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 /*
  14  * This disables KASAN checking when reading a value from another task's stack,
  15  * since the other task could be running on another CPU and could have poisoned
  16  * the stack in the meantime.
  17  */
  18 #define READ_ONCE_TASK_STACK(task, x)                   \
  19 ({                                                      \
  20         unsigned long val;                              \
  21         if (task == current)                            \
  22                 val = READ_ONCE(x);                     \
  23         else                                            \
  24                 val = READ_ONCE_NOCHECK(x);             \
  25         val;                                            \
  26 })
  27
  28 static void unwind_dump(struct unwind_state *state)
  29 {
  30         static bool dumped_before = false;
  31         bool prev_zero, zero = false;
  32         unsigned long word, *sp;
  33         struct stack_info stack_info = {0};
  34         unsigned long visit_mask = 0;
  35
  36         if (dumped_before)
  37                 return;
  38
  39         dumped_before = true;
  40
  41         printk_deferred("unwind stack type:%d next_sp:%p mask:0x%lx graph_idx:%d\n",
  42                         state->stack_info.type, state->stack_info.next_sp,
  43                         state->stack_mask, state->graph_idx);
  44
  45         for (sp = state->orig_sp; sp; sp = PTR_ALIGN(stack_info.next_sp, sizeof(long))) {
  46                 if (get_stack_info(sp, state->task, &stack_info, &visit_mask))
  47                         break;
  48
  49                 for (; sp < stack_info.end; sp++) {
  50
  51                         word = READ_ONCE_NOCHECK(*sp);
  52
  53                         prev_zero = zero;
  54                         zero = word == 0;
  55
  56                         if (zero) {
  57                                 if (!prev_zero)
  58                                         printk_deferred("%p: %0*x ...\n",
  59                                                         sp, BITS_PER_LONG/4, 0);
  60                                 continue;
  61                         }
  62
  63                         printk_deferred("%p: %0*lx (%pB)\n",
  64                                         sp, BITS_PER_LONG/4, word, (void *)word);
  65                 }
  66         }
  67 }
  68
  69 unsigned long unwind_get_return_address(struct unwind_state *state)
  70 {
  71         if (unwind_done(state))
  72                 return 0;
  73
  74         return __kernel_text_address(state->ip) ? state->ip : 0;
  75 }
  76 EXPORT_SYMBOL_GPL(unwind_get_return_address);
  77
  78 static size_t regs_size(struct pt_regs *regs)
  79 {
  80         /* x86_32 regs from kernel mode are two words shorter: */
  81         if (IS_ENABLED(CONFIG_X86_32) && !user_mode(regs))
  82                 return sizeof(*regs) - 2*sizeof(long);
  83
  84         return sizeof(*regs);
  85 }
  86
  87 static bool in_entry_code(unsigned long ip)
  88 {
  89         char *addr = (char *)ip;
  90
  91         if (addr >= __entry_text_start && addr < __entry_text_end)
  92                 return true;
  93
  94 #if defined(CONFIG_FUNCTION_GRAPH_TRACER) || defined(CONFIG_KASAN)
  95         if (addr >= __irqentry_text_start && addr < __irqentry_text_end)
  96                 return true;
  97 #endif
  98
  99         return false;
 100 }
 101
 102 static inline unsigned long *last_frame(struct unwind_state *state)
 103 {
 104         return (unsigned long *)task_pt_regs(state->task) - 2;
 105 }
 106
 107 static bool is_last_frame(struct unwind_state *state)
 108 {
 109         return state->bp == last_frame(state);
 110 }
 111
 112 #ifdef CONFIG_X86_32
 113 #define GCC_REALIGN_WORDS 3
 114 #else
 115 #define GCC_REALIGN_WORDS 1
 116 #endif
 117
 118 static inline unsigned long *last_aligned_frame(struct unwind_state *state)
 119 {
 120         return last_frame(state) - GCC_REALIGN_WORDS;
 121 }
 122
 123 static bool is_last_aligned_frame(struct unwind_state *state)
 124 {
 125         unsigned long *last_bp = last_frame(state);
 126         unsigned long *aligned_bp = last_aligned_frame(state);
 127
 128         /*
 129          * GCC can occasionally decide to realign the stack pointer and change
 130          * the offset of the stack frame in the prologue of a function called
 131          * by head/entry code.  Examples:
 132          *
 133          * <start_secondary>:
 134          *      push   %edi
 135          *      lea    0x8(%esp),%edi
 136          *      and    $0xfffffff8,%esp
 137          *      pushl  -0x4(%edi)
 138          *      push   %ebp
 139          *      mov    %esp,%ebp
 140          *
 141          * <x86_64_start_kernel>:
 142          *      lea    0x8(%rsp),%r10
 143          *      and    $0xfffffffffffffff0,%rsp
 144          *      pushq  -0x8(%r10)
 145          *      push   %rbp
 146          *      mov    %rsp,%rbp
 147          *
 148          * After aligning the stack, it pushes a duplicate copy of the return
 149          * address before pushing the frame pointer.
