xref: /linux-6.15/kernel/trace/fgraph.c (revision 474ec3e8)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Infrastructure to took into function calls and returns.
4  * Copyright (c) 2008-2009 Frederic Weisbecker <[email protected]>
5  * Mostly borrowed from function tracer which
6  * is Copyright (c) Steven Rostedt <[email protected]>
7  *
8  * Highly modified by Steven Rostedt (VMware).
9  */
10 #include <linux/bits.h>
11 #include <linux/jump_label.h>
12 #include <linux/suspend.h>
13 #include <linux/ftrace.h>
14 #include <linux/static_call.h>
15 #include <linux/slab.h>
16 
17 #include <trace/events/sched.h>
18 
19 #include "ftrace_internal.h"
20 #include "trace.h"
21 
22 /*
23  * FGRAPH_FRAME_SIZE:	Size in bytes of the meta data on the shadow stack
24  * FGRAPH_FRAME_OFFSET:	Size in long words of the meta data frame
25  */
26 #define FGRAPH_FRAME_SIZE	sizeof(struct ftrace_ret_stack)
27 #define FGRAPH_FRAME_OFFSET	DIV_ROUND_UP(FGRAPH_FRAME_SIZE, sizeof(long))
28 
29 /*
30  * On entry to a function (via function_graph_enter()), a new fgraph frame
31  * (ftrace_ret_stack) is pushed onto the stack as well as a word that
32  * holds a bitmask and a type (called "bitmap"). The bitmap is defined as:
33  *
34  * bits:  0 -  9	offset in words from the previous ftrace_ret_stack
35  *
36  * bits: 10 - 11	Type of storage
37  *			  0 - reserved
38  *			  1 - bitmap of fgraph_array index
39  *			  2 - reserved data
40  *
41  * For type with "bitmap of fgraph_array index" (FGRAPH_TYPE_BITMAP):
42  *  bits: 12 - 27	The bitmap of fgraph_ops fgraph_array index
43  *			That is, it's a bitmask of 0-15 (16 bits)
44  *			where if a corresponding ops in the fgraph_array[]
45  *			expects a callback from the return of the function
46  *			it's corresponding bit will be set.
47  *
48  *
49  * The top of the ret_stack (when not empty) will always have a reference
50  * word that points to the last fgraph frame that was saved.
51  *
52  * For reserved data:
53  *  bits: 12 - 17	The size in words that is stored
54  *  bits: 18 - 23	The index of fgraph_array, which shows who is stored
55  *
56  * That is, at the end of function_graph_enter, if the first and forth
57  * fgraph_ops on the fgraph_array[] (index 0 and 3) needs their retfunc called
58  * on the return of the function being traced, and the forth fgraph_ops
59  * stored two words of data, this is what will be on the task's shadow
60  * ret_stack: (the stack grows upward)
61  *
62  *  ret_stack[SHADOW_STACK_OFFSET]
63  * | SHADOW_STACK_TASK_VARS(ret_stack)[15]      |
64  * ...
65  * | SHADOW_STACK_TASK_VARS(ret_stack)[0]       |
66  *  ret_stack[SHADOW_STACK_MAX_OFFSET]
67  * ...
68  * |                                            | <- task->curr_ret_stack
69  * +--------------------------------------------+
70  * | (3 << 12) | (3 << 10) | FGRAPH_FRAME_OFFSET|
71  * |         *or put another way*               |
72  * | (3 << FGRAPH_DATA_INDEX_SHIFT)| \          | This is for fgraph_ops[3].
73  * | ((2 - 1) << FGRAPH_DATA_SHIFT)| \          | The data size is 2 words.
74  * | (FGRAPH_TYPE_DATA << FGRAPH_TYPE_SHIFT)| \ |
75  * | (offset2:FGRAPH_FRAME_OFFSET+3)            | <- the offset2 is from here
76  * +--------------------------------------------+ ( It is 4 words from the ret_stack)
77  * |            STORED DATA WORD 2              |
78  * |            STORED DATA WORD 1              |
79  * +--------------------------------------------+
80  * | (9 << 12) | (1 << 10) | FGRAPH_FRAME_OFFSET|
81  * |         *or put another way*               |
82  * | (BIT(3)|BIT(0)) << FGRAPH_INDEX_SHIFT | \  |
83  * | FGRAPH_TYPE_BITMAP << FGRAPH_TYPE_SHIFT| \ |
84  * | (offset1:FGRAPH_FRAME_OFFSET)              | <- the offset1 is from here
85  * +--------------------------------------------+
86  * | struct ftrace_ret_stack                    |
87  * |   (stores the saved ret pointer)           | <- the offset points here
88  * +--------------------------------------------+
89  * |                 (X) | (N)                  | ( N words away from
90  * |                                            |   previous ret_stack)
91  * ...
92  * ret_stack[0]
93  *
94  * If a backtrace is required, and the real return pointer needs to be
95  * fetched, then it looks at the task's curr_ret_stack offset, if it
96  * is greater than zero (reserved, or right before popped), it would mask
97  * the value by FGRAPH_FRAME_OFFSET_MASK to get the offset of the
98  * ftrace_ret_stack structure stored on the shadow stack.
99  */
100 
101 /*
102  * The following is for the top word on the stack:
103  *
104  *   FGRAPH_FRAME_OFFSET (0-9) holds the offset delta to the fgraph frame
105  *   FGRAPH_TYPE (10-11) holds the type of word this is.
106  *     (RESERVED or BITMAP)
107  */
108 #define FGRAPH_FRAME_OFFSET_BITS	10
109 #define FGRAPH_FRAME_OFFSET_MASK	GENMASK(FGRAPH_FRAME_OFFSET_BITS - 1, 0)
110 
111 #define FGRAPH_TYPE_BITS	2
112 #define FGRAPH_TYPE_MASK	GENMASK(FGRAPH_TYPE_BITS - 1, 0)
113 #define FGRAPH_TYPE_SHIFT	FGRAPH_FRAME_OFFSET_BITS
114 
115 enum {
116 	FGRAPH_TYPE_RESERVED	= 0,
117 	FGRAPH_TYPE_BITMAP	= 1,
118 	FGRAPH_TYPE_DATA	= 2,
119 };
120 
121 /*
122  * For BITMAP type:
123  *   FGRAPH_INDEX (12-27) bits holding the gops index wanting return callback called
124  */
125 #define FGRAPH_INDEX_BITS	16
126 #define FGRAPH_INDEX_MASK	GENMASK(FGRAPH_INDEX_BITS - 1, 0)
127 #define FGRAPH_INDEX_SHIFT	(FGRAPH_TYPE_SHIFT + FGRAPH_TYPE_BITS)
128 
129 /*
130  * For DATA type:
131  *  FGRAPH_DATA (12-17) bits hold the size of data (in words)
132  *  FGRAPH_INDEX (18-23) bits hold the index for which gops->idx the data is for
133  *
134  * Note:
135  *  data_size == 0 means 1 word, and 31 (=2^5 - 1) means 32 words.
