arch/powerpc/kernel/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 "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_OVERHEAD)
  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, regs->nip);
  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         if (!__get_user_inatomic(*ret, ptr))
 158                 return 0;
 159
 160         return read_user_stack_slow(ptr, ret, 8);
 161 }
 162
 163 static int read_user_stack_32(unsigned int __user *ptr, unsigned int *ret)
 164 {
 165         if ((unsigned long)ptr > TASK_SIZE - sizeof(unsigned int) ||
 166             ((unsigned long)ptr & 3))
 167                 return -EFAULT;
 168
 169         if (!__get_user_inatomic(*ret, ptr))
 170                 return 0;
 171
 172         return read_user_stack_slow(ptr, ret, 4);
 173 }
 174
 175 static inline int valid_user_sp(unsigned long sp, int is_64)
 176 {
 177         if (!sp || (sp & 7) || sp > (is_64 ? TASK_SIZE : 0x100000000UL) - 32)
 178                 return 0;
 179         return 1;
 180 }
 181
 182 /*
 183  * 64-bit user processes use the same stack frame for RT and non-RT signals.
 184  */
 185 struct signal_frame_64 {
 186         char            dummy[__SIGNAL_FRAMESIZE];
 187         struct ucontext uc;
 188         unsigned long   unused[2];
 189         unsigned int    tramp[6];
 190         struct siginfo  *pinfo;
 191         void            *puc;
 192         struct siginfo  info;
 193         char            abigap[288];
 194 };
 195
 196 static int is_sigreturn_64_address(unsigned long nip, unsigned long fp)
 197 {
 198         if (nip == fp + offsetof(struct signal_frame_64, tramp))
 199                 return 1;
 200         if (vdso64_rt_sigtramp && current->mm->context.vdso_base &&
 201             nip == current->mm->context.vdso_base + vdso64_rt_sigtramp)
 202                 return 1;
 203         return 0;
 204 }
 205
 206 /*
 207  * Do some sanity checking on the signal frame pointed to by sp.
 208  * We check the pinfo and puc pointers in the frame.
 209  */
 210 static int sane_signal_64_frame(unsigned long sp)
 211 {
 212         struct signal_frame_64 __user *sf;
 213         unsigned long pinfo, puc;
 214
 215         sf = (struct signal_frame_64 __user *) sp;
 216         if (read_user_stack_64((unsigned long __user *) &sf->pinfo, &pinfo) ||
 217             read_user_stack_64((unsigned long __user *) &sf->puc, &puc))
 218                 return 0;
 219         return pinfo == (unsigned long) &sf->info &&
 220                 puc == (unsigned long) &sf->uc;
 221 }
 222
 223 static void perf_callchain_user_64(struct perf_callchain_entry *entry,
 224                                    struct pt_regs *regs)
 225 {
 226         unsigned long sp, next_sp;
 227         unsigned long next_ip;
 228         unsigned long lr;
 229         long level = 0;
 230         struct signal_frame_64 __user *sigframe;
 231         unsigned long __user *fp, *uregs;
 232
 233         next_ip = regs->nip;
 234         lr = regs->link;
 235         sp = regs->gpr[1];
 236         perf_callchain_store(entry, next_ip);
 237
 238         for (;;) {
 239                 fp = (unsigned long __user *) sp;
 240                 if (!valid_user_sp(sp, 1) || read_user_stack_64(fp, &next_sp))
 241                         return;
 242                 if (level > 0 && read_user_stack_64(&fp[2], &next_ip))
 243                         return;
 244
 245                 /*
 246                  * Note: the next_sp - sp >= signal frame size check
 247                  * is true when next_sp < sp, which can happen when
 248                  * transitioning from an alternate signal stack to the
 249                  * normal stack.
