arch/powerpc/perf/callchain.c

   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 #include <asm/pte-walk.h>
  26
  27
  28 /*
  29  * Is sp valid as the address of the next kernel stack frame after prev_sp?
  30  * The next frame may be in a different stack area but should not go
  31  * back down in the same stack area.
  32  */
  33 static int valid_next_sp(unsigned long sp, unsigned long prev_sp)
  34 {
  35         if (sp & 0xf)
  36                 return 0;               /* must be 16-byte aligned */
  37         if (!validate_sp(sp, current, STACK_FRAME_OVERHEAD))
  38                 return 0;
  39         if (sp >= prev_sp + STACK_FRAME_MIN_SIZE)
  40                 return 1;
  41         /*
  42          * sp could decrease when we jump off an interrupt stack
  43          * back to the regular process stack.
  44          */
  45         if ((sp & ~(THREAD_SIZE - 1)) != (prev_sp & ~(THREAD_SIZE - 1)))
  46                 return 1;
  47         return 0;
  48 }
  49
  50 void
  51 perf_callchain_kernel(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs)
  52 {
  53         unsigned long sp, next_sp;
  54         unsigned long next_ip;
  55         unsigned long lr;
  56         long level = 0;
  57         unsigned long *fp;
  58
  59         lr = regs->link;
  60         sp = regs->gpr[1];
  61         perf_callchain_store(entry, perf_instruction_pointer(regs));
  62
  63         if (!validate_sp(sp, current, STACK_FRAME_OVERHEAD))
  64                 return;
  65
  66         for (;;) {
  67                 fp = (unsigned long *) sp;
  68                 next_sp = fp[0];
  69
  70                 if (next_sp == sp + STACK_INT_FRAME_SIZE &&
  71                     validate_sp(sp, current, STACK_INT_FRAME_SIZE) &&
  72                     fp[STACK_FRAME_MARKER] == STACK_FRAME_REGS_MARKER) {
  73                         /*
  74                          * This looks like an interrupt frame for an
  75                          * interrupt that occurred in the kernel
  76                          */
  77                         regs = (struct pt_regs *)(sp + STACK_FRAME_OVERHEAD);
  78                         next_ip = regs->nip;
  79                         lr = regs->link;
  80                         level = 0;
  81                         perf_callchain_store_context(entry, PERF_CONTEXT_KERNEL);
  82
  83                 } else {
  84                         if (level == 0)
  85                                 next_ip = lr;
  86                         else
  87                                 next_ip = fp[STACK_FRAME_LR_SAVE];
  88
  89                         /*
  90                          * We can't tell which of the first two addresses
  91                          * we get are valid, but we can filter out the
  92                          * obviously bogus ones here.  We replace them
  93                          * with 0 rather than removing them entirely so
  94                          * that userspace can tell which is which.
  95                          */
  96                         if ((level == 1 && next_ip == lr) ||
  97                             (level <= 1 && !kernel_text_address(next_ip)))
  98                                 next_ip = 0;
  99
 100                         ++level;
 101                 }
 102
 103                 perf_callchain_store(entry, next_ip);
 104                 if (!valid_next_sp(next_sp, sp))
 105                         return;
 106                 sp = next_sp;
 107         }
 108 }
 109
 110 #ifdef CONFIG_PPC64
 111 /*
 112  * On 64-bit we don't want to invoke hash_page on user addresses from
 113  * interrupt context, so if the access faults, we read the page tables
 114  * to find which page (if any) is mapped and access it directly.
