arch/x86/kernel/unwind_frame.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 #include <linux/sched.h>
   3 #include <linux/sched/task.h>
   4 #include <linux/sched/task_stack.h>
   5 #include <linux/interrupt.h>
   6 #include <asm/sections.h>
   7 #include <asm/ptrace.h>
   8 #include <asm/bitops.h>
   9 #include <asm/stacktrace.h>
  10 #include <asm/unwind.h>
  11
  12 #define FRAME_HEADER_SIZE (sizeof(long) * 2)
  13
  14 unsigned long unwind_get_return_address(struct unwind_state *state)
  15 {
  16         if (unwind_done(state))
  17                 return 0;
  18
  19         return __kernel_text_address(state->ip) ? state->ip : 0;
  20 }
  21 EXPORT_SYMBOL_GPL(unwind_get_return_address);
  22
  23 unsigned long *unwind_get_return_address_ptr(struct unwind_state *state)
  24 {
  25         if (unwind_done(state))
  26                 return NULL;
  27
  28         return state->regs ? &state->regs->ip : state->bp + 1;
  29 }
  30
  31 static void unwind_dump(struct unwind_state *state)
  32 {
  33         static bool dumped_before = false;
  34         bool prev_zero, zero = false;
  35         unsigned long word, *sp;
  36         struct stack_info stack_info = {0};
  37         unsigned long visit_mask = 0;
  38
  39         if (dumped_before)
  40                 return;
  41
  42         dumped_before = true;
  43
  44         printk_deferred("unwind stack type:%d next_sp:%p mask:0x%lx graph_idx:%d\n",
  45                         state->stack_info.type, state->stack_info.next_sp,
  46                         state->stack_mask, state->graph_idx);
  47
  48         for (sp = PTR_ALIGN(state->orig_sp, sizeof(long)); sp;
  49              sp = PTR_ALIGN(stack_info.next_sp, sizeof(long))) {
  50                 if (get_stack_info(sp, state->task, &stack_info, &visit_mask))
  51                         break;
  52
  53                 for (; sp < stack_info.end; sp++) {
  54
  55                         word = READ_ONCE_NOCHECK(*sp);
  56
  57                         prev_zero = zero;
  58                         zero = word == 0;
  59
  60                         if (zero) {
  61                                 if (!prev_zero)
  62                                         printk_deferred("%p: %0*x ...\n",
  63                                                         sp, BITS_PER_LONG/4, 0);
  64                                 continue;
  65                         }
  66
  67                         printk_deferred("%p: %0*lx (%pB)\n",
  68                                         sp, BITS_PER_LONG/4, word, (void *)word);
  69                 }
  70         }
  71 }
  72
  73 static bool in_entry_code(unsigned long ip)
  74 {
  75         char *addr = (char *)ip;
  76
  77         if (addr >= __entry_text_start && addr < __entry_text_end)
  78                 return true;
  79
  80         if (addr >= __irqentry_text_start && addr < __irqentry_text_end)
  81                 return true;
  82
  83         return false;
  84 }
  85
  86 static inline unsigned long *last_frame(struct unwind_state *state)
  87 {
  88         return (unsigned long *)task_pt_regs(state->task) - 2;
  89 }
  90
  91 static bool is_last_frame(struct unwind_state *state)
  92 {
  93         return state->bp == last_frame(state);
  94 }
  95
  96 #ifdef CONFIG_X86_32
  97 #define GCC_REALIGN_WORDS 3
  98 #else
  99 #define GCC_REALIGN_WORDS 1
 100 #endif
 101
 102 static inline unsigned long *last_aligned_frame(struct unwind_state *state)
 103 {
 104         return last_frame(state) - GCC_REALIGN_WORDS;
 105 }
 106
 107 static bool is_last_aligned_frame(struct unwind_state *state)
 108 {
 109         unsigned long *last_bp = last_frame(state);
 110         unsigned long *aligned_bp = last_aligned_frame(state);
 111
 112         /*
 113          * GCC can occasionally decide to realign the stack pointer and change
 114          * the offset of the stack frame in the prologue of a function called
 115          * by head/entry code.  Examples:
 116          *
 117          * <start_secondary>:
 118          *      push   %edi
 119          *      lea    0x8(%esp),%edi
 120          *      and    $0xfffffff8,%esp
 121          *      pushl  -0x4(%edi)
 122          *      push   %ebp
 123          *      mov    %esp,%ebp
 124          *
 125          * <x86_64_start_kernel>:
 126          *      lea    0x8(%rsp),%r10
 127          *      and    $0xfffffffffffffff0,%rsp
 128          *      pushq  -0x8(%r10)
 129          *      push   %rbp
 130          *      mov    %rsp,%rbp
 131          *
 132          * After aligning the stack, it pushes a duplicate copy of the return
 133          * address before pushing the frame pointer.
