Loading...
1// SPDX-License-Identifier: GPL-2.0
2// Copyright (c) 2019 Facebook
3#include <linux/sched.h>
4#include <linux/ptrace.h>
5#include <stdint.h>
6#include <stddef.h>
7#include <stdbool.h>
8#include <linux/bpf.h>
9#include <bpf/bpf_helpers.h>
10#include "bpf_misc.h"
11#include "bpf_compiler.h"
12
13#define FUNCTION_NAME_LEN 64
14#define FILE_NAME_LEN 128
15#define TASK_COMM_LEN 16
16
17typedef struct {
18 int PyThreadState_frame;
19 int PyThreadState_thread;
20 int PyFrameObject_back;
21 int PyFrameObject_code;
22 int PyFrameObject_lineno;
23 int PyCodeObject_filename;
24 int PyCodeObject_name;
25 int String_data;
26 int String_size;
27} OffsetConfig;
28
29typedef struct {
30 uintptr_t current_state_addr;
31 uintptr_t tls_key_addr;
32 OffsetConfig offsets;
33 bool use_tls;
34} PidData;
35
36typedef struct {
37 uint32_t success;
38} Stats;
39
40typedef struct {
41 char name[FUNCTION_NAME_LEN];
42 char file[FILE_NAME_LEN];
43} Symbol;
44
45typedef struct {
46 uint32_t pid;
47 uint32_t tid;
48 char comm[TASK_COMM_LEN];
49 int32_t kernel_stack_id;
50 int32_t user_stack_id;
51 bool thread_current;
52 bool pthread_match;
53 bool stack_complete;
54 int16_t stack_len;
55 int32_t stack[STACK_MAX_LEN];
56
57 int has_meta;
58 int metadata;
59 char dummy_safeguard;
60} Event;
61
62
63typedef int pid_t;
64
65typedef struct {
66 void* f_back; // PyFrameObject.f_back, previous frame
67 void* f_code; // PyFrameObject.f_code, pointer to PyCodeObject
68 void* co_filename; // PyCodeObject.co_filename
69 void* co_name; // PyCodeObject.co_name
70} FrameData;
71
72#ifdef SUBPROGS
73__noinline
74#else
75__always_inline
76#endif
77static void *get_thread_state(void *tls_base, PidData *pidData)
78{
79 void* thread_state;
80 int key;
81
82 bpf_probe_read_user(&key, sizeof(key), (void*)(long)pidData->tls_key_addr);
83 bpf_probe_read_user(&thread_state, sizeof(thread_state),
84 tls_base + 0x310 + key * 0x10 + 0x08);
85 return thread_state;
86}
87
88static __always_inline bool get_frame_data(void *frame_ptr, PidData *pidData,
89 FrameData *frame, Symbol *symbol)
90{
91 // read data from PyFrameObject
92 bpf_probe_read_user(&frame->f_back,
93 sizeof(frame->f_back),
94 frame_ptr + pidData->offsets.PyFrameObject_back);
95 bpf_probe_read_user(&frame->f_code,
96 sizeof(frame->f_code),
97 frame_ptr + pidData->offsets.PyFrameObject_code);
98
99 // read data from PyCodeObject
100 if (!frame->f_code)
101 return false;
102 bpf_probe_read_user(&frame->co_filename,
103 sizeof(frame->co_filename),
104 frame->f_code + pidData->offsets.PyCodeObject_filename);
105 bpf_probe_read_user(&frame->co_name,
106 sizeof(frame->co_name),
107 frame->f_code + pidData->offsets.PyCodeObject_name);
108 // read actual names into symbol
109 if (frame->co_filename)
110 bpf_probe_read_user_str(&symbol->file,
111 sizeof(symbol->file),
112 frame->co_filename +
113 pidData->offsets.String_data);
114 if (frame->co_name)
115 bpf_probe_read_user_str(&symbol->name,
116 sizeof(symbol->name),
117 frame->co_name +
118 pidData->offsets.String_data);
119 return true;
120}
121
122struct {
123 __uint(type, BPF_MAP_TYPE_HASH);
124 __uint(max_entries, 1);
