arch/x86/kernel/dumpstack_64.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/hardirq.h>
   9 #include <linux/kdebug.h>
  10 #include <linux/export.h>
  11 #include <linux/ptrace.h>
  12 #include <linux/kexec.h>
  13 #include <linux/sysfs.h>
  14 #include <linux/bug.h>
  15 #include <linux/nmi.h>
  16
  17 #include <asm/stacktrace.h>
  18
  19
  20 #define N_EXCEPTION_STACKS_END \
  21                 (N_EXCEPTION_STACKS + DEBUG_STKSZ/EXCEPTION_STKSZ - 2)
  22
  23 static char x86_stack_ids[][8] = {
  24                 [ DEBUG_STACK-1                 ]       = "#DB",
  25                 [ NMI_STACK-1                   ]       = "NMI",
  26                 [ DOUBLEFAULT_STACK-1           ]       = "#DF",
  27                 [ MCE_STACK-1                   ]       = "#MC",
  28 #if DEBUG_STKSZ > EXCEPTION_STKSZ
  29                 [ N_EXCEPTION_STACKS ...
  30                   N_EXCEPTION_STACKS_END        ]       = "#DB[?]"
  31 #endif
  32 };
  33
  34 static unsigned long *in_exception_stack(unsigned cpu, unsigned long stack,
  35                                          unsigned *usedp, char **idp)
  36 {
  37         unsigned k;
  38
  39         /*
  40          * Iterate over all exception stacks, and figure out whether
  41          * 'stack' is in one of them:
  42          */
  43         for (k = 0; k < N_EXCEPTION_STACKS; k++) {
  44                 unsigned long end = per_cpu(orig_ist, cpu).ist[k];
  45                 /*
  46                  * Is 'stack' above this exception frame's end?
  47                  * If yes then skip to the next frame.
  48                  */
  49                 if (stack >= end)
  50                         continue;
  51                 /*
  52                  * Is 'stack' above this exception frame's start address?
  53                  * If yes then we found the right frame.
  54                  */
  55                 if (stack >= end - EXCEPTION_STKSZ) {
  56                         /*
  57                          * Make sure we only iterate through an exception
  58                          * stack once. If it comes up for the second time
  59                          * then there's something wrong going on - just
  60                          * break out and return NULL:
  61                          */
  62                         if (*usedp & (1U << k))
  63                                 break;
  64                         *usedp |= 1U << k;
  65                         *idp = x86_stack_ids[k];
  66                         return (unsigned long *)end;
  67                 }
  68                 /*
  69                  * If this is a debug stack, and if it has a larger size than
  70                  * the usual exception stacks, then 'stack' might still
  71                  * be within the lower portion of the debug stack:
  72                  */
  73 #if DEBUG_STKSZ > EXCEPTION_STKSZ
  74                 if (k == DEBUG_STACK - 1 && stack >= end - DEBUG_STKSZ) {
  75                         unsigned j = N_EXCEPTION_STACKS - 1;
  76
  77                         /*
  78                          * Black magic. A large debug stack is composed of
  79                          * multiple exception stack entries, which we
  80                          * iterate through now. Dont look:
  81                          */
  82                         do {
  83                                 ++j;
  84                                 end -= EXCEPTION_STKSZ;
  85                                 x86_stack_ids[j][4] = '1' +
  86                                                 (j - N_EXCEPTION_STACKS);
  87                         } while (stack < end - EXCEPTION_STKSZ);
  88                         if (*usedp & (1U << j))
  89                                 break;
  90                         *usedp |= 1U << j;
  91                         *idp = x86_stack_ids[j];
  92                         return (unsigned long *)end;
  93                 }
  94 #endif
  95         }
  96         return NULL;
  97 }
  98
  99 static inline int
 100 in_irq_stack(unsigned long *stack, unsigned long *irq_stack,
 101              unsigned long *irq_stack_end)
 102 {
 103         return (stack >= irq_stack && stack < irq_stack_end);
 104 }
 105
 106 enum stack_type {
 107         STACK_IS_UNKNOWN,
 108         STACK_IS_NORMAL,
 109         STACK_IS_EXCEPTION,
 110         STACK_IS_IRQ,
 111 };
 112
 113 static enum stack_type
 114 analyze_stack(int cpu, struct task_struct *task, unsigned long *stack,
