1 /* 2 * Performance counter callchain support - powerpc architecture code 3 * 4 * Copyright © 2009 Paul Mackerras, IBM Corporation. 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 #include <linux/kernel.h> 12 #include <linux/sched.h> 13 #include <linux/perf_event.h> 14 #include <linux/percpu.h> 15 #include <linux/uaccess.h> 16 #include <linux/mm.h> 17 #include <asm/ptrace.h> 18 #include <asm/pgtable.h> 19 #include <asm/sigcontext.h> 20 #include <asm/ucontext.h> 21 #include <asm/vdso.h> 22 #ifdef CONFIG_PPC64 23 #include "../kernel/ppc32.h" 24 #endif 25 26 27 /* 28 * Is sp valid as the address of the next kernel stack frame after prev_sp? 29 * The next frame may be in a different stack area but should not go 30 * back down in the same stack area. 31 */ 32 static int valid_next_sp(unsigned long sp, unsigned long prev_sp) 33 { 34 if (sp & 0xf) 35 return 0; /* must be 16-byte aligned */ 36 if (!validate_sp(sp, current, STACK_FRAME_OVERHEAD)) 37 return 0; 38 if (sp >= prev_sp + STACK_FRAME_MIN_SIZE) 39 return 1; 40 /* 41 * sp could decrease when we jump off an interrupt stack 42 * back to the regular process stack. 43 */ 44 if ((sp & ~(THREAD_SIZE - 1)) != (prev_sp & ~(THREAD_SIZE - 1))) 45 return 1; 46 return 0; 47 } 48 49 void 50 perf_callchain_kernel(struct perf_callchain_entry *entry, struct pt_regs *regs) 51 { 52 unsigned long sp, next_sp; 53 unsigned long next_ip; 54 unsigned long lr; 55 long level = 0; 56 unsigned long *fp; 57 58 lr = regs->link; 59 sp = regs->gpr[1]; 60 perf_callchain_store(entry, perf_instruction_pointer(regs)); 61 62 if (!validate_sp(sp, current, STACK_FRAME_OVERHEAD)) 63 return; 64 65 for (;;) { 66 fp = (unsigned long *) sp; 67 next_sp = fp[0]; 68 69 if (next_sp == sp + STACK_INT_FRAME_SIZE && 70 fp[STACK_FRAME_MARKER] == STACK_FRAME_REGS_MARKER) { 71 /* 72 * This looks like an interrupt frame for an 73 * interrupt that occurred in the kernel 74 */ 75 regs = (struct pt_regs *)(sp + STACK_FRAME_OVERHEAD); 76 next_ip = regs->nip; 77 lr = regs->link; 78 level = 0; 79 perf_callchain_store(entry, PERF_CONTEXT_KERNEL); 80 81 } else { 82 if (level == 0) 83 next_ip = lr; 84 else 85 next_ip = fp[STACK_FRAME_LR_SAVE]; 86 87 /* 88 * We can't tell which of the first two addresses 89 * we get are valid, but we can filter out the 90 * obviously bogus ones here. We replace them 91 * with 0 rather than removing them entirely so 92 * that userspace can tell which is which. 93 */ 94 if ((level == 1 && next_ip == lr) || 95 (level <= 1 && !kernel_text_address(next_ip))) 96 next_ip = 0; 97 98 ++level; 99 } 100 101 perf_callchain_store(entry, next_ip); 102 if (!valid_next_sp(next_sp, sp)) 103 return; 104 sp = next_sp; 105 } 106 } 107 108 #ifdef CONFIG_PPC64 109 /* 110 * On 64-bit we don't want to invoke hash_page on user addresses from 111 * interrupt context, so if the access faults, we read the page tables 112 * to find which page (if any) is mapped and access it directly. 113 */ 114 static int read_user_stack_slow(void __user *ptr, void *ret, int nb) 115 { 116 pgd_t *pgdir; 117 pte_t *ptep, pte; 118 unsigned shift; 119 unsigned long addr = (unsigned long) ptr; 120 unsigned long offset; 121 unsigned long pfn; 122 void *kaddr; 123 124 pgdir = current->mm->pgd; 125 if (!pgdir) 126 return -EFAULT; 127 128 ptep = find_linux_pte_or_hugepte(pgdir, addr, &shift); 129 if (!shift) 130 shift = PAGE_SHIFT; 131 132 /* align address to page boundary */ 133 offset = addr & ((1UL << shift) - 1); 134 addr -= offset; 135 136 if (ptep == NULL) 137 return -EFAULT; 138 pte = *ptep; 139 if (!pte_present(pte) || !