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 <uprobe/swap_uprobes.h>
  38
  39 #include "swap_uprobes.h"
  40
  41
  42 /**
  43  * @struct uprobe_ctlblk
  44  * @brief Uprobe control block
  45  */
  46 struct uprobe_ctlblk {
  47         unsigned long flags;            /**< Flags */
  48         struct kprobe *p;               /**< Pointer to the uprobe's kprobe */
  49 };
  50
  51 static unsigned long trampoline_addr(struct uprobe *up)
  52 {
  53         return (unsigned long)(up->kp.ainsn.insn +
  54                                UPROBES_TRAMP_RET_BREAK_IDX);
  55 }
  56
  57 unsigned long arch_tramp_by_ri(struct uretprobe_instance *ri)
  58 {
  59         return trampoline_addr(&ri->rp->up);
  60 }
  61
  62 static struct uprobe_ctlblk *current_ucb(void)
  63 {
  64         /* FIXME hardcoded offset */
  65         return (struct uprobe_ctlblk *)(end_of_stack(current) + 20);
  66 }
  67
  68 static struct kprobe *get_current_probe(void)
  69 {
  70         return current_ucb()->p;
  71 }
  72
  73 static void set_current_probe(struct kprobe *p)
  74 {
  75         current_ucb()->p = p;
  76 }
  77
  78 static void save_current_flags(struct pt_regs *regs)
  79 {
  80         current_ucb()->flags = regs->flags;
  81 }
  82
  83 static void restore_current_flags(struct pt_regs *regs, unsigned long flags)
  84 {
  85         regs->flags &= ~IF_MASK;
  86         regs->flags |= flags & IF_MASK;
  87 }
  88
  89 /**
  90  * @brief Prepares uprobe for x86.
  91  *
  92  * @param up Pointer to the uprobe.
  93  * @return 0 on success,\n
  94  * -1 on error.
  95  */
  96 int arch_prepare_uprobe(struct uprobe *up)
  97 {
  98         struct kprobe *p = up2kp(up);
  99         struct task_struct *task = up->task;
 100         u8 *tramp = up->atramp.tramp;
 101         enum { call_relative_opcode = 0xe8 };
 102
 103         if (!read_proc_vm_atomic(task, (unsigned long)p->addr,
 104                                  tramp, MAX_INSN_SIZE)) {
 105                 printk(KERN_ERR "failed to read memory %p!\n", p->addr);
 106                 return -EINVAL;
 107         }
 108         /* TODO: this is a workaround */
 109         if (tramp[0] == call_relative_opcode) {
 110                 printk(KERN_INFO "cannot install probe: 1st instruction is call\n");
 111                 return -EINVAL;
 112         }
 113
 114         tramp[UPROBES_TRAMP_RET_BREAK_IDX] = BREAKPOINT_INSTRUCTION;
 115
 116         /* TODO: remove dual info */
 117         p->opcode = tramp[0];
 118
 119         p->ainsn.boostable = swap_can_boost(tramp) ? 0 : -1;
 120
 121         p->ainsn.insn = swap_slot_alloc(up->sm);
 122         if (p->ainsn.insn == NULL) {
 123                 printk(KERN_ERR "trampoline out of memory\n");
 124                 return -ENOMEM;
 125         }
 126
 127         if (!write_proc_vm_atomic(task, (unsigned long)p->ainsn.insn,
 128                                   tramp, sizeof(up->atramp.tramp))) {
 129                 swap_slot_free(up->sm, p->ainsn.insn);
 130                 printk(KERN_INFO "failed to write memory %p!\n", tramp);
 131                 return -EINVAL;
 132         }
 133
 134         /* for uretprobe */
 135         add_uprobe_table(p);
 136
 137         return 0;
 138 }
 139
 140 /**
 141  * @brief Jump pre-handler.
 142  *
 143  * @param p Pointer to the uprobe's kprobe.
 144  * @param regs Pointer to CPU register data.
 145  * @return 0.
 146  */
 147 int setjmp_upre_handler(struct kprobe *p, struct pt_regs *regs)
 148 {
 149         struct uprobe *up = container_of(p, struct uprobe, kp);
 150         struct ujprobe *jp = container_of(up, struct ujprobe, up);
 151         kprobe_pre_entry_handler_t pre_entry =
 152                 (kprobe_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()] = (kprobe_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.kp.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 = (kprobe_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's kprobe.
 271  * @param regs Pointer to CPU register data.
 272  * @return Trampoline address.
