arch/x86/kernel/kprobes/opt.c

   1 /*
   2  *  Kernel Probes Jump Optimization (Optprobes)
   3  *
   4  * This program is free software; you can redistribute it and/or modify
   5  * it under the terms of the GNU General Public License as published by
   6  * the Free Software Foundation; either version 2 of the License, or
   7  * (at your option) any later version.
   8  *
   9  * This program is distributed in the hope that it will be useful,
  10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12  * GNU General Public License for more details.
  13  *
  14  * You should have received a copy of the GNU General Public License
  15  * along with this program; if not, write to the Free Software
  16  * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
  17  *
  18  * Copyright (C) IBM Corporation, 2002, 2004
  19  * Copyright (C) Hitachi Ltd., 2012
  20  */
  21 #include <linux/kprobes.h>
  22 #include <linux/ptrace.h>
  23 #include <linux/string.h>
  24 #include <linux/slab.h>
  25 #include <linux/hardirq.h>
  26 #include <linux/preempt.h>
  27 #include <linux/extable.h>
  28 #include <linux/kdebug.h>
  29 #include <linux/kallsyms.h>
  30 #include <linux/ftrace.h>
  31
  32 #include <asm/text-patching.h>
  33 #include <asm/cacheflush.h>
  34 #include <asm/desc.h>
  35 #include <asm/pgtable.h>
  36 #include <linux/uaccess.h>
  37 #include <asm/alternative.h>
  38 #include <asm/insn.h>
  39 #include <asm/debugreg.h>
  40 #include <asm/set_memory.h>
  41
  42 #include "common.h"
  43
  44 unsigned long __recover_optprobed_insn(kprobe_opcode_t *buf, unsigned long addr)
  45 {
  46         struct optimized_kprobe *op;
  47         struct kprobe *kp;
  48         long offs;
  49         int i;
  50
  51         for (i = 0; i < RELATIVEJUMP_SIZE; i++) {
  52                 kp = get_kprobe((void *)addr - i);
  53                 /* This function only handles jump-optimized kprobe */
  54                 if (kp && kprobe_optimized(kp)) {
  55                         op = container_of(kp, struct optimized_kprobe, kp);
  56                         /* If op->list is not empty, op is under optimizing */
  57                         if (list_empty(&op->list))
  58                                 goto found;
  59                 }
  60         }
  61
  62         return addr;
  63 found:
  64         /*
  65          * If the kprobe can be optimized, original bytes which can be
  66          * overwritten by jump destination address. In this case, original
  67          * bytes must be recovered from op->optinsn.copied_insn buffer.
  68          */
  69         if (probe_kernel_read(buf, (void *)addr,
  70                 MAX_INSN_SIZE * sizeof(kprobe_opcode_t)))
  71                 return 0UL;
  72
  73         if (addr == (unsigned long)kp->addr) {
  74                 buf[0] = kp->opcode;
  75                 memcpy(buf + 1, op->optinsn.copied_insn, RELATIVE_ADDR_SIZE);
  76         } else {
  77                 offs = addr - (unsigned long)kp->addr - 1;
  78                 memcpy(buf, op->optinsn.copied_insn + offs, RELATIVE_ADDR_SIZE - offs);
  79         }
  80
  81         return (unsigned long)buf;
  82 }
  83
  84 /* Insert a move instruction which sets a pointer to eax/rdi (1st arg). */
  85 static void synthesize_set_arg1(kprobe_opcode_t *addr, unsigned long val)
  86 {
  87 #ifdef CONFIG_X86_64
  88         *addr++ = 0x48;
  89         *addr++ = 0xbf;
  90 #else
  91         *addr++ = 0xb8;
  92 #endif
  93         *(unsigned long *)addr = val;
  94 }
  95
  96 asm (
  97                         ".global optprobe_template_entry\n"
  98                         "optprobe_template_entry:\n"
  99 #ifdef CONFIG_X86_64
 100                         /* We don't bother saving the ss register */
 101                         "       pushq %rsp\n"
 102                         "       pushfq\n"
 103                         SAVE_REGS_STRING
 104                         "       movq %rsp, %rsi\n"
 105                         ".global optprobe_template_val\n"
 106                         "optprobe_template_val:\n"
 107                         ASM_NOP5
 108                         ASM_NOP5
 109                         ".global optprobe_template_call\n"
 110                         "optprobe_template_call:\n"
 111                         ASM_NOP5
 112                         /* Move flags to rsp */
 113                         "       movq 144(%rsp), %rdx\n"
 114                         "       movq %rdx, 152(%rsp)\n"
 115                         RESTORE_REGS_STRING
 116                         /* Skip flags entry */
 117                         "       addq $8, %rsp\n"
 118                         "       popfq\n"
 119 #else /* CONFIG_X86_32 */
 120                         "       pushf\n"
 121                         SAVE_REGS_STRING
 122                         "       movl %esp, %edx\n"
 123                         ".global optprobe_template_val\n"
