2 * Performance counter callchain support - powerpc architecture code
4 * Copyright © 2009 Paul Mackerras, IBM Corporation.
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.
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>
17 #include <asm/ptrace.h>
18 #include <asm/pgtable.h>
19 #include <asm/sigcontext.h>
20 #include <asm/ucontext.h>
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.
32 static int valid_next_sp(unsigned long sp
, unsigned long prev_sp
)
35 return 0; /* must be 16-byte aligned */
36 if (!validate_sp(sp
, current
, STACK_FRAME_OVERHEAD
))
38 if (sp
>= prev_sp
+ STACK_FRAME_OVERHEAD
)
41 * sp could decrease when we jump off an interrupt stack
42 * back to the regular process stack.
44 if ((sp
& ~(THREAD_SIZE
- 1)) != (prev_sp
& ~(THREAD_SIZE
- 1)))
50 perf_callchain_kernel(struct perf_callchain_entry
*entry
, struct pt_regs
*regs
)
52 unsigned long sp
, next_sp
;
53 unsigned long next_ip
;
60 perf_callchain_store(entry
, regs
->nip
);
62 if (!validate_sp(sp
, current
, STACK_FRAME_OVERHEAD
))
66 fp
= (unsigned long *) sp
;
69 if (next_sp
== sp
+ STACK_INT_FRAME_SIZE
&&
70 fp
[STACK_FRAME_MARKER
] == STACK_FRAME_REGS_MARKER
) {
72 * This looks like an interrupt frame for an
73 * interrupt that occurred in the kernel
75 regs
= (struct pt_regs
*)(sp
+ STACK_FRAME_OVERHEAD
);
79 perf_callchain_store(entry
, PERF_CONTEXT_KERNEL
);
85 next_ip
= fp
[STACK_FRAME_LR_SAVE
];
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.
94 if ((level
== 1 && next_ip
== lr
) ||
95 (level
<= 1 && !kernel_text_address(next_ip
)))
101 perf_callchain_store(entry
, next_ip
);
102 if (!valid_next_sp(next_sp
, sp
))
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.
114 static int read_user_stack_slow(void __user
*ptr
, void *ret
, int nb
)
119 unsigned long addr
= (unsigned long) ptr
;
120 unsigned long offset
;
124 pgdir
= current
->mm
->pgd
;
128 ptep
= find_linux_pte_or_hugepte(pgdir
, addr
, &shift
);
132 /* align address to page boundary */
133 offset
= addr
& ((1UL << shift
) - 1);
139 if (!pte_present(pte
) || !(pte_val(pte
) & _PAGE_USER
))
142 if (!page_is_ram(pfn
))
145 /* no highmem to worry about here */
146 kaddr
= pfn_to_kaddr(pfn
);
147 memcpy(ret
, kaddr
+ offset
, nb
);
151 static int read_user_stack_64(unsigned long __user
*ptr
, unsigned long *ret
)
153 if ((unsigned long)ptr
> TASK_SIZE
- sizeof(unsigned long) ||
154 ((unsigned long)ptr
& 7))
158 if (!__get_user_inatomic(*ret
, ptr
)) {
164 return read_user_stack_slow(ptr
, ret
, 8);
167 static int read_user_stack_32(unsigned int __user
*ptr
, unsigned int *ret
)
169 if ((unsigned long)ptr
> TASK_SIZE
- sizeof(unsigned int) ||
170 ((unsigned long)ptr
& 3))
174 if (!__get_user_inatomic(*ret
, ptr
)) {
180 return read_user_stack_slow(ptr
, ret
, 4);
183 static inline int valid_user_sp(unsigned long sp
, int is_64
)
185 if (!sp
|| (sp
& 7) || sp
> (is_64
? TASK_SIZE
: 0x100000000UL
) - 32)
191 * 64-bit user processes use the same stack frame for RT and non-RT signals.
193 struct signal_frame_64
{
194 char dummy
[__SIGNAL_FRAMESIZE
];
196 unsigned long unused
[2];
197 unsigned int tramp
[6];
198 struct siginfo
*pinfo
;
204 static int is_sigreturn_64_address(unsigned long nip
, unsigned long fp
)
206 if (nip
== fp
+ offsetof(struct signal_frame_64
, tramp
))
208 if (vdso64_rt_sigtramp
&& current
->mm
->context
.vdso_base
&&
209 nip
== current
->mm
->context
.vdso_base
+ vdso64_rt_sigtramp
)
215 * Do some sanity checking on the signal frame pointed to by sp.
216 * We check the pinfo and puc pointers in the frame.
