linux/arch/x86/kernel/dumpstack.c
<<
>>
Prefs
   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#include <linux/kasan.h>
  21
  22#include <asm/cpu_entry_area.h>
  23#include <asm/stacktrace.h>
  24#include <asm/unwind.h>
  25
  26int panic_on_unrecovered_nmi;
  27int panic_on_io_nmi;
  28static int die_counter;
  29
  30static struct pt_regs exec_summary_regs;
  31
  32bool noinstr in_task_stack(unsigned long *stack, struct task_struct *task,
  33                           struct stack_info *info)
  34{
  35        unsigned long *begin = task_stack_page(task);
  36        unsigned long *end   = task_stack_page(task) + THREAD_SIZE;
  37
  38        if (stack < begin || stack >= end)
  39                return false;
  40
  41        info->type      = STACK_TYPE_TASK;
  42        info->begin     = begin;
  43        info->end       = end;
  44        info->next_sp   = NULL;
  45
  46        return true;
  47}
  48
  49/* Called from get_stack_info_noinstr - so must be noinstr too */
  50bool noinstr in_entry_stack(unsigned long *stack, struct stack_info *info)
  51{
  52        struct entry_stack *ss = cpu_entry_stack(smp_processor_id());
  53
  54        void *begin = ss;
  55        void *end = ss + 1;
  56
  57        if ((void *)stack < begin || (void *)stack >= end)
  58                return false;
  59
  60        info->type      = STACK_TYPE_ENTRY;
  61        info->begin     = begin;
  62        info->end       = end;
  63        info->next_sp   = NULL;
  64
  65        return true;
  66}
  67
  68static void printk_stack_address(unsigned long address, int reliable,
  69                                 const char *log_lvl)
  70{
  71        touch_nmi_watchdog();
  72        printk("%s %s%pBb\n", log_lvl, reliable ? "" : "? ", (void *)address);
  73}
  74
  75static int copy_code(struct pt_regs *regs, u8 *buf, unsigned long src,
  76                     unsigned int nbytes)
  77{
  78        if (!user_mode(regs))
  79                return copy_from_kernel_nofault(buf, (u8 *)src, nbytes);
  80
  81        /* The user space code from other tasks cannot be accessed. */
  82        if (regs != task_pt_regs(current))
  83                return -EPERM;
  84        /*
  85         * Make sure userspace isn't trying to trick us into dumping kernel
  86         * memory by pointing the userspace instruction pointer at it.
  87         */
  88        if (__chk_range_not_ok(src, nbytes, TASK_SIZE_MAX))
  89                return -EINVAL;
  90
  91        /*
  92         * Even if named copy_from_user_nmi() this can be invoked from
  93         * other contexts and will not try to resolve a pagefault, which is
  94         * the correct thing to do here as this code can be called from any
  95         * context.
  96         */
  97        return copy_from_user_nmi(buf, (void __user *)src, nbytes);
  98}
  99
 100/*
 101 * There are a couple of reasons for the 2/3rd prologue, courtesy of Linus:
 102 *
 103 * In case where we don't have the exact kernel image (which, if we did, we can
 104 * simply disassemble and navigate to the RIP), the purpose of the bigger
 105 * prologue is to have more context and to be able to correlate the code from
 106 * the different toolchains better.
 107 *
 108 * In addition, it helps in recreating the register allocation of the failing
 109 * kernel and thus make sense of the register dump.
 110 *
 111 * What is more, the additional complication of a variable length insn arch like
 112 * x86 warrants having longer byte sequence before rIP so that the disassembler
 113 * can "sync" up properly and find instruction boundaries when decoding the
 114 * opcode bytes.
 115 *
 116 * Thus, the 2/3rds prologue and 64 byte OPCODE_BUFSIZE is just a random
 117 * guesstimate in attempt to achieve all of the above.