 115  */
 116 static int read_user_stack_slow(void __user *ptr, void *buf, int nb)
 117 {
 118         int ret = -EFAULT;
 119         pgd_t *pgdir;
 120         pte_t *ptep, pte;
 121         unsigned shift;
 122         unsigned long addr = (unsigned long) ptr;
 123         unsigned long offset;
 124         unsigned long pfn, flags;
 125         void *kaddr;
 126
 127         pgdir = current->mm->pgd;
 128         if (!pgdir)
 129                 return -EFAULT;
 130
 131         local_irq_save(flags);
 132         ptep = find_current_mm_pte(pgdir, addr, NULL, &shift);
 133         if (!ptep)
 134                 goto err_out;
 135         if (!shift)
 136                 shift = PAGE_SHIFT;
 137
 138         /* align address to page boundary */
 139         offset = addr & ((1UL << shift) - 1);
 140
 141         pte = READ_ONCE(*ptep);
 142         if (!pte_present(pte) || !pte_user(pte))
 143                 goto err_out;
 144         pfn = pte_pfn(pte);
 145         if (!page_is_ram(pfn))
 146                 goto err_out;
 147
 148         /* no highmem to worry about here */
 149         kaddr = pfn_to_kaddr(pfn);
 150         memcpy(buf, kaddr + offset, nb);
 151         ret = 0;
 152 err_out:
 153         local_irq_restore(flags);
 154         return ret;
 155 }
 156
 157 static int read_user_stack_64(unsigned long __user *ptr, unsigned long *ret)
 158 {
 159         if ((unsigned long)ptr > TASK_SIZE - sizeof(unsigned long) ||
 160             ((unsigned long)ptr & 7))
 161                 return -EFAULT;
 162
 163         pagefault_disable();
 164         if (!__get_user_inatomic(*ret, ptr)) {
 165                 pagefault_enable();
 166                 return 0;
 167         }
 168         pagefault_enable();
 169
 170         return read_user_stack_slow(ptr, ret, 8);
 171 }
 172
 173 static int read_user_stack_32(unsigned int __user *ptr, unsigned int *ret)
 174 {
 175         if ((unsigned long)ptr > TASK_SIZE - sizeof(unsigned int) ||
 176             ((unsigned long)ptr & 3))
 177                 return -EFAULT;
 178
 179         pagefault_disable();
 180         if (!__get_user_inatomic(*ret, ptr)) {
 181                 pagefault_enable();
 182                 return 0;
 183         }
 184         pagefault_enable();
 185
 186         return read_user_stack_slow(ptr, ret, 4);
 187 }
 188
 189 static inline int valid_user_sp(unsigned long sp, int is_64)
 190 {
 191         if (!sp || (sp & 7) || sp > (is_64 ? TASK_SIZE : 0x100000000UL) - 32)
 192                 return 0;
 193         return 1;
 194 }
 195
 196 /*
 197  * 64-bit user processes use the same stack frame for RT and non-RT signals.
 198  */
 199 struct signal_frame_64 {
 200         char            dummy[__SIGNAL_FRAMESIZE];
 201         struct ucontext uc;
 202         unsigned long   unused[2];
 203         unsigned int    tramp[6];
 204         struct siginfo  *pinfo;
 205         void            *puc;
 206         struct siginfo  info;
 207         char            abigap[288];
 208 };
 209
 210 static int is_sigreturn_64_address(unsigned long nip, unsigned long fp)
 211 {
 212         if (nip == fp + offsetof(struct signal_frame_64, tramp))
 213                 return 1;
 214         if (vdso64_rt_sigtramp && current->mm->context.vdso_base &&
 215             nip == current->mm->context.vdso_base + vdso64_rt_sigtramp)
 216                 return 1;
 217         return 0;
 218 }
 219
 220 /*
 221  * Do some sanity checking on the signal frame pointed to by sp.
 222  * We check the pinfo and puc pointers in the frame.
 223  */
 224 static int sane_signal_64_frame(unsigned long sp)
 225 {
 226         struct signal_frame_64 __user *sf;
 227         unsigned long pinfo, puc;
 228
 229         sf = (struct signal_frame_64 __user *) sp;
 230         if (read_user_stack_64((unsigned long __user *) &sf->pinfo, &pinfo) ||
 231             read_user_stack_64((unsigned long __user *) &sf->puc, &puc))
 232                 return 0;
 233         return pinfo == (unsigned long) &sf->info &&
 234                 puc == (unsigned long) &sf->uc;
 235 }
 236
 237 static void perf_callchain_user_64(struct perf_callchain_entry_ctx *entry,
 238                                    struct pt_regs *regs)
 239 {
 240         unsigned long sp, next_sp;
 241         unsigned long next_ip;
 242         unsigned long lr;
 243         long level = 0;
 244         struct signal_frame_64 __user *sigframe;
 245         unsigned long __user *fp, *uregs;
 246
 247         next_ip = perf_instruction_pointer(regs);
 248         lr = regs->link;
 249         sp = regs->gpr[1];
 250         perf_callchain_store(entry, next_ip);
 251
 252         while (entry->nr < entry->max_stack) {
 253                 fp = (unsigned long __user *) sp;
 254                 if (!valid_user_sp(sp, 1) || read_user_stack_64(fp, &next_sp))
 255                         return;
 256                 if (level > 0 && read_user_stack_64(&fp[2], &next_ip))
 257                         return;
 258
 259                 /*
 260                  * Note: the next_sp - sp >= signal frame size check
 261                  * is true when next_sp < sp, which can happen when
 262                  * transitioning from an alternate signal stack to the
 263                  * normal stack.
