uprobe/arch/asm-x86/swap_uprobes.c

   1 /*
   2  *  Dynamic Binary Instrumentation Module based on KProbes
   3  *  modules/uprobe/arch/asm-x86/swap_uprobes.c
   4  *
   5  * This program is free software; you can redistribute it and/or modify
   6  * it under the terms of the GNU General Public License as published by
   7  * the Free Software Foundation; either version 2 of the License, or
   8  * (at your option) any later version.
   9  *
  10  * This program is distributed in the hope that it will be useful,
  11  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  13  * GNU General Public License for more details.
  14  *
  15  * You should have received a copy of the GNU General Public License
  16  * along with this program; if not, write to the Free Software
  17  * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
  18  *
  19  * Copyright (C) Samsung Electronics, 2006-2010
  20  *
  21  * 2008-2009    Alexey Gerenkov <a.gerenkov@samsung.com> User-Space
  22  *              Probes initial implementation; Support x86/ARM/MIPS for both user and kernel spaces.
  23  * 2010         Ekaterina Gorelkina <e.gorelkina@samsung.com>: redesign module for separating core and arch parts
  24  *
  25  */
  26
  27 #include <linux/kdebug.h>
  28 #include <asm/dbi_kprobes.h>
  29 #include <swap_uprobes.h>
  30 #include <asm/swap_uprobes.h>
  31 #include <dbi_insn_slots.h>
  32
  33 struct uprobe_ctlblk {
  34         unsigned long flags;
  35         struct kprobe *p;
  36 };
  37
  38 static DEFINE_PER_CPU(struct uprobe_ctlblk, ucb) = { 0, NULL };
  39
  40 static struct kprobe *get_current_probe(void)
  41 {
  42         return __get_cpu_var(ucb).p;
  43 }
  44
  45 static void set_current_probe(struct kprobe *p)
  46 {
  47         __get_cpu_var(ucb).p = p;
  48 }
  49
  50 static void reset_current_probe(void)
  51 {
  52         set_current_probe(NULL);
  53 }
  54
  55 static void save_current_flags(struct pt_regs *regs)
  56 {
  57         __get_cpu_var(ucb).flags = regs->EREG(flags);
  58 }
  59
  60 static void restore_current_flags(struct pt_regs *regs)
  61 {
  62         regs->EREG(flags) &= ~IF_MASK;
  63         regs->EREG(flags) |= __get_cpu_var(ucb).flags & IF_MASK;
  64 }
  65
  66 int arch_prepare_uprobe(struct uprobe *up, struct hlist_head *page_list)
  67 {
  68         int ret = 0;
  69         struct kprobe *p = &up->kp;
  70         struct task_struct *task = up->task;
  71         kprobe_opcode_t insns[UPROBES_TRAMP_LEN];
  72
  73         if (!ret) {
  74                 kprobe_opcode_t insn[MAX_INSN_SIZE];
  75                 struct arch_specific_insn ainsn;
  76
  77                 if (!read_proc_vm_atomic(task, (unsigned long)p->addr, &insn, MAX_INSN_SIZE * sizeof(kprobe_opcode_t)))
  78                         panic("failed to read memory %p!\n", p->addr);
  79
  80                 ainsn.insn = insn;
  81                 ret = arch_check_insn(&ainsn);
  82                 if (!ret) {
  83                         p->opcode = insn[0];
  84                         p->ainsn.insn = alloc_insn_slot(up->sm);
  85                         if (!p->ainsn.insn)
  86                                 return -ENOMEM;
  87
  88                         if (can_boost(insn))
  89                                 p->ainsn.boostable = 0;
  90                         else
  91                                 p->ainsn.boostable = -1;
  92
  93                         memcpy(&insns[UPROBES_TRAMP_INSN_IDX], insn, MAX_INSN_SIZE*sizeof(kprobe_opcode_t));
  94                         insns[UPROBES_TRAMP_RET_BREAK_IDX] = BREAKPOINT_INSTRUCTION;
  95
  96                         if (!write_proc_vm_atomic(task, (unsigned long)p->ainsn.insn, insns, sizeof(insns))) {
  97                                 free_insn_slot(up->sm, p->ainsn.insn);
  98                                 panic("failed to write memory %p!\n", p->ainsn.insn);
  99                                 return -EINVAL;
 100                         }
 101                 }
 102         }
 103
 104         return ret;
 105 }
 106
 107 int setjmp_upre_handler(struct kprobe *p, struct pt_regs *regs)
 108 {
 109         struct uprobe *up = container_of(p, struct uprobe, kp);
 110         struct ujprobe *jp = container_of(up, struct ujprobe, up);
 111         kprobe_pre_entry_handler_t pre_entry = (kprobe_pre_entry_handler_t)jp->pre_entry;
 112         entry_point_t entry = (entry_point_t)jp->entry;
 113         unsigned long addr, args[6];
 114
 115         /* FIXME some user space apps crash if we clean interrupt bit */
 116         //regs->EREG(flags) &= ~IF_MASK;