 250                  */
 251                 if (next_sp - sp >= sizeof(struct signal_frame_64) &&
 252                     (is_sigreturn_64_address(next_ip, sp) ||
 253                      (level <= 1 && is_sigreturn_64_address(lr, sp))) &&
 254                     sane_signal_64_frame(sp)) {
 255                         /*
 256                          * This looks like an signal frame
 257                          */
 258                         sigframe = (struct signal_frame_64 __user *) sp;
 259                         uregs = sigframe->uc.uc_mcontext.gp_regs;
 260                         if (read_user_stack_64(&uregs[PT_NIP], &next_ip) ||
 261                             read_user_stack_64(&uregs[PT_LNK], &lr) ||
 262                             read_user_stack_64(&uregs[PT_R1], &sp))
 263                                 return;
 264                         level = 0;
 265                         perf_callchain_store(entry, PERF_CONTEXT_USER);
 266                         perf_callchain_store(entry, next_ip);
 267                         continue;
 268                 }
 269
 270                 if (level == 0)
 271                         next_ip = lr;
 272                 perf_callchain_store(entry, next_ip);
 273                 ++level;
 274                 sp = next_sp;
 275         }
 276 }
 277
 278 static inline int current_is_64bit(void)
 279 {
 280         /*
 281          * We can't use test_thread_flag() here because we may be on an
 282          * interrupt stack, and the thread flags don't get copied over
 283          * from the thread_info on the main stack to the interrupt stack.
 284          */
 285         return !test_ti_thread_flag(task_thread_info(current), TIF_32BIT);
 286 }
 287
 288 #else  /* CONFIG_PPC64 */
 289 /*
 290  * On 32-bit we just access the address and let hash_page create a
 291  * HPTE if necessary, so there is no need to fall back to reading
 292  * the page tables.  Since this is called at interrupt level,
 293  * do_page_fault() won't treat a DSI as a page fault.
 294  */
 295 static int read_user_stack_32(unsigned int __user *ptr, unsigned int *ret)
 296 {
 297         if ((unsigned long)ptr > TASK_SIZE - sizeof(unsigned int) ||
 298             ((unsigned long)ptr & 3))
 299                 return -EFAULT;
 300
 301         return __get_user_inatomic(*ret, ptr);
 302 }
 303
 304 static inline void perf_callchain_user_64(struct perf_callchain_entry *entry,
 305                                           struct pt_regs *regs)
 306 {
 307 }
 308
 309 static inline int current_is_64bit(void)
 310 {
 311         return 0;
 312 }
 313
 314 static inline int valid_user_sp(unsigned long sp, int is_64)
 315 {
 316         if (!sp || (sp & 7) || sp > TASK_SIZE - 32)
 317                 return 0;
 318         return 1;
 319 }
 320
 321 #define __SIGNAL_FRAMESIZE32    __SIGNAL_FRAMESIZE
 322 #define sigcontext32            sigcontext
 323 #define mcontext32              mcontext
 324 #define ucontext32              ucontext
 325 #define compat_siginfo_t        struct siginfo
 326
 327 #endif /* CONFIG_PPC64 */
 328
 329 /*
 330  * Layout for non-RT signal frames
 331  */
 332 struct signal_frame_32 {
 333         char                    dummy[__SIGNAL_FRAMESIZE32];
 334         struct sigcontext32     sctx;
 335         struct mcontext32       mctx;
 336         int                     abigap[56];
 337 };
 338
 339 /*
 340  * Layout for RT signal frames
 341  */
 342 struct rt_signal_frame_32 {
 343         char                    dummy[__SIGNAL_FRAMESIZE32 + 16];
 344         compat_siginfo_t        info;
 345         struct ucontext32       uc;
 346         int                     abigap[56];
 347 };
 348
 349 static int is_sigreturn_32_address(unsigned int nip, unsigned int fp)
 350 {
 351         if (nip == fp + offsetof(struct signal_frame_32, mctx.mc_pad))
 352                 return 1;
 353         if (vdso32_sigtramp && current->mm->context.vdso_base &&
 354             nip == current->mm->context.vdso_base + vdso32_sigtramp)
 355                 return 1;
 356         return 0;
 357 }
 358
 359 static int is_rt_sigreturn_32_address(unsigned int nip, unsigned int fp)
 360 {
 361         if (nip == fp + offsetof(struct rt_signal_frame_32,
 362                                  uc.uc_mcontext.mc_pad))
 363                 return 1;
 364         if (vdso32_rt_sigtramp && current->mm->context.vdso_base &&
 365             nip == current->mm->context.vdso_base + vdso32_rt_sigtramp)
 366                 return 1;
 367         return 0;
 368 }
 369
 370 static int sane_signal_32_frame(unsigned int sp)
 371 {
 372         struct signal_frame_32 __user *sf;
 373         unsigned int regs;
 374
 375         sf = (struct signal_frame_32 __user *) (unsigned long) sp;
 376         if (read_user_stack_32((unsigned int __user *) &sf->sctx.regs, &regs))
 377                 return 0;
 378         return regs == (unsigned long) &sf->mctx;
 379 }
 380
 381 static int sane_rt_signal_32_frame(unsigned int sp)
 382 {
 383         struct rt_signal_frame_32 __user *sf;
 384         unsigned int regs;
 385
 386         sf = (struct rt_signal_frame_32 __user *) (unsigned long) sp;
 387         if (read_user_stack_32((unsigned int __user *) &sf->uc.uc_regs, &regs))
 388                 return 0;
 389         return regs == (unsigned long) &sf->uc.uc_mcontext;
 390 }
 391
 392 static unsigned int __user *signal_frame_32_regs(unsigned int sp,
 393                                 unsigned int next_sp, unsigned int next_ip)
 394 {
 395         struct mcontext32 __user *mctx = NULL;
 396         struct signal_frame_32 __user *sf;
 397         struct rt_signal_frame_32 __user *rt_sf;
 398
 399         /*
 400          * Note: the next_sp - sp >= signal frame size check
 401          * is true when next_sp < sp, for example, when
 402          * transitioning from an alternate signal stack to the
 403          * normal stack.