218 static int sane_signal_64_frame(unsigned long sp
)
220 struct signal_frame_64 __user
*sf
;
221 unsigned long pinfo
, puc
;
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
))
227 return pinfo
== (unsigned long) &sf
->info
&&
228 puc
== (unsigned long) &sf
->uc
;
231 static void perf_callchain_user_64(struct perf_callchain_entry
*entry
,
232 struct pt_regs
*regs
)
234 unsigned long sp
, next_sp
;
235 unsigned long next_ip
;
238 struct signal_frame_64 __user
*sigframe
;
239 unsigned long __user
*fp
, *uregs
;
244 perf_callchain_store(entry
, next_ip
);
247 fp
= (unsigned long __user
*) sp
;
248 if (!valid_user_sp(sp
, 1) || read_user_stack_64(fp
, &next_sp
))
250 if (level
> 0 && read_user_stack_64(&fp
[2], &next_ip
))
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
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
)) {
264 * This looks like an signal frame
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
))
273 perf_callchain_store(entry
, PERF_CONTEXT_USER
);
274 perf_callchain_store(entry
, next_ip
);
280 perf_callchain_store(entry
, next_ip
);
286 static inline int current_is_64bit(void)
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.
293 return !test_ti_thread_flag(task_thread_info(current
), TIF_32BIT
);
296 #else /* CONFIG_PPC64 */
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.
303 static int read_user_stack_32(unsigned int __user
*ptr
, unsigned int *ret
)
307 if ((unsigned long)ptr
> TASK_SIZE
- sizeof(unsigned int) ||
308 ((unsigned long)ptr
& 3))
312 rc
= __get_user_inatomic(*ret
, ptr
);
318 static inline void perf_callchain_user_64(struct perf_callchain_entry
*entry
,
319 struct pt_regs
*regs
)
323 static inline int current_is_64bit(void)
328 static inline int valid_user_sp(unsigned long sp
, int is_64
)
330 if (!sp
|| (sp
& 7) || sp
> TASK_SIZE
- 32)
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
341 #endif /* CONFIG_PPC64 */
344 * Layout for non-RT signal frames
346 struct signal_frame_32
{
347 char dummy
[__SIGNAL_FRAMESIZE32
];
348 struct sigcontext32 sctx
;
349 struct mcontext32 mctx
;
354 * Layout for RT signal frames
356 struct rt_signal_frame_32
{
357 char dummy
[__SIGNAL_FRAMESIZE32
+ 16];
358 compat_siginfo_t info
;
359 struct ucontext32 uc
;
363 static int is_sigreturn_32_address(unsigned int nip
, unsigned int fp
)
365 if (nip
== fp
+ offsetof(struct signal_frame_32
, mctx
.mc_pad
))
367 if (vdso32_sigtramp
&& current
->mm
->context
.vdso_base
&&
368 nip
== current
->mm
->context
.vdso_base
+ vdso32_sigtramp
)
373 static int is_rt_sigreturn_32_address(unsigned int nip
, unsigned int fp
)
375 if (nip
== fp
+ offsetof(struct rt_signal_frame_32
,
376 uc
.uc_mcontext
.mc_pad
))
378 if (vdso32_rt_sigtramp
&& current
->mm
->context
.vdso_base
&&
379 nip
== current
->mm
->context
.vdso_base
+ vdso32_rt_sigtramp
)
384 static int sane_signal_32_frame(unsigned int sp
)
386 struct signal_frame_32 __user
*sf
;
389 sf
= (struct signal_frame_32 __user
*) (unsigned long) sp
;
390 if (read_user_stack_32((unsigned int __user
*) &sf
->sctx
.regs
, ®s
))
392 return regs
== (unsigned long) &sf
->mctx
;
395 static int sane_rt_signal_32_frame(unsigned int sp
)
397 struct rt_signal_frame_32 __user
*sf
;
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
))
403 return regs
== (unsigned long) &sf
->uc
.uc_mcontext
;
406 static unsigned int __user
*signal_frame_32_regs(unsigned int sp
,
407 unsigned int next_sp
, unsigned int next_ip
)
409 struct mcontext32 __user
*mctx
= NULL
;
410 struct signal_frame_32 __user
*sf
;
411 struct rt_signal_frame_32 __user
*rt_sf
;
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
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
;
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
;
435 return mctx
->mc_gregs
;
438 static void perf_callchain_user_32(struct perf_callchain_entry
*entry
,
439 struct pt_regs
*regs
)
441 unsigned int sp
, next_sp
;
442 unsigned int next_ip
;
445 unsigned int __user
*fp
, *uregs
;
450 perf_callchain_store(entry
, next_ip
);
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
))
456 if (level
> 0 && read_user_stack_32(&fp
[1], &next_ip
))
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
);
464 * This looks like an signal frame, so restart
465 * the stack trace with the values in it.
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
))
472 perf_callchain_store(entry
, PERF_CONTEXT_USER
);
473 perf_callchain_store(entry
, next_ip
);
479 perf_callchain_store(entry
, next_ip
);
486 perf_callchain_user(struct perf_callchain_entry
*entry
, struct pt_regs
*regs
)
488 if (current_is_64bit())
489 perf_callchain_user_64(entry
, regs
);
491 perf_callchain_user_32(entry
, regs
);