uprobe/arch/x86/swap-asm/swap_uprobes.c

   1 /**
   2  * uprobe/arch/asm-x86/swap_uprobes.c
   3  * @author Alexey Gerenkov <a.gerenkov@samsung.com> User-Space Probes initial
   4  * implementation; Support x86/ARM/MIPS for both user and kernel spaces.
   5  * @author Ekaterina Gorelkina <e.gorelkina@samsung.com>: redesign module for
   6  * separating core and arch parts
   7  *
   8  * @section LICENSE
   9  *
  10  * This program is free software; you can redistribute it and/or modify
  11  * it under the terms of the GNU General Public License as published by
  12  * the Free Software Foundation; either version 2 of the License, or
  13  * (at your option) any later version.
  14  *
  15  * This program is distributed in the hope that it will be useful,
  16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  18  * GNU General Public License for more details.
  19  *
  20  * You should have received a copy of the GNU General Public License
  21  * along with this program; if not, write to the Free Software
  22  * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
  23  *
  24  * @section COPYRIGHT
  25  *
  26  * Copyright (C) Samsung Electronics, 2006-2010
  27  *
  28  * @section DESCRIPTION
  29  *
  30  * Arch-dependent uprobe interface implementation for x86.
  31  */
  32
  33
  34 #include <linux/kdebug.h>
  35
  36 #include <kprobe/swap_slots.h>
  37 #include <kprobe/swap_td_raw.h>
  38 #include <uprobe/swap_uprobes.h>
  39
  40 #include "swap_uprobes.h"
  41
  42
  43 /**
  44  * @struct uprobe_ctlblk
  45  * @brief Uprobe control block
  46  */
  47 struct uprobe_ctlblk {
  48         unsigned long flags;            /**< Flags */
  49         struct uprobe *p;               /**< Pointer to the uprobe */
  50 };
  51
  52
  53 static struct td_raw td_raw;
  54
  55
  56 static unsigned long trampoline_addr(struct uprobe *up)
  57 {
  58         return (unsigned long)(up->ainsn.insn +
  59                                UPROBES_TRAMP_RET_BREAK_IDX);
  60 }
  61
  62 unsigned long arch_tramp_by_ri(struct uretprobe_instance *ri)
  63 {
  64         return trampoline_addr(&ri->rp->up);
  65 }
  66
  67 static struct uprobe_ctlblk *current_ucb(void)
  68 {
  69         return (struct uprobe_ctlblk *)swap_td_raw(&td_raw, current);
  70 }
  71
  72 static struct uprobe *get_current_probe(void)
  73 {
  74         return current_ucb()->p;
  75 }
  76
  77 static void set_current_probe(struct uprobe *p)
  78 {
  79         current_ucb()->p = p;
  80 }
  81
  82 static void save_current_flags(struct pt_regs *regs)
  83 {
  84         current_ucb()->flags = regs->flags;
  85 }
  86
  87 static void restore_current_flags(struct pt_regs *regs, unsigned long flags)
  88 {
  89         regs->flags &= ~IF_MASK;
  90         regs->flags |= flags & IF_MASK;
  91 }
  92
  93 /**
  94  * @brief Prepares uprobe for x86.
  95  *
  96  * @param up Pointer to the uprobe.
  97  * @return 0 on success,\n
  98  * -1 on error.
  99  */
 100 int arch_prepare_uprobe(struct uprobe *p)
 101 {
 102         struct task_struct *task = p->task;
 103         u8 tramp[UPROBES_TRAMP_LEN + BP_INSN_SIZE];     /* BP for uretprobe */
 104         enum { call_relative_opcode = 0xe8 };
 105
 106         if (!read_proc_vm_atomic(task, (unsigned long)p->addr,
 107                                  tramp, MAX_INSN_SIZE)) {
 108                 printk(KERN_ERR "failed to read memory %p!\n", p->addr);
 109                 return -EINVAL;
 110         }
 111         /* TODO: this is a workaround */
 112         if (tramp[0] == call_relative_opcode) {
 113                 printk(KERN_INFO "cannot install probe: 1st instruction is call\n");
 114                 return -EINVAL;
 115         }
 116
 117         tramp[UPROBES_TRAMP_RET_BREAK_IDX] = BREAKPOINT_INSTRUCTION;
 118
 119         p->opcode = tramp[0];
 120         p->ainsn.boostable = swap_can_boost(tramp) ? 0 : -1;
 121
 122         p->ainsn.insn = swap_slot_alloc(p->sm);
 123         if (p->ainsn.insn == NULL) {
 124                 printk(KERN_ERR "trampoline out of memory\n");
 125                 return -ENOMEM;
 126         }
 127
 128         if (!write_proc_vm_atomic(task, (unsigned long)p->ainsn.insn,
 129                                   tramp, sizeof(tramp))) {
 130                 swap_slot_free(p->sm, p->ainsn.insn);
 131                 printk(KERN_INFO "failed to write memory %p!\n", tramp);
 132                 return -EINVAL;
 133         }
 134
 135         /* for uretprobe */
 136         add_uprobe_table(p);
 137
 138         return 0;
 139 }
 140
 141 /**
 142  * @brief Jump pre-handler.
