1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * Augment the raw_syscalls tracepoints with the contents of the pointer arguments.
4 *
5 * This exactly matches what is marshalled into the raw_syscall:sys_enter
6 * payload expected by the 'perf trace' beautifiers.
7 */
8
9 #include <linux/bpf.h>
10 #include <bpf/bpf_helpers.h>
11 #include <linux/limits.h>
12
13 /**
14 * is_power_of_2() - check if a value is a power of two
15 * @n: the value to check
16 *
17 * Determine whether some value is a power of two, where zero is *not*
18 * considered a power of two. Return: true if @n is a power of 2, otherwise
19 * false.
20 */
21 #define is_power_of_2(n) (n != 0 && ((n & (n - 1)) == 0))
22
23 #define MAX_CPUS 4096
24
25 // FIXME: These should come from system headers
26 #ifndef bool
27 typedef char bool;
28 #endif
29 typedef int pid_t;
30 typedef long long int __s64;
31 typedef __s64 time64_t;
32
33 struct timespec64 {
34 time64_t tv_sec;
35 long int tv_nsec;
36 };
37
38 /* bpf-output associated map */
39 struct __augmented_syscalls__ {
40 __uint(type, BPF_MAP_TYPE_PERF_EVENT_ARRAY);
41 __type(key, int);
42 __type(value, __u32);
43 __uint(max_entries, MAX_CPUS);
44 } __augmented_syscalls__ SEC(".maps");
45
46 /*
47 * What to augment at entry?
48 *
49 * Pointer arg payloads (filenames, etc) passed from userspace to the kernel
50 */
51 struct syscalls_sys_enter {
52 __uint(type, BPF_MAP_TYPE_PROG_ARRAY);
53 __type(key, __u32);
54 __type(value, __u32);
55 __uint(max_entries, 512);
56 } syscalls_sys_enter SEC(".maps");
57
58 /*
59 * What to augment at exit?
60 *
61 * Pointer arg payloads returned from the kernel (struct stat, etc) to userspace.
62 */
63 struct syscalls_sys_exit {
64 __uint(type, BPF_MAP_TYPE_PROG_ARRAY);
65 __type(key, __u32);
66 __type(value, __u32);
67 __uint(max_entries, 512);
68 } syscalls_sys_exit SEC(".maps");
69
70 struct syscall_enter_args {
71 unsigned long long common_tp_fields;
72 long syscall_nr;
73 unsigned long args[6];
74 };
75
76 struct syscall_exit_args {
77 unsigned long long common_tp_fields;
78 long syscall_nr;
79 long ret;
80 };
81
82 struct augmented_arg {
83 unsigned int size;
84 int err;
85 char value[PATH_MAX];
86 };
87
88 struct pids_filtered {
89 __uint(type, BPF_MAP_TYPE_HASH);
90 __type(key, pid_t);
91 __type(value, bool);
92 __uint(max_entries, 64);
93 } pids_filtered SEC(".maps");
94
95 /*
96 * Desired design of maximum size and alignment (see RFC2553)
97 */
98 #define SS_MAXSIZE 128 /* Implementation specific max size */
99
100 typedef unsigned short sa_family_t;
101
102 /*
103 * FIXME: Should come from system headers
104 *
105 * The definition uses anonymous union and struct in order to control the
106 * default alignment.
