arch/powerpc/kernel/optprobes.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * Code for Kernel probes Jump optimization.
   4  *
   5  * Copyright 2017, Anju T, IBM Corp.
   6  */
   7
   8 #include <linux/kprobes.h>
   9 #include <linux/jump_label.h>
  10 #include <linux/types.h>
  11 #include <linux/slab.h>
  12 #include <linux/list.h>
  13 #include <asm/kprobes.h>
  14 #include <asm/ptrace.h>
  15 #include <asm/cacheflush.h>
  16 #include <asm/code-patching.h>
  17 #include <asm/sstep.h>
  18 #include <asm/ppc-opcode.h>
  19 #include <asm/inst.h>
  20
  21 #define TMPL_CALL_HDLR_IDX      \
  22         (optprobe_template_call_handler - optprobe_template_entry)
  23 #define TMPL_EMULATE_IDX        \
  24         (optprobe_template_call_emulate - optprobe_template_entry)
  25 #define TMPL_RET_IDX            \
  26         (optprobe_template_ret - optprobe_template_entry)
  27 #define TMPL_OP_IDX             \
  28         (optprobe_template_op_address - optprobe_template_entry)
  29 #define TMPL_INSN_IDX           \
  30         (optprobe_template_insn - optprobe_template_entry)
  31 #define TMPL_END_IDX            \
  32         (optprobe_template_end - optprobe_template_entry)
  33
  34 DEFINE_INSN_CACHE_OPS(ppc_optinsn);
  35
  36 static bool insn_page_in_use;
  37
  38 static void *__ppc_alloc_insn_page(void)
  39 {
  40         if (insn_page_in_use)
  41                 return NULL;
  42         insn_page_in_use = true;
  43         return &optinsn_slot;
  44 }
  45
  46 static void __ppc_free_insn_page(void *page __maybe_unused)
  47 {
  48         insn_page_in_use = false;
  49 }
  50
  51 struct kprobe_insn_cache kprobe_ppc_optinsn_slots = {
  52         .mutex = __MUTEX_INITIALIZER(kprobe_ppc_optinsn_slots.mutex),
  53         .pages = LIST_HEAD_INIT(kprobe_ppc_optinsn_slots.pages),
  54         /* insn_size initialized later */
  55         .alloc = __ppc_alloc_insn_page,
  56         .free = __ppc_free_insn_page,
  57         .nr_garbage = 0,
  58 };
  59
  60 /*
  61  * Check if we can optimize this probe. Returns NIP post-emulation if this can
  62  * be optimized and 0 otherwise.
  63  */
  64 static unsigned long can_optimize(struct kprobe *p)
  65 {
  66         struct pt_regs regs;
  67         struct instruction_op op;
  68         unsigned long nip = 0;
  69
  70         /*
  71          * kprobe placed for kretprobe during boot time
  72          * has a 'nop' instruction, which can be emulated.
  73          * So further checks can be skipped.
  74          */
  75         if (p->addr == (kprobe_opcode_t *)&kretprobe_trampoline)
  76                 return (unsigned long)p->addr + sizeof(kprobe_opcode_t);
  77
  78         /*
  79          * We only support optimizing kernel addresses, but not
  80          * module addresses.
  81          *
  82          * FIXME: Optimize kprobes placed in module addresses.
  83          */
  84         if (!is_kernel_addr((unsigned long)p->addr))
  85                 return 0;
  86
  87         memset(&regs, 0, sizeof(struct pt_regs));
  88         regs.nip = (unsigned long)p->addr;
  89         regs.trap = 0x0;
  90         regs.msr = MSR_KERNEL;
  91
  92         /*
  93          * Kprobe placed in conditional branch instructions are
  94          * not optimized, as we can't predict the nip prior with
  95          * dummy pt_regs and can not ensure that the return branch
  96          * from detour buffer falls in the range of address (i.e 32MB).
  97          * A branch back from trampoline is set up in the detour buffer
  98          * to the nip returned by the analyse_instr() here.
  99          *
 100          * Ensure that the instruction is not a conditional branch,
 101          * and that can be emulated.
