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_helpers.h"
 10
 11#define FUNCTION_NAME_LEN 64
 12#define FILE_NAME_LEN 128
 13#define TASK_COMM_LEN 16
 14
 15typedef struct {
 16	int PyThreadState_frame;
 17	int PyThreadState_thread;
 18	int PyFrameObject_back;
 19	int PyFrameObject_code;
 20	int PyFrameObject_lineno;
 21	int PyCodeObject_filename;
 22	int PyCodeObject_name;
 23	int String_data;
 24	int String_size;
 25} OffsetConfig;
 26
 27typedef struct {
 28	uintptr_t current_state_addr;
 29	uintptr_t tls_key_addr;
 30	OffsetConfig offsets;
 31	bool use_tls;
 32} PidData;
 33
 34typedef struct {
 35	uint32_t success;
 36} Stats;
 37
 38typedef struct {
 39	char name[FUNCTION_NAME_LEN];
 40	char file[FILE_NAME_LEN];
 41} Symbol;
 42
 43typedef struct {
 44	uint32_t pid;
 45	uint32_t tid;
 46	char comm[TASK_COMM_LEN];
 47	int32_t kernel_stack_id;
 48	int32_t user_stack_id;
 49	bool thread_current;
 50	bool pthread_match;
 51	bool stack_complete;
 52	int16_t stack_len;
 53	int32_t stack[STACK_MAX_LEN];
 54
 55	int has_meta;
 56	int metadata;
 57	char dummy_safeguard;
 58} Event;
 59
 60
 61typedef int pid_t;
 62
 63typedef struct {
 64	void* f_back; // PyFrameObject.f_back, previous frame
 65	void* f_code; // PyFrameObject.f_code, pointer to PyCodeObject
 66	void* co_filename; // PyCodeObject.co_filename
 67	void* co_name; // PyCodeObject.co_name
 68} FrameData;
 69
 70static __always_inline void *get_thread_state(void *tls_base, PidData *pidData)
 71{
 72	void* thread_state;
 73	int key;
 74
 75	bpf_probe_read(&key, sizeof(key), (void*)(long)pidData->tls_key_addr);
 76	bpf_probe_read(&thread_state, sizeof(thread_state),
 77		       tls_base + 0x310 + key * 0x10 + 0x08);
 78	return thread_state;
 79}
 80
 81static __always_inline bool get_frame_data(void *frame_ptr, PidData *pidData,
 82					   FrameData *frame, Symbol *symbol)
 83{
 84	// read data from PyFrameObject
 85	bpf_probe_read(&frame->f_back,
 86		       sizeof(frame->f_back),
 87		       frame_ptr + pidData->offsets.PyFrameObject_back);
 88	bpf_probe_read(&frame->f_code,
 89		       sizeof(frame->f_code),
 90		       frame_ptr + pidData->offsets.PyFrameObject_code);
 91
 92	// read data from PyCodeObject
 93	if (!frame->f_code)
 94		return false;
 95	bpf_probe_read(&frame->co_filename,
 96		       sizeof(frame->co_filename),
 97		       frame->f_code + pidData->offsets.PyCodeObject_filename);
 98	bpf_probe_read(&frame->co_name,
 99		       sizeof(frame->co_name),
100		       frame->f_code + pidData->offsets.PyCodeObject_name);
101	// read actual names into symbol
102	if (frame->co_filename)
103		bpf_probe_read_str(&symbol->file,
104				   sizeof(symbol->file),
105				   frame->co_filename + pidData->offsets.String_data);
106	if (frame->co_name)
107		bpf_probe_read_str(&symbol->name,
108				   sizeof(symbol->name),
109				   frame->co_name + pidData->offsets.String_data);
110	return true;
111}
112
113struct {
114	__uint(type, BPF_MAP_TYPE_HASH);
115	__uint(max_entries, 1);
116	__type(key, int);
117	__type(value, PidData);
118} pidmap SEC(".maps");
119
120struct {
121	__uint(type, BPF_MAP_TYPE_HASH);
122	__uint(max_entries, 1);
123	__type(key, int);
124	__type(value, Event);
125} eventmap SEC(".maps");
126
127struct {
128	__uint(type, BPF_MAP_TYPE_HASH);
129	__uint(max_entries, 1);
130	__type(key, Symbol);
131	__type(value, int);
132} symbolmap SEC(".maps");
133
134struct {
135	__uint(type, BPF_MAP_TYPE_ARRAY);
136	__uint(max_entries, 1);
137	__type(key, int);
138	__type(value, Stats);
139} statsmap SEC(".maps");
140
141struct {
142	__uint(type, BPF_MAP_TYPE_PERF_EVENT_ARRAY);
143	__uint(max_entries, 32);
144	__uint(key_size, sizeof(int));