 143  *
 144  * @param p Pointer to the uprobe.
 145  * @param regs Pointer to CPU register data.
 146  * @return 0.
 147  */
 148 int setjmp_upre_handler(struct uprobe *p, struct pt_regs *regs)
 149 {
 150         struct ujprobe *jp = container_of(p, struct ujprobe, up);
 151         uprobe_pre_entry_handler_t pre_entry =
 152                 (uprobe_pre_entry_handler_t)jp->pre_entry;
 153         entry_point_t entry = (entry_point_t)jp->entry;
 154         unsigned long args[6];
 155
 156         /* FIXME some user space apps crash if we clean interrupt bit */
 157         /* regs->EREG(flags) &= ~IF_MASK; */
 158 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 18)
 159         trace_hardirqs_off();
 160 #endif
 161
 162         /* read first 6 args from stack */
 163         if (!read_proc_vm_atomic(current, regs->EREG(sp) + 4,
 164                                  args, sizeof(args)))
 165                 printk(KERN_WARNING
 166                        "failed to read user space func arguments %lx!\n",
 167                        regs->sp + 4);
 168
 169         if (pre_entry)
 170                 p->ss_addr[smp_processor_id()] = (uprobe_opcode_t *)
 171                                                  pre_entry(jp->priv_arg, regs);
 172
 173         if (entry)
 174                 entry(args[0], args[1], args[2], args[3], args[4], args[5]);
 175         else
 176                 arch_ujprobe_return();
 177
 178         return 0;
 179 }
 180
 181 /**
 182  * @brief Prepares uretprobe for x86.
 183  *
 184  * @param ri Pointer to the uretprobe instance.
 185  * @param regs Pointer to CPU register data.
 186  * @return Void.
 187  */
 188 int arch_prepare_uretprobe(struct uretprobe_instance *ri, struct pt_regs *regs)
 189 {
 190         /* Replace the return addr with trampoline addr */
 191         unsigned long ra = trampoline_addr(&ri->rp->up);
 192         unsigned long ret_addr;
 193         ri->sp = (kprobe_opcode_t *)regs->sp;
 194
 195         if (get_user(ret_addr, (unsigned long *)regs->sp)) {
 196                 pr_err("failed to read user space func ra %lx addr=%p!\n",
 197                        regs->sp, ri->rp->up.addr);
 198                 return -EINVAL;
 199         }
 200
 201         if (put_user(ra, (unsigned long *)regs->sp)) {
 202                 pr_err("failed to write user space func ra %lx!\n", regs->sp);
 203                 return -EINVAL;
 204         }
 205
 206         ri->ret_addr = (uprobe_opcode_t *)ret_addr;
 207
 208         return 0;
 209 }
 210
 211 static bool get_long(struct task_struct *task,
 212                      unsigned long vaddr, unsigned long *val)
 213 {
 214         return sizeof(*val) != read_proc_vm_atomic(task, vaddr,
 215                                                    val, sizeof(*val));
 216 }
 217
 218 static bool put_long(struct task_struct *task,
 219                      unsigned long vaddr, unsigned long *val)
 220 {
 221         return sizeof(*val) != write_proc_vm_atomic(task, vaddr,
 222                                                     val, sizeof(*val));
 223 }
 224
 225 /**
 226  * @brief Disarms uretprobe on x86 arch.
 227  *
 228  * @param ri Pointer to the uretprobe instance.
 229  * @param task Pointer to the task for which the probe.
 230  * @return 0 on success,\n
 231  * negative error code on error.