(pte_val(pte) & _PAGE_USER)) 140 return -EFAULT; 141 pfn = pte_pfn(pte); 142 if (!page_is_ram(pfn)) 143 return -EFAULT; 144 145 /* no highmem to worry about here */ 146 kaddr = pfn_to_kaddr(pfn); 147 memcpy(ret, kaddr + offset, nb); 148 return 0; 149 } 150 151 static int read_user_stack_64(unsigned long __user *ptr, unsigned long *ret) 152 { 153 if ((unsigned long)ptr > TASK_SIZE - sizeof(unsigned long) || 154 ((unsigned long)ptr & 7)) 155 return -EFAULT; 156 157 pagefault_disable(); 158 if (!__get_user_inatomic(*ret, ptr)) { 159 pagefault_enable(); 160 return 0; 161 } 162 pagefault_enable(); 163 164 return read_user_stack_slow(ptr, ret, 8); 165 } 166 167 static int read_user_stack_32(unsigned int __user *ptr, unsigned int *ret) 168 { 169 if ((unsigned long)ptr > TASK_SIZE - sizeof(unsigned int) || 170 ((unsigned long)ptr & 3)) 171 return -EFAULT; 172 173 pagefault_disable(); 174 if (!__get_user_inatomic(*ret, ptr)) { 175 pagefault_enable(); 176 return 0; 177 } 178 pagefault_enable(); 179 180 return read_user_stack_slow(ptr, ret, 4); 181 } 182 183 static inline int valid_user_sp(unsigned long sp, int is_64) 184 { 185 if (!sp || (sp & 7) || sp > (is_64 ? TASK_SIZE : 0x100000000UL) - 32) 186 return 0; 187 return 1; 188 } 189 190 /* 191 * 64-bit user processes use the same stack frame for RT and non-RT signals. 192 */ 193 struct signal_frame_64 { 194 char dummy[__SIGNAL_FRAMESIZE]; 195 struct ucontext uc; 196 unsigned long unused[2]; 197 unsigned int tramp[6]; 198 struct siginfo *pinfo; 199 void *puc; 200 struct siginfo info; 201 char abigap[288]; 202 }; 203 204 static int is_sigreturn_64_address(unsigned long nip, unsigned long fp) 205 { 206 if (nip == fp + offsetof(struct signal_frame_64, tramp)) 207 return 1; 208 if (vdso64_rt_sigtramp && current->mm->context.vdso_base && 209 nip == current->mm->context.vdso_base + vdso64_rt_sigtramp) 210 return 1; 211 return 0; 212 } 213 214 /* 215 * Do some sanity checking on the signal frame pointed to by sp. 216 * We check the pinfo and puc pointers in the frame. 217 */ 218 static int sane_signal_64_frame(unsigned long sp) 219 { 220 struct signal_frame_64 __user *sf; 221 unsigned long pinfo, puc; 222 223 sf = (struct signal_frame_64 __user *) sp; 224 if (read_user_stack_64((unsigned long __user *) &sf->pinfo, &pinfo) || 225 read_user_stack_64((unsigned long __user *) &sf->puc, &puc)) 226 return 0; 227 return pinfo == (unsigned long) &sf->info && 228 puc == (unsigned long) &sf->uc; 229 } 230 231 static void perf_callchain_user_64(struct perf_callchain_entry *entry, 232 struct pt_regs *regs) 233 { 234 unsigned long sp, next_sp; 235 unsigned long next_ip; 236 unsigned long lr; 237 long level = 0; 238 struct signal_frame_64 __user *sigframe; 239 unsigned long __user *fp, *uregs; 240 241 next_ip = perf_instruction_pointer(regs); 242 lr = regs->link; 243 sp = regs->gpr[1]; 244 perf_callchain_store(entry, next_ip); 245 246 for (;;) { 247 fp = (unsigned long __user *) sp; 248 if (!valid_user_sp(sp, 1) || read_user_stack_64(fp, &next_sp)) 249 return; 250 if (level > 0 && read_user_stack_64(&fp[2], &next_ip)) 251 return; 252 253 /* 254 * Note: the next_sp - sp >= signal frame size check 255 * is true when next_sp < sp, which can happen when 256 * transitioning from an alternate signal stack to the 257 * normal stack. 