136  */
137 #define FGRAPH_DATA_BITS	5
138 #define FGRAPH_DATA_MASK	GENMASK(FGRAPH_DATA_BITS - 1, 0)
139 #define FGRAPH_DATA_SHIFT	(FGRAPH_TYPE_SHIFT + FGRAPH_TYPE_BITS)
140 #define FGRAPH_MAX_DATA_SIZE (sizeof(long) * (1 << FGRAPH_DATA_BITS))
141 
142 #define FGRAPH_DATA_INDEX_BITS	4
143 #define FGRAPH_DATA_INDEX_MASK	GENMASK(FGRAPH_DATA_INDEX_BITS - 1, 0)
144 #define FGRAPH_DATA_INDEX_SHIFT	(FGRAPH_DATA_SHIFT + FGRAPH_DATA_BITS)
145 
146 #define FGRAPH_MAX_INDEX	\
147 	((FGRAPH_INDEX_SIZE << FGRAPH_DATA_BITS) + FGRAPH_RET_INDEX)
148 
149 #define FGRAPH_ARRAY_SIZE	FGRAPH_INDEX_BITS
150 
151 /*
152  * SHADOW_STACK_SIZE:	The size in bytes of the entire shadow stack
153  * SHADOW_STACK_OFFSET:	The size in long words of the shadow stack
154  * SHADOW_STACK_MAX_OFFSET: The max offset of the stack for a new frame to be added
155  */
156 #define SHADOW_STACK_SIZE	(PAGE_SIZE)
157 #define SHADOW_STACK_OFFSET	(SHADOW_STACK_SIZE / sizeof(long))
158 /* Leave on a buffer at the end */
159 #define SHADOW_STACK_MAX_OFFSET				\
160 	(SHADOW_STACK_OFFSET - (FGRAPH_FRAME_OFFSET + 1 + FGRAPH_ARRAY_SIZE))
161 
162 /* RET_STACK():		Return the frame from a given @offset from task @t */
163 #define RET_STACK(t, offset) ((struct ftrace_ret_stack *)(&(t)->ret_stack[offset]))
164 
165 /*
166  * Each fgraph_ops has a reservered unsigned long at the end (top) of the
167  * ret_stack to store task specific state.
168  */
169 #define SHADOW_STACK_TASK_VARS(ret_stack) \
170 	((unsigned long *)(&(ret_stack)[SHADOW_STACK_OFFSET - FGRAPH_ARRAY_SIZE]))
171 
172 DEFINE_STATIC_KEY_FALSE(kill_ftrace_graph);
173 int ftrace_graph_active;
174 
175 static struct fgraph_ops *fgraph_array[FGRAPH_ARRAY_SIZE];
176 static unsigned long fgraph_array_bitmask;
177 
178 /* LRU index table for fgraph_array */
179 static int fgraph_lru_table[FGRAPH_ARRAY_SIZE];
180 static int fgraph_lru_next;
181 static int fgraph_lru_last;
182 
183 /* Initialize fgraph_lru_table with unused index */
184 static void fgraph_lru_init(void)
185 {
186 	int i;
187 
188 	for (i = 0; i < FGRAPH_ARRAY_SIZE; i++)
189 		fgraph_lru_table[i] = i;
190 }
191 
192 /* Release the used index to the LRU table */
193 static int fgraph_lru_release_index(int idx)
194 {
195 	if (idx < 0 || idx >= FGRAPH_ARRAY_SIZE ||
196 	    WARN_ON_ONCE(fgraph_lru_table[fgraph_lru_last] != -1))
197 		return -1;
198 
199 	fgraph_lru_table[fgraph_lru_last] = idx;
200 	fgraph_lru_last = (fgraph_lru_last + 1) % FGRAPH_ARRAY_SIZE;
201 
202 	clear_bit(idx, &fgraph_array_bitmask);
203 	return 0;
204 }
205 
206 /* Allocate a new index from LRU table */
207 static int fgraph_lru_alloc_index(void)
208 {
209 	int idx = fgraph_lru_table[fgraph_lru_next];
210 
211 	/* No id is available */
212 	if (idx == -1)
213 		return -1;
214 
215 	fgraph_lru_table[fgraph_lru_next] = -1;
216 	fgraph_lru_next = (fgraph_lru_next + 1) % FGRAPH_ARRAY_SIZE;
217 
218 	set_bit(idx, &fgraph_array_bitmask);
219 	return idx;
220 }
221 
222 /* Get the offset to the fgraph frame from a ret_stack value */
223 static inline int __get_offset(unsigned long val)
224 {
225 	return val & FGRAPH_FRAME_OFFSET_MASK;
226 }
227 
228 /* Get the type of word from a ret_stack value */
229 static inline int __get_type(unsigned long val)
230 {
231 	return (val >> FGRAPH_TYPE_SHIFT) & FGRAPH_TYPE_MASK;
232 }
233 
234 /* Get the data_index for a DATA type ret_stack word */
235 static inline int __get_data_index(unsigned long val)
236 {
237 	return (val >> FGRAPH_DATA_INDEX_SHIFT) & FGRAPH_DATA_INDEX_MASK;
238 }
239 
240 /* Get the data_size for a DATA type ret_stack word */
241 static inline int __get_data_size(unsigned long val)
242 {
243 	return ((val >> FGRAPH_DATA_SHIFT) & FGRAPH_DATA_MASK) + 1;
244 }
245 
246 /* Get the word from the ret_stack at @offset */
247 static inline unsigned long get_fgraph_entry(struct task_struct *t, int offset)
248 {
249 	return t->ret_stack[offset];
250 }
251 
252 /* Get the FRAME_OFFSET from the word from the @offset on ret_stack */
253 static inline int get_frame_offset(struct task_struct *t, int offset)
254 {
255 	return __get_offset(t->ret_stack[offset]);
256 }
257 
258 /* For BITMAP type: get the bitmask from the @offset at ret_stack */
259 static inline unsigned long
260 get_bitmap_bits(struct task_struct *t, int offset)
261 {
262 	return (t->ret_stack[offset] >> FGRAPH_INDEX_SHIFT) & FGRAPH_INDEX_MASK;
263 }
264 
265 /* Write the bitmap to the ret_stack at @offset (does index, offset and bitmask) */
266 static inline void
267 set_bitmap(struct task_struct *t, int offset, unsigned long bitmap)
268 {
269 	t->ret_stack[offset] = (bitmap << FGRAPH_INDEX_SHIFT) |
270 		(FGRAPH_TYPE_BITMAP << FGRAPH_TYPE_SHIFT) | FGRAPH_FRAME_OFFSET;
271 }
272 
273 /* For DATA type: get the data saved under the ret_stack word at @offset */
274 static inline void *get_data_type_data(struct task_struct *t, int offset)
275 {
276 	unsigned long val = t->ret_stack[offset];
277 
278 	if (__get_type(val) != FGRAPH_TYPE_DATA)
279 		return NULL;
280 	offset -= __get_data_size(val);
281 	return (void *)&t->ret_stack[offset];
282 }
283 
284 /* Create the ret_stack word for a DATA type */
285 static inline unsigned long make_data_type_val(int idx, int size, int offset)
286 {
287 	return (idx << FGRAPH_DATA_INDEX_SHIFT) |
288 		((size - 1) << FGRAPH_DATA_SHIFT) |
289 		(FGRAPH_TYPE_DATA << FGRAPH_TYPE_SHIFT) | offset;
290 }
291 
292 /* ftrace_graph_entry set to this to tell some archs to run function graph */
293 static int entry_run(struct ftrace_graph_ent *trace, struct fgraph_ops *ops,
294 		     struct fgraph_extras *extras)
295 {