 115               unsigned long **stack_end, unsigned long *irq_stack,
 116               unsigned *used, char **id)
 117 {
 118         unsigned long addr;
 119
 120         addr = ((unsigned long)stack & (~(THREAD_SIZE - 1)));
 121         if ((unsigned long)task_stack_page(task) == addr)
 122                 return STACK_IS_NORMAL;
 123
 124         *stack_end = in_exception_stack(cpu, (unsigned long)stack,
 125                                         used, id);
 126         if (*stack_end)
 127                 return STACK_IS_EXCEPTION;
 128
 129         if (!irq_stack)
 130                 return STACK_IS_NORMAL;
 131
 132         *stack_end = irq_stack;
 133         irq_stack -= (IRQ_STACK_SIZE / sizeof(long));
 134
 135         if (in_irq_stack(stack, irq_stack, *stack_end))
 136                 return STACK_IS_IRQ;
 137
 138         return STACK_IS_UNKNOWN;
 139 }
 140
 141 /*
 142  * x86-64 can have up to three kernel stacks:
 143  * process stack
 144  * interrupt stack
 145  * severe exception (double fault, nmi, stack fault, debug, mce) hardware stack
 146  */
 147
 148 void dump_trace(struct task_struct *task, struct pt_regs *regs,
 149                 unsigned long *stack, unsigned long bp,
 150                 const struct stacktrace_ops *ops, void *data)
 151 {
 152         const unsigned cpu = get_cpu();
 153         unsigned long *irq_stack = (unsigned long *)per_cpu(irq_stack_ptr, cpu);
 154         unsigned long dummy;
 155         unsigned used = 0;
 156         int graph = 0;
 157         int done = 0;
 158
 159         if (!task)
 160                 task = current;
 161
 162         if (!stack) {
 163                 if (regs)
 164                         stack = (unsigned long *)regs->sp;
 165                 else if (task != current)
 166                         stack = (unsigned long *)task->thread.sp;
 167                 else
 168                         stack = &dummy;
 169         }
 170
 171         if (!bp)
 172                 bp = stack_frame(task, regs);
 173         /*
 174          * Print function call entries in all stacks, starting at the
 175          * current stack address. If the stacks consist of nested
 176          * exceptions
 177          */
 178         while (!done) {
 179                 unsigned long *stack_end;
 180                 enum stack_type stype;
 181                 char *id;
 182
 183                 stype = analyze_stack(cpu, task, stack, &stack_end,
 184                                       irq_stack, &used, &id);
 185
 186                 /* Default finish unless specified to continue */
 187                 done = 1;
 188
 189                 switch (stype) {
 190
 191                 /* Break out early if we are on the thread stack */
 192                 case STACK_IS_NORMAL:
 193                         break;
 194
 195                 case STACK_IS_EXCEPTION:
 196
 197                         if (ops->stack(data, id) < 0)
 198                                 break;
 199
 200                         bp = ops->walk_stack(task, stack, bp, ops,
 201                                              data, stack_end, &graph);
 202                         ops->stack(data, "EOE");
 203                         /*
 204                          * We link to the next stack via the
 205                          * second-to-last pointer (index -2 to end) in the
 206                          * exception stack:
 207                          */
 208                         stack = (unsigned long *) stack_end[-2];
 209                         done = 0;
 210                         break;
 211
 212                 case STACK_IS_IRQ:
 213
 214                         if (ops->stack(data, "IRQ") < 0)
 215                                 break;
 216                         bp = ops->walk_stack(task, stack, bp,
 217                                      ops, data, stack_end, &graph);
 218                         /*
 219                          * We link to the next stack (which would be
 220                          * the process stack normally) the last
 221                          * pointer (index -1 to end) in the IRQ stack:
 222                          */
 223                         stack = (unsigned long *) (stack_end[-1]);
 224                         irq_stack = NULL;
 225                         ops->stack(data, "EOI");
 226                         done = 0;
 227                         break;
 228
 229                 case STACK_IS_UNKNOWN:
 230                         ops->stack(data, "UNK");
 231                         break;