 102          */
 103         if (!is_conditional_branch(ppc_inst_read((struct ppc_inst *)p->ainsn.insn)) &&
 104             analyse_instr(&op, &regs,
 105                           ppc_inst_read((struct ppc_inst *)p->ainsn.insn)) == 1) {
 106                 emulate_update_regs(&regs, &op);
 107                 nip = regs.nip;
 108         }
 109
 110         return nip;
 111 }
 112
 113 static void optimized_callback(struct optimized_kprobe *op,
 114                                struct pt_regs *regs)
 115 {
 116         /* This is possible if op is under delayed unoptimizing */
 117         if (kprobe_disabled(&op->kp))
 118                 return;
 119
 120         preempt_disable();
 121
 122         if (kprobe_running()) {
 123                 kprobes_inc_nmissed_count(&op->kp);
 124         } else {
 125                 __this_cpu_write(current_kprobe, &op->kp);
 126                 regs->nip = (unsigned long)op->kp.addr;
 127                 get_kprobe_ctlblk()->kprobe_status = KPROBE_HIT_ACTIVE;
 128                 opt_pre_handler(&op->kp, regs);
 129                 __this_cpu_write(current_kprobe, NULL);
 130         }
 131
 132         preempt_enable_no_resched();
 133 }
 134 NOKPROBE_SYMBOL(optimized_callback);
 135
 136 void arch_remove_optimized_kprobe(struct optimized_kprobe *op)
 137 {
 138         if (op->optinsn.insn) {
 139                 free_ppc_optinsn_slot(op->optinsn.insn, 1);
 140                 op->optinsn.insn = NULL;
 141         }
 142 }
 143
 144 static void patch_imm32_load_insns(unsigned long val, int reg, kprobe_opcode_t *addr)
 145 {
 146         patch_instruction((struct ppc_inst *)addr,
 147                           ppc_inst(PPC_RAW_LIS(reg, IMM_H(val))));
 148         addr++;
 149
 150         patch_instruction((struct ppc_inst *)addr,
 151                           ppc_inst(PPC_RAW_ORI(reg, reg, IMM_L(val))));
 152 }
 153
 154 /*
 155  * Generate instructions to load provided immediate 64-bit value
 156  * to register 'reg' and patch these instructions at 'addr'.
 157  */
 158 static void patch_imm64_load_insns(unsigned long long val, int reg, kprobe_opcode_t *addr)
 159 {
 160         /* lis reg,(op)@highest */
 161         patch_instruction((struct ppc_inst *)addr,
 162                           ppc_inst(PPC_INST_ADDIS | ___PPC_RT(reg) |
 163                                    ((val >> 48) & 0xffff)));
 164         addr++;
 165
 166         /* ori reg,reg,(op)@higher */
 167         patch_instruction((struct ppc_inst *)addr,
 168                           ppc_inst(PPC_INST_ORI | ___PPC_RA(reg) |
 169                                    ___PPC_RS(reg) | ((val >> 32) & 0xffff)));
 170         addr++;
 171
 172         /* rldicr reg,reg,32,31 */
 173         patch_instruction((struct ppc_inst *)addr,
 174                           ppc_inst(PPC_INST_RLDICR | ___PPC_RA(reg) |
 175                                    ___PPC_RS(reg) | __PPC_SH64(32) | __PPC_ME64(31)));
 176         addr++;
 177
 178         /* oris reg,reg,(op)@h */
 179         patch_instruction((struct ppc_inst *)addr,
 180                           ppc_inst(PPC_INST_ORIS | ___PPC_RA(reg) |
 181                                    ___PPC_RS(reg) | ((val >> 16) & 0xffff)));
 182         addr++;
 183
 184         /* ori reg,reg,(op)@l */
 185         patch_instruction((struct ppc_inst *)addr,
 186                           ppc_inst(PPC_INST_ORI | ___PPC_RA(reg) |
 187                                    ___PPC_RS(reg) | (val & 0xffff)));
 188 }
 189
 190 static void patch_imm_load_insns(unsigned long val, int reg, kprobe_opcode_t *addr)
 191 {
 192         if (IS_ENABLED(CONFIG_PPC64))
 193                 patch_imm64_load_insns(val, reg, addr);
 194         else
 195                 patch_imm32_load_insns(val, reg, addr);
 196 }
 197
 198 int arch_prepare_optimized_kprobe(struct optimized_kprobe *op, struct kprobe *p)
 199 {
 200         struct ppc_inst branch_op_callback, branch_emulate_step, temp;
 201         kprobe_opcode_t *op_callback_addr, *emulate_step_addr, *buff;
 202         long b_offset;
 203         unsigned long nip, size;
 204         int rc, i;
 205
 206         kprobe_ppc_optinsn_slots.insn_size = MAX_OPTINSN_SIZE;
 207
 208         nip = can_optimize(p);
 209         if (!nip)
 210                 return -EILSEQ;
 211
 212         /* Allocate instruction slot for detour buffer */
 213         buff = get_ppc_optinsn_slot();
 214         if (!buff)
 215                 return -ENOMEM;
 216
 217         /*
 218          * OPTPROBE uses 'b' instruction to branch to optinsn.insn.
 219          *
 220          * The target address has to be relatively nearby, to permit use
 221          * of branch instruction in powerpc, because the address is specified
 222          * in an immediate field in the instruction opcode itself, ie 24 bits
 223          * in the opcode specify the address. Therefore the address should
 224          * be within 32MB on either side of the current instruction.
