arch/x86/kernel/unwind_frame.c

   1 #include <linux/sched.h>
   2 #include <linux/sched/task.h>
   3 #include <linux/sched/task_stack.h>
   4 #include <linux/interrupt.h>
   5 #include <asm/sections.h>
   6 #include <asm/ptrace.h>
   7 #include <asm/bitops.h>
   8 #include <asm/stacktrace.h>
   9 #include <asm/unwind.h>
  10
  11 #define FRAME_HEADER_SIZE (sizeof(long) * 2)
  12
  13 unsigned long unwind_get_return_address(struct unwind_state *state)
  14 {
  15         if (unwind_done(state))
  16                 return 0;
  17
  18         return __kernel_text_address(state->ip) ? state->ip : 0;
  19 }
  20 EXPORT_SYMBOL_GPL(unwind_get_return_address);
  21
  22 unsigned long *unwind_get_return_address_ptr(struct unwind_state *state)
  23 {
  24         if (unwind_done(state))
  25                 return NULL;
  26
  27         return state->regs ? &state->regs->ip : state->bp + 1;
  28 }
  29
  30 static void unwind_dump(struct unwind_state *state)
  31 {
  32         static bool dumped_before = false;
  33         bool prev_zero, zero = false;
  34         unsigned long word, *sp;
  35         struct stack_info stack_info = {0};
  36         unsigned long visit_mask = 0;
  37
  38         if (dumped_before)
  39                 return;
  40
  41         dumped_before = true;
  42
  43         printk_deferred("unwind stack type:%d next_sp:%p mask:0x%lx graph_idx:%d\n",
  44                         state->stack_info.type, state->stack_info.next_sp,
  45                         state->stack_mask, state->graph_idx);
  46
  47         for (sp = state->orig_sp; sp; sp = PTR_ALIGN(stack_info.next_sp, sizeof(long))) {
  48                 if (get_stack_info(sp, state->task, &stack_info, &visit_mask))
  49                         break;
  50
  51                 for (; sp < stack_info.end; sp++) {
  52
  53                         word = READ_ONCE_NOCHECK(*sp);
  54
  55                         prev_zero = zero;
  56                         zero = word == 0;
  57
  58                         if (zero) {
  59                                 if (!prev_zero)
  60                                         printk_deferred("%p: %0*x ...\n",
  61                                                         sp, BITS_PER_LONG/4, 0);
  62                                 continue;
  63                         }
  64
  65                         printk_deferred("%p: %0*lx (%pB)\n",
  66                                         sp, BITS_PER_LONG/4, word, (void *)word);
  67                 }
  68         }
  69 }
  70
  71 static size_t regs_size(struct pt_regs *regs)
  72 {
  73         /* x86_32 regs from kernel mode are two words shorter: */
  74         if (IS_ENABLED(CONFIG_X86_32) && !user_mode(regs))
  75                 return sizeof(*regs) - 2*sizeof(long);
  76
  77         return sizeof(*regs);
  78 }
  79
  80 static bool in_entry_code(unsigned long ip)
  81 {
  82         char *addr = (char *)ip;
  83
  84         if (addr >= __entry_text_start && addr < __entry_text_end)
  85                 return true;
  86
  87         if (addr >= __irqentry_text_start && addr < __irqentry_text_end)
  88                 return true;
  89
  90         return false;
  91 }
  92
  93 static inline unsigned long *last_frame(struct unwind_state *state)
  94 {
  95         return (unsigned long *)task_pt_regs(state->task) - 2;
  96 }
  97
  98 static bool is_last_frame(struct unwind_state *state)
  99 {
 100         return state->bp == last_frame(state);
 101 }
 102
 103 #ifdef CONFIG_X86_32
 104 #define GCC_REALIGN_WORDS 3
 105 #else
 106 #define GCC_REALIGN_WORDS 1
 107 #endif
 108
 109 static inline unsigned long *last_aligned_frame(struct unwind_state *state)
 110 {
 111         return last_frame(state) - GCC_REALIGN_WORDS;
 112 }
 113
 114 static bool is_last_aligned_frame(struct unwind_state *state)
 115 {
 116         unsigned long *last_bp = last_frame(state);
 117         unsigned long *aligned_bp = last_aligned_frame(state);
 118
 119         /*
 120          * GCC can occasionally decide to realign the stack pointer and change
 121          * the offset of the stack frame in the prologue of a function called
 122          * by head/entry code.  Examples:
 123          *
 124          * <start_secondary>:
 125          *      push   %edi
 126          *      lea    0x8(%esp),%edi
 127          *      and    $0xfffffff8,%esp
 128          *      pushl  -0x4(%edi)
 129          *      push   %ebp
 130          *      mov    %esp,%ebp
 131          *
 132          * <x86_64_start_kernel>:
 133          *      lea    0x8(%rsp),%r10
 134          *      and    $0xfffffffffffffff0,%rsp
 135          *      pushq  -0x8(%r10)
 136          *      push   %rbp
 137          *      mov    %rsp,%rbp
 138          *
 139          * After aligning the stack, it pushes a duplicate copy of the return
 140          * address before pushing the frame pointer.