 118 */
 119void show_opcodes(struct pt_regs *regs, const char *loglvl)
 120{
 121#define PROLOGUE_SIZE 42
 122#define EPILOGUE_SIZE 21
 123#define OPCODE_BUFSIZE (PROLOGUE_SIZE + 1 + EPILOGUE_SIZE)
 124        u8 opcodes[OPCODE_BUFSIZE];
 125        unsigned long prologue = regs->ip - PROLOGUE_SIZE;
 126
 127        switch (copy_code(regs, opcodes, prologue, sizeof(opcodes))) {
 128        case 0:
 129                printk("%sCode: %" __stringify(PROLOGUE_SIZE) "ph <%02x> %"
 130                       __stringify(EPILOGUE_SIZE) "ph\n", loglvl, opcodes,
 131                       opcodes[PROLOGUE_SIZE], opcodes + PROLOGUE_SIZE + 1);
 132                break;
 133        case -EPERM:
 134                /* No access to the user space stack of other tasks. Ignore. */
 135                break;
 136        default:
 137                printk("%sCode: Unable to access opcode bytes at RIP 0x%lx.\n",
 138                       loglvl, prologue);
 139                break;
 140        }
 141}
 142
 143void show_ip(struct pt_regs *regs, const char *loglvl)
 144{
 145#ifdef CONFIG_X86_32
 146        printk("%sEIP: %pS\n", loglvl, (void *)regs->ip);
 147#else
 148        printk("%sRIP: %04x:%pS\n", loglvl, (int)regs->cs, (void *)regs->ip);
 149#endif
 150        show_opcodes(regs, loglvl);
 151}
 152
 153void show_iret_regs(struct pt_regs *regs, const char *log_lvl)
 154{
 155        show_ip(regs, log_lvl);
 156        printk("%sRSP: %04x:%016lx EFLAGS: %08lx", log_lvl, (int)regs->ss,
 157                regs->sp, regs->flags);
 158}
 159
 160static void show_regs_if_on_stack(struct stack_info *info, struct pt_regs *regs,
 161                                  bool partial, const char *log_lvl)
 162{
 163        /*
 164         * These on_stack() checks aren't strictly necessary: the unwind code
 165         * has already validated the 'regs' pointer.  The checks are done for
 166         * ordering reasons: if the registers are on the next stack, we don't
 167         * want to print them out yet.  Otherwise they'll be shown as part of
 168         * the wrong stack.  Later, when show_trace_log_lvl() switches to the
 169         * next stack, this function will be called again with the same regs so
 170         * they can be printed in the right context.
 171         */
 172        if (!partial && on_stack(info, regs, sizeof(*regs))) {
 173                __show_regs(regs, SHOW_REGS_SHORT, log_lvl);
 174
 175        } else if (partial && on_stack(info, (void *)regs + IRET_FRAME_OFFSET,
 176                                       IRET_FRAME_SIZE)) {
 177                /*
 178                 * When an interrupt or exception occurs in entry code, the
 179                 * full pt_regs might not have been saved yet.  In that case
 180                 * just print the iret frame.
 181                 */
 182                show_iret_regs(regs, log_lvl);
 183        }
 184}
 185
 186static void show_trace_log_lvl(struct task_struct *task, struct pt_regs *regs,
 187                        unsigned long *stack, const char *log_lvl)
 188{
 189        struct unwind_state state;
 190        struct stack_info stack_info = {0};
 191        unsigned long visit_mask = 0;
 192        int graph_idx = 0;
 193        bool partial = false;
 194
 195        printk("%sCall Trace:\n", log_lvl);
 196
 197        unwind_start(&state, task, regs, stack);
 198        stack = stack ? : get_stack_pointer(task, regs);
 199        regs = unwind_get_entry_regs(&state, &partial);
 200
 201        /*
 202         * Iterate through the stacks, starting with the current stack pointer.
 203         * Each stack has a pointer to the next one.
 204         *
 205         * x86-64 can have several stacks:
 206         * - task stack
 207         * - interrupt stack
 208         * - HW exception stacks (double fault, nmi, debug, mce)
 209         * - entry stack
 210         *
 211         * x86-32 can have up to four stacks:
 212         * - task stack
 213         * - softirq stack
 214         * - hardirq stack
 215         * - entry stack
 216         */
 217        for ( ; stack; stack = PTR_ALIGN(stack_info.next_sp, sizeof(long))) {
 218                const char *stack_name;
 219
 220                if (get_stack_info(stack, task, &stack_info, &visit_mask)) {
 221                        /*
 222                         * We weren't on a valid stack.  It's possible that
 223                         * we overflowed a valid stack into a guard page.
 224                         * See if the next page up is valid so that we can
 225                         * generate some kind of backtrace if this happens.
 226                         */
 227                        stack = (unsigned long *)PAGE_ALIGN((unsigned long)stack);
 228                        if (get_stack_info(stack, task, &stack_info, &visit_mask))
 229                                break;
 230                }
 231
 232                stack_name = stack_type_name(stack_info.type);
 233                if (stack_name)
 234                        printk("%s <%s>\n", log_lvl, stack_name);
 235
 236                if (regs)
 237                        show_regs_if_on_stack(&stack_info, regs, partial, log_lvl);
 238
 239                /*
 240                 * Scan the stack, printing any text addresses we find.  At the
 241                 * same time, follow proper stack frames with the unwinder.