 232  */
 233 int arch_disarm_urp_inst(struct uretprobe_instance *ri,
 234                          struct task_struct *task, unsigned long tr)
 235 {
 236         unsigned long ret_addr;
 237         unsigned long sp = (unsigned long)ri->sp;
 238         unsigned long tramp_addr;
 239
 240         if (tr == 0)
 241                 tramp_addr = arch_tramp_by_ri(ri);
 242         else
 243                 tramp_addr = tr; /* ri - invalid */
 244
 245         if (get_long(task, sp, &ret_addr)) {
 246                 printk(KERN_INFO "---> %s (%d/%d): failed to read stack from %08lx\n",
 247                        task->comm, task->tgid, task->pid, sp);
 248                 return -EFAULT;
 249         }
 250
 251         if (tramp_addr == ret_addr) {
 252                 if (put_long(task, sp, (unsigned long *)&ri->ret_addr)) {
 253                         printk(KERN_INFO "---> %s (%d/%d): failed to write "
 254                                "orig_ret_addr to %08lx",
 255                                task->comm, task->tgid, task->pid, sp);
 256                         return -EFAULT;
 257                 }
 258         } else {
 259                 printk(KERN_INFO "---> %s (%d/%d): trampoline NOT found at sp = %08lx\n",
 260                        task->comm, task->tgid, task->pid, sp);
 261                 return -ENOENT;
 262         }
 263
 264         return 0;
 265 }
 266
 267 /**
 268  * @brief Gets trampoline address.
 269  *
 270  * @param p Pointer to the uprobe.
 271  * @param regs Pointer to CPU register data.
 272  * @return Trampoline address.
 273  */
 274 unsigned long arch_get_trampoline_addr(struct uprobe *p, struct pt_regs *regs)
 275 {
 276         return trampoline_addr(p);
 277 }
 278
 279 /**
 280  * @brief Restores return address.
 281  *
 282  * @param orig_ret_addr Original return address.
 283  * @param regs Pointer to CPU register data.
 284  * @return Void.
 285  */
 286 void arch_set_orig_ret_addr(unsigned long orig_ret_addr, struct pt_regs *regs)
 287 {
 288         regs->EREG(ip) = orig_ret_addr;
 289 }
 290
 291 /**
 292  * @brief Removes uprobe.
 293  *
 294  * @param up Pointer to the target uprobe.
 295  * @return Void.
 296  */
 297 void arch_remove_uprobe(struct uprobe *p)
 298 {
 299         swap_slot_free(p->sm, p->ainsn.insn);
 300 }
 301
 302 int arch_arm_uprobe(struct uprobe *p)
 303 {
 304         int ret;
 305         uprobe_opcode_t insn = BREAKPOINT_INSTRUCTION;
 306         unsigned long vaddr = (unsigned long)p->addr;
 307
 308         ret = write_proc_vm_atomic(p->task, vaddr, &insn, sizeof(insn));
 309         if (!ret) {
 310                 pr_err("arch_arm_uprobe: failed to write memory tgid=%u vaddr=%08lx\n",
 311                        p->task->tgid, vaddr);
 312
 313                 return -EACCES;
 314         }
 315
 316         return 0;
 317 }
 318
 319 void arch_disarm_uprobe(struct uprobe *p, struct task_struct *task)
 320 {
 321         int ret;
 322         unsigned long vaddr = (unsigned long)p->addr;
 323
 324         ret = write_proc_vm_atomic(task, vaddr, &p->opcode, sizeof(p->opcode));
 325         if (!ret) {
 326                 pr_err("arch_disarm_uprobe: failed to write memory tgid=%u, vaddr=%08lx\n",
 327                        task->tgid, vaddr);
 328         }
 329 }
 330
 331 static void set_user_jmp_op(void *from, void *to)
 332 {
 333         struct __arch_jmp_op {
 334                 char op;
 335                 long raddr;
 336         } __packed jop;
 337
 338         jop.raddr = (long)(to) - ((long)(from) + 5);
 339         jop.op = RELATIVEJUMP_INSTRUCTION;
 340
 341         if (put_user(jop.op, (char *)from) ||
 342             put_user(jop.raddr, (long *)(from + 1)))
 343                 pr_err("failed to write jump opcode to user space %p\n", from);
 344 }
 345
 346 static void resume_execution(struct uprobe *p,
 347                              struct pt_regs *regs,