 264                  */
 265                 if (next_sp - sp >= sizeof(struct signal_frame_64) &&
 266                     (is_sigreturn_64_address(next_ip, sp) ||
 267                      (level <= 1 && is_sigreturn_64_address(lr, sp))) &&
 268                     sane_signal_64_frame(sp)) {
 269                         /*
 270                          * This looks like an signal frame
 271                          */
 272                         sigframe = (struct signal_frame_64 __user *) sp;
 273                         uregs = sigframe->uc.uc_mcontext.gp_regs;
 274                         if (read_user_stack_64(&uregs[PT_NIP], &next_ip) ||
 275                             read_user_stack_64(&uregs[PT_LNK], &lr) ||
 276                             read_user_stack_64(&uregs[PT_R1], &sp))
 277                                 return;
 278                         level = 0;
 279                         perf_callchain_store_context(entry, PERF_CONTEXT_USER);
 280                         perf_callchain_store(entry, next_ip);
 281                         continue;
 282                 }
 283
 284                 if (level == 0)
 285                         next_ip = lr;
 286                 perf_callchain_store(entry, next_ip);
 287                 ++level;
 288                 sp = next_sp;
 289         }
 290 }
 291
 292 static inline int current_is_64bit(void)
 293 {
 294         /*
 295          * We can't use test_thread_flag() here because we may be on an
 296          * interrupt stack, and the thread flags don't get copied over
 297          * from the thread_info on the main stack to the interrupt stack.
 298          */
 299         return !test_ti_thread_flag(task_thread_info(current), TIF_32BIT);
 300 }
 301
 302 #else  /* CONFIG_PPC64 */
 303 /*
 304  * On 32-bit we just access the address and let hash_page create a
 305  * HPTE if necessary, so there is no need to fall back to reading
 306  * the page tables.  Since this is called at interrupt level,
 307  * do_page_fault() won't treat a DSI as a page fault.
 308  */
 309 static int read_user_stack_32(unsigned int __user *ptr, unsigned int *ret)
 310 {
 311         int rc;
 312
 313         if ((unsigned long)ptr > TASK_SIZE - sizeof(unsigned int) ||
 314             ((unsigned long)ptr & 3))
 315                 return -EFAULT;
 316
 317         pagefault_disable();
 318         rc = __get_user_inatomic(*ret, ptr);
 319         pagefault_enable();
 320
 321         return rc;
 322 }
 323
 324 static inline void perf_callchain_user_64(struct perf_callchain_entry_ctx *entry,
 325                                           struct pt_regs *regs)
 326 {
 327 }
 328
 329 static inline int current_is_64bit(void)
 330 {
 331         return 0;
 332 }
 333
 334 static inline int valid_user_sp(unsigned long sp, int is_64)
 335 {
 336         if (!sp || (sp & 7) || sp > TASK_SIZE - 32)
 337                 return 0;
 338         return 1;
 339 }
 340
 341 #define __SIGNAL_FRAMESIZE32    __SIGNAL_FRAMESIZE
 342 #define sigcontext32            sigcontext
 343 #define mcontext32              mcontext
 344 #define ucontext32              ucontext
 345 #define compat_siginfo_t        struct siginfo
 346
 347 #endif /* CONFIG_PPC64 */
 348
 349 /*
 350  * Layout for non-RT signal frames
 351  */
 352 struct signal_frame_32 {
 353         char                    dummy[__SIGNAL_FRAMESIZE32];
 354         struct sigcontext32     sctx;
 355         struct mcontext32       mctx;
 356         int                     abigap[56];
 357 };
 358
 359 /*
 360  * Layout for RT signal frames
 361  */
 362 struct rt_signal_frame_32 {
 363         char                    dummy[__SIGNAL_FRAMESIZE32 + 16];
 364         compat_siginfo_t        info;
 365         struct ucontext32       uc;
 366         int                     abigap[56];
 367 };
 368
 369 static int is_sigreturn_32_address(unsigned int nip, unsigned int fp)
 370 {
 371         if (nip == fp + offsetof(struct signal_frame_32, mctx.mc_pad))
 372                 return 1;
 373         if (vdso32_sigtramp && current->mm->context.vdso_base &&
 374             nip == current->mm->context.vdso_base + vdso32_sigtramp)
 375                 return 1;
 376         return 0;
 377 }
 378
 379 static int is_rt_sigreturn_32_address(unsigned int nip, unsigned int fp)
 380 {
 381         if (nip == fp + offsetof(struct rt_signal_frame_32,
 382                                  uc.uc_mcontext.mc_pad))
 383                 return 1;
 384         if (vdso32_rt_sigtramp && current->mm->context.vdso_base &&
 385             nip == current->mm->context.vdso_base + vdso32_rt_sigtramp)