296 	return 0;
297 }
298 
299 /* ftrace_graph_return set to this to tell some archs to run function graph */
300 static void return_run(struct ftrace_graph_ret *trace, struct fgraph_ops *ops)
301 {
302 }
303 
304 static void ret_stack_set_task_var(struct task_struct *t, int idx, long val)
305 {
306 	unsigned long *gvals = SHADOW_STACK_TASK_VARS(t->ret_stack);
307 
308 	gvals[idx] = val;
309 }
310 
311 static unsigned long *
312 ret_stack_get_task_var(struct task_struct *t, int idx)
313 {
314 	unsigned long *gvals = SHADOW_STACK_TASK_VARS(t->ret_stack);
315 
316 	return &gvals[idx];
317 }
318 
319 static void ret_stack_init_task_vars(unsigned long *ret_stack)
320 {
321 	unsigned long *gvals = SHADOW_STACK_TASK_VARS(ret_stack);
322 
323 	memset(gvals, 0, sizeof(*gvals) * FGRAPH_ARRAY_SIZE);
324 }
325 
326 /**
327  * fgraph_reserve_data - Reserve storage on the task's ret_stack
328  * @idx:	The index of fgraph_array
329  * @size_bytes: The size in bytes to reserve
330  *
331  * Reserves space of up to FGRAPH_MAX_DATA_SIZE bytes on the
332  * task's ret_stack shadow stack, for a given fgraph_ops during
333  * the entryfunc() call. If entryfunc() returns zero, the storage
334  * is discarded. An entryfunc() can only call this once per iteration.
335  * The fgraph_ops retfunc() can retrieve this stored data with
336  * fgraph_retrieve_data().
337  *
338  * Returns: On success, a pointer to the data on the stack.
339  *   Otherwise, NULL if there's not enough space left on the
340  *   ret_stack for the data, or if fgraph_reserve_data() was called
341  *   more than once for a single entryfunc() call.
342  */
343 void *fgraph_reserve_data(int idx, int size_bytes)
344 {
345 	unsigned long val;
346 	void *data;
347 	int curr_ret_stack = current->curr_ret_stack;
348 	int data_size;
349 
350 	if (size_bytes > FGRAPH_MAX_DATA_SIZE)
351 		return NULL;
352 
353 	/* Convert the data size to number of longs. */
354 	data_size = (size_bytes + sizeof(long) - 1) >> (sizeof(long) == 4 ? 2 : 3);
355 
356 	val = get_fgraph_entry(current, curr_ret_stack - 1);
357 	data = &current->ret_stack[curr_ret_stack];
358 
359 	curr_ret_stack += data_size + 1;
360 	if (unlikely(curr_ret_stack >= SHADOW_STACK_MAX_OFFSET))
361 		return NULL;
362 
363 	val = make_data_type_val(idx, data_size, __get_offset(val) + data_size + 1);
364 
365 	/* Set the last word to be reserved */
366 	current->ret_stack[curr_ret_stack - 1] = val;
367 
368 	/* Make sure interrupts see this */
369 	barrier();
370 	current->curr_ret_stack = curr_ret_stack;
371 	/* Again sync with interrupts, and reset reserve */
372 	current->ret_stack[curr_ret_stack - 1] = val;
373 
374 	return data;
375 }
376 
377 /**
378  * fgraph_retrieve_data - Retrieve stored data from fgraph_reserve_data()
379  * @idx:	the index of fgraph_array (fgraph_ops::idx)
380  * @size_bytes: pointer to retrieved data size.
381  *
382  * This is to be called by a fgraph_ops retfunc(), to retrieve data that
383  * was stored by the fgraph_ops entryfunc() on the function entry.
384  * That is, this will retrieve the data that was reserved on the
385  * entry of the function that corresponds to the exit of the function
386  * that the fgraph_ops retfunc() is called on.
387  *
388  * Returns: The stored data from fgraph_reserve_data() called by the
389  *    matching entryfunc() for the retfunc() this is called from.
390  *   Or NULL if there was nothing stored.
391  */
392 void *fgraph_retrieve_data(int idx, int *size_bytes)
393 {
394 	return fgraph_retrieve_parent_data(idx, size_bytes, 0);
395 }
396 
397 /**
398  * fgraph_get_task_var - retrieve a task specific state variable
399  * @gops: The ftrace_ops that owns the task specific variable
400  *
401  * Every registered fgraph_ops has a task state variable
402  * reserved on the task's ret_stack. This function returns the
403  * address to that variable.
404  *
405  * Returns the address to the fgraph_ops @gops tasks specific
406  * unsigned long variable.
407  */
408 unsigned long *fgraph_get_task_var(struct fgraph_ops *gops)
409 {
410 	return ret_stack_get_task_var(current, gops->idx);
411 }
412 
413 /*
414  * @offset: The offset into @t->ret_stack to find the ret_stack entry
415  * @frame_offset: Where to place the offset into @t->ret_stack of that entry
416  *
417  * Returns a pointer to the previous ret_stack below @offset or NULL
418  *   when it reaches the bottom of the stack.
419  *
420  * Calling this with:
421  *
422  *   offset = task->curr_ret_stack;
423  *   do {
424  *	ret_stack = get_ret_stack(task, offset, &offset);
425  *   } while (ret_stack);
426  *
427  * Will iterate through all the ret_stack entries from curr_ret_stack
428  * down to the first one.
429  */
430 static inline struct ftrace_ret_stack *
431 get_ret_stack(struct task_struct *t, int offset, int *frame_offset)
432 {
433 	int offs;
434 
435 	BUILD_BUG_ON(FGRAPH_FRAME_SIZE % sizeof(long));
436 
437 	if (unlikely(offset <= 0))
438 		return NULL;
439 
440 	offs = get_frame_offset(t, --offset);
441 	if (WARN_ON_ONCE(offs <= 0 || offs > offset))
442 		return NULL;
443 
444 	offset -= offs;
445 
446 	*frame_offset = offset;
447 	return RET_STACK(t, offset);
448 }
449 
450 /**
451  * fgraph_retrieve_parent_data - get data from a parent function
452  * @idx: The index into the fgraph_array (fgraph_ops::idx)
453  * @size_bytes: A pointer to retrieved data size
454  * @depth: The depth to find the parent (0 is the current function)
455  *
456  * This is similar to fgraph_retrieve_data() but can be used to retrieve
457  * data from a parent caller function.