 117 #if LINUX_VERSION_CODE >= KERNEL_VERSION(2, 6, 18)
 118         trace_hardirqs_off();
 119 #endif
 120
 121         /* read first 6 args from stack */
 122         if (!read_proc_vm_atomic(current, regs->EREG(sp) + 4, args, sizeof(args)))
 123                 panic("failed to read user space func arguments %lx!\n", regs->EREG(sp) + 4);
 124
 125         if (pre_entry)
 126                 p->ss_addr = pre_entry(jp->priv_arg, regs);
 127
 128         if (entry)
 129                 entry(args[0], args[1], args[2], args[3], args[4], args[5]);
 130         else
 131                 arch_ujprobe_return();
 132
 133         return 0;
 134 }
 135
 136 void arch_prepare_uretprobe(struct uretprobe_instance *ri, struct pt_regs *regs)
 137 {
 138         /* Replace the return addr with trampoline addr */
 139         unsigned long ra = (unsigned long)(ri->rp->up.kp.ainsn.insn + UPROBES_TRAMP_RET_BREAK_IDX);
 140
 141         if (!read_proc_vm_atomic(current, regs->EREG(sp), &(ri->ret_addr), sizeof(ri->ret_addr)))
 142                 panic("failed to read user space func ra %lx!\n", regs->EREG(sp));
 143
 144         if (!write_proc_vm_atomic(current, regs->EREG(sp), &ra, sizeof(ra)))
 145                 panic("failed to write user space func ra %lx!\n", regs->EREG(sp));
 146 }
 147
 148 unsigned long arch_get_trampoline_addr(struct kprobe *p, struct pt_regs *regs)
 149 {
 150         return (unsigned long)(p->ainsn.insn + UPROBES_TRAMP_RET_BREAK_IDX);
 151 }
 152
 153 void arch_set_orig_ret_addr(unsigned long orig_ret_addr, struct pt_regs *regs)
 154 {
 155         regs->EREG(ip) = orig_ret_addr;
 156 }
 157
 158 static void set_user_jmp_op(void *from, void *to)
 159 {
 160         struct __arch_jmp_op
 161         {
 162                 char op;
 163                 long raddr;
 164         } __attribute__ ((packed)) jop;
 165
 166         jop.raddr = (long)(to) - ((long)(from) + 5);
 167         jop.op = RELATIVEJUMP_INSTRUCTION;
 168
 169         if (!write_proc_vm_atomic(current, (unsigned long)from, &jop, sizeof(jop)))
 170                 panic("failed to write jump opcode to user space %p!\n", from);
 171 }
 172
 173 static void resume_execution(struct kprobe *p, struct pt_regs *regs, unsigned long flags)
 174 {
 175         unsigned long *tos, tos_dword = 0;
 176         unsigned long copy_eip = (unsigned long)p->ainsn.insn;
 177         unsigned long orig_eip = (unsigned long)p->addr;
 178         kprobe_opcode_t insns[2];
 179
 180         regs->EREG(flags) &= ~TF_MASK;
 181
 182         tos = (unsigned long *)&tos_dword;
 183         if (!read_proc_vm_atomic(current, regs->EREG(sp), &tos_dword, sizeof(tos_dword)))
 184                 panic("failed to read dword from top of the user space stack %lx!\n", regs->EREG(sp));
 185
 186         if (!read_proc_vm_atomic(current, (unsigned long)p->ainsn.insn, insns, 2 * sizeof(kprobe_opcode_t)))
 187                 panic("failed to read first 2 opcodes of instruction copy from user space %p!\n", p->ainsn.insn);
 188
 189         switch (insns[0]) {
 190                 case 0x9c:              /* pushfl */
 191                         *tos &= ~(TF_MASK | IF_MASK);
 192                         *tos |= flags & (TF_MASK | IF_MASK);
 193                         break;
 194                 case 0xc2:              /* iret/ret/lret */
 195                 case 0xc3:
 196                 case 0xca:
 197                 case 0xcb:
 198                 case 0xcf:
 199                 case 0xea:              /* jmp absolute -- eip is correct */
 200                         /* eip is already adjusted, no more changes required */
 201                         p->ainsn.boostable = 1;
 202                         goto no_change;
 203                 case 0xe8:              /* call relative - Fix return addr */
 204                         *tos = orig_eip + (*tos - copy_eip);
 205                         break;
 206                 case 0x9a:              /* call absolute -- same as call absolute, indirect */
 207                         *tos = orig_eip + (*tos - copy_eip);
 208
 209                         if (!write_proc_vm_atomic(current, regs->EREG (sp), &tos_dword, sizeof(tos_dword)))
 210                                 panic("failed to write dword to top of the user space stack %lx!\n", regs->EREG (sp));
 211
 212                         goto no_change;
 213                 case 0xff:
 214                         if ((insns[1] & 0x30) == 0x10) {
 215                                 /*
 216                                  * call absolute, indirect
 217                                  * Fix return addr; eip is correct.