 150          */
 151         return (state->bp == aligned_bp && *(aligned_bp + 1) == *(last_bp + 1));
 152 }
 153
 154 static bool is_last_ftrace_frame(struct unwind_state *state)
 155 {
 156         unsigned long *last_bp = last_frame(state);
 157         unsigned long *last_ftrace_bp = last_bp - 3;
 158
 159         /*
 160          * When unwinding from an ftrace handler of a function called by entry
 161          * code, the stack layout of the last frame is:
 162          *
 163          *   bp
 164          *   parent ret addr
 165          *   bp
 166          *   function ret addr
 167          *   parent ret addr
 168          *   pt_regs
 169          *   -----------------
 170          */
 171         return (state->bp == last_ftrace_bp &&
 172                 *state->bp == *(state->bp + 2) &&
 173                 *(state->bp + 1) == *(state->bp + 4));
 174 }
 175
 176 static bool is_last_task_frame(struct unwind_state *state)
 177 {
 178         return is_last_frame(state) || is_last_aligned_frame(state) ||
 179                is_last_ftrace_frame(state);
 180 }
 181
 182 /*
 183  * This determines if the frame pointer actually contains an encoded pointer to
 184  * pt_regs on the stack.  See ENCODE_FRAME_POINTER.
 185  */
 186 static struct pt_regs *decode_frame_pointer(unsigned long *bp)
 187 {
 188         unsigned long regs = (unsigned long)bp;
 189
 190         if (!(regs & 0x1))
 191                 return NULL;
 192
 193         return (struct pt_regs *)(regs & ~0x1);
 194 }
 195
 196 static bool update_stack_state(struct unwind_state *state,
 197                                unsigned long *next_bp)
 198 {
 199         struct stack_info *info = &state->stack_info;
 200         enum stack_type prev_type = info->type;
 201         struct pt_regs *regs;
 202         unsigned long *frame, *prev_frame_end, *addr_p, addr;
 203         size_t len;
 204
 205         if (state->regs)
 206                 prev_frame_end = (void *)state->regs + regs_size(state->regs);
 207         else
 208                 prev_frame_end = (void *)state->bp + FRAME_HEADER_SIZE;
 209
 210         /* Is the next frame pointer an encoded pointer to pt_regs? */
 211         regs = decode_frame_pointer(next_bp);
 212         if (regs) {
 213                 frame = (unsigned long *)regs;
 214                 len = regs_size(regs);
 215                 state->got_irq = true;
 216         } else {
 217                 frame = next_bp;
 218                 len = FRAME_HEADER_SIZE;
 219         }
 220
 221         /*
 222          * If the next bp isn't on the current stack, switch to the next one.
 223          *
 224          * We may have to traverse multiple stacks to deal with the possibility
 225          * that info->next_sp could point to an empty stack and the next bp
 226          * could be on a subsequent stack.
 227          */
 228         while (!on_stack(info, frame, len))
 229                 if (get_stack_info(info->next_sp, state->task, info,
 230                                    &state->stack_mask))
 231                         return false;
 232
 233         /* Make sure it only unwinds up and doesn't overlap the prev frame: */
 234         if (state->orig_sp && state->stack_info.type == prev_type &&
 235             frame < prev_frame_end)
 236                 return false;
 237
 238         /* Move state to the next frame: */
 239         if (regs) {
 240                 state->regs = regs;
 241                 state->bp = NULL;
 242         } else {
 243                 state->bp = next_bp;
 244                 state->regs = NULL;
 245         }
 246
 247         /* Save the return address: */
 248         if (state->regs && user_mode(state->regs))
 249                 state->ip = 0;
 250         else {
 251                 addr_p = unwind_get_return_address_ptr(state);
 252                 addr = READ_ONCE_TASK_STACK(state->task, *addr_p);
 253                 state->ip = ftrace_graph_ret_addr(state->task, &state->graph_idx,
 254                                                   addr, addr_p);
 255         }
 256
 257         /* Save the original stack pointer for unwind_dump(): */
 258         if (!state->orig_sp)
 259                 state->orig_sp = frame;
 260
 261         return true;
 262 }
 263
 264 bool unwind_next_frame(struct unwind_state *state)
 265 {
 266         struct pt_regs *regs;
 267         unsigned long *next_bp;
 268
 269         if (unwind_done(state))
 270                 return false;
 271
 272         /* Have we reached the end? */
 273         if (state->regs && user_mode(state->regs))
 274                 goto the_end;
 275
 276         if (is_last_task_frame(state)) {
 277                 regs = task_pt_regs(state->task);
 278
 279                 /*
 280                  * kthreads (other than the boot CPU's idle thread) have some
 281                  * partial regs at the end of their stack which were placed
 282                  * there by copy_thread_tls().  But the regs don't have any
 283                  * useful information, so we can skip them.