258 */ 259 if (next_sp - sp >= sizeof(struct signal_frame_64) && 260 (is_sigreturn_64_address(next_ip, sp) || 261 (level <= 1 && is_sigreturn_64_address(lr, sp))) && 262 sane_signal_64_frame(sp)) { 263 /* 264 * This looks like an signal frame 265 */ 266 sigframe = (struct signal_frame_64 __user *) sp; 267 uregs = sigframe->uc.uc_mcontext.gp_regs; 268 if (read_user_stack_64(&uregs[PT_NIP], &next_ip) || 269 read_user_stack_64(&uregs[PT_LNK], &lr) || 270 read_user_stack_64(&uregs[PT_R1], &sp)) 271 return; 272 level = 0; 273 perf_callchain_store(entry, PERF_CONTEXT_USER); 274 perf_callchain_store(entry, next_ip); 275 continue; 276 } 277 278 if (level == 0) 279 next_ip = lr; 280 perf_callchain_store(entry, next_ip); 281 ++level; 282 sp = next_sp; 283 } 284 } 285 286 static inline int current_is_64bit(void) 287 { 288 /* 289 * We can't use test_thread_flag() here because we may be on an 290 * interrupt stack, and the thread flags don't get copied over 291 * from the thread_info on the main stack to the interrupt stack. 292 */ 293 return !test_ti_thread_flag(task_thread_info(current), TIF_32BIT); 294 } 295 296 #else /* CONFIG_PPC64 */ 297 /* 298 * On 32-bit we just access the address and let hash_page create a 299 * HPTE if necessary, so there is no need to fall back to reading 300 * the page tables. Since this is called at interrupt level, 301 * do_page_fault() won't treat a DSI as a page fault. 302 */ 303 static int read_user_stack_32(unsigned int __user *ptr, unsigned int *ret) 304 { 305 int rc; 306 307 if ((unsigned long)ptr > TASK_SIZE - sizeof(unsigned int) || 308 ((unsigned long)ptr & 3)) 309 return -EFAULT; 310 311 pagefault_disable(); 312 rc = __get_user_inatomic(*ret, ptr); 313 pagefault_enable(); 314 315 return rc; 316 } 317 318 static inline void perf_callchain_user_64(struct perf_callchain_entry *entry, 319 struct pt_regs *regs) 320 { 321 } 322 323 static inline int current_is_64bit(void) 324 { 325 return 0; 326 } 327 328 static inline int valid_user_sp(unsigned long sp, int is_64) 329 { 330 if (!sp || (sp & 7) || sp > TASK_SIZE - 32) 331 return 0; 332 return 1; 333 } 334 335 #define __SIGNAL_FRAMESIZE32 __SIGNAL_FRAMESIZE 336 #define sigcontext32 sigcontext 337 #define mcontext32 mcontext 338 #define ucontext32 ucontext 339 #define compat_siginfo_t struct siginfo 340 341 #endif /* CONFIG_PPC64 */ 342 343 /* 344 * Layout for non-RT signal frames 345 */ 346 struct signal_frame_32 { 347 char dummy[__SIGNAL_FRAMESIZE32]; 348 struct sigcontext32 sctx; 349 struct mcontext32 mctx; 350 int abigap[56]; 351 }; 352 353 /* 354 * Layout for RT signal frames 355 */ 356 struct rt_signal_frame_32 { 357 char dummy[__SIGNAL_FRAMESIZE32 + 16]; 358 compat_siginfo_t info; 359 struct ucontext32 uc; 360 int abigap[56]; 361 }; 362 363 static int is_sigreturn_32_address(unsigned int nip, unsigned int fp) 364 { 365 if (nip == fp + offsetof(struct signal_frame_32, mctx.mc_pad)) 366 return 1; 367 if (vdso32_sigtramp && current->mm->context.vdso_base && 368 nip == current->mm->context.vdso_base + vdso32_sigtramp) 369 return 1; 370 return 0; 371 } 372 373 static int is_rt_sigreturn_32_address(unsigned int nip, unsigned int fp) 374 { 375 if (nip == fp + offsetof(struct rt_signal_frame_32, 376 uc.uc_mcontext.mc_pad)) 377 return 1; 378 if (vdso32_rt_sigtramp && current->mm->context.vdso_base && 379 nip == current->mm->context.vdso_base + vdso32_rt_sigtramp) 380 return 1; 381 return 0; 382 } 383 384 static int sane_signal_32_frame(unsigned int sp) 385 { 386 struct signal_frame_32 __user *sf; 387 unsigned int regs; 388 389 sf = (struct signal_frame_32 __user *) (unsigned long) sp; 390 if (read_user_stack_32((unsigned int __user *) &sf->sctx.regs, ®s)) 391 return 0; 392 return regs == (unsigned long) &sf->mctx; 393 } 394 395 static int sane_rt_signal_32_frame(unsigned int sp) 396 { 397 struct rt_signal_frame_32 __user *sf; 398 unsigned int regs; 399 400 sf = (struct rt_signal_frame_32 __user *) (unsigned long) sp; 401 if (read_user_stack_32((unsigned int __user *) &sf->uc.uc_regs, ®s)) 402 return 0; 403 return regs == (unsigned long) &sf->uc.uc_mcontext; 404 } 405 406 static unsigned int __user *signal_frame_32_regs(unsigned int sp, 407 unsigned int next_sp, unsigned int next_ip) 408 { 409 struct mcontext32 __user *mctx = NULL; 410 struct signal_frame_32 __user *sf; 411 struct rt_signal_frame_32 __user *rt_sf; 412 413 /* 414 * Note: the next_sp - sp >= signal frame size check 415 * is true when next_sp < sp, for example, when 416 * transitioning from an alternate signal stack to the 417 * normal stack. 418 */ 419 if (next_sp - sp >= sizeof(struct signal_frame_32) && 420 is_sigreturn_32_address(next_ip, sp) && 421 sane_signal_32_frame(sp)) { 422 sf = (struct signal_frame_32 __user *) (unsigned long) sp; 423 mctx = &sf->mctx; 424 } 425 426 if (!mctx && next_sp - sp >= sizeof(struct rt_signal_frame_32) && 427 is_rt_sigreturn_32_address(next_ip, sp) && 428 sane_rt_signal_32_frame(sp)) { 429 rt_sf = (struct rt_signal_frame_32 __user *) (unsigned long) sp; 430 mctx = &rt_sf->uc.uc_mcontext; 431 } 432 433 if (!mctx) 434 return NULL; 435 return mctx->mc_gregs; 436 } 437 438 static void perf_callchain_user_32(struct perf_callchain_entry *entry, 439 struct pt_regs *regs) 440 { 441 unsigned int sp, next_sp; 442 unsigned int next_ip; 443 unsigned int lr; 444 long level = 0; 445 unsigned int __user *fp, *uregs; 446 447 next_ip = perf_instruction_pointer(regs); 448 lr = regs->link; 449 sp = regs->gpr[1]; 450 perf_callchain_store(entry, next_ip); 451 452 while (entry->nr < PERF_MAX_STACK_DEPTH) { 453 fp = (unsigned int __user *) (unsigned long) sp; 454 if (!valid_user_sp(sp, 0) || read_user_stack_32(fp, &next_sp)) 455 return; 456 if (level > 0 && read_user_stack_32(&fp[1], &next_ip)) 457 return; 458 459 uregs = signal_frame_32_regs(sp, next_sp, next_ip); 460 if (!uregs && level <= 1) 461 uregs = signal_frame_32_regs(sp, next_sp, lr); 462 if (uregs) { 463 /* 464 * This looks like an signal frame, so restart 465 * the stack trace with the values in it. 466 */ 467 if (read_user_stack_32(&uregs[PT_NIP], &next_ip) || 468 read_user_stack_32(&uregs[PT_LNK], &lr) || 469 read_user_stack_32(&uregs[PT_R1], &sp)) 470 return; 471 level = 0; 472 perf_callchain_store(entry, PERF_CONTEXT_USER); 473 perf_callchain_store(entry, next_ip); 474 continue; 475 } 476 477 if (level == 0) 478 next_ip = lr; 479 perf_callchain_store(entry, next_ip); 480 ++level; 481 sp = next_sp; 482 } 483 } 484 485 void 486 perf_callchain_user(struct perf_callchain_entry *entry, struct pt_regs *regs) 487 { 488 if (current_is_64bit()) 489 perf_callchain_user_64(entry, regs); 490 else 491 perf_callchain_user_32(entry, regs); 492 } 493