arch/powerpc/kernel/optprobes.c

   1 /*
   2  * Code for Kernel probes Jump optimization.
   3  *
   4  * Copyright 2017, Anju T, IBM Corp.
   5  *
   6  * This program is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU General Public License
   8  * as published by the Free Software Foundation; either version
   9  * 2 of the License, or (at your option) any later version.
  10  */
  11
  12 #include <linux/kprobes.h>
  13 #include <linux/jump_label.h>
  14 #include <linux/types.h>
  15 #include <linux/slab.h>
  16 #include <linux/list.h>
  17 #include <asm/kprobes.h>
  18 #include <asm/ptrace.h>
  19 #include <asm/cacheflush.h>
  20 #include <asm/code-patching.h>
  21 #include <asm/sstep.h>
  22 #include <asm/ppc-opcode.h>
  23
  24 #define TMPL_CALL_HDLR_IDX      \
  25         (optprobe_template_call_handler - optprobe_template_entry)
  26 #define TMPL_EMULATE_IDX        \
  27         (optprobe_template_call_emulate - optprobe_template_entry)
  28 #define TMPL_RET_IDX            \
  29         (optprobe_template_ret - optprobe_template_entry)
  30 #define TMPL_OP_IDX             \
  31         (optprobe_template_op_address - optprobe_template_entry)
  32 #define TMPL_INSN_IDX           \
  33         (optprobe_template_insn - optprobe_template_entry)
  34 #define TMPL_END_IDX            \
  35         (optprobe_template_end - optprobe_template_entry)
  36
  37 DEFINE_INSN_CACHE_OPS(ppc_optinsn);
  38
  39 static bool insn_page_in_use;
  40
  41 static void *__ppc_alloc_insn_page(void)
  42 {
  43         if (insn_page_in_use)
  44                 return NULL;
  45         insn_page_in_use = true;
  46         return &optinsn_slot;
  47 }
  48
  49 static void __ppc_free_insn_page(void *page __maybe_unused)
  50 {
  51         insn_page_in_use = false;
  52 }
  53
  54 struct kprobe_insn_cache kprobe_ppc_optinsn_slots = {
  55         .mutex = __MUTEX_INITIALIZER(kprobe_ppc_optinsn_slots.mutex),
  56         .pages = LIST_HEAD_INIT(kprobe_ppc_optinsn_slots.pages),
  57         /* insn_size initialized later */
  58         .alloc = __ppc_alloc_insn_page,
  59         .free = __ppc_free_insn_page,
  60         .nr_garbage = 0,
  61 };
  62
  63 /*
  64  * Check if we can optimize this probe. Returns NIP post-emulation if this can
  65  * be optimized and 0 otherwise.
  66  */
  67 static unsigned long can_optimize(struct kprobe *p)
  68 {
  69         struct pt_regs regs;
  70         struct instruction_op op;
  71         unsigned long nip = 0;
  72
  73         /*
  74          * kprobe placed for kretprobe during boot time
  75          * has a 'nop' instruction, which can be emulated.
  76          * So further checks can be skipped.
  77          */
  78         if (p->addr == (kprobe_opcode_t *)&kretprobe_trampoline)
  79                 return (unsigned long)p->addr + sizeof(kprobe_opcode_t);
  80
  81         /*
  82          * We only support optimizing kernel addresses, but not
  83          * module addresses.
  84          *
  85          * FIXME: Optimize kprobes placed in module addresses.
  86          */
  87         if (!is_kernel_addr((unsigned long)p->addr))
  88                 return 0;
  89
  90         memset(&regs, 0, sizeof(struct pt_regs));
  91         regs.nip = (unsigned long)p->addr;
  92         regs.trap = 0x0;
  93         regs.msr = MSR_KERNEL;
  94
  95         /*
  96          * Kprobe placed in conditional branch instructions are
  97          * not optimized, as we can't predict the nip prior with
  98          * dummy pt_regs and can not ensure that the return branch
  99          * from detour buffer falls in the range of address (i.e 32MB).
 100          * A branch back from trampoline is set up in the detour buffer
 101          * to the nip returned by the analyse_instr() here.
 102          *
 103          * Ensure that the instruction is not a conditional branch,
 104          * and that can be emulated.
 105          */
 106         if (!is_conditional_branch(*p->ainsn.insn) &&
 107                         analyse_instr(&op, &regs, *p->ainsn.insn) == 1) {
 108                 emulate_update_regs(&regs, &op);
 109                 nip = regs.nip;
 110         }
 111
 112         return nip;
 113 }
 114
 115 static void optimized_callback(struct optimized_kprobe *op,
 116                                struct pt_regs *regs)
 117 {
 118         unsigned long flags;
 119
 120         /* This is possible if op is under delayed unoptimizing */
 121         if (kprobe_disabled(&op->kp))
 122                 return;
 123
 124         local_irq_save(flags);
 125         hard_irq_disable();
 126         preempt_disable();
 127
 128         if (kprobe_running()) {
 129                 kprobes_inc_nmissed_count(&op->kp);
 130         } else {
 131                 __this_cpu_write(current_kprobe, &op->kp);
 132                 regs->nip = (unsigned long)op->kp.addr;
 133                 get_kprobe_ctlblk()->kprobe_status = KPROBE_HIT_ACTIVE;
 134                 opt_pre_handler(&op->kp, regs);
 135                 __this_cpu_write(current_kprobe, NULL);
 136         }
 137
 138         /*
 139          * No need for an explicit __hard_irq_enable() here.