 124                         "optprobe_template_val:\n"
 125                         ASM_NOP5
 126                         ".global optprobe_template_call\n"
 127                         "optprobe_template_call:\n"
 128                         ASM_NOP5
 129                         RESTORE_REGS_STRING
 130                         "       addl $4, %esp\n"        /* skip cs */
 131                         "       popf\n"
 132 #endif
 133                         ".global optprobe_template_end\n"
 134                         "optprobe_template_end:\n");
 135
 136 #define TMPL_MOVE_IDX \
 137         ((long)&optprobe_template_val - (long)&optprobe_template_entry)
 138 #define TMPL_CALL_IDX \
 139         ((long)&optprobe_template_call - (long)&optprobe_template_entry)
 140 #define TMPL_END_IDX \
 141         ((long)&optprobe_template_end - (long)&optprobe_template_entry)
 142
 143 #define INT3_SIZE sizeof(kprobe_opcode_t)
 144
 145 /* Optimized kprobe call back function: called from optinsn */
 146 static void
 147 optimized_callback(struct optimized_kprobe *op, struct pt_regs *regs)
 148 {
 149         struct kprobe_ctlblk *kcb = get_kprobe_ctlblk();
 150         unsigned long flags;
 151
 152         /* This is possible if op is under delayed unoptimizing */
 153         if (kprobe_disabled(&op->kp))
 154                 return;
 155
 156         local_irq_save(flags);
 157         if (kprobe_running()) {
 158                 kprobes_inc_nmissed_count(&op->kp);
 159         } else {
 160                 /* Save skipped registers */
 161 #ifdef CONFIG_X86_64
 162                 regs->cs = __KERNEL_CS;
 163 #else
 164                 regs->cs = __KERNEL_CS | get_kernel_rpl();
 165                 regs->gs = 0;
 166 #endif
 167                 regs->ip = (unsigned long)op->kp.addr + INT3_SIZE;
 168                 regs->orig_ax = ~0UL;
 169
 170                 __this_cpu_write(current_kprobe, &op->kp);
 171                 kcb->kprobe_status = KPROBE_HIT_ACTIVE;
 172                 opt_pre_handler(&op->kp, regs);
 173                 __this_cpu_write(current_kprobe, NULL);
 174         }
 175         local_irq_restore(flags);
 176 }
 177 NOKPROBE_SYMBOL(optimized_callback);
 178
 179 static int copy_optimized_instructions(u8 *dest, u8 *src)
 180 {
 181         struct insn insn;
 182         int len = 0, ret;
 183
 184         while (len < RELATIVEJUMP_SIZE) {
 185                 ret = __copy_instruction(dest + len, src + len, &insn);
 186                 if (!ret || !can_boost(&insn, src + len))
 187                         return -EINVAL;
 188                 len += ret;
 189         }
 190         /* Check whether the address range is reserved */
 191         if (ftrace_text_reserved(src, src + len - 1) ||
 192             alternatives_text_reserved(src, src + len - 1) ||
 193             jump_label_text_reserved(src, src + len - 1))
 194                 return -EBUSY;
 195
 196         return len;
 197 }
 198
 199 /* Check whether insn is indirect jump */
 200 static int insn_is_indirect_jump(struct insn *insn)
 201 {
 202         return ((insn->opcode.bytes[0] == 0xff &&
 203                 (X86_MODRM_REG(insn->modrm.value) & 6) == 4) || /* Jump */
 204                 insn->opcode.bytes[0] == 0xea); /* Segment based jump */
 205 }
 206
 207 /* Check whether insn jumps into specified address range */
 208 static int insn_jump_into_range(struct insn *insn, unsigned long start, int len)
 209 {
 210         unsigned long target = 0;
 211
 212         switch (insn->opcode.bytes[0]) {
 213         case 0xe0:      /* loopne */
 214         case 0xe1:      /* loope */
 215         case 0xe2:      /* loop */
 216         case 0xe3:      /* jcxz */
 217         case 0xe9:      /* near relative jump */
 218         case 0xeb:      /* short relative jump */
 219                 break;
 220         case 0x0f:
 221                 if ((insn->opcode.bytes[1] & 0xf0) == 0x80) /* jcc near */
 222                         break;
 223                 return 0;
 224         default:
 225                 if ((insn->opcode.bytes[0] & 0xf0) == 0x70) /* jcc short */
 226                         break;
 227                 return 0;
 228         }
 229         target = (unsigned long)insn->next_byte + insn->immediate.value;
 230
 231         return (start <= target && target <= start + len);
 232 }
 233
 234 /* Decode whole function to ensure any instructions don't jump into target */
 235 static int can_optimize(unsigned long paddr)
 236 {
 237         unsigned long addr, size = 0, offset = 0;
 238         struct insn insn;
 239         kprobe_opcode_t buf[MAX_INSN_SIZE];
 240
 241         /* Lookup symbol including addr */
 242         if (!kallsyms_lookup_size_offset(paddr, &size, &offset))
 243                 return 0;
 244
 245         /*
 246          * Do not optimize in the entry code due to the unstable
 247          * stack handling.