 404          */
 405         if (next_sp - sp >= sizeof(struct signal_frame_32) &&
 406             is_sigreturn_32_address(next_ip, sp) &&
 407             sane_signal_32_frame(sp)) {
 408                 sf = (struct signal_frame_32 __user *) (unsigned long) sp;
 409                 mctx = &sf->mctx;
 410         }
 411
 412         if (!mctx && next_sp - sp >= sizeof(struct rt_signal_frame_32) &&
 413             is_rt_sigreturn_32_address(next_ip, sp) &&
 414             sane_rt_signal_32_frame(sp)) {
 415                 rt_sf = (struct rt_signal_frame_32 __user *) (unsigned long) sp;
 416                 mctx = &rt_sf->uc.uc_mcontext;
 417         }
 418
 419         if (!mctx)
 420                 return NULL;
 421         return mctx->mc_gregs;
 422 }
 423
 424 static void perf_callchain_user_32(struct perf_callchain_entry *entry,
 425                                    struct pt_regs *regs)
 426 {
 427         unsigned int sp, next_sp;
 428         unsigned int next_ip;
 429         unsigned int lr;
 430         long level = 0;
 431         unsigned int __user *fp, *uregs;
 432
 433         next_ip = regs->nip;
 434         lr = regs->link;
 435         sp = regs->gpr[1];
 436         perf_callchain_store(entry, next_ip);
 437
 438         while (entry->nr < PERF_MAX_STACK_DEPTH) {
 439                 fp = (unsigned int __user *) (unsigned long) sp;
 440                 if (!valid_user_sp(sp, 0) || read_user_stack_32(fp, &next_sp))
 441                         return;
 442                 if (level > 0 && read_user_stack_32(&fp[1], &next_ip))
 443                         return;
 444
 445                 uregs = signal_frame_32_regs(sp, next_sp, next_ip);
 446                 if (!uregs && level <= 1)
 447                         uregs = signal_frame_32_regs(sp, next_sp, lr);
 448                 if (uregs) {
 449                         /*
 450                          * This looks like an signal frame, so restart
 451                          * the stack trace with the values in it.
 452                          */
 453                         if (read_user_stack_32(&uregs[PT_NIP], &next_ip) ||
 454                             read_user_stack_32(&uregs[PT_LNK], &lr) ||
 455                             read_user_stack_32(&uregs[PT_R1], &sp))
 456                                 return;
 457                         level = 0;
 458                         perf_callchain_store(entry, PERF_CONTEXT_USER);
 459                         perf_callchain_store(entry, next_ip);
 460                         continue;
 461                 }
 462
 463                 if (level == 0)
 464                         next_ip = lr;
 465                 perf_callchain_store(entry, next_ip);
 466                 ++level;
 467                 sp = next_sp;
 468         }
 469 }
 470
 471 void
 472 perf_callchain_user(struct perf_callchain_entry *entry, struct pt_regs *regs)
 473 {
 474         if (current_is_64bit())
 475                 perf_callchain_user_64(entry, regs);
 476         else
 477                 perf_callchain_user_32(entry, regs);
 478 }