 140          * local_irq_restore() will re-enable interrupts,
 141          * if they were hard disabled.
 142          */
 143         preempt_enable_no_resched();
 144         local_irq_restore(flags);
 145 }
 146 NOKPROBE_SYMBOL(optimized_callback);
 147
 148 void arch_remove_optimized_kprobe(struct optimized_kprobe *op)
 149 {
 150         if (op->optinsn.insn) {
 151                 free_ppc_optinsn_slot(op->optinsn.insn, 1);
 152                 op->optinsn.insn = NULL;
 153         }
 154 }
 155
 156 /*
 157  * emulate_step() requires insn to be emulated as
 158  * second parameter. Load register 'r4' with the
 159  * instruction.
 160  */
 161 void patch_imm32_load_insns(unsigned int val, kprobe_opcode_t *addr)
 162 {
 163         /* addis r4,0,(insn)@h */
 164         patch_instruction(addr, PPC_INST_ADDIS | ___PPC_RT(4) |
 165                           ((val >> 16) & 0xffff));
 166         addr++;
 167
 168         /* ori r4,r4,(insn)@l */
 169         patch_instruction(addr, PPC_INST_ORI | ___PPC_RA(4) |
 170                           ___PPC_RS(4) | (val & 0xffff));
 171 }
 172
 173 /*
 174  * Generate instructions to load provided immediate 64-bit value
 175  * to register 'r3' and patch these instructions at 'addr'.
 176  */
 177 void patch_imm64_load_insns(unsigned long val, kprobe_opcode_t *addr)
 178 {
 179         /* lis r3,(op)@highest */
 180         patch_instruction(addr, PPC_INST_ADDIS | ___PPC_RT(3) |
 181                           ((val >> 48) & 0xffff));
 182         addr++;
 183
 184         /* ori r3,r3,(op)@higher */
 185         patch_instruction(addr, PPC_INST_ORI | ___PPC_RA(3) |
 186                           ___PPC_RS(3) | ((val >> 32) & 0xffff));
 187         addr++;
 188
 189         /* rldicr r3,r3,32,31 */
 190         patch_instruction(addr, PPC_INST_RLDICR | ___PPC_RA(3) |
 191                           ___PPC_RS(3) | __PPC_SH64(32) | __PPC_ME64(31));
 192         addr++;
 193
 194         /* oris r3,r3,(op)@h */
 195         patch_instruction(addr, PPC_INST_ORIS | ___PPC_RA(3) |
 196                           ___PPC_RS(3) | ((val >> 16) & 0xffff));
 197         addr++;
 198
 199         /* ori r3,r3,(op)@l */
 200         patch_instruction(addr, PPC_INST_ORI | ___PPC_RA(3) |
 201                           ___PPC_RS(3) | (val & 0xffff));
 202 }
 203
 204 int arch_prepare_optimized_kprobe(struct optimized_kprobe *op, struct kprobe *p)
 205 {
 206         kprobe_opcode_t *buff, branch_op_callback, branch_emulate_step;
 207         kprobe_opcode_t *op_callback_addr, *emulate_step_addr;
 208         long b_offset;
 209         unsigned long nip, size;
 210         int rc, i;
 211
 212         kprobe_ppc_optinsn_slots.insn_size = MAX_OPTINSN_SIZE;
 213
 214         nip = can_optimize(p);
 215         if (!nip)
 216                 return -EILSEQ;
 217
 218         /* Allocate instruction slot for detour buffer */
 219         buff = get_ppc_optinsn_slot();
 220         if (!buff)
 221                 return -ENOMEM;
 222
 223         /*
 224          * OPTPROBE uses 'b' instruction to branch to optinsn.insn.
 225          *
 226          * The target address has to be relatively nearby, to permit use
 227          * of branch instruction in powerpc, because the address is specified
 228          * in an immediate field in the instruction opcode itself, ie 24 bits
 229          * in the opcode specify the address. Therefore the address should
 230          * be within 32MB on either side of the current instruction.