458  *
459  * Return: a pointer to the specified parent data or NULL if not found
460  */
461 void *fgraph_retrieve_parent_data(int idx, int *size_bytes, int depth)
462 {
463 	struct ftrace_ret_stack *ret_stack = NULL;
464 	int offset = current->curr_ret_stack;
465 	unsigned long val;
466 
467 	if (offset <= 0)
468 		return NULL;
469 
470 	for (;;) {
471 		int next_offset;
472 
473 		ret_stack = get_ret_stack(current, offset, &next_offset);
474 		if (!ret_stack || --depth < 0)
475 			break;
476 		offset = next_offset;
477 	}
478 
479 	if (!ret_stack)
480 		return NULL;
481 
482 	offset--;
483 
484 	val = get_fgraph_entry(current, offset);
485 	while (__get_type(val) == FGRAPH_TYPE_DATA) {
486 		if (__get_data_index(val) == idx)
487 			goto found;
488 		offset -= __get_data_size(val) + 1;
489 		val = get_fgraph_entry(current, offset);
490 	}
491 	return NULL;
492 found:
493 	if (size_bytes)
494 		*size_bytes = __get_data_size(val) * sizeof(long);
495 	return get_data_type_data(current, offset);
496 }
497 
498 /* Both enabled by default (can be cleared by function_graph tracer flags */
499 bool fgraph_sleep_time = true;
500 
501 #ifdef CONFIG_DYNAMIC_FTRACE
502 /*
503  * archs can override this function if they must do something
504  * to enable hook for graph tracer.
505  */
506 int __weak ftrace_enable_ftrace_graph_caller(void)
507 {
508 	return 0;
509 }
510 
511 /*
512  * archs can override this function if they must do something
513  * to disable hook for graph tracer.
514  */
515 int __weak ftrace_disable_ftrace_graph_caller(void)
516 {
517 	return 0;
518 }
519 #endif
520 
521 int ftrace_graph_entry_stub(struct ftrace_graph_ent *trace,
522 			    struct fgraph_ops *gops,
523 			    struct fgraph_extras *extras)
524 {
525 	return 0;
526 }
527 
528 static void ftrace_graph_ret_stub(struct ftrace_graph_ret *trace,
529 				  struct fgraph_ops *gops)
530 {
531 }
532 
533 static struct fgraph_ops fgraph_stub = {
534 	.entryfunc = ftrace_graph_entry_stub,
535 	.retfunc = ftrace_graph_ret_stub,
536 };
537 
538 static struct fgraph_ops *fgraph_direct_gops = &fgraph_stub;
539 DEFINE_STATIC_CALL(fgraph_func, ftrace_graph_entry_stub);
540 DEFINE_STATIC_CALL(fgraph_retfunc, ftrace_graph_ret_stub);
541 static DEFINE_STATIC_KEY_TRUE(fgraph_do_direct);
542 
543 /**
544  * ftrace_graph_stop - set to permanently disable function graph tracing
545  *
546  * In case of an error int function graph tracing, this is called
547  * to try to keep function graph tracing from causing any more harm.
548  * Usually this is pretty severe and this is called to try to at least
549  * get a warning out to the user.
550  */
551 void ftrace_graph_stop(void)
552 {
553 	static_branch_enable(&kill_ftrace_graph);
554 }
555 
556 /* Add a function return address to the trace stack on thread info.*/
557 static int
558 ftrace_push_return_trace(unsigned long ret, unsigned long func,
559 			 unsigned long frame_pointer, unsigned long *retp,
560 			 int fgraph_idx)
561 {
562 	struct ftrace_ret_stack *ret_stack;
563 	unsigned long val;
564 	int offset;
565 
566 	if (unlikely(ftrace_graph_is_dead()))
567 		return -EBUSY;
568 
569 	if (!current->ret_stack)
570 		return -EBUSY;
571 
572 	BUILD_BUG_ON(SHADOW_STACK_SIZE % sizeof(long));
573 
574 	/* Set val to "reserved" with the delta to the new fgraph frame */
575 	val = (FGRAPH_TYPE_RESERVED << FGRAPH_TYPE_SHIFT) | FGRAPH_FRAME_OFFSET;
576 
577 	/*
578 	 * We must make sure the ret_stack is tested before we read
579 	 * anything else.
580 	 */
581 	smp_rmb();
582 
583 	/*
584 	 * Check if there's room on the shadow stack to fit a fraph frame
585 	 * and a bitmap word.
586 	 */
587 	if (current->curr_ret_stack + FGRAPH_FRAME_OFFSET + 1 >= SHADOW_STACK_MAX_OFFSET) {
588 		atomic_inc(&current->trace_overrun);
589 		return -EBUSY;
590 	}
591 
592 	offset = READ_ONCE(current->curr_ret_stack);
593 	ret_stack = RET_STACK(current, offset);
594 	offset += FGRAPH_FRAME_OFFSET;
595 
596 	/* ret offset = FGRAPH_FRAME_OFFSET ; type = reserved */
597 	current->ret_stack[offset] = val;
598 	ret_stack->ret = ret;
599 	/*
600 	 * The unwinders expect curr_ret_stack to point to either zero
601 	 * or an offset where to find the next ret_stack. Even though the
602 	 * ret stack might be bogus, we want to write the ret and the
603 	 * offset to find the ret_stack before we increment the stack point.
604 	 * If an interrupt comes in now before we increment the curr_ret_stack
605 	 * it may blow away what we wrote. But that's fine, because the
606 	 * offset will still be correct (even though the 'ret' won't be).
607 	 * What we worry about is the offset being correct after we increment
608 	 * the curr_ret_stack and before we update that offset, as if an
609 	 * interrupt comes in and does an unwind stack dump, it will need
610 	 * at least a correct offset!
611 	 */
612 	barrier();
613 	WRITE_ONCE(current->curr_ret_stack, offset + 1);
614 	/*
615 	 * This next barrier is to ensure that an interrupt coming in
616 	 * will not corrupt what we are about to write.
617 	 */
618 	barrier();
619 
620 	/* Still keep it reserved even if an interrupt came in */
621 	current->ret_stack[offset] = val;
622 
623 	ret_stack->ret = ret;
624 	ret_stack->func = func;
625 #ifdef HAVE_FUNCTION_GRAPH_FP_TEST
626 	ret_stack->fp = frame_pointer;
627 #endif
628 	ret_stack->retp = retp;
629 	return offset;
630 }
631 
632 /*
633  * Not all archs define MCOUNT_INSN_SIZE which is used to look for direct
634  * functions. But those archs currently don't support direct functions
635  * anyway, and ftrace_find_rec_direct() is just a stub for them.
636  * Define MCOUNT_INSN_SIZE to keep those archs compiling.
