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 static DEFINE_PER_CPU(struct uprobe_ctlblk, ucb) = { 0, NULL };
  58
  59 static struct kprobe *get_current_probe(void)
  60 {
  61         return __get_cpu_var(ucb).p;
  62 }
  63
  64 static void set_current_probe(struct kprobe *p)
  65 {
  66         __get_cpu_var(ucb).p = p;
  67 }
  68
  69 static void reset_current_probe(void)
  70 {
  71         set_current_probe(NULL);
  72 }
  73
  74 static void save_current_flags(struct pt_regs *regs)
  75 {
  76         __get_cpu_var(ucb).flags = regs->EREG(flags);
  77 }
  78
  79 static void restore_current_flags(struct pt_regs *regs)
  80 {
  81         regs->EREG(flags) &= ~IF_MASK;
  82         regs->EREG(flags) |= __get_cpu_var(ucb).flags & IF_MASK;
  83 }
  84
  85 /**
  86  * @brief Prepares uprobe for x86.
  87  *
  88  * @param up Pointer to the uprobe.
  89  * @return 0 on success,\n
  90  * -1 on error.
  91  */
  92 int arch_prepare_uprobe(struct uprobe *up)
  93 {
  94         struct kprobe *p = up2kp(up);
  95         struct task_struct *task = up->task;
  96         u8 *tramp = up->atramp.tramp;
  97         enum { call_relative_opcode = 0xe8 };
  98
  99         if (!read_proc_vm_atomic(task, (unsigned long)p->addr,
 100                                  tramp, MAX_INSN_SIZE)) {
 101                 printk(KERN_ERR "failed to read memory %p!\n", p->addr);
 102                 return -EINVAL;
 103         }
 104         /* TODO: this is a workaround */
 105         if (tramp[0] == call_relative_opcode) {
 106                 printk(KERN_INFO "cannot install probe: 1st instruction is call\n");
 107                 return -EINVAL;
 108         }
 109
 110         tramp[UPROBES_TRAMP_RET_BREAK_IDX] = BREAKPOINT_INSTRUCTION;
 111
 112         /* TODO: remove dual info */
 113         p->opcode = tramp[0];
 114
 115         p->ainsn.boostable = swap_can_boost(tramp) ? 0 : -1;
 116
 117         p->ainsn.insn = swap_slot_alloc(up->sm);
 118         if (p->ainsn.insn == NULL) {
 119                 printk(KERN_ERR "trampoline out of memory\n");
 120                 return -ENOMEM;
 121         }
 122
 123         if (!write_proc_vm_atomic(task, (unsigned long)p->ainsn.insn,
 124                                   tramp, sizeof(up->atramp.tramp))) {
 125                 swap_slot_free(up->sm, p->ainsn.insn);
 126                 printk(KERN_INFO "failed to write memory %p!\n", tramp);
 127                 return -EINVAL;
 128         }
 129
 130         return 0;
 131 }
 132
 133 /**
 134  * @brief Jump pre-handler.
 135  *
 136  * @param p Pointer to the uprobe's kprobe.
 137  * @param regs Pointer to CPU register data.
 138  * @return 0.
 139  */
 140 int setjmp_upre_handler(struct kprobe *p, struct pt_regs *regs)
 141 {
 142         struct uprobe *up = container_of(p, struct uprobe, kp);
 143         struct ujprobe *jp = container_of(up, struct ujprobe, up);
 144         kprobe_pre_entry_handler_t pre_entry =
 145                 (kprobe_pre_entry_handler_t)jp->pre_entry;
 146         entry_point_t entry = (entry_point_t)jp->entry;
 147         unsigned long args[6];
 148
 149         /* FIXME some user space apps crash if we clean interrupt bit */
 150         /* regs->EREG(flags) &= ~IF_MASK; */
 151 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 18)
 152         trace_hardirqs_off();
 153 #endif
 154
 155         /* read first 6 args from stack */
 156         if (!read_proc_vm_atomic(current, regs->EREG(sp) + 4,
 157                                  args, sizeof(args)))
 158                 printk(KERN_WARN "failed to read user space func arguments %lx!\n",
 159                        regs->EREG(sp) + 4);
 160
 161         if (pre_entry)
 162                 p->ss_addr[smp_processor_id()] = (kprobe_opcode_t *)
 163                                                  pre_entry(jp->priv_arg, regs);
 164
 165         if (entry)
 166                 entry(args[0], args[1], args[2], args[3], args[4], args[5]);
 167         else
 168                 arch_ujprobe_return();
 169
 170         return 0;
 171 }
 172
 173 /**
 174  * @brief Prepares uretprobe for x86.