 273  */
 274 unsigned long arch_get_trampoline_addr(struct kprobe *p, struct pt_regs *regs)
 275 {
 276         return trampoline_addr(kp2up(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 *up)
 298 {
 299         struct kprobe *p = up2kp(up);
 300
 301         swap_slot_free(up->sm, p->ainsn.insn);
 302 }
 303
 304 int arch_arm_uprobe(struct uprobe *p)
 305 {
 306         int ret;
 307         kprobe_opcode_t insn = BREAKPOINT_INSTRUCTION;
 308         unsigned long vaddr = (unsigned long)p->kp.addr;
 309
 310         ret = write_proc_vm_atomic(p->task, vaddr, &insn, sizeof(insn));
 311         if (!ret) {
 312                 pr_err("arch_arm_uprobe: failed to write memory tgid=%u vaddr=%08lx\n",
 313                        p->task->tgid, vaddr);
 314
 315                 return -EACCES;
 316         }
 317
 318         return 0;
 319 }
 320
 321 void arch_disarm_uprobe(struct kprobe *p, struct task_struct *task)
 322 {
 323         int ret;
 324         unsigned long vaddr = (unsigned long)p->addr;
 325
 326         ret = write_proc_vm_atomic(task, vaddr, &p->opcode, sizeof(p->opcode));
 327         if (!ret) {
 328                 pr_err("arch_disarm_uprobe: failed to write memory tgid=%u, vaddr=%08lx\n",
 329                        task->tgid, vaddr);
 330         }
 331 }
 332
 333 static void set_user_jmp_op(void *from, void *to)
 334 {
 335         struct __arch_jmp_op {
 336                 char op;
 337                 long raddr;
 338         } __packed jop;
 339
 340         jop.raddr = (long)(to) - ((long)(from) + 5);
 341         jop.op = RELATIVEJUMP_INSTRUCTION;
 342
 343         if (put_user(jop.op, (char *)from) ||
 344             put_user(jop.raddr, (long *)(from + 1)))
 345                 pr_err("failed to write jump opcode to user space %p\n", from);
 346 }
 347
 348 static void resume_execution(struct kprobe *p,
 349                              struct pt_regs *regs,
 350                              unsigned long flags)
 351 {
 352         unsigned long *tos, tos_dword = 0;
 353         unsigned long copy_eip = (unsigned long)p->ainsn.insn;
 354         unsigned long orig_eip = (unsigned long)p->addr;
 355         kprobe_opcode_t insns[2];
 356
 357         regs->EREG(flags) &= ~TF_MASK;
 358
 359         tos = (unsigned long *)&tos_dword;
 360         if (get_user(tos_dword, (unsigned long *)regs->sp)) {
 361                 pr_err("failed to read from user space sp=%lx!\n", regs->sp);
 362                 return;
 363         }
 364
 365         if (get_user(*(unsigned short *)insns, (unsigned short *)p->ainsn.insn)) {
 366                 pr_err("failed to read first 2 opcodes %p!\n", p->ainsn.insn);
 367                 return;
 368         }
 369
 370         switch (insns[0]) {
 371         case 0x9c: /* pushfl */
 372                 *tos &= ~(TF_MASK | IF_MASK);
 373                 *tos |= flags & (TF_MASK | IF_MASK);
 374                 break;
 375         case 0xc2: /* iret/ret/lret */
 376         case 0xc3:
 377         case 0xca:
 378         case 0xcb:
 379         case 0xcf:
 380         case 0xea: /* jmp absolute -- eip is correct */
 381                 /* eip is already adjusted, no more changes required */
 382                 p->ainsn.boostable = 1;
 383                 goto no_change;
 384         case 0xe8: /* call relative - Fix return addr */
 385                 *tos = orig_eip + (*tos - copy_eip);
 386                 break;
 387         case 0x9a: /* call absolute -- same as call absolute, indirect */
 388                 *tos = orig_eip + (*tos - copy_eip);
 389
 390                 if (put_user(tos_dword, (unsigned long *)regs->sp)) {
 391                         pr_err("failed to write dword to sp=%lx\n", regs->sp);
 392                         return;
 393                 }
 394
 395                 goto no_change;
 396         case 0xff:
 397                 if ((insns[1] & 0x30) == 0x10) {
 398                         /*
 399                          * call absolute, indirect
 400                          * Fix return addr; eip is correct.
 401                          * But this is not boostable
 402                          */
 403                         *tos = orig_eip + (*tos - copy_eip);
 404
 405                         if (put_user(tos_dword, (unsigned long *)regs->sp)) {
 406                                 pr_err("failed to write dword to sp=%lx\n", regs->sp);
 407                                 return;
 408                         }
 409
 410                         goto no_change;
 411                 } else if (((insns[1] & 0x31) == 0x20) || /* jmp near, absolute
 412                                                            * indirect */
 413                            ((insns[1] & 0x31) == 0x21)) {
 414                         /* jmp far, absolute indirect */
 415                         /* eip is correct. And this is boostable */
 416                         p->ainsn.boostable = 1;
 417                         goto no_change;
 418                 }
 419         case 0xf3:
 420                 if (insns[1] == 0xc3)
 421                         /* repz ret special handling: no more changes */
 422                         goto no_change;
 423                 break;
 424         default:
 425                 break;
 426         }
 427
 428         if (put_user(tos_dword, (unsigned long *)regs->sp)) {
 429                 pr_err("failed to write dword to sp=%lx\n", regs->sp);
 430                 return;
 431         }
 432
 433         if (p->ainsn.boostable == 0) {
 434                 if ((regs->EREG(ip) > copy_eip) && (regs->EREG(ip) - copy_eip) +
 435                     5 < MAX_INSN_SIZE) {
 436                         /*
 437                          * These instructions can be executed directly if it
 438                          * jumps back to correct address.