637  */
638 #ifndef MCOUNT_INSN_SIZE
639 /* Make sure this only works without direct calls */
640 # ifdef CONFIG_DYNAMIC_FTRACE_WITH_DIRECT_CALLS
641 #  error MCOUNT_INSN_SIZE not defined with direct calls enabled
642 # endif
643 # define MCOUNT_INSN_SIZE 0
644 #endif
645 
646 /* If the caller does not use ftrace, call this function. */
647 int function_graph_enter(unsigned long ret, unsigned long func,
648 			 unsigned long frame_pointer, unsigned long *retp)
649 {
650 	struct ftrace_graph_ent trace;
651 	struct fgraph_extras extras;
652 	unsigned long bitmap = 0;
653 	int offset;
654 	int i;
655 	int idx = 0;
656 
657 	trace.func = func;
658 	trace.depth = ++current->curr_ret_depth;
659 
660 	extras.flags = graph_tracer_flags_get(TRACE_GRAPH_PRINT_RETADDR);
661 	if (IS_ENABLED(CONFIG_FUNCTION_GRAPH_RETADDR)
662 		&& extras.flags & TRACE_GRAPH_PRINT_RETADDR)
663 		extras.retaddr = ftrace_graph_ret_addr(current, &idx, ret, retp);
664 
665 	offset = ftrace_push_return_trace(ret, func, frame_pointer, retp, 0);
666 	if (offset < 0)
667 		goto out;
668 
669 #ifdef CONFIG_HAVE_STATIC_CALL
670 	if (static_branch_likely(&fgraph_do_direct)) {
671 		int save_curr_ret_stack = current->curr_ret_stack;
672 
673 		if (static_call(fgraph_func)(&trace, fgraph_direct_gops, &extras))
674 			bitmap |= BIT(fgraph_direct_gops->idx);
675 		else
676 			/* Clear out any saved storage */
677 			current->curr_ret_stack = save_curr_ret_stack;
678 	} else
679 #endif
680 	{
681 		for_each_set_bit(i, &fgraph_array_bitmask,
682 					 sizeof(fgraph_array_bitmask) * BITS_PER_BYTE) {
683 			struct fgraph_ops *gops = READ_ONCE(fgraph_array[i]);
684 			int save_curr_ret_stack;
685 
686 			if (gops == &fgraph_stub)
687 				continue;
688 
689 			save_curr_ret_stack = current->curr_ret_stack;
690 			if (ftrace_ops_test(&gops->ops, func, NULL) &&
691 			    gops->entryfunc(&trace, gops, &extras))
692 				bitmap |= BIT(i);
693 			else
694 				/* Clear out any saved storage */
695 				current->curr_ret_stack = save_curr_ret_stack;
696 		}
697 	}
698 
699 	if (!bitmap)
700 		goto out_ret;
701 
702 	/*
703 	 * Since this function uses fgraph_idx = 0 as a tail-call checking
704 	 * flag, set that bit always.
705 	 */
706 	set_bitmap(current, offset, bitmap | BIT(0));
707 
708 	return 0;
709  out_ret:
710 	current->curr_ret_stack -= FGRAPH_FRAME_OFFSET + 1;
711  out:
712 	current->curr_ret_depth--;
713 	return -EBUSY;
714 }
715 
716 /* Retrieve a function return address to the trace stack on thread info.*/
717 static struct ftrace_ret_stack *
718 ftrace_pop_return_trace(struct ftrace_graph_ret *trace, unsigned long *ret,
719 			unsigned long frame_pointer, int *offset)
720 {
721 	struct ftrace_ret_stack *ret_stack;
722 
723 	ret_stack = get_ret_stack(current, current->curr_ret_stack, offset);
724 
725 	if (unlikely(!ret_stack)) {
726 		ftrace_graph_stop();
727 		WARN(1, "Bad function graph ret_stack pointer: %d",
728 		     current->curr_ret_stack);
729 		/* Might as well panic, otherwise we have no where to go */
730 		*ret = (unsigned long)panic;
731 		return NULL;
732 	}
733 
734 #ifdef HAVE_FUNCTION_GRAPH_FP_TEST
735 	/*
736 	 * The arch may choose to record the frame pointer used
737 	 * and check it here to make sure that it is what we expect it
738 	 * to be. If gcc does not set the place holder of the return
739 	 * address in the frame pointer, and does a copy instead, then
740 	 * the function graph trace will fail. This test detects this
741 	 * case.
742 	 *
743 	 * Currently, x86_32 with optimize for size (-Os) makes the latest
744 	 * gcc do the above.
745 	 *
746 	 * Note, -mfentry does not use frame pointers, and this test
747 	 *  is not needed if CC_USING_FENTRY is set.
748 	 */
749 	if (unlikely(ret_stack->fp != frame_pointer)) {
750 		ftrace_graph_stop();
751 		WARN(1, "Bad frame pointer: expected %lx, received %lx\n"
752 		     "  from func %ps return to %lx\n",
753 		     ret_stack->fp,
754 		     frame_pointer,
755 		     (void *)ret_stack->func,
756 		     ret_stack->ret);
757 		*ret = (unsigned long)panic;
758 		return NULL;
759 	}
760 #endif
761 
762 	*offset += FGRAPH_FRAME_OFFSET;
763 	*ret = ret_stack->ret;
764 	trace->func = ret_stack->func;
765 	trace->overrun = atomic_read(&current->trace_overrun);
766 	trace->depth = current->curr_ret_depth;
767 	/*
768 	 * We still want to trace interrupts coming in if
769 	 * max_depth is set to 1. Make sure the decrement is
770 	 * seen before ftrace_graph_return.
771 	 */
772 	barrier();
773 
774 	return ret_stack;
775 }
776 
777 /*
778  * Hibernation protection.
779  * The state of the current task is too much unstable during
780  * suspend/restore to disk. We want to protect against that.
781  */
782 static int
783 ftrace_suspend_notifier_call(struct notifier_block *bl, unsigned long state,
784 							void *unused)
785 {
786 	switch (state) {
787 	case PM_HIBERNATION_PREPARE:
788 		pause_graph_tracing();
789 		break;
790 
791 	case PM_POST_HIBERNATION:
792 		unpause_graph_tracing();
793 		break;
794 	}
795 	return NOTIFY_DONE;
796 }
797 
798 static struct notifier_block ftrace_suspend_notifier = {
799 	.notifier_call = ftrace_suspend_notifier_call,
800 };
801 
802 /* fgraph_ret_regs is not defined without CONFIG_FUNCTION_GRAPH_RETVAL */
803 struct fgraph_ret_regs;
804 
805 /*
806  * Send the trace to the ring-buffer.
807  * @return the original return address.