 284                  *
 285                  * This user_mode() check is slightly broader than a PF_KTHREAD
 286                  * check because it also catches the awkward situation where a
 287                  * newly forked kthread transitions into a user task by calling
 288                  * do_execve(), which eventually clears PF_KTHREAD.
 289                  */
 290                 if (!user_mode(regs))
 291                         goto the_end;
 292
 293                 /*
 294                  * We're almost at the end, but not quite: there's still the
 295                  * syscall regs frame.  Entry code doesn't encode the regs
 296                  * pointer for syscalls, so we have to set it manually.
 297                  */
 298                 state->regs = regs;
 299                 state->bp = NULL;
 300                 state->ip = 0;
 301                 return true;
 302         }
 303
 304         /* Get the next frame pointer: */
 305         if (state->regs)
 306                 next_bp = (unsigned long *)state->regs->bp;
 307         else
 308                 next_bp = (unsigned long *)READ_ONCE_TASK_STACK(state->task, *state->bp);
 309
 310         /* Move to the next frame if it's safe: */
 311         if (!update_stack_state(state, next_bp))
 312                 goto bad_address;
 313
 314         return true;
 315
 316 bad_address:
 317         state->error = true;
 318
 319         /*
 320          * When unwinding a non-current task, the task might actually be
 321          * running on another CPU, in which case it could be modifying its
 322          * stack while we're reading it.  This is generally not a problem and
 323          * can be ignored as long as the caller understands that unwinding
 324          * another task will not always succeed.
 325          */
 326         if (state->task != current)
 327                 goto the_end;
 328
 329         /*
 330          * Don't warn if the unwinder got lost due to an interrupt in entry
 331          * code or in the C handler before the first frame pointer got set up:
 332          */
 333         if (state->got_irq && in_entry_code(state->ip))
 334                 goto the_end;
 335         if (state->regs &&
 336             state->regs->sp >= (unsigned long)last_aligned_frame(state) &&
 337             state->regs->sp < (unsigned long)task_pt_regs(state->task))
 338                 goto the_end;
 339
 340         if (state->regs) {
 341                 printk_deferred_once(KERN_WARNING
 342                         "WARNING: kernel stack regs at %p in %s:%d has bad 'bp' value %p\n",
 343                         state->regs, state->task->comm,
 344                         state->task->pid, next_bp);
 345                 unwind_dump(state);
 346         } else {
 347                 printk_deferred_once(KERN_WARNING
 348                         "WARNING: kernel stack frame pointer at %p in %s:%d has bad value %p\n",
 349                         state->bp, state->task->comm,
 350                         state->task->pid, next_bp);
 351                 unwind_dump(state);
 352         }
 353 the_end:
 354         state->stack_info.type = STACK_TYPE_UNKNOWN;
 355         return false;
 356 }
 357 EXPORT_SYMBOL_GPL(unwind_next_frame);
 358
 359 void __unwind_start(struct unwind_state *state, struct task_struct *task,
 360                     struct pt_regs *regs, unsigned long *first_frame)
 361 {
 362         unsigned long *bp;
 363
 364         memset(state, 0, sizeof(*state));
 365         state->task = task;
 366         state->got_irq = (regs);
 367
 368         /* Don't even attempt to start from user mode regs: */
 369         if (regs && user_mode(regs)) {
 370                 state->stack_info.type = STACK_TYPE_UNKNOWN;
 371                 return;
 372         }
 373
 374         bp = get_frame_pointer(task, regs);
 375
 376         /* Initialize stack info and make sure the frame data is accessible: */
 377         get_stack_info(bp, state->task, &state->stack_info,
 378                        &state->stack_mask);
 379         update_stack_state(state, bp);
 380
 381         /*
 382          * The caller can provide the address of the first frame directly
 383          * (first_frame) or indirectly (regs->sp) to indicate which stack frame
 384          * to start unwinding at.  Skip ahead until we reach it.
 385          */
 386         while (!unwind_done(state) &&
 387                (!on_stack(&state->stack_info, first_frame, sizeof(long)) ||
 388                         state->bp < first_frame))
 389                 unwind_next_frame(state);
 390 }
 391 EXPORT_SYMBOL_GPL(__unwind_start);