1 // SPDX-License-Identifier: (GPL-2.0-only OR BSD-2-Clause)
2 // Copyright (c) 2022 Google
3 #include "vmlinux.h"
4 #include <bpf/bpf_helpers.h>
5 #include <bpf/bpf_tracing.h>
6 #include <bpf/bpf_core_read.h>
7 #include <asm-generic/errno-base.h>
8 
9 #include "lock_data.h"
10 
11 /* for collect_lock_syms().  4096 was rejected by the verifier */
12 #define MAX_CPUS  1024
13 
14 /* lock contention flags from include/trace/events/lock.h */
15 #define LCB_F_SPIN	(1U << 0)
16 #define LCB_F_READ	(1U << 1)
17 #define LCB_F_WRITE	(1U << 2)
18 #define LCB_F_RT	(1U << 3)
19 #define LCB_F_PERCPU	(1U << 4)
20 #define LCB_F_MUTEX	(1U << 5)
21 
22 struct tstamp_data {
23 	__u64 timestamp;
24 	__u64 lock;
25 	__u32 flags;
26 	__s32 stack_id;
27 };
28 
29 /* callstack storage  */
30 struct {
31 	__uint(type, BPF_MAP_TYPE_STACK_TRACE);
32 	__uint(key_size, sizeof(__u32));
33 	__uint(value_size, sizeof(__u64));
34 	__uint(max_entries, MAX_ENTRIES);
35 } stacks SEC(".maps");
36 
37 /* maintain timestamp at the beginning of contention */
38 struct {
39 	__uint(type, BPF_MAP_TYPE_HASH);
40 	__type(key, int);
41 	__type(value, struct tstamp_data);
42 	__uint(max_entries, MAX_ENTRIES);
43 } tstamp SEC(".maps");
44 
45 /* actual lock contention statistics */
46 struct {
47 	__uint(type, BPF_MAP_TYPE_HASH);
48 	__uint(key_size, sizeof(struct contention_key));
49 	__uint(value_size, sizeof(struct contention_data));
50 	__uint(max_entries, MAX_ENTRIES);
51 } lock_stat SEC(".maps");
52 
53 struct {
54 	__uint(type, BPF_MAP_TYPE_HASH);
55 	__uint(key_size, sizeof(__u32));
56 	__uint(value_size, sizeof(struct contention_task_data));
57 	__uint(max_entries, MAX_ENTRIES);
58 } task_data SEC(".maps");
59 
60 struct {
61 	__uint(type, BPF_MAP_TYPE_HASH);
62 	__uint(key_size, sizeof(__u64));
63 	__uint(value_size, sizeof(__u32));
64 	__uint(max_entries, MAX_ENTRIES);
65 } lock_syms SEC(".maps");
66 
67 struct {
68 	__uint(type, BPF_MAP_TYPE_HASH);
69 	__uint(key_size, sizeof(__u32));
70 	__uint(value_size, sizeof(__u8));
71 	__uint(max_entries, 1);
72 } cpu_filter SEC(".maps");
73 
74 struct {
75 	__uint(type, BPF_MAP_TYPE_HASH);
76 	__uint(key_size, sizeof(__u32));
77 	__uint(value_size, sizeof(__u8));
78 	__uint(max_entries, 1);
79 } task_filter SEC(".maps");
80 
81 struct {
82 	__uint(type, BPF_MAP_TYPE_HASH);
83 	__uint(key_size, sizeof(__u32));
84 	__uint(value_size, sizeof(__u8));
85 	__uint(max_entries, 1);
86 } type_filter SEC(".maps");
87 
88 struct {
89 	__uint(type, BPF_MAP_TYPE_HASH);
90 	__uint(key_size, sizeof(__u64));
91 	__uint(value_size, sizeof(__u8));
92 	__uint(max_entries, 1);
93 } addr_filter SEC(".maps");
94 
95 struct rw_semaphore___old {
96 	struct task_struct *owner;
97 } __attribute__((preserve_access_index));