125 __type(key, int);
126 __type(value, PidData);
127} pidmap SEC(".maps");
128
129struct {
130 __uint(type, BPF_MAP_TYPE_HASH);
131 __uint(max_entries, 1);
132 __type(key, int);
133 __type(value, Event);
134} eventmap SEC(".maps");
135
136struct {
137 __uint(type, BPF_MAP_TYPE_HASH);
138 __uint(max_entries, 1);
139 __type(key, Symbol);
140 __type(value, int);
141} symbolmap SEC(".maps");
142
143struct {
144 __uint(type, BPF_MAP_TYPE_ARRAY);
145 __uint(max_entries, 1);
146 __type(key, int);
147 __type(value, Stats);
148} statsmap SEC(".maps");
149
150struct {
151 __uint(type, BPF_MAP_TYPE_PERF_EVENT_ARRAY);
152 __uint(max_entries, 32);
153 __uint(key_size, sizeof(int));
154 __uint(value_size, sizeof(int));
155} perfmap SEC(".maps");
156
157struct {
158 __uint(type, BPF_MAP_TYPE_STACK_TRACE);
159 __uint(max_entries, 1000);
160 __uint(key_size, sizeof(int));
161 __uint(value_size, sizeof(long long) * 127);
162} stackmap SEC(".maps");
163
164#ifdef USE_BPF_LOOP
165struct process_frame_ctx {
166 int cur_cpu;
167 int32_t *symbol_counter;
168 void *frame_ptr;
169 FrameData *frame;
170 PidData *pidData;
171 Symbol *sym;
172 Event *event;
173 bool done;
174};
175
176static int process_frame_callback(__u32 i, struct process_frame_ctx *ctx)
177{
178 int zero = 0;
179 void *frame_ptr = ctx->frame_ptr;
180 PidData *pidData = ctx->pidData;
181 FrameData *frame = ctx->frame;
182 int32_t *symbol_counter = ctx->symbol_counter;
183 int cur_cpu = ctx->cur_cpu;
184 Event *event = ctx->event;
185 Symbol *sym = ctx->sym;
186
187 if (frame_ptr && get_frame_data(frame_ptr, pidData, frame, sym)) {
188 int32_t new_symbol_id = *symbol_counter * 64 + cur_cpu;
189 int32_t *symbol_id = bpf_map_lookup_elem(&symbolmap, sym);
190
191 if (!symbol_id) {
192 bpf_map_update_elem(&symbolmap, sym, &zero, 0);
193 symbol_id = bpf_map_lookup_elem(&symbolmap, sym);
194 if (!symbol_id) {
195 ctx->done = true;
196 return 1;
197 }
198 }
199 if (*symbol_id == new_symbol_id)
200 (*symbol_counter)++;
201
202 barrier_var(i);
203 if (i >= STACK_MAX_LEN)
204 return 1;
205
206 event->stack[i] = *symbol_id;
207
208 event->stack_len = i + 1;
209 frame_ptr = frame->f_back;
210 }
211 return 0;
212}
213#endif /* USE_BPF_LOOP */
214
215#ifdef GLOBAL_FUNC
216__noinline
217#elif defined(SUBPROGS)
218static __noinline
219#else
220static __always_inline
221#endif
222int __on_event(struct bpf_raw_tracepoint_args *ctx)
223{
224 uint64_t pid_tgid = bpf_get_current_pid_tgid();
225 pid_t pid = (pid_t)(pid_tgid >> 32);
226 PidData* pidData = bpf_map_lookup_elem(&pidmap, &pid);
227 if (!pidData)
228 return 0;
229
230 int zero = 0;
231 Event* event = bpf_map_lookup_elem(&eventmap, &zero);
232 if (!event)
233 return 0;
234
235 event->pid = pid;
236
237 event->tid = (pid_t)pid_tgid;
238 bpf_get_current_comm(&event->comm, sizeof(event->comm));
239
240 event->user_stack_id = bpf_get_stackid(ctx, &stackmap, BPF_F_USER_STACK);
241 event->kernel_stack_id = bpf_get_stackid(ctx, &stackmap, 0);
242
243 void* thread_state_current = (void*)0;
244 bpf_probe_read_user(&thread_state_current,
245 sizeof(thread_state_current),
