include <asm/hypervisor.h>
 #include <asm/processor.h>
@@ -1814,8 +1815,6 @@ static inline void tss_setup_ist(struct tss_struct *tss)
        tss->x86_tss.ist[IST_INDEX_MCE] = __this_cpu_ist_top_va(MCE);
 }
 
-static inline void gdt_setup_doublefault_tss(int cpu) { }
-
 #else /* CONFIG_X86_64 */
 
 static inline void setup_getcpu(int cpu) { }
@@ -1827,13 +1826,6 @@ static inline void ucode_cpu_init(int cpu)
 
 static inline void tss_setup_ist(struct tss_struct *tss) { }
 
-static inline void gdt_setup_doublefault_tss(int cpu)
-{
-#ifdef CONFIG_DOUBLEFAULT
-       /* Set up the doublefault TSS pointer in the GDT */
-       __set_tss_desc(cpu, GDT_ENTRY_DOUBLEFAULT_TSS, &doublefault_tss);
-#endif
-}
 #endif /* !CONFIG_X86_64 */
 
 static inline void tss_setup_io_bitmap(struct tss_struct *tss)
@@ -1923,7 +1915,7 @@ void cpu_init(void)
        clear_all_debug_regs();
        dbg_restore_debug_regs();
 
-       gdt_setup_doublefault_tss(cpu);
+       doublefault_init_cpu_tss();
 
        fpu__init_cpu();
 
index 61c707c..4eecfe4 100644 (file)
 #include <asm/processor.h>
 #include <asm/desc.h>
 
-#define DOUBLEFAULT_STACKSIZE (1024)
-static unsigned long doublefault_stack[DOUBLEFAULT_STACKSIZE];
-#define STACK_START (unsigned long)(doublefault_stack+DOUBLEFAULT_STACKSIZE)
-
 #define ptr_ok(x) ((x) > PAGE_OFFSET && (x) < PAGE_OFFSET + MAXMEM)
 
 static void doublefault_fn(void)
@@ -21,6 +17,8 @@ static void doublefault_fn(void)
        struct desc_ptr gdt_desc = {0, 0};
        unsigned long gdt, tss;
 
+       BUILD_BUG_ON(sizeof(struct doublefault_stack) != PAGE_SIZE);
+
        native_store_gdt(&gdt_desc);
        gdt = gdt_desc.address;
 
@@ -48,24 +46,46 @@ static void doublefault_fn(void)
                cpu_relax();
 }
 
-struct x86_hw_tss doublefault_tss __cacheline_aligned = {
-       .sp0            = STACK_START,
-       .ss0            = __KERNEL_DS,
-       .ldt            = 0,
+DEFINE_PER_CPU_PAGE_ALIGNED(struct doublefault_stack, doublefault_stack) = {
+       .tss = {
+                /*
+                 * No sp0 or ss0 -- we never run CPL != 0 with this TSS
+                 * active.  sp is filled in later.
+                 */
+               .ldt            = 0,
        .io_bitmap_base = IO_BITMAP_OFFSET_INVALID,
 
-       .ip             = (unsigned long) doublefault_fn,
-       /* 0x2 bit is always set */
-       .flags          = X86_EFLAGS_SF | 0x2,
-       .sp             = STACK_START,
-       .es             = __USER_DS,
-       .cs             = __KERNEL_CS,
-       .ss             = __KERNEL_DS,
-       .ds             = __USER_DS,
-       .fs             = __KERNEL_PERCPU,
+               .ip             = (unsigned long) doublefault_fn,
+               /* 0x2 bit is always set */
+               .flags          = X86_EFLAGS_SF | 0x2,
+               .es             = __USER_DS,
+               .cs             = __KERNEL_CS,
+               .ss             = __KERNEL_DS,
+               .ds             = __USER_DS,
+               .fs             = __KERNEL_PERCPU,
 #ifndef CONFIG_X86_32_LAZY_GS
-       .gs             = __KERNEL_STACK_CANARY,
+               .gs             = __KERNEL_STACK_CANARY,
 #endif
 
