arch/x86/kernel/dumpstack.c

   1 /*
   2  *  Copyright (C) 1991, 1992  Linus Torvalds
   3  *  Copyright (C) 2000, 2001, 2002 Andi Kleen, SuSE Labs
   4  */
   5 #include <linux/kallsyms.h>
   6 #include <linux/kprobes.h>
   7 #include <linux/uaccess.h>
   8 #include <linux/utsname.h>
   9 #include <linux/hardirq.h>
  10 #include <linux/kdebug.h>
  11 #include <linux/module.h>
  12 #include <linux/ptrace.h>
  13 #include <linux/sched/debug.h>
  14 #include <linux/sched/task_stack.h>
  15 #include <linux/ftrace.h>
  16 #include <linux/kexec.h>
  17 #include <linux/bug.h>
  18 #include <linux/nmi.h>
  19 #include <linux/sysfs.h>
  20
  21 #include <asm/stacktrace.h>
  22 #include <asm/unwind.h>
  23
  24 int panic_on_unrecovered_nmi;
  25 int panic_on_io_nmi;
  26 unsigned int code_bytes = 64;
  27 static int die_counter;
  28
  29 bool in_task_stack(unsigned long *stack, struct task_struct *task,
  30                    struct stack_info *info)
  31 {
  32         unsigned long *begin = task_stack_page(task);
  33         unsigned long *end   = task_stack_page(task) + THREAD_SIZE;
  34
  35         if (stack < begin || stack >= end)
  36                 return false;
  37
  38         info->type      = STACK_TYPE_TASK;
  39         info->begin     = begin;
  40         info->end       = end;
  41         info->next_sp   = NULL;
  42
  43         return true;
  44 }
  45
  46 static void printk_stack_address(unsigned long address, int reliable,
  47                                  char *log_lvl)
  48 {
  49         touch_nmi_watchdog();
  50         printk("%s %s%pB\n", log_lvl, reliable ? "" : "? ", (void *)address);
  51 }
  52
  53 void show_trace_log_lvl(struct task_struct *task, struct pt_regs *regs,
  54                         unsigned long *stack, char *log_lvl)
  55 {
  56         struct unwind_state state;
  57         struct stack_info stack_info = {0};
  58         unsigned long visit_mask = 0;
  59         int graph_idx = 0;
  60
  61         printk("%sCall Trace:\n", log_lvl);
  62
  63         unwind_start(&state, task, regs, stack);
  64         stack = stack ? : get_stack_pointer(task, regs);
  65
  66         /*
  67          * Iterate through the stacks, starting with the current stack pointer.
  68          * Each stack has a pointer to the next one.
  69          *
  70          * x86-64 can have several stacks:
  71          * - task stack
  72          * - interrupt stack
  73          * - HW exception stacks (double fault, nmi, debug, mce)
  74          *
  75          * x86-32 can have up to three stacks:
  76          * - task stack
  77          * - softirq stack
  78          * - hardirq stack
  79          */
  80         for (regs = NULL; stack; stack = PTR_ALIGN(stack_info.next_sp, sizeof(long))) {
  81                 const char *stack_name;
  82
  83                 /*
  84                  * If we overflowed the task stack into a guard page, jump back
  85                  * to the bottom of the usable stack.
  86                  */
  87                 if (task_stack_page(task) - (void *)stack < PAGE_SIZE)
  88                         stack = task_stack_page(task);
  89
  90                 if (get_stack_info(stack, task, &stack_info, &visit_mask))
  91                         break;
  92
  93                 stack_name = stack_type_name(stack_info.type);
  94                 if (stack_name)
  95                         printk("%s <%s>\n", log_lvl, stack_name);
  96
  97                 /*
  98                  * Scan the stack, printing any text addresses we find.  At the
  99                  * same time, follow proper stack frames with the unwinder.