 175  *
 176  * @param ri Pointer to the uretprobe instance.
 177  * @param regs Pointer to CPU register data.
 178  * @return Void.
 179  */
 180 int arch_prepare_uretprobe(struct uretprobe_instance *ri, struct pt_regs *regs)
 181 {
 182         /* Replace the return addr with trampoline addr */
 183         unsigned long ra = trampoline_addr(&ri->rp->up);
 184         ri->sp = (kprobe_opcode_t *)regs->sp;
 185
 186         if (!read_proc_vm_atomic(current, regs->EREG(sp), &(ri->ret_addr),
 187                                  sizeof(ri->ret_addr))) {
 188                 printk(KERN_ERR "failed to read user space func ra %lx addr=%p!\n",
 189                                 regs->EREG(sp), ri->rp->up.kp.addr);
 190                 return -EINVAL;
 191         }
 192
 193         if (!write_proc_vm_atomic(current, regs->EREG(sp), &ra, sizeof(ra))) {
 194                 printk(KENR_ERR "failed to write user space func ra %lx!\n",
 195                        regs->EREG(sp));
 196                 return -EINVAL;
 197         }
 198
 199         add_uprobe_table(&ri->rp->up.kp);
 200
 201         return 0;
 202 }
 203
 204 /**
 205  * @brief Disarms uretprobe on x86 arch.
 206  *
 207  * @param ri Pointer to the uretprobe instance.
 208  * @param task Pointer to the task for which the probe.
 209  * @return 0 on success,\n
 210  * negative error code on error.
 211  */
 212 int arch_disarm_urp_inst(struct uretprobe_instance *ri,
 213                          struct task_struct *task)
 214 {
 215         int len;
 216         unsigned long ret_addr;
 217         unsigned long sp = (unsigned long)ri->sp;
 218         unsigned long tramp_addr = trampoline_addr(&ri->rp->up);
 219         len = read_proc_vm_atomic(task, sp, &ret_addr, sizeof(ret_addr));
 220         if (len != sizeof(ret_addr)) {
 221                 printk(KERN_INFO "---> %s (%d/%d): failed to read stack from %08lx\n",
 222                        task->comm, task->tgid, task->pid, sp);
 223                 return -EFAULT;
 224         }
 225
 226         if (tramp_addr == ret_addr) {
 227                 len = write_proc_vm_atomic(task, sp, &ri->ret_addr,
 228                                            sizeof(ri->ret_addr));
 229                 if (len != sizeof(ri->ret_addr)) {
 230                         printk(KERN_INFO "---> %s (%d/%d): failed to write "
 231                                "orig_ret_addr to %08lx",
 232                                task->comm, task->tgid, task->pid, sp);
 233                         return -EFAULT;
 234                 }
 235         } else {
 236                 printk(KERN_INFO "---> %s (%d/%d): trampoline NOT found at sp = %08lx\n",
 237                        task->comm, task->tgid, task->pid, sp);
 238                 return -ENOENT;
 239         }
 240
 241         return 0;
 242 }
 243
 244 /**
 245  * @brief Gets trampoline address.
 246  *
 247  * @param p Pointer to the uprobe's kprobe.
 248  * @param regs Pointer to CPU register data.
 249  * @return Trampoline address.
 250  */
 251 unsigned long arch_get_trampoline_addr(struct kprobe *p, struct pt_regs *regs)
 252 {
 253         return trampoline_addr(kp2up(p));
 254 }
 255
 256 /**
 257  * @brief Restores return address.
 258  *
 259  * @param orig_ret_addr Original return address.
 260  * @param regs Pointer to CPU register data.
 261  * @return Void.
 262  */
 263 void arch_set_orig_ret_addr(unsigned long orig_ret_addr, struct pt_regs *regs)
 264 {
 265         regs->EREG(ip) = orig_ret_addr;
 266 }
 267
 268 /**
 269  * @brief Removes uprobe.
 270  *
 271  * @param up Pointer to the target uprobe.
 272  * @return Void.
