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