 100                  *
 101                  * Addresses found during the scan which are not reported by
 102                  * the unwinder are considered to be additional clues which are
 103                  * sometimes useful for debugging and are prefixed with '?'.
 104                  * This also serves as a failsafe option in case the unwinder
 105                  * goes off in the weeds.
 106                  */
 107                 for (; stack < stack_info.end; stack++) {
 108                         unsigned long real_addr;
 109                         int reliable = 0;
 110                         unsigned long addr = READ_ONCE_NOCHECK(*stack);
 111                         unsigned long *ret_addr_p =
 112                                 unwind_get_return_address_ptr(&state);
 113
 114                         if (!__kernel_text_address(addr))
 115                                 continue;
 116
 117                         /*
 118                          * Don't print regs->ip again if it was already printed
 119                          * by __show_regs() below.
 120                          */
 121                         if (regs && stack == &regs->ip) {
 122                                 unwind_next_frame(&state);
 123                                 continue;
 124                         }
 125
 126                         if (stack == ret_addr_p)
 127                                 reliable = 1;
 128
 129                         /*
 130                          * When function graph tracing is enabled for a
 131                          * function, its return address on the stack is
 132                          * replaced with the address of an ftrace handler
 133                          * (return_to_handler).  In that case, before printing
 134                          * the "real" address, we want to print the handler
 135                          * address as an "unreliable" hint that function graph
 136                          * tracing was involved.
 137                          */
 138                         real_addr = ftrace_graph_ret_addr(task, &graph_idx,
 139                                                           addr, stack);
 140                         if (real_addr != addr)
 141                                 printk_stack_address(addr, 0, log_lvl);
 142                         printk_stack_address(real_addr, reliable, log_lvl);
 143
 144                         if (!reliable)
 145                                 continue;
 146
 147                         /*
 148                          * Get the next frame from the unwinder.  No need to
 149                          * check for an error: if anything goes wrong, the rest
 150                          * of the addresses will just be printed as unreliable.
 151                          */
 152                         unwind_next_frame(&state);
 153
 154                         /* if the frame has entry regs, print them */
 155                         regs = unwind_get_entry_regs(&state);
 156                         if (regs)
 157                                 __show_regs(regs, 0);
 158                 }
 159
 160                 if (stack_name)
 161                         printk("%s </%s>\n", log_lvl, stack_name);
 162         }
 163 }
 164
 165 void show_stack(struct task_struct *task, unsigned long *sp)
 166 {
 167         task = task ? : current;
 168
 169         /*
 170          * Stack frames below this one aren't interesting.  Don't show them
 171          * if we're printing for %current.
 172          */
 173         if (!sp && task == current)
 174                 sp = get_stack_pointer(current, NULL);
 175
 176         show_trace_log_lvl(task, NULL, sp, KERN_DEFAULT);
 177 }
 178
 179 void show_stack_regs(struct pt_regs *regs)
 180 {
 181         show_trace_log_lvl(current, regs, NULL, KERN_DEFAULT);
 182 }
 183
 184 static arch_spinlock_t die_lock = __ARCH_SPIN_LOCK_UNLOCKED;
 185 static int die_owner = -1;
 186 static unsigned int die_nest_count;
 187
 188 unsigned long oops_begin(void)
 189 {
 190         int cpu;
 191         unsigned long flags;
 192
 193         oops_enter();
 194
 195         /* racy, but better than risking deadlock. */
 196         raw_local_irq_save(flags);
 197         cpu = smp_processor_id();
 198         if (!arch_spin_trylock(&die_lock)) {
 199                 if (cpu == die_owner)
 200                         /* nested oops. should stop eventually */;
 201                 else
 202                         arch_spin_lock(&die_lock);
 203         }
 204         die_nest_count++;
 205         die_owner = cpu;
 206         console_verbose();
 207         bust_spinlocks(1);
 208         return flags;
 209 }
 210 EXPORT_SYMBOL_GPL(oops_begin);
 211 NOKPROBE_SYMBOL(oops_begin);
 212
 213 void __noreturn rewind_stack_do_exit(int signr);
 214
 215 void oops_end(unsigned long flags, struct pt_regs *regs, int signr)
 216 {
 217         if (regs && kexec_should_crash(current))
 218                 crash_kexec(regs);
 219
 220         bust_spinlocks(0);
 221         die_owner = -1;
 222         add_taint(TAINT_DIE, LOCKDEP_NOW_UNRELIABLE);
 223         die_nest_count--;
 224         if (!die_nest_count)
 225                 /* Nest count reaches zero, release the lock. */
 226                 arch_spin_unlock(&die_lock);
 227         raw_local_irq_restore(flags);
 228         oops_exit();
 229
 230         if (!signr)
 231                 return;
 232         if (in_interrupt())
 233                 panic("Fatal exception in interrupt");
 234         if (panic_on_oops)
 235                 panic("Fatal exception");
 236
 237         /*
 238          * We're not going to return, but we might be on an IST stack or
 239          * have very little stack space left.  Rewind the stack and kill
 240          * the task.