 141          */
 142         return (state->bp == aligned_bp && *(aligned_bp + 1) == *(last_bp + 1));
 143 }
 144
 145 static bool is_last_ftrace_frame(struct unwind_state *state)
 146 {
 147         unsigned long *last_bp = last_frame(state);
 148         unsigned long *last_ftrace_bp = last_bp - 3;
 149
 150         /*
 151          * When unwinding from an ftrace handler of a function called by entry
 152          * code, the stack layout of the last frame is:
 153          *
 154          *   bp
 155          *   parent ret addr
 156          *   bp
 157          *   function ret addr
 158          *   parent ret addr
 159          *   pt_regs
 160          *   -----------------
 161          */
 162         return (state->bp == last_ftrace_bp &&
 163                 *state->bp == *(state->bp + 2) &&
 164                 *(state->bp + 1) == *(state->bp + 4));
 165 }
 166
 167 static bool is_last_task_frame(struct unwind_state *state)
 168 {
 169         return is_last_frame(state) || is_last_aligned_frame(state) ||
 170                is_last_ftrace_frame(state);
 171 }
 172
 173 /*
 174  * This determines if the frame pointer actually contains an encoded pointer to
 175  * pt_regs on the stack.  See ENCODE_FRAME_POINTER.
 176  */
 177 static struct pt_regs *decode_frame_pointer(unsigned long *bp)
 178 {
 179         unsigned long regs = (unsigned long)bp;
 180
 181         if (!(regs & 0x1))
 182                 return NULL;
 183
 184         return (struct pt_regs *)(regs & ~0x1);
 185 }
 186
 187 static bool update_stack_state(struct unwind_state *state,
 188                                unsigned long *next_bp)
 189 {
 190         struct stack_info *info = &state->stack_info;
 191         enum stack_type prev_type = info->type;
 192         struct pt_regs *regs;
 193         unsigned long *frame, *prev_frame_end, *addr_p, addr;
 194         size_t len;
 195
 196         if (state->regs)
 197                 prev_frame_end = (void *)state->regs + regs_size(state->regs);
 198         else
 199                 prev_frame_end = (void *)state->bp + FRAME_HEADER_SIZE;
 200
 201         /* Is the next frame pointer an encoded pointer to pt_regs? */
 202         regs = decode_frame_pointer(next_bp);
 203         if (regs) {
 204                 frame = (unsigned long *)regs;
 205                 len = regs_size(regs);
 206                 state->got_irq = true;
 207         } else {
 208                 frame = next_bp;
 209                 len = FRAME_HEADER_SIZE;
 210         }
 211
 212         /*
 213          * If the next bp isn't on the current stack, switch to the next one.
 214          *
 215          * We may have to traverse multiple stacks to deal with the possibility
 216          * that info->next_sp could point to an empty stack and the next bp
 217          * could be on a subsequent stack.
 218          */
 219         while (!on_stack(info, frame, len))
 220                 if (get_stack_info(info->next_sp, state->task, info,
 221                                    &state->stack_mask))
 222                         return false;
 223
 224         /* Make sure it only unwinds up and doesn't overlap the prev frame: */
 225         if (state->orig_sp && state->stack_info.type == prev_type &&
 226             frame < prev_frame_end)
 227                 return false;
 228
 229         /* Move state to the next frame: */
 230         if (regs) {
 231                 state->regs = regs;
 232                 state->bp = NULL;
 233         } else {
 234                 state->bp = next_bp;
 235                 state->regs = NULL;
 236         }
 237
 238         /* Save the return address: */
 239         if (state->regs && user_mode(state->regs))
 240                 state->ip = 0;
 241         else {
 242                 addr_p = unwind_get_return_address_ptr(state);
 243                 addr = READ_ONCE_TASK_STACK(state->task, *addr_p);
 244                 state->ip = ftrace_graph_ret_addr(state->task, &state->graph_idx,
 245                                                   addr, addr_p);
 246         }
 247
 248         /* Save the original stack pointer for unwind_dump(): */
 249         if (!state->orig_sp)
 250                 state->orig_sp = frame;
 251
 252         return true;
 253 }
 254
 255 bool unwind_next_frame(struct unwind_state *state)
 256 {
 257         struct pt_regs *regs;
 258         unsigned long *next_bp;
 259
 260         if (unwind_done(state))
 261                 return false;
 262
 263         /* Have we reached the end? */
 264         if (state->regs && user_mode(state->regs))
 265                 goto the_end;
 266
 267         if (is_last_task_frame(state)) {
 268                 regs = task_pt_regs(state->task);
 269
 270                 /*
 271                  * kthreads (other than the boot CPU's idle thread) have some
 272                  * partial regs at the end of their stack which were placed
 273                  * there by copy_thread_tls().  But the regs don't have any
 274                  * useful information, so we can skip them.
