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