 273  */
 274 void arch_remove_uprobe(struct uprobe *up)
 275 {
 276         struct kprobe *p = up2kp(up);
 277
 278         swap_slot_free(up->sm, p->ainsn.insn);
 279 }
 280
 281 static void set_user_jmp_op(void *from, void *to)
 282 {
 283         struct __arch_jmp_op {
 284                 char op;
 285                 long raddr;
 286         } __packed jop;
 287
 288         jop.raddr = (long)(to) - ((long)(from) + 5);
 289         jop.op = RELATIVEJUMP_INSTRUCTION;
 290
 291         if (!write_proc_vm_atomic(current, (unsigned long)from, &jop,
 292                                   sizeof(jop)))
 293                 printk(KERN_WARN "failed to write jump opcode to user space %p\n",
 294                        from);
 295 }
 296
 297 static void resume_execution(struct kprobe *p,
 298                              struct pt_regs *regs,
 299                              unsigned long flags)
 300 {
 301         unsigned long *tos, tos_dword = 0;
 302         unsigned long copy_eip = (unsigned long)p->ainsn.insn;
 303         unsigned long orig_eip = (unsigned long)p->addr;
 304         kprobe_opcode_t insns[2];
 305
 306         regs->EREG(flags) &= ~TF_MASK;
 307
 308         tos = (unsigned long *)&tos_dword;
 309         if (!read_proc_vm_atomic(current, regs->EREG(sp), &tos_dword,
 310                                  sizeof(tos_dword))) {
 311                 printk(KERN_WARN "failed to read dword from top of the user space stack "
 312                        "%lx!\n", regs->EREG(sp));
 313                 return;
 314         }
 315
 316         if (!read_proc_vm_atomic(current, (unsigned long)p->ainsn.insn, insns,
 317                                  2 * sizeof(kprobe_opcode_t))) {
 318                 printk(KERN_WARN "failed to read first 2 opcodes of instruction copy "
 319                        "from user space %p!\n", p->ainsn.insn);
 320                 return;
 321         }
 322
 323         switch (insns[0]) {
 324         case 0x9c: /* pushfl */
 325                 *tos &= ~(TF_MASK | IF_MASK);
 326                 *tos |= flags & (TF_MASK | IF_MASK);
 327                 break;
 328         case 0xc2: /* iret/ret/lret */
 329         case 0xc3:
 330         case 0xca:
 331         case 0xcb:
 332         case 0xcf:
 333         case 0xea: /* jmp absolute -- eip is correct */
 334                 /* eip is already adjusted, no more changes required */
 335                 p->ainsn.boostable = 1;
 336                 goto no_change;
 337         case 0xe8: /* call relative - Fix return addr */
 338                 *tos = orig_eip + (*tos - copy_eip);
 339                 break;
 340         case 0x9a: /* call absolute -- same as call absolute, indirect */
 341                 *tos = orig_eip + (*tos - copy_eip);
 342
 343                 if (!write_proc_vm_atomic(current,
 344                                           regs->EREG(sp),
 345                                           &tos_dword,
 346                                           sizeof(tos_dword))) {
 347                         printk(KERN_WARN "failed to write dword to top of the"
 348                                " user space stack %lx!\n", regs->EREG(sp));
 349                         return;
 350                 }
 351
 352                 goto no_change;
 353         case 0xff:
 354                 if ((insns[1] & 0x30) == 0x10) {
 355                         /*
 356                          * call absolute, indirect
 357                          * Fix return addr; eip is correct.
 358                          * But this is not boostable
 359                          */
 360                         *tos = orig_eip + (*tos - copy_eip);
 361
 362                         if (!write_proc_vm_atomic(current, regs->EREG(sp),
 363                                                   &tos_dword,
 364                                                   sizeof(tos_dword))) {
 365                                 printk(KERN_WARN "failed to write dword to top of the user space stack %lx!\n",
 366                                        regs->EREG(sp));
 367                                 return;
 368                         }
 369
 370                         goto no_change;
 371                 } else if (((insns[1] & 0x31) == 0x20) || /* jmp near, absolute
 372                                                            * indirect */
 373                            ((insns[1] & 0x31) == 0x21)) {
 374                         /* jmp far, absolute indirect */
 375                         /* eip is correct. And this is boostable */
 376                         p->ainsn.boostable = 1;
 377                         goto no_change;
 378                 }
 379         case 0xf3:
 380                 if (insns[1] == 0xc3)
 381                         /* repz ret special handling: no more changes */
 382                         goto no_change;
 383                 break;
 384         default:
 385                 break;
 386         }
 387
 388         if (!write_proc_vm_atomic(current, regs->EREG(sp), &tos_dword,
 389                                   sizeof(tos_dword))) {
 390                 printk(KERN_WARN "failed to write dword to top of the user space stack %lx!\n",
 391                        regs->EREG(sp));
 392                 return;
 393         }
 394
 395         if (p->ainsn.boostable == 0) {
 396                 if ((regs->EREG(ip) > copy_eip) && (regs->EREG(ip) - copy_eip) +
 397                     5 < MAX_INSN_SIZE) {
 398                         /*
 399                          * These instructions can be executed directly if it
 400                          * jumps back to correct address.