 275                  *
 276                  * This user_mode() check is slightly broader than a PF_KTHREAD
 277                  * check because it also catches the awkward situation where a
 278                  * newly forked kthread transitions into a user task by calling
 279                  * do_execve(), which eventually clears PF_KTHREAD.
 280                  */
 281                 if (!user_mode(regs))
 282                         goto the_end;
 283
 284                 /*
 285                  * We're almost at the end, but not quite: there's still the
 286                  * syscall regs frame.  Entry code doesn't encode the regs
 287                  * pointer for syscalls, so we have to set it manually.
 288                  */
 289                 state->regs = regs;
 290                 state->bp = NULL;
 291                 state->ip = 0;
 292                 return true;
 293         }
 294
 295         /* Get the next frame pointer: */
 296         if (state->regs)
 297                 next_bp = (unsigned long *)state->regs->bp;
 298         else
 299                 next_bp = (unsigned long *)READ_ONCE_TASK_STACK(state->task, *state->bp);
 300
 301         /* Move to the next frame if it's safe: */
 302         if (!update_stack_state(state, next_bp))
 303                 goto bad_address;
 304
 305         return true;
 306
 307 bad_address:
 308         state->error = true;
 309
 310         /*
 311          * When unwinding a non-current task, the task might actually be
 312          * running on another CPU, in which case it could be modifying its
 313          * stack while we're reading it.  This is generally not a problem and
 314          * can be ignored as long as the caller understands that unwinding
 315          * another task will not always succeed.
 316          */
 317         if (state->task != current)
 318                 goto the_end;
 319
 320         /*
 321          * Don't warn if the unwinder got lost due to an interrupt in entry
 322          * code or in the C handler before the first frame pointer got set up:
 323          */
 324         if (state->got_irq && in_entry_code(state->ip))
 325                 goto the_end;
 326         if (state->regs &&
 327             state->regs->sp >= (unsigned long)last_aligned_frame(state) &&
 328             state->regs->sp < (unsigned long)task_pt_regs(state->task))
 329                 goto the_end;
 330
 331         if (state->regs) {
 332                 printk_deferred_once(KERN_WARNING
 333                         "WARNING: kernel stack regs at %p in %s:%d has bad 'bp' value %p\n",
 334                         state->regs, state->task->comm,
 335                         state->task->pid, next_bp);
 336                 unwind_dump(state);
 337         } else {
 338                 printk_deferred_once(KERN_WARNING
 339                         "WARNING: kernel stack frame pointer at %p in %s:%d has bad value %p\n",
 340                         state->bp, state->task->comm,
 341                         state->task->pid, next_bp);
 342                 unwind_dump(state);
 343         }
 344 the_end:
 345         state->stack_info.type = STACK_TYPE_UNKNOWN;
 346         return false;
 347 }
 348 EXPORT_SYMBOL_GPL(unwind_next_frame);
 349
 350 void __unwind_start(struct unwind_state *state, struct task_struct *task,
 351                     struct pt_regs *regs, unsigned long *first_frame)
 352 {
 353         unsigned long *bp;
 354
 355         memset(state, 0, sizeof(*state));
 356         state->task = task;
 357         state->got_irq = (regs);
 358
 359         /* Don't even attempt to start from user mode regs: */
 360         if (regs && user_mode(regs)) {
 361                 state->stack_info.type = STACK_TYPE_UNKNOWN;
 362                 return;
 363         }
 364
 365         bp = get_frame_pointer(task, regs);
 366
 367         /* Initialize stack info and make sure the frame data is accessible: */
 368         get_stack_info(bp, state->task, &state->stack_info,
 369                        &state->stack_mask);
 370         update_stack_state(state, bp);
 371
 372         /*
 373          * The caller can provide the address of the first frame directly
 374          * (first_frame) or indirectly (regs->sp) to indicate which stack frame
 375          * to start unwinding at.  Skip ahead until we reach it.
 376          */
 377         while (!unwind_done(state) &&
 378                (!on_stack(&state->stack_info, first_frame, sizeof(long)) ||
 379                         state->bp < first_frame))
 380                 unwind_next_frame(state);
 381 }
 382 EXPORT_SYMBOL_GPL(__unwind_start);