 348                              unsigned long flags)
 349 {
 350         unsigned long *tos, tos_dword = 0;
 351         unsigned long copy_eip = (unsigned long)p->ainsn.insn;
 352         unsigned long orig_eip = (unsigned long)p->addr;
 353         uprobe_opcode_t insns[2];
 354
 355         regs->EREG(flags) &= ~TF_MASK;
 356
 357         tos = (unsigned long *)&tos_dword;
 358         if (get_user(tos_dword, (unsigned long *)regs->sp)) {
 359                 pr_err("failed to read from user space sp=%lx!\n", regs->sp);
 360                 return;
 361         }
 362
 363         if (get_user(*(unsigned short *)insns, (unsigned short *)p->ainsn.insn)) {
 364                 pr_err("failed to read first 2 opcodes %p!\n", p->ainsn.insn);
 365                 return;
 366         }
 367
 368         switch (insns[0]) {
 369         case 0x9c: /* pushfl */
 370                 *tos &= ~(TF_MASK | IF_MASK);
 371                 *tos |= flags & (TF_MASK | IF_MASK);
 372                 break;
 373         case 0xc2: /* iret/ret/lret */
 374         case 0xc3:
 375         case 0xca:
 376         case 0xcb:
 377         case 0xcf:
 378         case 0xea: /* jmp absolute -- eip is correct */
 379                 /* eip is already adjusted, no more changes required */
 380                 p->ainsn.boostable = 1;
 381                 goto no_change;
 382         case 0xe8: /* call relative - Fix return addr */
 383                 *tos = orig_eip + (*tos - copy_eip);
 384                 break;
 385         case 0x9a: /* call absolute -- same as call absolute, indirect */
 386                 *tos = orig_eip + (*tos - copy_eip);
 387
 388                 if (put_user(tos_dword, (unsigned long *)regs->sp)) {
 389                         pr_err("failed to write dword to sp=%lx\n", regs->sp);
 390                         return;
 391                 }
 392
 393                 goto no_change;
 394         case 0xff:
 395                 if ((insns[1] & 0x30) == 0x10) {
 396                         /*
 397                          * call absolute, indirect
 398                          * Fix return addr; eip is correct.
 399                          * But this is not boostable
 400                          */
 401                         *tos = orig_eip + (*tos - copy_eip);
 402
 403                         if (put_user(tos_dword, (unsigned long *)regs->sp)) {
 404                                 pr_err("failed to write dword to sp=%lx\n", regs->sp);
 405                                 return;
 406                         }
 407
 408                         goto no_change;
 409                 } else if (((insns[1] & 0x31) == 0x20) || /* jmp near, absolute
 410                                                            * indirect */
 411                            ((insns[1] & 0x31) == 0x21)) {
 412                         /* jmp far, absolute indirect */
 413                         /* eip is correct. And this is boostable */
 414                         p->ainsn.boostable = 1;
 415                         goto no_change;
 416                 }
 417         case 0xf3:
 418                 if (insns[1] == 0xc3)
 419                         /* repz ret special handling: no more changes */
 420                         goto no_change;
 421                 break;
 422         default:
 423                 break;
 424         }
 425
 426         if (put_user(tos_dword, (unsigned long *)regs->sp)) {
 427                 pr_err("failed to write dword to sp=%lx\n", regs->sp);
 428                 return;
 429         }
 430
 431         if (p->ainsn.boostable == 0) {
 432                 if ((regs->EREG(ip) > copy_eip) && (regs->EREG(ip) - copy_eip) +
 433                     5 < MAX_INSN_SIZE) {
 434                         /*
 435                          * These instructions can be executed directly if it
 436                          * jumps back to correct address.