107 */
108 struct sockaddr_storage {
109 union {
110 struct {
111 sa_family_t ss_family; /* address family */
112 /* Following field(s) are implementation specific */
113 char __data[SS_MAXSIZE - sizeof(unsigned short)];
114 /* space to achieve desired size, */
115 /* _SS_MAXSIZE value minus size of ss_family */
116 };
117 void *__align; /* implementation specific desired alignment */
118 };
119 };
120
121 struct augmented_args_payload {
122 struct syscall_enter_args args;
123 union {
124 struct {
125 struct augmented_arg arg, arg2;
126 };
127 struct sockaddr_storage saddr;
128 char __data[sizeof(struct augmented_arg)];
129 };
130 };
131
132 // We need more tmp space than the BPF stack can give us
133 struct augmented_args_tmp {
134 __uint(type, BPF_MAP_TYPE_PERCPU_ARRAY);
135 __type(key, int);
136 __type(value, struct augmented_args_payload);
137 __uint(max_entries, 1);
138 } augmented_args_tmp SEC(".maps");
139
augmented_args_payload(void)140 static inline struct augmented_args_payload *augmented_args_payload(void)
141 {
142 int key = 0;
143 return bpf_map_lookup_elem(&augmented_args_tmp, &key);
144 }
145
augmented__output(void * ctx,struct augmented_args_payload * args,int len)146 static inline int augmented__output(void *ctx, struct augmented_args_payload *args, int len)
147 {
148 /* If perf_event_output fails, return non-zero so that it gets recorded unaugmented */
149 return bpf_perf_event_output(ctx, &__augmented_syscalls__, BPF_F_CURRENT_CPU, args, len);
150 }
151
152 static inline
augmented_arg__read_str(struct augmented_arg * augmented_arg,const void * arg,unsigned int arg_len)153 unsigned int augmented_arg__read_str(struct augmented_arg *augmented_arg, const void *arg, unsigned int arg_len)
154 {
155 unsigned int augmented_len = sizeof(*augmented_arg);
156 int string_len = bpf_probe_read_user_str(&augmented_arg->value, arg_len, arg);
157
158 augmented_arg->size = augmented_arg->err = 0;
159 /*
160 * probe_read_str may return < 0, e.g. -EFAULT
161 * So we leave that in the augmented_arg->size that userspace will
162 */
163 if (string_len > 0) {
164 augmented_len -= sizeof(augmented_arg->value) - string_len;
165 _Static_assert(is_power_of_2(sizeof(augmented_arg->value)), "sizeof(augmented_arg->value) needs to be a power of two");
166 augmented_len &= sizeof(augmented_arg->value) - 1;
167 augmented_arg->size = string_len;
168 } else {
169 /*
170 * So that username notice the error while still being able
171 * to skip this augmented arg record
172 */
173 augmented_arg->err = string_len;
174 augmented_len = offsetof(struct augmented_arg, value);
175 }
176
177 return augmented_len;
178 }
179
180 SEC("tp/raw_syscalls/sys_enter")
syscall_unaugmented(struct syscall_enter_args * args)181 int syscall_unaugmented(struct syscall_enter_args *args)
182 {
183 return 1;
184 }
185
186 /*
187 * These will be tail_called from SEC("raw_syscalls:sys_enter"), so will find in
188 * augmented_args_tmp what was read by that raw_syscalls:sys_enter and go
189 * on from there, reading the first syscall arg as a string, i.e. open's
190 * filename.
191 */
192 SEC("tp/syscalls/sys_enter_connect")
sys_enter_connect(struct syscall_enter_args * args)193 int sys_enter_connect(struct syscall_enter_args *args)
194 {
195 struct augmented_args_payload *augmented_args = augmented_args_payload();
196 const void *sockaddr_arg = (const void *)args->args[1];
197 unsigned int socklen = args->args[2];
198 unsigned int len = sizeof(augmented_args->args);
199
200 if (augmented_args == NULL)
201 return 1; /* Failure: don't filter */
202
203 _Static_assert(is_power_of_2(sizeof(augmented_args->saddr)), "sizeof(augmented_args->saddr) needs to be a power of two");
204 socklen &= sizeof(augmented_args->saddr) - 1;
205
206 bpf_probe_read_user(&augmented_args->saddr, socklen, sockaddr_arg);
207