 232                 }
 233         }
 234
 235         /*
 236          * This handles the process stack:
 237          */
 238         bp = ops->walk_stack(task, stack, bp, ops, data, NULL, &graph);
 239         put_cpu();
 240 }
 241 EXPORT_SYMBOL(dump_trace);
 242
 243 void
 244 show_stack_log_lvl(struct task_struct *task, struct pt_regs *regs,
 245                    unsigned long *sp, unsigned long bp, char *log_lvl)
 246 {
 247         unsigned long *irq_stack_end;
 248         unsigned long *irq_stack;
 249         unsigned long *stack;
 250         int cpu;
 251         int i;
 252
 253         preempt_disable();
 254         cpu = smp_processor_id();
 255
 256         irq_stack_end = (unsigned long *)(per_cpu(irq_stack_ptr, cpu));
 257         irq_stack     = irq_stack_end - (IRQ_STACK_SIZE / sizeof(long));
 258
 259         /*
 260          * Debugging aid: "show_stack(NULL, NULL);" prints the
 261          * back trace for this cpu:
 262          */
 263         if (sp == NULL) {
 264                 if (regs)
 265                         sp = (unsigned long *)regs->sp;
 266                 else if (task)
 267                         sp = (unsigned long *)task->thread.sp;
 268                 else
 269                         sp = (unsigned long *)&sp;
 270         }
 271
 272         stack = sp;
 273         for (i = 0; i < kstack_depth_to_print; i++) {
 274                 unsigned long word;
 275
 276                 if (stack >= irq_stack && stack <= irq_stack_end) {
 277                         if (stack == irq_stack_end) {
 278                                 stack = (unsigned long *) (irq_stack_end[-1]);
 279                                 pr_cont(" <EOI> ");
 280                         }
 281                 } else {
 282                 if (kstack_end(stack))
 283                         break;
 284                 }
 285
 286                 if (probe_kernel_address(stack, word))
 287                         break;
 288
 289                 if ((i % STACKSLOTS_PER_LINE) == 0) {
 290                         if (i != 0)
 291                                 pr_cont("\n");
 292                         printk("%s %016lx", log_lvl, word);
 293                 } else
 294                         pr_cont(" %016lx", word);
 295
 296                 stack++;
 297                 touch_nmi_watchdog();
 298         }
 299         preempt_enable();
 300
 301         pr_cont("\n");
 302         show_trace_log_lvl(task, regs, sp, bp, log_lvl);
 303 }
 304
 305 void show_regs(struct pt_regs *regs)
 306 {
 307         int i;
 308         unsigned long sp;
 309
 310         sp = regs->sp;
 311         show_regs_print_info(KERN_DEFAULT);
 312         __show_regs(regs, 1);
 313
 314         /*
 315          * When in-kernel, we also print out the stack and code at the
 316          * time of the fault..
 317          */
 318         if (!user_mode(regs)) {
 319                 unsigned int code_prologue = code_bytes * 43 / 64;
 320                 unsigned int code_len = code_bytes;
 321                 unsigned char c;
 322                 u8 *ip;
 323
 324                 printk(KERN_DEFAULT "Stack:\n");
 325                 show_stack_log_lvl(NULL, regs, (unsigned long *)sp,
 326                                    0, KERN_DEFAULT);
 327
 328                 printk(KERN_DEFAULT "Code: ");
 329
 330                 ip = (u8 *)regs->ip - code_prologue;
 331                 if (ip < (u8 *)PAGE_OFFSET || probe_kernel_address(ip, c)) {
 332                         /* try starting at IP */
 333                         ip = (u8 *)regs->ip;
 334                         code_len = code_len - code_prologue + 1;
 335                 }
 336                 for (i = 0; i < code_len; i++, ip++) {
 337                         if (ip < (u8 *)PAGE_OFFSET ||
 338                                         probe_kernel_address(ip, c)) {
 339                                 pr_cont(" Bad RIP value.");
 340                                 break;
 341                         }
 342                         if (ip == (u8 *)regs->ip)
 343                                 pr_cont("<%02x> ", c);
 344                         else
 345                                 pr_cont("%02x ", c);
 346                 }
 347         }
 348         pr_cont("\n");
 349 }
 350
 351 int is_valid_bugaddr(unsigned long ip)
 352 {
 353         unsigned short ud2;
 354
 355         if (__copy_from_user(&ud2, (const void __user *) ip, sizeof(ud2)))
 356                 return 0;
 357
 358         return ud2 == 0x0b0f;
 359 }