 386                 return 1;
 387         return 0;
 388 }
 389
 390 static int sane_signal_32_frame(unsigned int sp)
 391 {
 392         struct signal_frame_32 __user *sf;
 393         unsigned int regs;
 394
 395         sf = (struct signal_frame_32 __user *) (unsigned long) sp;
 396         if (read_user_stack_32((unsigned int __user *) &sf->sctx.regs, &regs))
 397                 return 0;
 398         return regs == (unsigned long) &sf->mctx;
 399 }
 400
 401 static int sane_rt_signal_32_frame(unsigned int sp)
 402 {
 403         struct rt_signal_frame_32 __user *sf;
 404         unsigned int regs;
 405
 406         sf = (struct rt_signal_frame_32 __user *) (unsigned long) sp;
 407         if (read_user_stack_32((unsigned int __user *) &sf->uc.uc_regs, &regs))
 408                 return 0;
 409         return regs == (unsigned long) &sf->uc.uc_mcontext;
 410 }
 411
 412 static unsigned int __user *signal_frame_32_regs(unsigned int sp,
 413                                 unsigned int next_sp, unsigned int next_ip)
 414 {
 415         struct mcontext32 __user *mctx = NULL;
 416         struct signal_frame_32 __user *sf;
 417         struct rt_signal_frame_32 __user *rt_sf;
 418
 419         /*
 420          * Note: the next_sp - sp >= signal frame size check
 421          * is true when next_sp < sp, for example, when
 422          * transitioning from an alternate signal stack to the
 423          * normal stack.
 424          */
 425         if (next_sp - sp >= sizeof(struct signal_frame_32) &&
 426             is_sigreturn_32_address(next_ip, sp) &&
 427             sane_signal_32_frame(sp)) {
 428                 sf = (struct signal_frame_32 __user *) (unsigned long) sp;
 429                 mctx = &sf->mctx;
 430         }
 431
 432         if (!mctx && next_sp - sp >= sizeof(struct rt_signal_frame_32) &&
 433             is_rt_sigreturn_32_address(next_ip, sp) &&
 434             sane_rt_signal_32_frame(sp)) {
 435                 rt_sf = (struct rt_signal_frame_32 __user *) (unsigned long) sp;
 436                 mctx = &rt_sf->uc.uc_mcontext;
 437         }
 438
 439         if (!mctx)
 440                 return NULL;
 441         return mctx->mc_gregs;
 442 }
 443
 444 static void perf_callchain_user_32(struct perf_callchain_entry_ctx *entry,
 445                                    struct pt_regs *regs)
 446 {
 447         unsigned int sp, next_sp;
 448         unsigned int next_ip;
 449         unsigned int lr;
 450         long level = 0;
 451         unsigned int __user *fp, *uregs;
 452
 453         next_ip = perf_instruction_pointer(regs);
 454         lr = regs->link;
 455         sp = regs->gpr[1];
 456         perf_callchain_store(entry, next_ip);
 457
 458         while (entry->nr < entry->max_stack) {
 459                 fp = (unsigned int __user *) (unsigned long) sp;
 460                 if (!valid_user_sp(sp, 0) || read_user_stack_32(fp, &next_sp))
 461                         return;
 462                 if (level > 0 && read_user_stack_32(&fp[1], &next_ip))
 463                         return;
 464
 465                 uregs = signal_frame_32_regs(sp, next_sp, next_ip);
 466                 if (!uregs && level <= 1)
 467                         uregs = signal_frame_32_regs(sp, next_sp, lr);
 468                 if (uregs) {
 469                         /*
 470                          * This looks like an signal frame, so restart
 471                          * the stack trace with the values in it.
 472                          */
 473                         if (read_user_stack_32(&uregs[PT_NIP], &next_ip) ||
 474                             read_user_stack_32(&uregs[PT_LNK], &lr) ||
 475                             read_user_stack_32(&uregs[PT_R1], &sp))
 476                                 return;
 477                         level = 0;
 478                         perf_callchain_store_context(entry, PERF_CONTEXT_USER);
 479                         perf_callchain_store(entry, next_ip);
 480                         continue;
 481                 }
 482
 483                 if (level == 0)
 484                         next_ip = lr;
 485                 perf_callchain_store(entry, next_ip);
 486                 ++level;
 487                 sp = next_sp;
 488         }
 489 }
 490
 491 void
 492 perf_callchain_user(struct perf_callchain_entry_ctx *entry, struct pt_regs *regs)
 493 {
 494         if (current_is_64bit())
 495                 perf_callchain_user_64(entry, regs);
 496         else
 497                 perf_callchain_user_32(entry, regs);
 498 }