 437                          */
 438                         set_user_jmp_op((void *) regs->EREG(ip),
 439                                         (void *)orig_eip +
 440                                         (regs->EREG(ip) - copy_eip));
 441                         p->ainsn.boostable = 1;
 442                 } else {
 443                         p->ainsn.boostable = -1;
 444                 }
 445         }
 446
 447         regs->EREG(ip) = orig_eip + (regs->EREG(ip) - copy_eip);
 448
 449 no_change:
 450         return;
 451 }
 452
 453 static bool prepare_ss_addr(struct uprobe *p, struct pt_regs *regs)
 454 {
 455         unsigned long *ss_addr = (long *)&p->ss_addr[smp_processor_id()];
 456
 457         if (*ss_addr) {
 458                 regs->ip = *ss_addr;
 459                 *ss_addr = 0;
 460                 return true;
 461         } else {
 462                 regs->ip = (unsigned long)p->ainsn.insn;
 463                 return false;
 464         }
 465 }
 466
 467 static void prepare_ss(struct pt_regs *regs)
 468 {
 469         /* set single step mode */
 470         regs->flags |= TF_MASK;
 471         regs->flags &= ~IF_MASK;
 472 }
 473
 474 static int uprobe_handler(struct pt_regs *regs)
 475 {
 476         struct uprobe *p;
 477         uprobe_opcode_t *addr;
 478         struct task_struct *task = current;
 479         pid_t tgid = task->tgid;
 480
 481         save_current_flags(regs);
 482
 483         addr = (uprobe_opcode_t *)(regs->EREG(ip) - sizeof(uprobe_opcode_t));
 484         p = get_uprobe(addr, tgid);
 485
 486         if (p == NULL) {
 487                 void *tramp_addr = (void *)addr - UPROBES_TRAMP_RET_BREAK_IDX;
 488
 489                 p = get_uprobe_by_insn_slot(tramp_addr, tgid, regs);
 490                 if (p == NULL) {
 491                         printk(KERN_INFO "no_uprobe\n");
 492                         return 0;
 493                 }
 494
 495                 trampoline_uprobe_handler(p, regs);
 496                 return 1;
 497         } else {
 498                 if (!p->pre_handler || !p->pre_handler(p, regs)) {
 499                         if (p->ainsn.boostable == 1 && !p->post_handler) {
 500                                 prepare_ss_addr(p, regs);
 501                                 return 1;
 502                         }
 503
 504                         if (prepare_ss_addr(p, regs) == false) {
 505                                 set_current_probe(p);
 506                                 prepare_ss(regs);
 507                         }
 508                 }
 509         }
 510
 511         return 1;
 512 }
 513
 514 static int post_uprobe_handler(struct pt_regs *regs)
 515 {
 516         struct uprobe *p = get_current_probe();
 517         unsigned long flags = current_ucb()->flags;
 518
 519         if (p == NULL) {
 520                 printk("task[%u %u %s] current uprobe is not found\n",
 521                        current->tgid, current->pid, current->comm);
 522                 return 0;
 523         }
 524
 525         resume_execution(p, regs, flags);
 526         restore_current_flags(regs, flags);
 527
 528         /* clean stack */
 529         current_ucb()->p = 0;
 530         current_ucb()->flags = 0;
 531
 532         return 1;
 533 }
 534
 535 static int uprobe_exceptions_notify(struct notifier_block *self,
 536                                     unsigned long val, void *data)
 537 {
 538         struct die_args *args = (struct die_args *)data;
 539         int ret = NOTIFY_DONE;
 540
 541         if (args->regs == NULL || !user_mode_vm(args->regs))
 542                 return ret;
 543
 544         switch (val) {
 545 #ifdef CONFIG_KPROBES
 546         case DIE_INT3:
 547 #else
 548         case DIE_TRAP:
 549 #endif
 550                 if (uprobe_handler(args->regs))
 551                         ret = NOTIFY_STOP;
 552                 break;
 553         case DIE_DEBUG:
 554                 if (post_uprobe_handler(args->regs))
 555                         ret = NOTIFY_STOP;
 556                 break;
 557         default:
 558                 break;
 559         }
 560
 561         return ret;
 562 }
 563
 564 static struct notifier_block uprobe_exceptions_nb = {
 565         .notifier_call = uprobe_exceptions_notify,
 566         .priority = INT_MAX
 567 };
 568
 569 /**
 570  * @brief Registers notify.
 571  *
 572  * @return register_die_notifier result.
 573  */
 574 int swap_arch_init_uprobes(void)
 575 {
 576         int ret;
 577
 578         ret = swap_td_raw_reg(&td_raw, sizeof(struct uprobe_ctlblk));
 579         if (ret)
 580                 return ret;
 581
 582         ret = register_die_notifier(&uprobe_exceptions_nb);
 583         if (ret)
 584                 swap_td_raw_unreg(&td_raw);
 585
 586         return ret;
 587 }
 588
 589 /**
 590  * @brief Unregisters notify.
 591  *
 592  * @return Void.
 593  */
 594 void swap_arch_exit_uprobes(void)
 595 {
 596         unregister_die_notifier(&uprobe_exceptions_nb);
 597         swap_td_raw_unreg(&td_raw);
 598 }
 599