246 (void*)(long)pidData->current_state_addr);
247
248 struct task_struct* task = (struct task_struct*)bpf_get_current_task();
249 void* tls_base = (void*)task;
250
251 void* thread_state = pidData->use_tls ? get_thread_state(tls_base, pidData)
252 : thread_state_current;
253 event->thread_current = thread_state == thread_state_current;
254
255 if (pidData->use_tls) {
256 uint64_t pthread_created;
257 uint64_t pthread_self;
258 bpf_probe_read_user(&pthread_self, sizeof(pthread_self),
259 tls_base + 0x10);
260
261 bpf_probe_read_user(&pthread_created,
262 sizeof(pthread_created),
263 thread_state +
264 pidData->offsets.PyThreadState_thread);
265 event->pthread_match = pthread_created == pthread_self;
266 } else {
267 event->pthread_match = 1;
268 }
269
270 if (event->pthread_match || !pidData->use_tls) {
271 void* frame_ptr;
272 FrameData frame;
273 Symbol sym = {};
274 int cur_cpu = bpf_get_smp_processor_id();
275
276 bpf_probe_read_user(&frame_ptr,
277 sizeof(frame_ptr),
278 thread_state +
279 pidData->offsets.PyThreadState_frame);
280
281 int32_t* symbol_counter = bpf_map_lookup_elem(&symbolmap, &sym);
282 if (symbol_counter == NULL)
283 return 0;
284#ifdef USE_BPF_LOOP
285 struct process_frame_ctx ctx = {
286 .cur_cpu = cur_cpu,
287 .symbol_counter = symbol_counter,
288 .frame_ptr = frame_ptr,
289 .frame = &frame,
290 .pidData = pidData,
291 .sym = &sym,
292 .event = event,
293 };
294
295 bpf_loop(STACK_MAX_LEN, process_frame_callback, &ctx, 0);
296 if (ctx.done)
297 return 0;
298#else
299#if defined(USE_ITER)
300/* no for loop, no unrolling */
301#elif defined(NO_UNROLL)
302 __pragma_loop_no_unroll
303#elif defined(UNROLL_COUNT)
304 __pragma_loop_unroll_count(UNROLL_COUNT)
305#else
306 __pragma_loop_unroll_full
307#endif /* NO_UNROLL */
308 /* Unwind python stack */
309#ifdef USE_ITER
310 int i;
311 bpf_for(i, 0, STACK_MAX_LEN) {
312#else /* !USE_ITER */
313 for (int i = 0; i < STACK_MAX_LEN; ++i) {
314#endif
315 if (frame_ptr && get_frame_data(frame_ptr, pidData, &frame, &sym)) {
316 int32_t new_symbol_id = *symbol_counter * 64 + cur_cpu;
317 int32_t *symbol_id = bpf_map_lookup_elem(&symbolmap, &sym);
318 if (!symbol_id) {
319 bpf_map_update_elem(&symbolmap, &sym, &zero, 0);
320 symbol_id = bpf_map_lookup_elem(&symbolmap, &sym);
321 if (!symbol_id)
322 return 0;
323 }
324 if (*symbol_id == new_symbol_id)
325 (*symbol_counter)++;
326 event->stack[i] = *symbol_id;
327 event->stack_len = i + 1;
328 frame_ptr = frame.f_back;
329 }
330 }
331#endif /* USE_BPF_LOOP */
332 event->stack_complete = frame_ptr == NULL;
333 } else {
334 event->stack_complete = 1;
335 }
336
337 Stats* stats = bpf_map_lookup_elem(&statsmap, &zero);
338 if (stats)
339 stats->success++;
340
341 event->has_meta = 0;
342 bpf_perf_event_output(ctx, &perfmap, 0, event, offsetof(Event, metadata));
343 return 0;
344}
345
346SEC("raw_tracepoint/kfree_skb")
347int on_event(struct bpf_raw_tracepoint_args* ctx)
348{
349 int ret = 0;
350 ret |= __on_event(ctx);
351 ret |= __on_event(ctx);
352 ret |= __on_event(ctx);
353 ret |= __on_event(ctx);
354 ret |= __on_event(ctx);
355 return ret;
356}
357
358char _license[] SEC("license") = "GPL";