 241          */
 242         rewind_stack_do_exit(signr);
 243 }
 244 NOKPROBE_SYMBOL(oops_end);
 245
 246 int __die(const char *str, struct pt_regs *regs, long err)
 247 {
 248 #ifdef CONFIG_X86_32
 249         unsigned short ss;
 250         unsigned long sp;
 251 #endif
 252         printk(KERN_DEFAULT
 253                "%s: %04lx [#%d]%s%s%s%s\n", str, err & 0xffff, ++die_counter,
 254                IS_ENABLED(CONFIG_PREEMPT) ? " PREEMPT"         : "",
 255                IS_ENABLED(CONFIG_SMP)     ? " SMP"             : "",
 256                debug_pagealloc_enabled()  ? " DEBUG_PAGEALLOC" : "",
 257                IS_ENABLED(CONFIG_KASAN)   ? " KASAN"           : "");
 258
 259         if (notify_die(DIE_OOPS, str, regs, err,
 260                         current->thread.trap_nr, SIGSEGV) == NOTIFY_STOP)
 261                 return 1;
 262
 263         print_modules();
 264         show_regs(regs);
 265 #ifdef CONFIG_X86_32
 266         if (user_mode(regs)) {
 267                 sp = regs->sp;
 268                 ss = regs->ss & 0xffff;
 269         } else {
 270                 sp = kernel_stack_pointer(regs);
 271                 savesegment(ss, ss);
 272         }
 273         printk(KERN_EMERG "EIP: %pS SS:ESP: %04x:%08lx\n",
 274                (void *)regs->ip, ss, sp);
 275 #else
 276         /* Executive summary in case the oops scrolled away */
 277         printk(KERN_ALERT "RIP: %pS RSP: %016lx\n", (void *)regs->ip, regs->sp);
 278 #endif
 279         return 0;
 280 }
 281 NOKPROBE_SYMBOL(__die);
 282
 283 /*
 284  * This is gone through when something in the kernel has done something bad
 285  * and is about to be terminated:
 286  */
 287 void die(const char *str, struct pt_regs *regs, long err)
 288 {
 289         unsigned long flags = oops_begin();
 290         int sig = SIGSEGV;
 291
 292         if (__die(str, regs, err))
 293                 sig = 0;
 294         oops_end(flags, regs, sig);
 295 }
 296
 297 static int __init code_bytes_setup(char *s)
 298 {
 299         ssize_t ret;
 300         unsigned long val;
 301
 302         if (!s)
 303                 return -EINVAL;
 304
 305         ret = kstrtoul(s, 0, &val);
 306         if (ret)
 307                 return ret;
 308
 309         code_bytes = val;
 310         if (code_bytes > 8192)
 311                 code_bytes = 8192;
 312
 313         return 1;
 314 }
 315 __setup("code_bytes=", code_bytes_setup);