 242                 *
 243                 * Addresses found during the scan which are not reported by
 244                 * the unwinder are considered to be additional clues which are
 245                 * sometimes useful for debugging and are prefixed with '?'.
 246                 * This also serves as a failsafe option in case the unwinder
 247                 * goes off in the weeds.
 248                 */
 249                for (; stack < stack_info.end; stack++) {
 250                        unsigned long real_addr;
 251                        int reliable = 0;
 252                        unsigned long addr = READ_ONCE_NOCHECK(*stack);
 253                        unsigned long *ret_addr_p =
 254                                unwind_get_return_address_ptr(&state);
 255
 256                        if (!__kernel_text_address(addr))
 257                                continue;
 258
 259                        /*
 260                         * Don't print regs->ip again if it was already printed
 261                         * by show_regs_if_on_stack().
 262                         */
 263                        if (regs && stack == &regs->ip)
 264                                goto next;
 265
 266                        if (stack == ret_addr_p)
 267                                reliable = 1;
 268
 269                        /*
 270                         * When function graph tracing is enabled for a
 271                         * function, its return address on the stack is
 272                         * replaced with the address of an ftrace handler
 273                         * (return_to_handler).  In that case, before printing
 274                         * the "real" address, we want to print the handler
 275                         * address as an "unreliable" hint that function graph
 276                         * tracing was involved.
 277                         */
 278                        real_addr = ftrace_graph_ret_addr(task, &graph_idx,
 279                                                          addr, stack);
 280                        if (real_addr != addr)
 281                                printk_stack_address(addr, 0, log_lvl);
 282                        printk_stack_address(real_addr, reliable, log_lvl);
 283
 284                        if (!reliable)
 285                                continue;
 286
 287next:
 288                        /*
 289                         * Get the next frame from the unwinder.  No need to
 290                         * check for an error: if anything goes wrong, the rest
 291                         * of the addresses will just be printed as unreliable.
 292                         */
 293                        unwind_next_frame(&state);
 294
 295                        /* if the frame has entry regs, print them */
 296                        regs = unwind_get_entry_regs(&state, &partial);
 297                        if (regs)
 298                                show_regs_if_on_stack(&stack_info, regs, partial, log_lvl);
 299                }
 300
 301                if (stack_name)
 302                        printk("%s </%s>\n", log_lvl, stack_name);
 303        }
 304}
 305
 306void show_stack(struct task_struct *task, unsigned long *sp,
 307                       const char *loglvl)
 308{
 309        task = task ? : current;
 310
 311        /*
 312         * Stack frames below this one aren't interesting.  Don't show them
 313         * if we're printing for %current.
 314         */
 315        if (!sp && task == current)
 316                sp = get_stack_pointer(current, NULL);
 317
 318        show_trace_log_lvl(task, NULL, sp, loglvl);
 319}
 320
 321void show_stack_regs(struct pt_regs *regs)
 322{
 323        show_trace_log_lvl(current, regs, NULL, KERN_DEFAULT);
 324}
 325
 326static arch_spinlock_t die_lock = __ARCH_SPIN_LOCK_UNLOCKED;
 327static int die_owner = -1;
 328static unsigned int die_nest_count;
 329
 330unsigned long oops_begin(void)
 331{
 332        int cpu;
 333        unsigned long flags;
 334
 335        oops_enter();
 336
 337        /* racy, but better than risking deadlock. */
 338        raw_local_irq_save(flags);
 339        cpu = smp_processor_id();
 340        if (!arch_spin_trylock(&die_lock)) {
 341                if (cpu == die_owner)
 342                        /* nested oops. should stop eventually */;
 343                else
 344                        arch_spin_lock(&die_lock);
 345        }
 346        die_nest_count++;
 347        die_owner = cpu;
 348        console_verbose();
 349        bust_spinlocks(1);
 350        return flags;
 351}
 352NOKPROBE_SYMBOL(oops_begin);
 353
 354void __noreturn rewind_stack_do_exit(int signr);
 355
 356void oops_end(unsigned long flags, struct pt_regs *regs, int signr)
 357{
 358        if (regs && kexec_should_crash(current))
 359                crash_kexec(regs);
 360
 361        bust_spinlocks(0);
 362        die_owner = -1;
 363        add_taint(TAINT_DIE, LOCKDEP_NOW_UNRELIABLE);
 364        die_nest_count--;
 365        if (!die_nest_count)
 366                /* Nest count reaches zero, release the lock. */
 367                arch_spin_unlock(&die_lock);
 368        raw_local_irq_restore(flags);
 369        oops_exit();
 370
 371        /* Executive summary in case the oops scrolled away */
 372        __show_regs(&exec_summary_regs, SHOW_REGS_ALL, KERN_DEFAULT);
 373
 374        if (!signr)
 375                return;
 376        if (in_interrupt())
 377                panic("Fatal exception in interrupt");
 378        if (panic_on_oops)
 379                panic("Fatal exception");
 380
 381        /*
 382         * We're not going to return, but we might be on an IST stack or
 383         * have very little stack space left.  Rewind the stack and kill
 384         * the task.