 401                          */
 402                         set_user_jmp_op((void *) regs->EREG(ip),
 403                                         (void *)orig_eip +
 404                                         (regs->EREG(ip) - copy_eip));
 405                         p->ainsn.boostable = 1;
 406                 } else {
 407                         p->ainsn.boostable = -1;
 408                 }
 409         }
 410
 411         regs->EREG(ip) = orig_eip + (regs->EREG(ip) - copy_eip);
 412
 413 no_change:
 414         return;
 415 }
 416
 417 static int uprobe_handler(struct pt_regs *regs)
 418 {
 419         struct kprobe *p;
 420         kprobe_opcode_t *addr;
 421         struct task_struct *task = current;
 422         pid_t tgid = task->tgid;
 423
 424         save_current_flags(regs);
 425
 426         addr = (kprobe_opcode_t *)(regs->EREG(ip) - sizeof(kprobe_opcode_t));
 427         p = get_ukprobe(addr, tgid);
 428
 429         if (p == NULL) {
 430                 void *tramp_addr = (void *)addr - UPROBES_TRAMP_RET_BREAK_IDX;
 431
 432                 p = get_ukprobe_by_insn_slot(tramp_addr, tgid, regs);
 433                 if (p == NULL) {
 434                         printk(KERN_INFO "no_uprobe\n");
 435                         return 0;
 436                 }
 437
 438                 trampoline_uprobe_handler(p, regs);
 439                 return 1;
 440         } else {
 441                 if (!p->pre_handler || !p->pre_handler(p, regs)) {
 442
 443                         if (p->ainsn.boostable == 1 && !p->post_handler) {
 444                                 if (p->ss_addr[smp_processor_id()]) {
 445                                         regs->EREG(ip) = (unsigned long)p->ss_addr[smp_processor_id()];
 446                                         p->ss_addr[smp_processor_id()] = NULL;
 447                                 } else {
 448                                         regs->EREG(ip) = (unsigned long)p->ainsn.insn;
 449                                 }
 450                                 return 1;
 451                         }
 452
 453                         prepare_singlestep(p, regs);
 454                 }
 455         }
 456
 457         set_current_probe(p);
 458
 459         return 1;
 460 }
 461
 462 static int post_uprobe_handler(struct pt_regs *regs)
 463 {
 464         struct kprobe *p = get_current_probe();
 465         unsigned long flags = __get_cpu_var(ucb).flags;
 466
 467         if (p == NULL)
 468                 return 0;
 469
 470         resume_execution(p, regs, flags);
 471         restore_current_flags(regs);
 472
 473         reset_current_probe();
 474
 475         return 1;
 476 }
 477
 478 static int uprobe_exceptions_notify(struct notifier_block *self,
 479                                     unsigned long val, void *data)
 480 {
 481         struct die_args *args = (struct die_args *)data;
 482         int ret = NOTIFY_DONE;
 483
 484         if (args->regs == NULL || !user_mode_vm(args->regs))
 485                 return ret;
 486
 487         switch (val) {
 488 #ifdef CONFIG_KPROBES
 489         case DIE_INT3:
 490 #else
 491         case DIE_TRAP:
 492 #endif
 493                 if (uprobe_handler(args->regs))
 494                         ret = NOTIFY_STOP;
 495                 break;
 496         case DIE_DEBUG:
 497                 if (post_uprobe_handler(args->regs))
 498                         ret = NOTIFY_STOP;
 499                 break;
 500         default:
 501                 break;
 502         }
 503
 504         return ret;
 505 }
 506
 507 static struct notifier_block uprobe_exceptions_nb = {
 508         .notifier_call = uprobe_exceptions_notify,
 509         .priority = INT_MAX
 510 };
 511
 512 /**
 513  * @brief Registers notify.
 514  *
 515  * @return register_die_notifier result.
 516  */
 517 int swap_arch_init_uprobes(void)
 518 {
 519         return register_die_notifier(&uprobe_exceptions_nb);
 520 }
 521
 522 /**
 523  * @brief Unregisters notify.
 524  *
 525  * @return Void.
 526  */
 527 void swap_arch_exit_uprobes(void)
 528 {
 529         unregister_die_notifier(&uprobe_exceptions_nb);
 530 }
 531