 439                          */
 440                         set_user_jmp_op((void *) regs->EREG(ip),
 441                                         (void *)orig_eip +
 442                                         (regs->EREG(ip) - copy_eip));
 443                         p->ainsn.boostable = 1;
 444                 } else {
 445                         p->ainsn.boostable = -1;
 446                 }
 447         }
 448
 449         regs->EREG(ip) = orig_eip + (regs->EREG(ip) - copy_eip);
 450
 451 no_change:
 452         return;
 453 }
 454
 455 static bool prepare_ss_addr(struct kprobe *p, struct pt_regs *regs)
 456 {
 457         unsigned long *ss_addr = (long *)&p->ss_addr[smp_processor_id()];
 458
 459         if (*ss_addr) {
 460                 regs->ip = *ss_addr;
 461                 *ss_addr = 0;
 462                 return true;
 463         } else {
 464                 regs->ip = (unsigned long)p->ainsn.insn;
 465                 return false;
 466         }
 467 }
 468
 469 static void prepare_ss(struct pt_regs *regs)
 470 {
 471         /* set single step mode */
 472         regs->flags |= TF_MASK;
 473         regs->flags &= ~IF_MASK;
 474 }
 475
 476 static int uprobe_handler(struct pt_regs *regs)
 477 {
 478         struct kprobe *p;
 479         kprobe_opcode_t *addr;
 480         struct task_struct *task = current;
 481         pid_t tgid = task->tgid;
 482
 483         save_current_flags(regs);
 484
 485         addr = (kprobe_opcode_t *)(regs->EREG(ip) - sizeof(kprobe_opcode_t));
 486         p = get_ukprobe(addr, tgid);
 487
 488         if (p == NULL) {
 489                 void *tramp_addr = (void *)addr - UPROBES_TRAMP_RET_BREAK_IDX;
 490
 491                 p = get_ukprobe_by_insn_slot(tramp_addr, tgid, regs);
 492                 if (p == NULL) {
 493                         printk(KERN_INFO "no_uprobe\n");
 494                         return 0;
 495                 }
 496
 497                 trampoline_uprobe_handler(p, regs);
 498                 return 1;
 499         } else {
 500                 if (!p->pre_handler || !p->pre_handler(p, regs)) {
 501                         if (p->ainsn.boostable == 1 && !p->post_handler) {
 502                                 prepare_ss_addr(p, regs);
 503                                 return 1;
 504                         }
 505
 506                         if (prepare_ss_addr(p, regs) == false) {
 507                                 set_current_probe(p);
 508                                 prepare_ss(regs);
 509                         }
 510                 }
 511         }
 512
 513         return 1;
 514 }
 515
 516 static int post_uprobe_handler(struct pt_regs *regs)
 517 {
 518         struct kprobe *p = get_current_probe();
 519         unsigned long flags = current_ucb()->flags;
 520
 521         if (p == NULL) {
 522                 printk("task[%u %u %s] current uprobe is not found\n",
 523                        current->tgid, current->pid, current->comm);
 524                 return 0;
 525         }
 526
 527         resume_execution(p, regs, flags);
 528         restore_current_flags(regs, flags);
 529
 530         /* clean stack */
 531         current_ucb()->p = 0;
 532         current_ucb()->flags = 0;
 533
 534         return 1;
 535 }
 536
 537 static int uprobe_exceptions_notify(struct notifier_block *self,
 538                                     unsigned long val, void *data)
 539 {
 540         struct die_args *args = (struct die_args *)data;
 541         int ret = NOTIFY_DONE;
 542
 543         if (args->regs == NULL || !user_mode_vm(args->regs))
 544                 return ret;
 545
 546         switch (val) {
 547 #ifdef CONFIG_KPROBES
 548         case DIE_INT3:
 549 #else
 550         case DIE_TRAP:
 551 #endif
 552                 if (uprobe_handler(args->regs))
 553                         ret = NOTIFY_STOP;
 554                 break;
 555         case DIE_DEBUG:
 556                 if (post_uprobe_handler(args->regs))
 557                         ret = NOTIFY_STOP;
 558                 break;
 559         default:
 560                 break;
 561         }
 562
 563         return ret;
 564 }
 565
 566 static struct notifier_block uprobe_exceptions_nb = {
 567         .notifier_call = uprobe_exceptions_notify,
 568         .priority = INT_MAX
 569 };
 570
 571 /**
 572  * @brief Registers notify.
 573  *
 574  * @return register_die_notifier result.
 575  */
 576 int swap_arch_init_uprobes(void)
 577 {
 578         return register_die_notifier(&uprobe_exceptions_nb);
 579 }
 580
 581 /**
 582  * @brief Unregisters notify.
 583  *
 584  * @return Void.
 585  */
 586 void swap_arch_exit_uprobes(void)
 587 {
 588         unregister_die_notifier(&uprobe_exceptions_nb);
 589 }
 590