808  */
809 static unsigned long __ftrace_return_to_handler(struct fgraph_ret_regs *ret_regs,
810 						unsigned long frame_pointer)
811 {
812 	struct ftrace_ret_stack *ret_stack;
813 	struct ftrace_graph_ret trace;
814 	unsigned long bitmap;
815 	unsigned long ret;
816 	int offset;
817 	int i;
818 
819 	ret_stack = ftrace_pop_return_trace(&trace, &ret, frame_pointer, &offset);
820 
821 	if (unlikely(!ret_stack)) {
822 		ftrace_graph_stop();
823 		WARN_ON(1);
824 		/* Might as well panic. What else to do? */
825 		return (unsigned long)panic;
826 	}
827 
828 	trace.rettime = trace_clock_local();
829 #ifdef CONFIG_FUNCTION_GRAPH_RETVAL
830 	trace.retval = fgraph_ret_regs_return_value(ret_regs);
831 #endif
832 
833 	bitmap = get_bitmap_bits(current, offset);
834 
835 #ifdef CONFIG_HAVE_STATIC_CALL
836 	if (static_branch_likely(&fgraph_do_direct)) {
837 		if (test_bit(fgraph_direct_gops->idx, &bitmap))
838 			static_call(fgraph_retfunc)(&trace, fgraph_direct_gops);
839 	} else
840 #endif
841 	{
842 		for_each_set_bit(i, &bitmap, sizeof(bitmap) * BITS_PER_BYTE) {
843 			struct fgraph_ops *gops = fgraph_array[i];
844 
845 			if (gops == &fgraph_stub)
846 				continue;
847 
848 			gops->retfunc(&trace, gops);
849 		}
850 	}
851 
852 	/*
853 	 * The ftrace_graph_return() may still access the current
854 	 * ret_stack structure, we need to make sure the update of
855 	 * curr_ret_stack is after that.
856 	 */
857 	barrier();
858 	current->curr_ret_stack = offset - FGRAPH_FRAME_OFFSET;
859 
860 	current->curr_ret_depth--;
861 	return ret;
862 }
863 
864 /*
865  * After all architecures have selected HAVE_FUNCTION_GRAPH_RETVAL, we can
866  * leave only ftrace_return_to_handler(ret_regs).
867  */
868 #ifdef CONFIG_HAVE_FUNCTION_GRAPH_RETVAL
869 unsigned long ftrace_return_to_handler(struct fgraph_ret_regs *ret_regs)
870 {
871 	return __ftrace_return_to_handler(ret_regs,
872 				fgraph_ret_regs_frame_pointer(ret_regs));
873 }
874 #else
875 unsigned long ftrace_return_to_handler(unsigned long frame_pointer)
876 {
877 	return __ftrace_return_to_handler(NULL, frame_pointer);
878 }
879 #endif
880 
881 /**
882  * ftrace_graph_get_ret_stack - return the entry of the shadow stack
883  * @task: The task to read the shadow stack from.
884  * @idx: Index down the shadow stack
885  *
886  * Return the ret_struct on the shadow stack of the @task at the
887  * call graph at @idx starting with zero. If @idx is zero, it
888  * will return the last saved ret_stack entry. If it is greater than
889  * zero, it will return the corresponding ret_stack for the depth
890  * of saved return addresses.
891  */
892 struct ftrace_ret_stack *
893 ftrace_graph_get_ret_stack(struct task_struct *task, int idx)
894 {
895 	struct ftrace_ret_stack *ret_stack = NULL;
896 	int offset = task->curr_ret_stack;
897 
898 	if (offset < 0)
899 		return NULL;
900 
901 	do {
902 		ret_stack = get_ret_stack(task, offset, &offset);
903 	} while (ret_stack && --idx >= 0);
904 
905 	return ret_stack;
906 }
907 
908 /**
909  * ftrace_graph_ret_addr - return the original value of the return address
910  * @task: The task the unwinder is being executed on
911  * @idx: An initialized pointer to the next stack index to use
912  * @ret: The current return address (likely pointing to return_handler)
913  * @retp: The address on the stack of the current return location
914  *
915  * This function can be called by stack unwinding code to convert a found stack
916  * return address (@ret) to its original value, in case the function graph
917  * tracer has modified it to be 'return_to_handler'.  If the address hasn't
918  * been modified, the unchanged value of @ret is returned.
919  *
920  * @idx holds the last index used to know where to start from. It should be
921  * initialized to zero for the first iteration as that will mean to start
922  * at the top of the shadow stack. If the location is found, this pointer
923  * will be assigned that location so that if called again, it will continue
924  * where it left off.
925  *
926  * @retp is a pointer to the return address on the stack.
927  */
928 unsigned long ftrace_graph_ret_addr(struct task_struct *task, int *idx,
929 				    unsigned long ret, unsigned long *retp)
930 {
931 	struct ftrace_ret_stack *ret_stack;
932 	unsigned long return_handler = (unsigned long)dereference_kernel_function_descriptor(return_to_handler);
933 	int i;
934 
935 	if (ret != return_handler)
936 		return ret;
937 
938 	if (!idx)
939 		return ret;
940 
941 	i = *idx ? : task->curr_ret_stack;
942 	while (i > 0) {
943 		ret_stack = get_ret_stack(task, i, &i);
944 		if (!ret_stack)
945 			break;
946 		/*
947 		 * For the tail-call, there would be 2 or more ftrace_ret_stacks on
948 		 * the ret_stack, which records "return_to_handler" as the return
949 		 * address except for the last one.
950 		 * But on the real stack, there should be 1 entry because tail-call
951 		 * reuses the return address on the stack and jump to the next function.
952 		 * Thus we will continue to find real return address.
953 		 */
954 		if (ret_stack->retp == retp &&
955 		    ret_stack->ret != return_handler) {
956 			*idx = i;
957 			return ret_stack->ret;
958 		}
959 	}
960 
961 	return ret;
962 }
963 
964 static struct ftrace_ops graph_ops = {
965 	.func			= ftrace_graph_func,
966 	.flags			= FTRACE_OPS_GRAPH_STUB,
967 #ifdef FTRACE_GRAPH_TRAMP_ADDR
968 	.trampoline		= FTRACE_GRAPH_TRAMP_ADDR,
969 	/* trampoline_size is only needed for dynamically allocated tramps */
970 #endif
971 };
972 
973 void fgraph_init_ops(struct ftrace_ops *dst_ops,
974 		     struct ftrace_ops *src_ops)
975 {
976 	dst_ops->flags = FTRACE_OPS_FL_PID | FTRACE_OPS_GRAPH_STUB;
977 
978 #ifdef CONFIG_DYNAMIC_FTRACE
979 	if (src_ops) {
980 		dst_ops->func_hash = &src_ops->local_hash;
981 		mutex_init(&dst_ops->local_hash.regex_lock);
982 		INIT_LIST_HEAD(&dst_ops->subop_list);
983 		dst_ops->flags |= FTRACE_OPS_FL_INITIALIZED;
984 	}
985 #endif
986 }
987 
988 void ftrace_graph_sleep_time_control(bool enable)
989 {
990 	fgraph_sleep_time = enable;
991 }
992 
993 /*
994  * Simply points to ftrace_stub, but with the proper protocol.