1// SPDX-License-Identifier: GPL-2.0
2// Copyright (c) 2019 Facebook
3#include <linux/sched.h>
4#include <linux/ptrace.h>
5#include <stdint.h>
6#include <stddef.h>
7#include <stdbool.h>
8#include <linux/bpf.h>
9#include <bpf/bpf_helpers.h>
10#include "bpf_misc.h"
11
12#define FUNCTION_NAME_LEN 64
13#define FILE_NAME_LEN 128
14#define TASK_COMM_LEN 16
15
16typedef struct {
17 int PyThreadState_frame;
18 int PyThreadState_thread;
19 int PyFrameObject_back;
20 int PyFrameObject_code;
21 int PyFrameObject_lineno;
22 int PyCodeObject_filename;
23 int PyCodeObject_name;
24 int String_data;
25 int String_size;
26} OffsetConfig;
27
28typedef struct {
29 uintptr_t current_state_addr;
30 uintptr_t tls_key_addr;
31 OffsetConfig offsets;
32 bool use_tls;
33} PidData;
34
35typedef struct {
36 uint32_t success;
37} Stats;
38
39typedef struct {
40 char name[FUNCTION_NAME_LEN];
41 char file[FILE_NAME_LEN];
42} Symbol;
43
44typedef struct {
45 uint32_t pid;
46 uint32_t tid;
47 char comm[TASK_COMM_LEN];
48 int32_t kernel_stack_id;
49 int32_t user_stack_id;
50 bool thread_current;
51 bool pthread_match;
52 bool stack_complete;
53 int16_t stack_len;
54 int32_t stack[STACK_MAX_LEN];
55
56 int has_meta;
57 int metadata;
58 char dummy_safeguard;
59} Event;
60
61
62typedef int pid_t;
63
64typedef struct {
65 void* f_back; // PyFrameObject.f_back, previous frame
66 void* f_code; // PyFrameObject.f_code, pointer to PyCodeObject
67 void* co_filename; // PyCodeObject.co_filename
68 void* co_name; // PyCodeObject.co_name
69} FrameData;
70
71#ifdef SUBPROGS
72__noinline
73#else
74__always_inline
75#endif
76static void *get_thread_state(void *tls_base, PidData *pidData)
77{
78 void* thread_state;
79 int key;
80
81 bpf_probe_read_user(&key, sizeof(key), (void*)(long)pidData->tls_key_addr);
82 bpf_probe_read_user(&thread_state, sizeof(thread_state),
83 tls_base + 0x310 + key * 0x10 + 0x08);
84 return thread_state;
85}
86
87static __always_inline bool get_frame_data(void *frame_ptr, PidData *pidData,
88 FrameData *frame, Symbol *symbol)
89{
90 // read data from PyFrameObject
91 bpf_probe_read_user(&frame->f_back,
92 sizeof(frame->f_back),
93 frame_ptr + pidData->offsets.PyFrameObject_back);
94 bpf_probe_read_user(&frame->f_code,
95 sizeof(frame->f_code),
96 frame_ptr + pidData->offsets.PyFrameObject_code);
97
98 // read data from PyCodeObject
99 if (!frame->f_code)
100 return false;
101 bpf_probe_read_user(&frame->co_filename,
102 sizeof(frame->co_filename),
103 frame->f_code + pidData->offsets.PyCodeObject_filename);
104 bpf_probe_read_user(&frame->co_name,
105 sizeof(frame->co_name),
106 frame->f_code + pidData->offsets.PyCodeObject_name);
107 // read actual names into symbol
108 if (frame->co_filename)
109 bpf_probe_read_user_str(&symbol->file,
110 sizeof(symbol->file),
111 frame->co_filename +
112 pidData->offsets.String_data);
113 if (frame->co_name)
114 bpf_probe_read_user_str(&symbol->name,
115 sizeof(symbol->name),
116 frame->co_name +
117 pidData->offsets.String_data);
118 return true;
119}
120
121struct {
122 __uint(type, BPF_MAP_TYPE_HASH);
123 __uint(max_entries, 1);
124 __type(key, int);