98 
99 struct rw_semaphore___new {
100 	atomic_long_t owner;
101 } __attribute__((preserve_access_index));
102 
103 struct mm_struct___old {
104 	struct rw_semaphore mmap_sem;
105 } __attribute__((preserve_access_index));
106 
107 struct mm_struct___new {
108 	struct rw_semaphore mmap_lock;
109 } __attribute__((preserve_access_index));
110 
111 /* control flags */
112 int enabled;
113 int has_cpu;
114 int has_task;
115 int has_type;
116 int has_addr;
117 int needs_callstack;
118 int stack_skip;
119 int lock_owner;
120 
121 /* determine the key of lock stat */
122 int aggr_mode;
123 
124 /* error stat */
125 int task_fail;
126 int stack_fail;
127 int time_fail;
128 int data_fail;
129 
130 int task_map_full;
131 int data_map_full;
132 
133 static inline int can_record(u64 *ctx)
134 {
135 	if (has_cpu) {
136 		__u32 cpu = bpf_get_smp_processor_id();
137 		__u8 *ok;
138 
139 		ok = bpf_map_lookup_elem(&cpu_filter, &cpu);
140 		if (!ok)
141 			return 0;
142 	}
143 
144 	if (has_task) {
145 		__u8 *ok;
146 		__u32 pid = bpf_get_current_pid_tgid();
147 
148 		ok = bpf_map_lookup_elem(&task_filter, &pid);
149 		if (!ok)
150 			return 0;
151 	}
152 
153 	if (has_type) {
154 		__u8 *ok;
155 		__u32 flags = (__u32)ctx[1];
156 
157 		ok = bpf_map_lookup_elem(&type_filter, &flags);
158 		if (!ok)
159 			return 0;
160 	}
161 
162 	if (has_addr) {
163 		__u8 *ok;
164 		__u64 addr = ctx[0];
165 
166 		ok = bpf_map_lookup_elem(&addr_filter, &addr);
167 		if (!ok)
168 			return 0;
169 	}
170 
171 	return 1;
172 }
173 
174 static inline int update_task_data(struct task_struct *task)
175 {
176 	struct contention_task_data *p;
177 	int pid, err;
178 
179 	err = bpf_core_read(&pid, sizeof(pid), &task->pid);
180 	if (err)
181 		return -1;
182 
183 	p = bpf_map_lookup_elem(&task_data, &pid);
184 	if (p == NULL && !task_map_full) {
185 		struct contention_task_data data = {};
186 
187 		BPF_CORE_READ_STR_INTO(&data.comm, task, comm);
188 		if (bpf_map_update_elem(&task_data, &pid, &data, BPF_NOEXIST) == -E2BIG)
189 			task_map_full = 1;
190 	}
191 
192 	return 0;
193 }
194 
195 #ifndef __has_builtin
196 # define __has_builtin(x) 0
197 #endif
198 
199 static inline struct task_struct *get_lock_owner(__u64 lock, __u32 flags)
200 {
201 	struct task_struct *task;
202 	__u64 owner = 0;
203 
204 	if (flags & LCB_F_MUTEX) {
205 		struct mutex *mutex = (void *)lock;
206 		owner = BPF_CORE_READ(mutex, owner.counter);
207 	} else if (flags == LCB_F_READ || flags == LCB_F_WRITE) {
208 	/*
209 	 * Support for the BPF_TYPE_MATCHES argument to the
210 	 * __builtin_preserve_type_info builtin was added at some point during
211 	 * development of clang 15 and it's what is needed for
212 	 * bpf_core_type_matches.