 134          */
 135         return (state->bp == aligned_bp && *(aligned_bp + 1) == *(last_bp + 1));
 136 }
 137
 138 static bool is_last_ftrace_frame(struct unwind_state *state)
 139 {
 140         unsigned long *last_bp = last_frame(state);
 141         unsigned long *last_ftrace_bp = last_bp - 3;
 142
 143         /*
 144          * When unwinding from an ftrace handler of a function called by entry
 145          * code, the stack layout of the last frame is:
 146          *
 147          *   bp
 148          *   parent ret addr
 149          *   bp
 150          *   function ret addr
 151          *   parent ret addr
 152          *   pt_regs
 153          *   -----------------
 154          */
 155         return (state->bp == last_ftrace_bp &&
 156                 *state->bp == *(state->bp + 2) &&
 157                 *(state->bp + 1) == *(state->bp + 4));
 158 }
 159
 160 static bool is_last_task_frame(struct unwind_state *state)
 161 {
 162         return is_last_frame(state) || is_last_aligned_frame(state) ||
 163                is_last_ftrace_frame(state);
 164 }
 165
 166 /*
 167  * This determines if the frame pointer actually contains an encoded pointer to
 168  * pt_regs on the stack.  See ENCODE_FRAME_POINTER.
 169  */
 170 #ifdef CONFIG_X86_64
 171 static struct pt_regs *decode_frame_pointer(unsigned long *bp)
 172 {
 173         unsigned long regs = (unsigned long)bp;
 174
 175         if (!(regs & 0x1))
 176                 return NULL;
 177
 178         return (struct pt_regs *)(regs & ~0x1);
 179 }
 180 #else
 181 static struct pt_regs *decode_frame_pointer(unsigned long *bp)
 182 {
 183         unsigned long regs = (unsigned long)bp;
 184
 185         if (regs & 0x80000000)
 186                 return NULL;
 187
 188         return (struct pt_regs *)(regs | 0x80000000);
 189 }
 190 #endif
 191
 192 static bool update_stack_state(struct unwind_state *state,
 193                                unsigned long *next_bp)
 194 {
 195         struct stack_info *info = &state->stack_info;
 196         enum stack_type prev_type = info->type;
 197         struct pt_regs *regs;
 198         unsigned long *frame, *prev_frame_end, *addr_p, addr;
 199         size_t len;
 200
 201         if (state->regs)
 202                 prev_frame_end = (void *)state->regs + sizeof(*state->regs);
 203         else
 204                 prev_frame_end = (void *)state->bp + FRAME_HEADER_SIZE;
 205
 206         /* Is the next frame pointer an encoded pointer to pt_regs? */
 207         regs = decode_frame_pointer(next_bp);
 208         if (regs) {
 209                 frame = (unsigned long *)regs;
 210                 len = sizeof(*regs);
 211                 state->got_irq = true;
 212         } else {
 213                 frame = next_bp;
 214                 len = FRAME_HEADER_SIZE;
 215         }
 216
 217         /*
 218          * If the next bp isn't on the current stack, switch to the next one.
 219          *
 220          * We may have to traverse multiple stacks to deal with the possibility
 221          * that info->next_sp could point to an empty stack and the next bp
 222          * could be on a subsequent stack.
 223          */
 224         while (!on_stack(info, frame, len))
 225                 if (get_stack_info(info->next_sp, state->task, info,
 226                                    &state->stack_mask))
 227                         return false;
 228
 229         /* Make sure it only unwinds up and doesn't overlap the prev frame: */
 230         if (state->orig_sp && state->stack_info.type == prev_type &&
 231             frame < prev_frame_end)
 232                 return false;
 233
 234         /* Move state to the next frame: */
 235         if (regs) {
 236                 state->regs = regs;
 237                 state->bp = NULL;
 238         } else {
 239                 state->bp = next_bp;
 240                 state->regs = NULL;
 241         }
 242
 243         /* Save the return address: */
 244         if (state->regs && user_mode(state->regs))
 245                 state->ip = 0;
 246         else {
 247                 addr_p = unwind_get_return_address_ptr(state);
 248                 addr = READ_ONCE_TASK_STACK(state->task, *addr_p);
 249                 state->ip = ftrace_graph_ret_addr(state->task, &state->graph_idx,
 250                                                   addr, addr_p);
 251         }
 252
 253         /* Save the original stack pointer for unwind_dump(): */
 254         if (!state->orig_sp)
 255                 state->orig_sp = frame;
 256
 257         return true;
 258 }
 259
 260 bool unwind_next_frame(struct unwind_state *state)
 261 {
 262         struct pt_regs *regs;
 263         unsigned long *next_bp;
 264
 265         if (unwind_done(state))
 266                 return false;
 267
 268         /* Have we reached the end? */
 269         if (state->regs && user_mode(state->regs))
 270                 goto the_end;
 271
 272         if (is_last_task_frame(state)) {
 273                 regs = task_pt_regs(state->task);
 274
 275                 /*
 276                  * kthreads (other than the boot CPU's idle thread) have some
 277                  * partial regs at the end of their stack which were placed
 278                  * there by copy_thread_tls().  But the regs don't have any
 279                  * useful information, so we can skip them.