125 __type(value, PidData);
126} pidmap SEC(".maps");
127
128struct {
129 __uint(type, BPF_MAP_TYPE_HASH);
130 __uint(max_entries, 1);
131 __type(key, int);
132 __type(value, Event);
133} eventmap SEC(".maps");
134
135struct {
136 __uint(type, BPF_MAP_TYPE_HASH);
137 __uint(max_entries, 1);
138 __type(key, Symbol);
139 __type(value, int);
140} symbolmap SEC(".maps");
141
142struct {
143 __uint(type, BPF_MAP_TYPE_ARRAY);
144 __uint(max_entries, 1);
145 __type(key, int);
146 __type(value, Stats);
147} statsmap SEC(".maps");
148
149struct {
150 __uint(type, BPF_MAP_TYPE_PERF_EVENT_ARRAY);
151 __uint(max_entries, 32);
152 __uint(key_size, sizeof(int));
153 __uint(value_size, sizeof(int));
154} perfmap SEC(".maps");
155
156struct {
157 __uint(type, BPF_MAP_TYPE_STACK_TRACE);
158 __uint(max_entries, 1000);
159 __uint(key_size, sizeof(int));
160 __uint(value_size, sizeof(long long) * 127);
161} stackmap SEC(".maps");
162
163#ifdef USE_BPF_LOOP
164struct process_frame_ctx {
165 int cur_cpu;
166 int32_t *symbol_counter;
167 void *frame_ptr;
168 FrameData *frame;
169 PidData *pidData;
170 Symbol *sym;
171 Event *event;
172 bool done;
173};
174
175static int process_frame_callback(__u32 i, struct process_frame_ctx *ctx)
176{
177 int zero = 0;
178 void *frame_ptr = ctx->frame_ptr;
179 PidData *pidData = ctx->pidData;
180 FrameData *frame = ctx->frame;
181 int32_t *symbol_counter = ctx->symbol_counter;
182 int cur_cpu = ctx->cur_cpu;
183 Event *event = ctx->event;
184 Symbol *sym = ctx->sym;
185
186 if (frame_ptr && get_frame_data(frame_ptr, pidData, frame, sym)) {
187 int32_t new_symbol_id = *symbol_counter * 64 + cur_cpu;
188 int32_t *symbol_id = bpf_map_lookup_elem(&symbolmap, sym);
189
190 if (!symbol_id) {
191 bpf_map_update_elem(&symbolmap, sym, &zero, 0);
192 symbol_id = bpf_map_lookup_elem(&symbolmap, sym);
193 if (!symbol_id) {
194 ctx->done = true;
195 return 1;
196 }
197 }
198 if (*symbol_id == new_symbol_id)
199 (*symbol_counter)++;
200
201 barrier_var(i);
202 if (i >= STACK_MAX_LEN)
203 return 1;
204
205 event->stack[i] = *symbol_id;
206
207 event->stack_len = i + 1;
208 frame_ptr = frame->f_back;
209 }
210 return 0;
211}
212#endif /* USE_BPF_LOOP */
213
214#ifdef GLOBAL_FUNC
215__noinline
216#elif defined(SUBPROGS)
217static __noinline
218#else
219static __always_inline
220#endif
221int __on_event(struct bpf_raw_tracepoint_args *ctx)
222{
223 uint64_t pid_tgid = bpf_get_current_pid_tgid();
224 pid_t pid = (pid_t)(pid_tgid >> 32);
225 PidData* pidData = bpf_map_lookup_elem(&pidmap, &pid);
226 if (!pidData)
227 return 0;
228
229 int zero = 0;
230 Event* event = bpf_map_lookup_elem(&eventmap, &zero);
231 if (!event)
232 return 0;
233
234 event->pid = pid;
235
236 event->tid = (pid_t)pid_tgid;
237 bpf_get_current_comm(&event->comm, sizeof(event->comm));
238
239 event->user_stack_id = bpf_get_stackid(ctx, &stackmap, BPF_F_USER_STACK);
240 event->kernel_stack_id = bpf_get_stackid(ctx, &stackmap, 0);
241
242 void* thread_state_current = (void*)0;
243 bpf_probe_read_user(&thread_state_current,
244 sizeof(thread_state_current),
245 (void*)(long)pidData->current_state_addr);