213 	 */
214 #if __has_builtin(__builtin_preserve_type_info) && __clang_major__ >= 15
215 		if (bpf_core_type_matches(struct rw_semaphore___old)) {
216 			struct rw_semaphore___old *rwsem = (void *)lock;
217 			owner = (unsigned long)BPF_CORE_READ(rwsem, owner);
218 		} else if (bpf_core_type_matches(struct rw_semaphore___new)) {
219 			struct rw_semaphore___new *rwsem = (void *)lock;
220 			owner = BPF_CORE_READ(rwsem, owner.counter);
221 		}
222 #else
223 		/* assume new struct */
224 		struct rw_semaphore *rwsem = (void *)lock;
225 		owner = BPF_CORE_READ(rwsem, owner.counter);
226 #endif
227 	}
228 
229 	if (!owner)
230 		return NULL;
231 
232 	task = (void *)(owner & ~7UL);
233 	return task;
234 }
235 
236 static inline __u32 check_lock_type(__u64 lock, __u32 flags)
237 {
238 	struct task_struct *curr;
239 	struct mm_struct___old *mm_old;
240 	struct mm_struct___new *mm_new;
241 
242 	switch (flags) {
243 	case LCB_F_READ:  /* rwsem */
244 	case LCB_F_WRITE:
245 		curr = bpf_get_current_task_btf();
246 		if (curr->mm == NULL)
247 			break;
248 		mm_new = (void *)curr->mm;
249 		if (bpf_core_field_exists(mm_new->mmap_lock)) {
250 			if (&mm_new->mmap_lock == (void *)lock)
251 				return LCD_F_MMAP_LOCK;
252 			break;
253 		}
254 		mm_old = (void *)curr->mm;
255 		if (bpf_core_field_exists(mm_old->mmap_sem)) {
256 			if (&mm_old->mmap_sem == (void *)lock)
257 				return LCD_F_MMAP_LOCK;
258 		}
259 		break;
260 	case LCB_F_SPIN:  /* spinlock */
261 		curr = bpf_get_current_task_btf();
262 		if (&curr->sighand->siglock == (void *)lock)
263 			return LCD_F_SIGHAND_LOCK;
264 		break;
265 	default:
266 		break;
267 	}
268 	return 0;
269 }
270 
271 SEC("tp_btf/contention_begin")
272 int contention_begin(u64 *ctx)
273 {
274 	__u32 pid;
275 	struct tstamp_data *pelem;
276 
277 	if (!enabled || !can_record(ctx))
278 		return 0;
279 
280 	pid = bpf_get_current_pid_tgid();
281 	pelem = bpf_map_lookup_elem(&tstamp, &pid);
282 	if (pelem && pelem->lock)
283 		return 0;
284 
285 	if (pelem == NULL) {
286 		struct tstamp_data zero = {};
287 
288 		bpf_map_update_elem(&tstamp, &pid, &zero, BPF_ANY);
289 		pelem = bpf_map_lookup_elem(&tstamp, &pid);
290 		if (pelem == NULL) {
291 			__sync_fetch_and_add(&task_fail, 1);
292 			return 0;
293 		}
294 	}
295 
296 	pelem->timestamp = bpf_ktime_get_ns();
297 	pelem->lock = (__u64)ctx[0];
298 	pelem->flags = (__u32)ctx[1];
299 
300 	if (needs_callstack) {
301 		pelem->stack_id = bpf_get_stackid(ctx, &stacks,
302 						  BPF_F_FAST_STACK_CMP | stack_skip);
303 		if (pelem->stack_id < 0)
304 			__sync_fetch_and_add(&stack_fail, 1);
305 	} else if (aggr_mode == LOCK_AGGR_TASK) {
306 		struct task_struct *task;
307 
308 		if (lock_owner) {
309 			task = get_lock_owner(pelem->lock, pelem->flags);
310 
311 			/* The flags is not used anymore.  Pass the owner pid. */
312 			if (task)
313 				pelem->flags = BPF_CORE_READ(task, pid);
314 			else
315 				pelem->flags = -1U;
316 
317 		} else {
318 			task = bpf_get_current_task_btf();
319 		}
320 
321 		if (task) {
322 			if (update_task_data(task) < 0 && lock_owner)
323 				pelem->flags = -1U;
324 		}
325 	}
326 
327 	return 0;
328 }
329 
330 SEC("tp_btf/contention_end")