 248          */
 249         if ((paddr >= (unsigned long)__entry_text_start) &&
 250             (paddr <  (unsigned long)__entry_text_end))
 251                 return 0;
 252
 253         /* Check there is enough space for a relative jump. */
 254         if (size - offset < RELATIVEJUMP_SIZE)
 255                 return 0;
 256
 257         /* Decode instructions */
 258         addr = paddr - offset;
 259         while (addr < paddr - offset + size) { /* Decode until function end */
 260                 unsigned long recovered_insn;
 261                 if (search_exception_tables(addr))
 262                         /*
 263                          * Since some fixup code will jumps into this function,
 264                          * we can't optimize kprobe in this function.
 265                          */
 266                         return 0;
 267                 recovered_insn = recover_probed_instruction(buf, addr);
 268                 if (!recovered_insn)
 269                         return 0;
 270                 kernel_insn_init(&insn, (void *)recovered_insn, MAX_INSN_SIZE);
 271                 insn_get_length(&insn);
 272                 /* Another subsystem puts a breakpoint */
 273                 if (insn.opcode.bytes[0] == BREAKPOINT_INSTRUCTION)
 274                         return 0;
 275                 /* Recover address */
 276                 insn.kaddr = (void *)addr;
 277                 insn.next_byte = (void *)(addr + insn.length);
 278                 /* Check any instructions don't jump into target */
 279                 if (insn_is_indirect_jump(&insn) ||
 280                     insn_jump_into_range(&insn, paddr + INT3_SIZE,
 281                                          RELATIVE_ADDR_SIZE))
 282                         return 0;
 283                 addr += insn.length;
 284         }
 285
 286         return 1;
 287 }
 288
 289 /* Check optimized_kprobe can actually be optimized. */
 290 int arch_check_optimized_kprobe(struct optimized_kprobe *op)
 291 {
 292         int i;
 293         struct kprobe *p;
 294
 295         for (i = 1; i < op->optinsn.size; i++) {
 296                 p = get_kprobe(op->kp.addr + i);
 297                 if (p && !kprobe_disabled(p))
 298                         return -EEXIST;
 299         }
 300
 301         return 0;
 302 }
 303
 304 /* Check the addr is within the optimized instructions. */
 305 int arch_within_optimized_kprobe(struct optimized_kprobe *op,
 306                                  unsigned long addr)
 307 {
 308         return ((unsigned long)op->kp.addr <= addr &&
 309                 (unsigned long)op->kp.addr + op->optinsn.size > addr);
 310 }
 311
 312 /* Free optimized instruction slot */
 313 static
 314 void __arch_remove_optimized_kprobe(struct optimized_kprobe *op, int dirty)
 315 {
 316         if (op->optinsn.insn) {
 317                 free_optinsn_slot(op->optinsn.insn, dirty);
 318                 op->optinsn.insn = NULL;
 319                 op->optinsn.size = 0;
 320         }
 321 }
 322
 323 void arch_remove_optimized_kprobe(struct optimized_kprobe *op)
 324 {
 325         __arch_remove_optimized_kprobe(op, 1);
 326 }
 327
 328 /*
 329  * Copy replacing target instructions
 330  * Target instructions MUST be relocatable (checked inside)
 331  * This is called when new aggr(opt)probe is allocated or reused.
 332  */
 333 int arch_prepare_optimized_kprobe(struct optimized_kprobe *op,
 334                                   struct kprobe *__unused)
 335 {
 336         u8 *buf;
 337         int ret;
 338         long rel;
 339
 340         if (!can_optimize((unsigned long)op->kp.addr))
 341                 return -EILSEQ;
 342
 343         op->optinsn.insn = get_optinsn_slot();
 344         if (!op->optinsn.insn)
 345                 return -ENOMEM;
 346
 347         /*
 348          * Verify if the address gap is in 2GB range, because this uses
 349          * a relative jump.