 225          */
 226         b_offset = (unsigned long)buff - (unsigned long)p->addr;
 227         if (!is_offset_in_branch_range(b_offset))
 228                 goto error;
 229
 230         /* Check if the return address is also within 32MB range */
 231         b_offset = (unsigned long)(buff + TMPL_RET_IDX) -
 232                         (unsigned long)nip;
 233         if (!is_offset_in_branch_range(b_offset))
 234                 goto error;
 235
 236         /* Setup template */
 237         /* We can optimize this via patch_instruction_window later */
 238         size = (TMPL_END_IDX * sizeof(kprobe_opcode_t)) / sizeof(int);
 239         pr_devel("Copying template to %p, size %lu\n", buff, size);
 240         for (i = 0; i < size; i++) {
 241                 rc = patch_instruction((struct ppc_inst *)(buff + i),
 242                                        ppc_inst(*(optprobe_template_entry + i)));
 243                 if (rc < 0)
 244                         goto error;
 245         }
 246
 247         /*
 248          * Fixup the template with instructions to:
 249          * 1. load the address of the actual probepoint
 250          */
 251         patch_imm_load_insns((unsigned long)op, 3, buff + TMPL_OP_IDX);
 252
 253         /*
 254          * 2. branch to optimized_callback() and emulate_step()
 255          */
 256         op_callback_addr = (kprobe_opcode_t *)ppc_kallsyms_lookup_name("optimized_callback");
 257         emulate_step_addr = (kprobe_opcode_t *)ppc_kallsyms_lookup_name("emulate_step");
 258         if (!op_callback_addr || !emulate_step_addr) {
 259                 WARN(1, "Unable to lookup optimized_callback()/emulate_step()\n");
 260                 goto error;
 261         }
 262
 263         rc = create_branch(&branch_op_callback,
 264                            (struct ppc_inst *)(buff + TMPL_CALL_HDLR_IDX),
 265                            (unsigned long)op_callback_addr,
 266                            BRANCH_SET_LINK);
 267
 268         rc |= create_branch(&branch_emulate_step,
 269                             (struct ppc_inst *)(buff + TMPL_EMULATE_IDX),
 270                             (unsigned long)emulate_step_addr,
 271                             BRANCH_SET_LINK);
 272
 273         if (rc)
 274                 goto error;
 275
 276         patch_instruction((struct ppc_inst *)(buff + TMPL_CALL_HDLR_IDX),
 277                           branch_op_callback);
 278         patch_instruction((struct ppc_inst *)(buff + TMPL_EMULATE_IDX),
 279                           branch_emulate_step);
 280
 281         /*
 282          * 3. load instruction to be emulated into relevant register, and
 283          */
 284         temp = ppc_inst_read((struct ppc_inst *)p->ainsn.insn);
 285         patch_imm_load_insns(ppc_inst_as_ulong(temp), 4, buff + TMPL_INSN_IDX);
 286
 287         /*
 288          * 4. branch back from trampoline
 289          */
 290         patch_branch((struct ppc_inst *)(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 ppc_inst instr;
 323         struct optimized_kprobe *op;
 324         struct optimized_kprobe *tmp;
 325
 326         list_for_each_entry_safe(op, tmp, oplist, list) {
 327                 /*
 328                  * Backup instructions which will be replaced
 329                  * by jump address
 330                  */
 331                 memcpy(op->optinsn.copied_insn, op->kp.addr,
 332                                                RELATIVEJUMP_SIZE);
 333                 create_branch(&instr,
 334                               (struct ppc_inst *)op->kp.addr,
 335                               (unsigned long)op->optinsn.insn, 0);
 336                 patch_instruction((struct ppc_inst *)op->kp.addr, instr);
 337                 list_del_init(&op->list);
 338         }
 339 }
 340
 341 void arch_unoptimize_kprobe(struct optimized_kprobe *op)
 342 {
 343         arch_arm_kprobe(&op->kp);
 344 }
 345
 346 void arch_unoptimize_kprobes(struct list_head *oplist,
 347                              struct list_head *done_list)
 348 {
 349         struct optimized_kprobe *op;
 350         struct optimized_kprobe *tmp;
 351
 352         list_for_each_entry_safe(op, tmp, oplist, list) {
 353                 arch_unoptimize_kprobe(op);
 354                 list_move(&op->list, done_list);
 355         }
 356 }
 357
 358 int arch_within_optimized_kprobe(struct optimized_kprobe *op,
 359                                  unsigned long addr)
 360 {
 361         return ((unsigned long)op->kp.addr <= addr &&
 362                 (unsigned long)op->kp.addr + RELATIVEJUMP_SIZE > addr);
 363 }