331 int contention_end(u64 *ctx)
332 {
333 	__u32 pid;
334 	struct tstamp_data *pelem;
335 	struct contention_key key = {};
336 	struct contention_data *data;
337 	__u64 duration;
338 
339 	if (!enabled)
340 		return 0;
341 
342 	pid = bpf_get_current_pid_tgid();
343 	pelem = bpf_map_lookup_elem(&tstamp, &pid);
344 	if (!pelem || pelem->lock != ctx[0])
345 		return 0;
346 
347 	duration = bpf_ktime_get_ns() - pelem->timestamp;
348 	if ((__s64)duration < 0) {
349 		bpf_map_delete_elem(&tstamp, &pid);
350 		__sync_fetch_and_add(&time_fail, 1);
351 		return 0;
352 	}
353 
354 	switch (aggr_mode) {
355 	case LOCK_AGGR_CALLER:
356 		key.stack_id = pelem->stack_id;
357 		break;
358 	case LOCK_AGGR_TASK:
359 		if (lock_owner)
360 			key.pid = pelem->flags;
361 		else
362 			key.pid = pid;
363 		if (needs_callstack)
364 			key.stack_id = pelem->stack_id;
365 		break;
366 	case LOCK_AGGR_ADDR:
367 		key.lock_addr = pelem->lock;
368 		if (needs_callstack)
369 			key.stack_id = pelem->stack_id;
370 		break;
371 	default:
372 		/* should not happen */
373 		return 0;
374 	}
375 
376 	data = bpf_map_lookup_elem(&lock_stat, &key);
377 	if (!data) {
378 		if (data_map_full) {
379 			bpf_map_delete_elem(&tstamp, &pid);
380 			__sync_fetch_and_add(&data_fail, 1);
381 			return 0;
382 		}
383 
384 		struct contention_data first = {
385 			.total_time = duration,
386 			.max_time = duration,
387 			.min_time = duration,
388 			.count = 1,
389 			.flags = pelem->flags,
390 		};
391 		int err;
392 
393 		if (aggr_mode == LOCK_AGGR_ADDR)
394 			first.flags |= check_lock_type(pelem->lock, pelem->flags);
395 
396 		err = bpf_map_update_elem(&lock_stat, &key, &first, BPF_NOEXIST);
397 		if (err < 0) {
398 			if (err == -E2BIG)
399 				data_map_full = 1;
400 			__sync_fetch_and_add(&data_fail, 1);
401 		}
402 		bpf_map_delete_elem(&tstamp, &pid);
403 		return 0;
404 	}
405 
406 	__sync_fetch_and_add(&data->total_time, duration);
407 	__sync_fetch_and_add(&data->count, 1);
408 
409 	/* FIXME: need atomic operations */
410 	if (data->max_time < duration)
411 		data->max_time = duration;
412 	if (data->min_time > duration)
413 		data->min_time = duration;
414 
415 	bpf_map_delete_elem(&tstamp, &pid);
416 	return 0;
417 }
418 
419 struct rq {};
420 
421 extern struct rq runqueues __ksym;
422 
423 struct rq___old {
424 	raw_spinlock_t lock;
425 } __attribute__((preserve_access_index));
426 
427 struct rq___new {
428 	raw_spinlock_t __lock;
429 } __attribute__((preserve_access_index));
430 
431 SEC("raw_tp/bpf_test_finish")
432 int BPF_PROG(collect_lock_syms)
433 {
434 	__u64 lock_addr, lock_off;
435 	__u32 lock_flag;
436 
437 	if (bpf_core_field_exists(struct rq___new, __lock))
438 		lock_off = offsetof(struct rq___new, __lock);
439 	else
440 		lock_off = offsetof(struct rq___old, lock);
441 
442 	for (int i = 0; i < MAX_CPUS; i++) {
443 		struct rq *rq = bpf_per_cpu_ptr(&runqueues, i);
444 
445 		if (rq == NULL)
446 			break;
447 
448 		lock_addr = (__u64)(void *)rq + lock_off;
449 		lock_flag = LOCK_CLASS_RQLOCK;
450 		bpf_map_update_elem(&lock_syms, &lock_addr, &lock_flag, BPF_ANY);
451 	}
452 	return 0;
453 }
454 
455 char LICENSE[] SEC("license") = "Dual BSD/GPL";
456