145	__uint(value_size, sizeof(int));
146} perfmap SEC(".maps");
147
148struct {
149	__uint(type, BPF_MAP_TYPE_STACK_TRACE);
150	__uint(max_entries, 1000);
151	__uint(key_size, sizeof(int));
152	__uint(value_size, sizeof(long long) * 127);
153} stackmap SEC(".maps");
154
155static __always_inline int __on_event(struct pt_regs *ctx)
156{
157	uint64_t pid_tgid = bpf_get_current_pid_tgid();
158	pid_t pid = (pid_t)(pid_tgid >> 32);
159	PidData* pidData = bpf_map_lookup_elem(&pidmap, &pid);
160	if (!pidData)
161		return 0;
162
163	int zero = 0;
164	Event* event = bpf_map_lookup_elem(&eventmap, &zero);
165	if (!event)
166		return 0;
167
168	event->pid = pid;
169
170	event->tid = (pid_t)pid_tgid;
171	bpf_get_current_comm(&event->comm, sizeof(event->comm));
172
173	event->user_stack_id = bpf_get_stackid(ctx, &stackmap, BPF_F_USER_STACK);
174	event->kernel_stack_id = bpf_get_stackid(ctx, &stackmap, 0);
175
176	void* thread_state_current = (void*)0;
177	bpf_probe_read(&thread_state_current,
178		       sizeof(thread_state_current),
179		       (void*)(long)pidData->current_state_addr);
180
181	struct task_struct* task = (struct task_struct*)bpf_get_current_task();
182	void* tls_base = (void*)task;
183
184	void* thread_state = pidData->use_tls ? get_thread_state(tls_base, pidData)
185		: thread_state_current;
186	event->thread_current = thread_state == thread_state_current;
187
188	if (pidData->use_tls) {
189		uint64_t pthread_created;
190		uint64_t pthread_self;
191		bpf_probe_read(&pthread_self, sizeof(pthread_self), tls_base + 0x10);
192
193		bpf_probe_read(&pthread_created,
194			       sizeof(pthread_created),
195			       thread_state + pidData->offsets.PyThreadState_thread);
196		event->pthread_match = pthread_created == pthread_self;
197	} else {
198		event->pthread_match = 1;
199	}
200
201	if (event->pthread_match || !pidData->use_tls) {
202		void* frame_ptr;
203		FrameData frame;
204		Symbol sym = {};
205		int cur_cpu = bpf_get_smp_processor_id();
206
207		bpf_probe_read(&frame_ptr,
208			       sizeof(frame_ptr),
209			       thread_state + pidData->offsets.PyThreadState_frame);
210
211		int32_t* symbol_counter = bpf_map_lookup_elem(&symbolmap, &sym);
212		if (symbol_counter == NULL)
213			return 0;
214#ifdef NO_UNROLL
215#pragma clang loop unroll(disable)
216#else
217#pragma clang loop unroll(full)
218#endif
219		/* Unwind python stack */
220		for (int i = 0; i < STACK_MAX_LEN; ++i) {
221			if (frame_ptr && get_frame_data(frame_ptr, pidData, &frame, &sym)) {
222				int32_t new_symbol_id = *symbol_counter * 64 + cur_cpu;
223				int32_t *symbol_id = bpf_map_lookup_elem(&symbolmap, &sym);
224				if (!symbol_id) {
225					bpf_map_update_elem(&symbolmap, &sym, &zero, 0);
226					symbol_id = bpf_map_lookup_elem(&symbolmap, &sym);
227					if (!symbol_id)
228						return 0;
229				}
230				if (*symbol_id == new_symbol_id)
231					(*symbol_counter)++;
232				event->stack[i] = *symbol_id;
233				event->stack_len = i + 1;
234				frame_ptr = frame.f_back;
235			}
236		}
237		event->stack_complete = frame_ptr == NULL;
238	} else {
239		event->stack_complete = 1;
240	}
241
242	Stats* stats = bpf_map_lookup_elem(&statsmap, &zero);
243	if (stats)
244		stats->success++;
245
246	event->has_meta = 0;
247	bpf_perf_event_output(ctx, &perfmap, 0, event, offsetof(Event, metadata));
248	return 0;
249}
250
251SEC("raw_tracepoint/kfree_skb")
252int on_event(struct pt_regs* ctx)
253{
254	int i, ret = 0;
255	ret |= __on_event(ctx);
256	ret |= __on_event(ctx);
257	ret |= __on_event(ctx);
258	ret |= __on_event(ctx);
259	ret |= __on_event(ctx);
260	return ret;
261}
262
263char _license[] SEC("license") = "GPL";