 280                  *
 281                  * This user_mode() check is slightly broader than a PF_KTHREAD
 282                  * check because it also catches the awkward situation where a
 283                  * newly forked kthread transitions into a user task by calling
 284                  * do_execve(), which eventually clears PF_KTHREAD.
 285                  */
 286                 if (!user_mode(regs))
 287                         goto the_end;
 288
 289                 /*
 290                  * We're almost at the end, but not quite: there's still the
 291                  * syscall regs frame.  Entry code doesn't encode the regs
 292                  * pointer for syscalls, so we have to set it manually.
 293                  */
 294                 state->regs = regs;
 295                 state->bp = NULL;
 296                 state->ip = 0;
 297                 return true;
 298         }
 299
 300         /* Get the next frame pointer: */
 301         if (state->next_bp) {
 302                 next_bp = state->next_bp;
 303                 state->next_bp = NULL;
 304         } else if (state->regs) {
 305                 next_bp = (unsigned long *)state->regs->bp;
 306         } else {
 307                 next_bp = (unsigned long *)READ_ONCE_TASK_STACK(state->task, *state->bp);
 308         }
 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         /*
 341          * There are some known frame pointer issues on 32-bit.  Disable
 342          * unwinder warnings on 32-bit until it gets objtool support.
 343          */
 344         if (IS_ENABLED(CONFIG_X86_32))
 345                 goto the_end;
 346
 347         if (state->regs) {
 348                 printk_deferred_once(KERN_WARNING
 349                         "WARNING: kernel stack regs at %p in %s:%d has bad 'bp' value %p\n",
 350                         state->regs, state->task->comm,
 351                         state->task->pid, next_bp);
 352                 unwind_dump(state);
 353         } else {
 354                 printk_deferred_once(KERN_WARNING
 355                         "WARNING: kernel stack frame pointer at %p in %s:%d has bad value %p\n",
 356                         state->bp, state->task->comm,
 357                         state->task->pid, next_bp);
 358                 unwind_dump(state);
 359         }
 360 the_end:
 361         state->stack_info.type = STACK_TYPE_UNKNOWN;
 362         return false;
 363 }
 364 EXPORT_SYMBOL_GPL(unwind_next_frame);
 365
 366 void __unwind_start(struct unwind_state *state, struct task_struct *task,
 367                     struct pt_regs *regs, unsigned long *first_frame)
 368 {
 369         unsigned long *bp;
 370
 371         memset(state, 0, sizeof(*state));
 372         state->task = task;
 373         state->got_irq = (regs);
 374
 375         /* Don't even attempt to start from user mode regs: */
 376         if (regs && user_mode(regs)) {
 377                 state->stack_info.type = STACK_TYPE_UNKNOWN;
 378                 return;
 379         }
 380
 381         bp = get_frame_pointer(task, regs);
 382
 383         /*
 384          * If we crash with IP==0, the last successfully executed instruction
 385          * was probably an indirect function call with a NULL function pointer.
 386          * That means that SP points into the middle of an incomplete frame:
 387          * *SP is a return pointer, and *(SP-sizeof(unsigned long)) is where we
 388          * would have written a frame pointer if we hadn't crashed.
 389          * Pretend that the frame is complete and that BP points to it, but save
 390          * the real BP so that we can use it when looking for the next frame.
 391          */
 392         if (regs && regs->ip == 0 && (unsigned long *)regs->sp >= first_frame) {
 393                 state->next_bp = bp;
 394                 bp = ((unsigned long *)regs->sp) - 1;
 395         }
 396
 397         /* Initialize stack info and make sure the frame data is accessible: */
 398         get_stack_info(bp, state->task, &state->stack_info,
 399                        &state->stack_mask);
 400         update_stack_state(state, bp);
 401
 402         /*
 403          * The caller can provide the address of the first frame directly
 404          * (first_frame) or indirectly (regs->sp) to indicate which stack frame
 405          * to start unwinding at.  Skip ahead until we reach it.
 406          */
 407         while (!unwind_done(state) &&
 408                (!on_stack(&state->stack_info, first_frame, sizeof(long)) ||
 409                         (state->next_bp == NULL && state->bp < first_frame)))
 410                 unwind_next_frame(state);
 411 }
 412 EXPORT_SYMBOL_GPL(__unwind_start);