 231          */
 232         b_offset = (unsigned long)buff - (unsigned long)p->addr;
 233         if (!is_offset_in_branch_range(b_offset))
 234                 goto error;
 235
 236         /* Check if the return address is also within 32MB range */
 237         b_offset = (unsigned long)(buff + TMPL_RET_IDX) -
 238                         (unsigned long)nip;
 239         if (!is_offset_in_branch_range(b_offset))
 240                 goto error;
 241
 242         /* Setup template */
 243         /* We can optimize this via patch_instruction_window later */
 244         size = (TMPL_END_IDX * sizeof(kprobe_opcode_t)) / sizeof(int);
 245         pr_devel("Copying template to %p, size %lu\n", buff, size);
 246         for (i = 0; i < size; i++) {
 247                 rc = patch_instruction(buff + i, *(optprobe_template_entry + i));
 248                 if (rc < 0)
 249                         goto error;
 250         }
 251
 252         /*
 253          * Fixup the template with instructions to:
 254          * 1. load the address of the actual probepoint
 255          */
 256         patch_imm64_load_insns((unsigned long)op, buff + TMPL_OP_IDX);
 257
 258         /*
 259          * 2. branch to optimized_callback() and emulate_step()
 260          */
 261         op_callback_addr = (kprobe_opcode_t *)ppc_kallsyms_lookup_name("optimized_callback");
 262         emulate_step_addr = (kprobe_opcode_t *)ppc_kallsyms_lookup_name("emulate_step");
 263         if (!op_callback_addr || !emulate_step_addr) {
 264                 WARN(1, "Unable to lookup optimized_callback()/emulate_step()\n");
 265                 goto error;
 266         }
 267
 268         branch_op_callback = create_branch((unsigned int *)buff + TMPL_CALL_HDLR_IDX,
 269                                 (unsigned long)op_callback_addr,
 270                                 BRANCH_SET_LINK);
 271
 272         branch_emulate_step = create_branch((unsigned int *)buff + TMPL_EMULATE_IDX,
 273                                 (unsigned long)emulate_step_addr,
 274                                 BRANCH_SET_LINK);
 275
 276         if (!branch_op_callback || !branch_emulate_step)
 277                 goto error;
 278
 279         patch_instruction(buff + TMPL_CALL_HDLR_IDX, branch_op_callback);
 280         patch_instruction(buff + TMPL_EMULATE_IDX, branch_emulate_step);
 281
 282         /*
 283          * 3. load instruction to be emulated into relevant register, and
 284          */
 285         patch_imm32_load_insns(*p->ainsn.insn, buff + TMPL_INSN_IDX);
 286
 287         /*
 288          * 4. branch back from trampoline
 289          */
 290         patch_branch(buff + TMPL_RET_IDX, (unsigned long)nip, 0);
 291
 292         flush_icache_range((unsigned long)buff,
 293                            (unsigned long)(&buff[TMPL_END_IDX]));
 294
 295         op->optinsn.insn = buff;
 296
 297         return 0;
 298
 299 error:
 300         free_ppc_optinsn_slot(buff, 0);
 301         return -ERANGE;
 302
 303 }
 304
 305 int arch_prepared_optinsn(struct arch_optimized_insn *optinsn)
 306 {
 307         return optinsn->insn != NULL;
 308 }
 309
 310 /*
 311  * On powerpc, Optprobes always replaces one instruction (4 bytes
 312  * aligned and 4 bytes long). It is impossible to encounter another
 313  * kprobe in this address range. So always return 0.
 314  */
 315 int arch_check_optimized_kprobe(struct optimized_kprobe *op)
 316 {
 317         return 0;
 318 }
 319
 320 void arch_optimize_kprobes(struct list_head *oplist)
 321 {
 322         struct optimized_kprobe *op;
 323         struct optimized_kprobe *tmp;
 324
 325         list_for_each_entry_safe(op, tmp, oplist, list) {
 326                 /*
 327                  * Backup instructions which will be replaced
 328                  * by jump address
 329                  */
 330                 memcpy(op->optinsn.copied_insn, op->kp.addr,
 331                                                RELATIVEJUMP_SIZE);
 332                 patch_instruction(op->kp.addr,
 333                         create_branch((unsigned int *)op->kp.addr,
 334                                       (unsigned long)op->optinsn.insn, 0));
 335                 list_del_init(&op->list);
 336         }
 337 }
 338
 339 void arch_unoptimize_kprobe(struct optimized_kprobe *op)
 340 {
 341         arch_arm_kprobe(&op->kp);
 342 }
 343
 344 void arch_unoptimize_kprobes(struct list_head *oplist,
 345                              struct list_head *done_list)
 346 {
 347         struct optimized_kprobe *op;
 348         struct optimized_kprobe *tmp;
 349
 350         list_for_each_entry_safe(op, tmp, oplist, list) {
 351                 arch_unoptimize_kprobe(op);
 352                 list_move(&op->list, done_list);
 353         }
 354 }
 355
 356 int arch_within_optimized_kprobe(struct optimized_kprobe *op,
 357                                  unsigned long addr)
 358 {
 359         return ((unsigned long)op->kp.addr <= addr &&
 360                 (unsigned long)op->kp.addr + RELATIVEJUMP_SIZE > addr);
 361 }