 218                                  * But this is not boostable
 219                                  */
 220                                 *tos = orig_eip + (*tos - copy_eip);
 221
 222                                 if (!write_proc_vm_atomic(current, regs->EREG(sp), &tos_dword, sizeof(tos_dword)))
 223                                         panic("failed to write dword to top of the user space stack %lx!\n", regs->EREG(sp));
 224
 225                                 goto no_change;
 226                         } else if (((insns[1] & 0x31) == 0x20) || /* jmp near, absolute indirect */
 227                                    ((insns[1] & 0x31) == 0x21)) {
 228                                 /* jmp far, absolute indirect */
 229                                 /* eip is correct. And this is boostable */
 230                                 p->ainsn.boostable = 1;
 231                                 goto no_change;
 232                         }
 233                 default:
 234                         break;
 235         }
 236
 237         if (!write_proc_vm_atomic(current, regs->EREG(sp), &tos_dword, sizeof(tos_dword)))
 238                 panic("failed to write dword to top of the user space stack %lx!\n", regs->EREG(sp));
 239
 240         if (p->ainsn.boostable == 0) {
 241                 if ((regs->EREG(ip) > copy_eip) && (regs->EREG(ip) - copy_eip) + 5 < MAX_INSN_SIZE) {
 242                         /*
 243                          * These instructions can be executed directly if it
 244                          * jumps back to correct address.
 245                          */
 246                         set_user_jmp_op((void *) regs->EREG(ip), (void *)orig_eip + (regs->EREG(ip) - copy_eip));
 247                         p->ainsn.boostable = 1;
 248                 } else {
 249                         p->ainsn.boostable = -1;
 250                 }
 251         }
 252
 253         regs->EREG(ip) = orig_eip + (regs->EREG(ip) - copy_eip);
 254
 255 no_change:
 256         return;
 257 }
 258
 259 static int uprobe_handler(struct pt_regs *regs)
 260 {
 261         struct kprobe *p;
 262         kprobe_opcode_t *addr;
 263         struct task_struct *task = current;
 264         pid_t tgid = task->tgid;
 265
 266         save_current_flags(regs);
 267
 268         addr = (kprobe_opcode_t *)(regs->EREG(ip) - sizeof(kprobe_opcode_t));
 269         p = get_ukprobe(addr, tgid);
 270
 271         if (p == NULL) {
 272                 p = get_ukprobe_by_insn_slot(addr, tgid, regs);
 273
 274                 if (p == NULL) {
 275                         printk("no_uprobe\n");
 276                         return 0;
 277                 }
 278
 279                 trampoline_uprobe_handler(p, regs);
 280                 return 1;
 281         } else {
 282                 if (!p->pre_handler || !p->pre_handler(p, regs)) {
 283                         if (p->ainsn.boostable == 1 && !p->post_handler) {
 284                                 regs->EREG(ip) = (unsigned long)p->ainsn.insn;
 285                                 return 1;
 286                         }
 287
 288                         prepare_singlestep(p, regs);
 289                 }
 290         }
 291
 292         set_current_probe(p);
 293
 294         return 1;
 295 }
 296
 297 static int post_uprobe_handler(struct pt_regs *regs)
 298 {
 299         struct kprobe *p = get_current_probe();
 300         unsigned long flags = __get_cpu_var(ucb).flags;
 301
 302         if (p == NULL)
 303                 return 0;
 304
 305         resume_execution(p, regs, flags);
 306         restore_current_flags(regs);
 307
 308         reset_current_probe();
 309
 310         return 1;
 311 }
 312
 313 static int uprobe_exceptions_notify(struct notifier_block *self, unsigned long val, void *data)
 314 {
 315         struct die_args *args = (struct die_args *)data;
 316         int ret = NOTIFY_DONE;
 317
 318         if (args->regs && !user_mode_vm(args->regs))
 319                 return ret;
 320
 321         switch (val) {
 322 #ifdef CONFIG_KPROBES
 323                 case DIE_INT3:
 324 #else
 325                 case DIE_TRAP:
 326 #endif
 327                         if (uprobe_handler(args->regs))
 328                                 ret = NOTIFY_STOP;
 329                         break;
 330                 case DIE_DEBUG:
 331                         if (post_uprobe_handler(args->regs))
 332                                 ret = NOTIFY_STOP;
 333                         break;
 334                 default:
 335                         break;
 336         }
 337
 338         return ret;
 339 }
 340
 341 static struct notifier_block uprobe_exceptions_nb = {
 342         .notifier_call = uprobe_exceptions_notify,
 343         .priority = INT_MAX
 344 };
 345
 346 int swap_arch_init_uprobes(void)
 347 {
 348         return register_die_notifier(&uprobe_exceptions_nb);
 349 }
 350
 351 void swap_arch_exit_uprobes(void)
 352 {
 353         unregister_die_notifier(&uprobe_exceptions_nb);
 354 }
 355