995  * Defined by the linker script in linux/vmlinux.lds.h
996  */
997 void ftrace_stub_graph(struct ftrace_graph_ret *trace, struct fgraph_ops *gops);
998 
999 /* The callbacks that hook a function */
1000 trace_func_graph_ret_t ftrace_graph_return = ftrace_stub_graph;
1001 trace_func_graph_ent_t ftrace_graph_entry = ftrace_graph_entry_stub;
1002 
1003 /* Try to assign a return stack array on FTRACE_RETSTACK_ALLOC_SIZE tasks. */
1004 static int alloc_retstack_tasklist(unsigned long **ret_stack_list)
1005 {
1006 	int i;
1007 	int ret = 0;
1008 	int start = 0, end = FTRACE_RETSTACK_ALLOC_SIZE;
1009 	struct task_struct *g, *t;
1010 
1011 	for (i = 0; i < FTRACE_RETSTACK_ALLOC_SIZE; i++) {
1012 		ret_stack_list[i] = kmalloc(SHADOW_STACK_SIZE, GFP_KERNEL);
1013 		if (!ret_stack_list[i]) {
1014 			start = 0;
1015 			end = i;
1016 			ret = -ENOMEM;
1017 			goto free;
1018 		}
1019 	}
1020 
1021 	rcu_read_lock();
1022 	for_each_process_thread(g, t) {
1023 		if (start == end) {
1024 			ret = -EAGAIN;
1025 			goto unlock;
1026 		}
1027 
1028 		if (t->ret_stack == NULL) {
1029 			atomic_set(&t->trace_overrun, 0);
1030 			ret_stack_init_task_vars(ret_stack_list[start]);
1031 			t->curr_ret_stack = 0;
1032 			t->curr_ret_depth = -1;
1033 			/* Make sure the tasks see the 0 first: */
1034 			smp_wmb();
1035 			t->ret_stack = ret_stack_list[start++];
1036 		}
1037 	}
1038 
1039 unlock:
1040 	rcu_read_unlock();
1041 free:
1042 	for (i = start; i < end; i++)
1043 		kfree(ret_stack_list[i]);
1044 	return ret;
1045 }
1046 
1047 static void
1048 ftrace_graph_probe_sched_switch(void *ignore, bool preempt,
1049 				struct task_struct *prev,
1050 				struct task_struct *next,
1051 				unsigned int prev_state)
1052 {
1053 	unsigned long long timestamp;
1054 
1055 	/*
1056 	 * Does the user want to count the time a function was asleep.
1057 	 * If so, do not update the time stamps.
1058 	 */
1059 	if (fgraph_sleep_time)
1060 		return;
1061 
1062 	timestamp = trace_clock_local();
1063 
1064 	prev->ftrace_timestamp = timestamp;
1065 
1066 	/* only process tasks that we timestamped */
1067 	if (!next->ftrace_timestamp)
1068 		return;
1069 
1070 	next->ftrace_sleeptime += timestamp - next->ftrace_timestamp;
1071 }
1072 
1073 static DEFINE_PER_CPU(unsigned long *, idle_ret_stack);
1074 
1075 static void
1076 graph_init_task(struct task_struct *t, unsigned long *ret_stack)
1077 {
1078 	atomic_set(&t->trace_overrun, 0);
1079 	ret_stack_init_task_vars(ret_stack);
1080 	t->ftrace_timestamp = 0;
1081 	t->curr_ret_stack = 0;
1082 	t->curr_ret_depth = -1;
1083 	/* make curr_ret_stack visible before we add the ret_stack */
1084 	smp_wmb();
1085 	t->ret_stack = ret_stack;
1086 }
1087 
1088 /*
1089  * Allocate a return stack for the idle task. May be the first
1090  * time through, or it may be done by CPU hotplug online.
1091  */
1092 void ftrace_graph_init_idle_task(struct task_struct *t, int cpu)
1093 {
1094 	t->curr_ret_stack = 0;
1095 	t->curr_ret_depth = -1;
1096 	/*
1097 	 * The idle task has no parent, it either has its own
1098 	 * stack or no stack at all.
1099 	 */
1100 	if (t->ret_stack)
1101 		WARN_ON(t->ret_stack != per_cpu(idle_ret_stack, cpu));
1102 
1103 	if (ftrace_graph_active) {
1104 		unsigned long *ret_stack;
1105 
1106 		ret_stack = per_cpu(idle_ret_stack, cpu);
1107 		if (!ret_stack) {
1108 			ret_stack = kmalloc(SHADOW_STACK_SIZE, GFP_KERNEL);
1109 			if (!ret_stack)
1110 				return;
1111 			per_cpu(idle_ret_stack, cpu) = ret_stack;
1112 		}
1113 		graph_init_task(t, ret_stack);
1114 	}
1115 }
1116 
1117 /* Allocate a return stack for newly created task */
1118 void ftrace_graph_init_task(struct task_struct *t)
1119 {
1120 	/* Make sure we do not use the parent ret_stack */
1121 	t->ret_stack = NULL;
1122 	t->curr_ret_stack = 0;
1123 	t->curr_ret_depth = -1;
1124 
1125 	if (ftrace_graph_active) {
1126 		unsigned long *ret_stack;
1127 
1128 		ret_stack = kmalloc(SHADOW_STACK_SIZE, GFP_KERNEL);
1129 		if (!ret_stack)
1130 			return;
1131 		graph_init_task(t, ret_stack);
1132 	}
1133 }
1134 
1135 void ftrace_graph_exit_task(struct task_struct *t)
1136 {
1137 	unsigned long *ret_stack = t->ret_stack;
1138 
1139 	t->ret_stack = NULL;
1140 	/* NULL must become visible to IRQs before we free it: */
1141 	barrier();
1142 
1143 	kfree(ret_stack);
1144 }
1145 
1146 #ifdef CONFIG_DYNAMIC_FTRACE
1147 static int fgraph_pid_func(struct ftrace_graph_ent *trace,
1148 			   struct fgraph_ops *gops,
1149 			   struct fgraph_extras *extras)
1150 {
1151 	struct trace_array *tr = gops->ops.private;
1152 	int pid;
1153 
1154 	if (tr) {
1155 		pid = this_cpu_read(tr->array_buffer.data->ftrace_ignore_pid);
1156 		if (pid == FTRACE_PID_IGNORE)
1157 			return 0;
1158 		if (pid != FTRACE_PID_TRACE &&
1159 		    pid != current->pid)
1160 			return 0;
1161 	}
1162 
1163 	return gops->saved_func(trace, gops, NULL);
1164 }
1165 
1166 void fgraph_update_pid_func(void)
1167 {
1168 	struct fgraph_ops *gops;
1169 	struct ftrace_ops *op;
1170 
1171 	if (!(graph_ops.flags & FTRACE_OPS_FL_INITIALIZED))
1172 		return;
1173 
1174 	list_for_each_entry(op, &graph_ops.subop_list, list) {
1175 		if (op->flags & FTRACE_OPS_FL_PID) {
1176 			gops = container_of(op, struct fgraph_ops, ops);
1177 			gops->entryfunc = ftrace_pids_enabled(op) ?