208 return augmented__output(args, augmented_args, len + socklen);
209 }
210
211 SEC("tp/syscalls/sys_enter_sendto")
sys_enter_sendto(struct syscall_enter_args * args)212 int sys_enter_sendto(struct syscall_enter_args *args)
213 {
214 struct augmented_args_payload *augmented_args = augmented_args_payload();
215 const void *sockaddr_arg = (const void *)args->args[4];
216 unsigned int socklen = args->args[5];
217 unsigned int len = sizeof(augmented_args->args);
218
219 if (augmented_args == NULL)
220 return 1; /* Failure: don't filter */
221
222 socklen &= sizeof(augmented_args->saddr) - 1;
223
224 bpf_probe_read_user(&augmented_args->saddr, socklen, sockaddr_arg);
225
226 return augmented__output(args, augmented_args, len + socklen);
227 }
228
229 SEC("tp/syscalls/sys_enter_open")
sys_enter_open(struct syscall_enter_args * args)230 int sys_enter_open(struct syscall_enter_args *args)
231 {
232 struct augmented_args_payload *augmented_args = augmented_args_payload();
233 const void *filename_arg = (const void *)args->args[0];
234 unsigned int len = sizeof(augmented_args->args);
235
236 if (augmented_args == NULL)
237 return 1; /* Failure: don't filter */
238
239 len += augmented_arg__read_str(&augmented_args->arg, filename_arg, sizeof(augmented_args->arg.value));
240
241 return augmented__output(args, augmented_args, len);
242 }
243
244 SEC("tp/syscalls/sys_enter_openat")
sys_enter_openat(struct syscall_enter_args * args)245 int sys_enter_openat(struct syscall_enter_args *args)
246 {
247 struct augmented_args_payload *augmented_args = augmented_args_payload();
248 const void *filename_arg = (const void *)args->args[1];
249 unsigned int len = sizeof(augmented_args->args);
250
251 if (augmented_args == NULL)
252 return 1; /* Failure: don't filter */
253
254 len += augmented_arg__read_str(&augmented_args->arg, filename_arg, sizeof(augmented_args->arg.value));
255
256 return augmented__output(args, augmented_args, len);
257 }
258
259 SEC("tp/syscalls/sys_enter_rename")
sys_enter_rename(struct syscall_enter_args * args)260 int sys_enter_rename(struct syscall_enter_args *args)
261 {
262 struct augmented_args_payload *augmented_args = augmented_args_payload();
263 const void *oldpath_arg = (const void *)args->args[0],
264 *newpath_arg = (const void *)args->args[1];
265 unsigned int len = sizeof(augmented_args->args), oldpath_len;
266
267 if (augmented_args == NULL)
268 return 1; /* Failure: don't filter */
269
270 oldpath_len = augmented_arg__read_str(&augmented_args->arg, oldpath_arg, sizeof(augmented_args->arg.value));
271 len += oldpath_len + augmented_arg__read_str((void *)(&augmented_args->arg) + oldpath_len, newpath_arg, sizeof(augmented_args->arg.value));
272
273 return augmented__output(args, augmented_args, len);
274 }
275
276 SEC("tp/syscalls/sys_enter_renameat")
sys_enter_renameat(struct syscall_enter_args * args)277 int sys_enter_renameat(struct syscall_enter_args *args)
278 {
279 struct augmented_args_payload *augmented_args = augmented_args_payload();
280 const void *oldpath_arg = (const void *)args->args[1],
281 *newpath_arg = (const void *)args->args[3];
282 unsigned int len = sizeof(augmented_args->args), oldpath_len;
283
284 if (augmented_args == NULL)
285 return 1; /* Failure: don't filter */
286
287 oldpath_len = augmented_arg__read_str(&augmented_args->arg, oldpath_arg, sizeof(augmented_args->arg.value));
288 len += oldpath_len + augmented_arg__read_str((void *)(&augmented_args->arg) + oldpath_len, newpath_arg, sizeof(augmented_args->arg.value));
289
290 return augmented__output(args, augmented_args, len);
291 }
292
293 #define PERF_ATTR_SIZE_VER0 64 /* sizeof first published struct */
294
295 // we need just the start, get the size to then copy it
296 struct perf_event_attr_size {
297 __u32 type;
298 /*
299 * Size of the attr structure, for fwd/bwd compat.