-       .__cr3          = __pa_nodebug(swapper_pg_dir),
+               .__cr3          = __pa_nodebug(swapper_pg_dir),
+       },
 };
+
+void doublefault_init_cpu_tss(void)
+{
+       unsigned int cpu = smp_processor_id();
+       struct cpu_entry_area *cea = get_cpu_entry_area(cpu);
+
+       /*
+        * The linker isn't smart enough to initialize percpu variables that
+        * point to other places in percpu space.
+        */
+        this_cpu_write(doublefault_stack.tss.sp,
+                       (unsigned long)&cea->doublefault_stack.stack +
+                       sizeof(doublefault_stack.stack));
+
+       /* Set up doublefault TSS pointer in the GDT */
+       __set_tss_desc(cpu, GDT_ENTRY_DOUBLEFAULT_TSS,
+                      &get_cpu_entry_area(cpu)->doublefault_stack.tss);
+
+}
index 64a59d7..8e3a8fe 100644 (file)
@@ -29,6 +29,9 @@ const char *stack_type_name(enum stack_type type)
        if (type == STACK_TYPE_ENTRY)
                return "ENTRY_TRAMPOLINE";
 
+       if (type == STACK_TYPE_EXCEPTION)
+               return "#DF";
+
        return NULL;
 }
 
@@ -82,6 +85,30 @@ static bool in_softirq_stack(unsigned long *stack, struct stack_info *info)
        return true;
 }
 
+static bool in_doublefault_stack(unsigned long *stack, struct stack_info *info)
+{
+#ifdef CONFIG_DOUBLEFAULT
+       struct cpu_entry_area *cea = get_cpu_entry_area(raw_smp_processor_id());
+       struct doublefault_stack *ss = &cea->doublefault_stack;
+
+       void *begin = ss->stack;
+       void *end = begin + sizeof(ss->stack);
+
+       if ((void *)stack < begin || (void *)stack >= end)
+               return false;
+
+       info->type      = STACK_TYPE_EXCEPTION;
+       info->begin     = begin;
+       info->end       = end;
+       info->next_sp   = (unsigned long *)this_cpu_read(cpu_tss_rw.x86_tss.sp);
+
+       return true;
+#else
+       return false;
+#endif
+}
+
+
 int get_stack_info(unsigned long *stack, struct task_struct *task,
                   struct stack_info *info, unsigned long *visit_mask)
 {
@@ -105,6 +132,9 @@ int get_stack_info(unsigned long *stack, struct task_struct *task,
        if (in_softirq_stack(stack, info))
                goto recursion_check;
 
+       if (in_doublefault_stack(stack, info))
+               goto recursion_check;
+
        goto unknown;
 
 recursion_check:
index 82ead8e..56f9189 100644 (file)
@@ -17,6 +17,10 @@ static DEFINE_PER_CPU_PAGE_ALIGNED(struct exception_stacks, exception_stacks);
 DEFINE_PER_CPU(struct cea_exception_stacks*, cea_exception_stacks);
 #endif
 
+#if defined(CONFIG_X86_32) && defined(CONFIG_DOUBLEFAULT)
+DECLARE_PER_CPU_PAGE_ALIGNED(struct doublefault_stack, doublefault_stack);
+#endif
+
 struct cpu_entry_area *get_cpu_entry_area(int cpu)
 {
        unsigned long va = CPU_ENTRY_AREA_PER_CPU + cpu * CPU_ENTRY_AREA_SIZE;
@@ -108,7 +112,15 @@ static void __init percpu_setup_exception_stacks(unsigned int cpu)
        cea_map_stack(MCE);
 }
 #else
-static inline void percpu_setup_exception_stacks(unsigned int cpu) {}
+static inline void percpu_setup_exception_stacks(unsigned int cpu)
+{
+#ifdef CONFIG_DOUBLEFAULT
+       struct cpu_entry_area *cea = get_cpu_entry_area(cpu);
+
+       cea_map_percpu_pages(&cea->doublefault_stack,
+                            &per_cpu(doublefault_stack, cpu), 1, PAGE_KERNEL);
+#endif
+}
 #endif
 
 /* Setup the fixmap mappings only once per-processor */