1178 				fgraph_pid_func : gops->saved_func;
1179 			if (ftrace_graph_active == 1)
1180 				static_call_update(fgraph_func, gops->entryfunc);
1181 		}
1182 	}
1183 }
1184 #endif
1185 
1186 /* Allocate a return stack for each task */
1187 static int start_graph_tracing(void)
1188 {
1189 	unsigned long **ret_stack_list;
1190 	int ret, cpu;
1191 
1192 	ret_stack_list = kmalloc(SHADOW_STACK_SIZE, GFP_KERNEL);
1193 
1194 	if (!ret_stack_list)
1195 		return -ENOMEM;
1196 
1197 	/* The cpu_boot init_task->ret_stack will never be freed */
1198 	for_each_online_cpu(cpu) {
1199 		if (!idle_task(cpu)->ret_stack)
1200 			ftrace_graph_init_idle_task(idle_task(cpu), cpu);
1201 	}
1202 
1203 	do {
1204 		ret = alloc_retstack_tasklist(ret_stack_list);
1205 	} while (ret == -EAGAIN);
1206 
1207 	if (!ret) {
1208 		ret = register_trace_sched_switch(ftrace_graph_probe_sched_switch, NULL);
1209 		if (ret)
1210 			pr_info("ftrace_graph: Couldn't activate tracepoint"
1211 				" probe to kernel_sched_switch\n");
1212 	}
1213 
1214 	kfree(ret_stack_list);
1215 	return ret;
1216 }
1217 
1218 static void init_task_vars(int idx)
1219 {
1220 	struct task_struct *g, *t;
1221 	int cpu;
1222 
1223 	for_each_online_cpu(cpu) {
1224 		if (idle_task(cpu)->ret_stack)
1225 			ret_stack_set_task_var(idle_task(cpu), idx, 0);
1226 	}
1227 
1228 	read_lock(&tasklist_lock);
1229 	for_each_process_thread(g, t) {
1230 		if (t->ret_stack)
1231 			ret_stack_set_task_var(t, idx, 0);
1232 	}
1233 	read_unlock(&tasklist_lock);
1234 }
1235 
1236 static void ftrace_graph_enable_direct(bool enable_branch, struct fgraph_ops *gops)
1237 {
1238 	trace_func_graph_ent_t func = NULL;
1239 	trace_func_graph_ret_t retfunc = NULL;
1240 	int i;
1241 
1242 	if (gops) {
1243 		func = gops->entryfunc;
1244 		retfunc = gops->retfunc;
1245 		fgraph_direct_gops = gops;
1246 	} else {
1247 		for_each_set_bit(i, &fgraph_array_bitmask,
1248 				 sizeof(fgraph_array_bitmask) * BITS_PER_BYTE) {
1249 			func = fgraph_array[i]->entryfunc;
1250 			retfunc = fgraph_array[i]->retfunc;
1251 			fgraph_direct_gops = fgraph_array[i];
1252 		}
1253 	}
1254 	if (WARN_ON_ONCE(!func))
1255 		return;
1256 
1257 	static_call_update(fgraph_func, func);
1258 	static_call_update(fgraph_retfunc, retfunc);
1259 	if (enable_branch)
1260 		static_branch_disable(&fgraph_do_direct);
1261 }
1262 
1263 static void ftrace_graph_disable_direct(bool disable_branch)
1264 {
1265 	if (disable_branch)
1266 		static_branch_disable(&fgraph_do_direct);
1267 	static_call_update(fgraph_func, ftrace_graph_entry_stub);
1268 	static_call_update(fgraph_retfunc, ftrace_graph_ret_stub);
1269 	fgraph_direct_gops = &fgraph_stub;
1270 }
1271 
1272 int register_ftrace_graph(struct fgraph_ops *gops)
1273 {
1274 	int command = 0;
1275 	int ret = 0;
1276 	int i = -1;
1277 
1278 	mutex_lock(&ftrace_lock);
1279 
1280 	if (!fgraph_array[0]) {
1281 		/* The array must always have real data on it */
1282 		for (i = 0; i < FGRAPH_ARRAY_SIZE; i++)
1283 			fgraph_array[i] = &fgraph_stub;
1284 		fgraph_lru_init();
1285 	}
1286 
1287 	i = fgraph_lru_alloc_index();
1288 	if (i < 0 || WARN_ON_ONCE(fgraph_array[i] != &fgraph_stub)) {
1289 		ret = -ENOSPC;
1290 		goto out;
1291 	}
1292 	gops->idx = i;
1293 
1294 	ftrace_graph_active++;
1295 
1296 	if (ftrace_graph_active == 2)
1297 		ftrace_graph_disable_direct(true);
1298 
1299 	if (ftrace_graph_active == 1) {
1300 		ftrace_graph_enable_direct(false, gops);
1301 		register_pm_notifier(&ftrace_suspend_notifier);
1302 		ret = start_graph_tracing();
1303 		if (ret)
1304 			goto error;
1305 		/*
1306 		 * Some archs just test to see if these are not
1307 		 * the default function
1308 		 */
1309 		ftrace_graph_return = return_run;
1310 		ftrace_graph_entry = entry_run;
1311 		command = FTRACE_START_FUNC_RET;
1312 	} else {
1313 		init_task_vars(gops->idx);
1314 	}
1315 	/* Always save the function, and reset at unregistering */
1316 	gops->saved_func = gops->entryfunc;
1317 
1318 	ret = ftrace_startup_subops(&graph_ops, &gops->ops, command);
1319 	if (!ret)
1320 		fgraph_array[i] = gops;
1321 
1322 error:
1323 	if (ret) {
1324 		ftrace_graph_active--;
1325 		gops->saved_func = NULL;
1326 		fgraph_lru_release_index(i);
1327 	}
1328 out:
1329 	mutex_unlock(&ftrace_lock);
1330 	return ret;
1331 }
1332 
1333 void unregister_ftrace_graph(struct fgraph_ops *gops)
1334 {
1335 	int command = 0;
1336 
1337 	mutex_lock(&ftrace_lock);
1338 
1339 	if (unlikely(!ftrace_graph_active))
1340 		goto out;
1341 
1342 	if (unlikely(gops->idx < 0 || gops->idx >= FGRAPH_ARRAY_SIZE ||
1343 		     fgraph_array[gops->idx] != gops))
1344 		goto out;
1345 
1346 	if (fgraph_lru_release_index(gops->idx) < 0)
1347 		goto out;
1348 
1349 	fgraph_array[gops->idx] = &fgraph_stub;
1350 
1351 	ftrace_graph_active--;
1352 
1353 	if (!ftrace_graph_active)
1354 		command = FTRACE_STOP_FUNC_RET;
1355 
1356 	ftrace_shutdown_subops(&graph_ops, &gops->ops, command);
1357 
1358 	if (ftrace_graph_active == 1)
1359 		ftrace_graph_enable_direct(true, NULL);
1360 	else if (!ftrace_graph_active)
1361 		ftrace_graph_disable_direct(false);
1362 
1363 	if (!ftrace_graph_active) {
1364 		ftrace_graph_return = ftrace_stub_graph;
1365 		ftrace_graph_entry = ftrace_graph_entry_stub;
1366 		unregister_pm_notifier(&ftrace_suspend_notifier);
1367 		unregister_trace_sched_switch(ftrace_graph_probe_sched_switch, NULL);
1368 	}
1369  out:
1370 	gops->saved_func = NULL;
1371 	mutex_unlock(&ftrace_lock);
1372 }
1373