300 */
301 __u32 size;
302 };
303
304 SEC("tp/syscalls/sys_enter_perf_event_open")
sys_enter_perf_event_open(struct syscall_enter_args * args)305 int sys_enter_perf_event_open(struct syscall_enter_args *args)
306 {
307 struct augmented_args_payload *augmented_args = augmented_args_payload();
308 const struct perf_event_attr_size *attr = (const struct perf_event_attr_size *)args->args[0], *attr_read;
309 unsigned int len = sizeof(augmented_args->args);
310
311 if (augmented_args == NULL)
312 goto failure;
313
314 if (bpf_probe_read_user(&augmented_args->__data, sizeof(*attr), attr) < 0)
315 goto failure;
316
317 attr_read = (const struct perf_event_attr_size *)augmented_args->__data;
318
319 __u32 size = attr_read->size;
320
321 if (!size)
322 size = PERF_ATTR_SIZE_VER0;
323
324 if (size > sizeof(augmented_args->__data))
325 goto failure;
326
327 // Now that we read attr->size and tested it against the size limits, read it completely
328 if (bpf_probe_read_user(&augmented_args->__data, size, attr) < 0)
329 goto failure;
330
331 return augmented__output(args, augmented_args, len + size);
332 failure:
333 return 1; /* Failure: don't filter */
334 }
335
336 SEC("tp/syscalls/sys_enter_clock_nanosleep")
sys_enter_clock_nanosleep(struct syscall_enter_args * args)337 int sys_enter_clock_nanosleep(struct syscall_enter_args *args)
338 {
339 struct augmented_args_payload *augmented_args = augmented_args_payload();
340 const void *rqtp_arg = (const void *)args->args[2];
341 unsigned int len = sizeof(augmented_args->args);
342 __u32 size = sizeof(struct timespec64);
343
344 if (augmented_args == NULL)
345 goto failure;
346
347 if (size > sizeof(augmented_args->__data))
348 goto failure;
349
350 bpf_probe_read_user(&augmented_args->__data, size, rqtp_arg);
351
352 return augmented__output(args, augmented_args, len + size);
353 failure:
354 return 1; /* Failure: don't filter */
355 }
356
getpid(void)357 static pid_t getpid(void)
358 {
359 return bpf_get_current_pid_tgid();
360 }
361
pid_filter__has(struct pids_filtered * pids,pid_t pid)362 static bool pid_filter__has(struct pids_filtered *pids, pid_t pid)
363 {
364 return bpf_map_lookup_elem(pids, &pid) != NULL;
365 }
366
367 SEC("tp/raw_syscalls/sys_enter")
sys_enter(struct syscall_enter_args * args)368 int sys_enter(struct syscall_enter_args *args)
369 {
370 struct augmented_args_payload *augmented_args;
371 /*
372 * We start len, the amount of data that will be in the perf ring
373 * buffer, if this is not filtered out by one of pid_filter__has(),
374 * syscall->enabled, etc, with the non-augmented raw syscall payload,
375 * i.e. sizeof(augmented_args->args).
376 *
377 * We'll add to this as we add augmented syscalls right after that
378 * initial, non-augmented raw_syscalls:sys_enter payload.
379 */
380
381 if (pid_filter__has(&pids_filtered, getpid()))
382 return 0;
383
384 augmented_args = augmented_args_payload();
385 if (augmented_args == NULL)
386 return 1;
387
388 bpf_probe_read_kernel(&augmented_args->args, sizeof(augmented_args->args), args);
389
390 /*
391 * Jump to syscall specific augmenter, even if the default one,
392 * "!raw_syscalls:unaugmented" that will just return 1 to return the
393 * unaugmented tracepoint payload.
394 */
395 bpf_tail_call(args, &syscalls_sys_enter, augmented_args->args.syscall_nr);
396
397 // If not found on the PROG_ARRAY syscalls map, then we're filtering it:
398 return 0;
399 }
400
401 SEC("tp/raw_syscalls/sys_exit")
sys_exit(struct syscall_exit_args * args)402 int sys_exit(struct syscall_exit_args *args)
403 {
404 struct syscall_exit_args exit_args;
405
406 if (pid_filter__has(&pids_filtered, getpid()))
407 return 0;
408
409 bpf_probe_read_kernel(&exit_args, sizeof(exit_args), args);
410 /*
411 * Jump to syscall specific return augmenter, even if the default one,
412 * "!raw_syscalls:unaugmented" that will just return 1 to return the
413 * unaugmented tracepoint payload.
414 */
415 bpf_tail_call(args, &syscalls_sys_exit, exit_args.syscall_nr);
416 /*
417 * If not found on the PROG_ARRAY syscalls map, then we're filtering it:
418 */
419 return 0;
420 }
421
422 char _license[] SEC("license") = "GPL";
423