246
247 struct task_struct* task = (struct task_struct*)bpf_get_current_task();
248 void* tls_base = (void*)task;
249
250 void* thread_state = pidData->use_tls ? get_thread_state(tls_base, pidData)
251 : thread_state_current;
252 event->thread_current = thread_state == thread_state_current;
253
254 if (pidData->use_tls) {
255 uint64_t pthread_created;
256 uint64_t pthread_self;
257 bpf_probe_read_user(&pthread_self, sizeof(pthread_self),
258 tls_base + 0x10);
259
260 bpf_probe_read_user(&pthread_created,
261 sizeof(pthread_created),
262 thread_state +
263 pidData->offsets.PyThreadState_thread);
264 event->pthread_match = pthread_created == pthread_self;
265 } else {
266 event->pthread_match = 1;
267 }
268
269 if (event->pthread_match || !pidData->use_tls) {
270 void* frame_ptr;
271 FrameData frame;
272 Symbol sym = {};
273 int cur_cpu = bpf_get_smp_processor_id();
274
275 bpf_probe_read_user(&frame_ptr,
276 sizeof(frame_ptr),
277 thread_state +
278 pidData->offsets.PyThreadState_frame);
279
280 int32_t* symbol_counter = bpf_map_lookup_elem(&symbolmap, &sym);
281 if (symbol_counter == NULL)
282 return 0;
283#ifdef USE_BPF_LOOP
284 struct process_frame_ctx ctx = {
285 .cur_cpu = cur_cpu,
286 .symbol_counter = symbol_counter,
287 .frame_ptr = frame_ptr,
288 .frame = &frame,
289 .pidData = pidData,
290 .sym = &sym,
291 .event = event,
292 };
293
294 bpf_loop(STACK_MAX_LEN, process_frame_callback, &ctx, 0);
295 if (ctx.done)
296 return 0;
297#else
298#if defined(USE_ITER)
299/* no for loop, no unrolling */
300#elif defined(NO_UNROLL)
301#pragma clang loop unroll(disable)
302#elif defined(UNROLL_COUNT)
303#pragma clang loop unroll_count(UNROLL_COUNT)
304#else
305#pragma clang loop unroll(full)
306#endif /* NO_UNROLL */
307 /* Unwind python stack */
308#ifdef USE_ITER
309 int i;
310 bpf_for(i, 0, STACK_MAX_LEN) {
311#else /* !USE_ITER */
312 for (int i = 0; i < STACK_MAX_LEN; ++i) {
313#endif
314 if (frame_ptr && get_frame_data(frame_ptr, pidData, &frame, &sym)) {
315 int32_t new_symbol_id = *symbol_counter * 64 + cur_cpu;
316 int32_t *symbol_id = bpf_map_lookup_elem(&symbolmap, &sym);
317 if (!symbol_id) {
318 bpf_map_update_elem(&symbolmap, &sym, &zero, 0);
319 symbol_id = bpf_map_lookup_elem(&symbolmap, &sym);
320 if (!symbol_id)
321 return 0;
322 }
323 if (*symbol_id == new_symbol_id)
324 (*symbol_counter)++;
325 event->stack[i] = *symbol_id;
326 event->stack_len = i + 1;
327 frame_ptr = frame.f_back;
328 }
329 }
330#endif /* USE_BPF_LOOP */
331 event->stack_complete = frame_ptr == NULL;
332 } else {
333 event->stack_complete = 1;
334 }
335
336 Stats* stats = bpf_map_lookup_elem(&statsmap, &zero);
337 if (stats)
338 stats->success++;
339
340 event->has_meta = 0;
341 bpf_perf_event_output(ctx, &perfmap, 0, event, offsetof(Event, metadata));
342 return 0;
343}
344
345SEC("raw_tracepoint/kfree_skb")
346int on_event(struct bpf_raw_tracepoint_args* ctx)
347{
348 int ret = 0;
349 ret |= __on_event(ctx);
350 ret |= __on_event(ctx);
351 ret |= __on_event(ctx);
352 ret |= __on_event(ctx);
353 ret |= __on_event(ctx);
354 return ret;
355}
356
357char _license[] SEC("license") = "GPL";