 350          */
 351         rel = (long)op->optinsn.insn - (long)op->kp.addr + RELATIVEJUMP_SIZE;
 352         if (abs(rel) > 0x7fffffff) {
 353                 __arch_remove_optimized_kprobe(op, 0);
 354                 return -ERANGE;
 355         }
 356
 357         buf = (u8 *)op->optinsn.insn;
 358         set_memory_rw((unsigned long)buf & PAGE_MASK, 1);
 359
 360         /* Copy instructions into the out-of-line buffer */
 361         ret = copy_optimized_instructions(buf + TMPL_END_IDX, op->kp.addr);
 362         if (ret < 0) {
 363                 __arch_remove_optimized_kprobe(op, 0);
 364                 return ret;
 365         }
 366         op->optinsn.size = ret;
 367
 368         /* Copy arch-dep-instance from template */
 369         memcpy(buf, &optprobe_template_entry, TMPL_END_IDX);
 370
 371         /* Set probe information */
 372         synthesize_set_arg1(buf + TMPL_MOVE_IDX, (unsigned long)op);
 373
 374         /* Set probe function call */
 375         synthesize_relcall(buf + TMPL_CALL_IDX, optimized_callback);
 376
 377         /* Set returning jmp instruction at the tail of out-of-line buffer */
 378         synthesize_reljump(buf + TMPL_END_IDX + op->optinsn.size,
 379                            (u8 *)op->kp.addr + op->optinsn.size);
 380
 381         set_memory_ro((unsigned long)buf & PAGE_MASK, 1);
 382
 383         flush_icache_range((unsigned long) buf,
 384                            (unsigned long) buf + TMPL_END_IDX +
 385                            op->optinsn.size + RELATIVEJUMP_SIZE);
 386         return 0;
 387 }
 388
 389 /*
 390  * Replace breakpoints (int3) with relative jumps.
 391  * Caller must call with locking kprobe_mutex and text_mutex.
 392  */
 393 void arch_optimize_kprobes(struct list_head *oplist)
 394 {
 395         struct optimized_kprobe *op, *tmp;
 396         u8 insn_buf[RELATIVEJUMP_SIZE];
 397
 398         list_for_each_entry_safe(op, tmp, oplist, list) {
 399                 s32 rel = (s32)((long)op->optinsn.insn -
 400                         ((long)op->kp.addr + RELATIVEJUMP_SIZE));
 401
 402                 WARN_ON(kprobe_disabled(&op->kp));
 403
 404                 /* Backup instructions which will be replaced by jump address */
 405                 memcpy(op->optinsn.copied_insn, op->kp.addr + INT3_SIZE,
 406                        RELATIVE_ADDR_SIZE);
 407
 408                 insn_buf[0] = RELATIVEJUMP_OPCODE;
 409                 *(s32 *)(&insn_buf[1]) = rel;
 410
 411                 text_poke_bp(op->kp.addr, insn_buf, RELATIVEJUMP_SIZE,
 412                              op->optinsn.insn);
 413
 414                 list_del_init(&op->list);
 415         }
 416 }
 417
 418 /* Replace a relative jump with a breakpoint (int3).  */
 419 void arch_unoptimize_kprobe(struct optimized_kprobe *op)
 420 {
 421         u8 insn_buf[RELATIVEJUMP_SIZE];
 422
 423         /* Set int3 to first byte for kprobes */
 424         insn_buf[0] = BREAKPOINT_INSTRUCTION;
 425         memcpy(insn_buf + 1, op->optinsn.copied_insn, RELATIVE_ADDR_SIZE);
 426         text_poke_bp(op->kp.addr, insn_buf, RELATIVEJUMP_SIZE,
 427                      op->optinsn.insn);
 428 }
 429
 430 /*
 431  * Recover original instructions and breakpoints from relative jumps.
 432  * Caller must call with locking kprobe_mutex.
 433  */
 434 extern void arch_unoptimize_kprobes(struct list_head *oplist,
 435                                     struct list_head *done_list)
 436 {
 437         struct optimized_kprobe *op, *tmp;
 438
 439         list_for_each_entry_safe(op, tmp, oplist, list) {
 440                 arch_unoptimize_kprobe(op);
 441                 list_move(&op->list, done_list);
 442         }
 443 }
 444
 445 int setup_detour_execution(struct kprobe *p, struct pt_regs *regs, int reenter)
 446 {
 447         struct optimized_kprobe *op;
 448
 449         if (p->flags & KPROBE_FLAG_OPTIMIZED) {
 450                 /* This kprobe is really able to run optimized path. */
 451                 op = container_of(p, struct optimized_kprobe, kp);
 452                 /* Detour through copied instructions */
 453                 regs->ip = (unsigned long)op->optinsn.insn + TMPL_END_IDX;
 454                 if (!reenter)
 455                         reset_current_kprobe();
 456                 preempt_enable_no_resched();
 457                 return 1;
 458         }
 459         return 0;
 460 }
 461 NOKPROBE_SYMBOL(setup_detour_execution);