 385         * Before we rewind the stack, we have to tell KASAN that we're going to
 386         * reuse the task stack and that existing poisons are invalid.
 387         */
 388        kasan_unpoison_task_stack(current);
 389        rewind_stack_do_exit(signr);
 390}
 391NOKPROBE_SYMBOL(oops_end);
 392
 393static void __die_header(const char *str, struct pt_regs *regs, long err)
 394{
 395        const char *pr = "";
 396
 397        /* Save the regs of the first oops for the executive summary later. */
 398        if (!die_counter)
 399                exec_summary_regs = *regs;
 400
 401        if (IS_ENABLED(CONFIG_PREEMPTION))
 402                pr = IS_ENABLED(CONFIG_PREEMPT_RT) ? " PREEMPT_RT" : " PREEMPT";
 403
 404        printk(KERN_DEFAULT
 405               "%s: %04lx [#%d]%s%s%s%s%s\n", str, err & 0xffff, ++die_counter,
 406               pr,
 407               IS_ENABLED(CONFIG_SMP)     ? " SMP"             : "",
 408               debug_pagealloc_enabled()  ? " DEBUG_PAGEALLOC" : "",
 409               IS_ENABLED(CONFIG_KASAN)   ? " KASAN"           : "",
 410               IS_ENABLED(CONFIG_PAGE_TABLE_ISOLATION) ?
 411               (boot_cpu_has(X86_FEATURE_PTI) ? " PTI" : " NOPTI") : "");
 412}
 413NOKPROBE_SYMBOL(__die_header);
 414
 415static int __die_body(const char *str, struct pt_regs *regs, long err)
 416{
 417        show_regs(regs);
 418        print_modules();
 419
 420        if (notify_die(DIE_OOPS, str, regs, err,
 421                        current->thread.trap_nr, SIGSEGV) == NOTIFY_STOP)
 422                return 1;
 423
 424        return 0;
 425}
 426NOKPROBE_SYMBOL(__die_body);
 427
 428int __die(const char *str, struct pt_regs *regs, long err)
 429{
 430        __die_header(str, regs, err);
 431        return __die_body(str, regs, err);
 432}
 433NOKPROBE_SYMBOL(__die);
 434
 435/*
 436 * This is gone through when something in the kernel has done something bad
 437 * and is about to be terminated:
 438 */
 439void die(const char *str, struct pt_regs *regs, long err)
 440{
 441        unsigned long flags = oops_begin();
 442        int sig = SIGSEGV;
 443
 444        if (__die(str, regs, err))
 445                sig = 0;
 446        oops_end(flags, regs, sig);
 447}
 448
 449void die_addr(const char *str, struct pt_regs *regs, long err, long gp_addr)
 450{
 451        unsigned long flags = oops_begin();
 452        int sig = SIGSEGV;
 453
 454        __die_header(str, regs, err);
 455        if (gp_addr)
 456                kasan_non_canonical_hook(gp_addr);
 457        if (__die_body(str, regs, err))
 458                sig = 0;
 459        oops_end(flags, regs, sig);
 460}
 461
 462void show_regs(struct pt_regs *regs)
 463{
 464        enum show_regs_mode print_kernel_regs;
 465
 466        show_regs_print_info(KERN_DEFAULT);
 467
 468        print_kernel_regs = user_mode(regs) ? SHOW_REGS_USER : SHOW_REGS_ALL;
 469        __show_regs(regs, print_kernel_regs, KERN_DEFAULT);
 470
 471        /*
 472         * When in-kernel, we also print out the stack at the time of the fault..
 473         */
 474        if (!user_mode(regs))
 475                show_trace_log_lvl(current, regs, NULL, KERN_DEFAULT);
 476}
 477