1 /* SPDX-License-Identifier: LGPL-2.1 OR MIT */
2 /*
3 * rseq.h
4 *
5 * (C) Copyright 2016-2018 - Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
6 */
7
8 #ifndef RSEQ_H
9 #define RSEQ_H
10
11 #include <stdint.h>
12 #include <stdbool.h>
13 #include <pthread.h>
14 #include <signal.h>
15 #include <sched.h>
16 #include <errno.h>
17 #include <stdio.h>
18 #include <stdlib.h>
19 #include <stddef.h>
20 #include "rseq-abi.h"
21 #include "compiler.h"
22
23 #ifndef rseq_sizeof_field
24 #define rseq_sizeof_field(TYPE, MEMBER) sizeof((((TYPE *)0)->MEMBER))
25 #endif
26
27 #ifndef rseq_offsetofend
28 #define rseq_offsetofend(TYPE, MEMBER) \
29 (offsetof(TYPE, MEMBER) + rseq_sizeof_field(TYPE, MEMBER))
30 #endif
31
32 /*
33 * Empty code injection macros, override when testing.
34 * It is important to consider that the ASM injection macros need to be
35 * fully reentrant (e.g. do not modify the stack).
36 */
37 #ifndef RSEQ_INJECT_ASM
38 #define RSEQ_INJECT_ASM(n)
39 #endif
40
41 #ifndef RSEQ_INJECT_C
42 #define RSEQ_INJECT_C(n)
43 #endif
44
45 #ifndef RSEQ_INJECT_INPUT
46 #define RSEQ_INJECT_INPUT
47 #endif
48
49 #ifndef RSEQ_INJECT_CLOBBER
50 #define RSEQ_INJECT_CLOBBER
51 #endif
52
53 #ifndef RSEQ_INJECT_FAILED
54 #define RSEQ_INJECT_FAILED
55 #endif
56
57 #include "rseq-thread-pointer.h"
58
59 /* Offset from the thread pointer to the rseq area. */
60 extern ptrdiff_t rseq_offset;
61
62 /*
63 * The rseq ABI is composed of extensible feature fields. The extensions
64 * are done by appending additional fields at the end of the structure.
65 * The rseq_size defines the size of the active feature set which can be
66 * used by the application for the current rseq registration. Features
67 * starting at offset >= rseq_size are inactive and should not be used.
68 *
69 * The rseq_size is the intersection between the available allocation
70 * size for the rseq area and the feature size supported by the kernel.
71 * unsuccessful.
72 */
73 extern unsigned int rseq_size;
74
75 /* Flags used during rseq registration. */
76 extern unsigned int rseq_flags;
77
78 enum rseq_mo {
79 RSEQ_MO_RELAXED = 0,
80 RSEQ_MO_CONSUME = 1, /* Unused */
81 RSEQ_MO_ACQUIRE = 2, /* Unused */
82 RSEQ_MO_RELEASE = 3,
83 RSEQ_MO_ACQ_REL = 4, /* Unused */
84 RSEQ_MO_SEQ_CST = 5, /* Unused */
85 };
86
87 enum rseq_percpu_mode {
88 RSEQ_PERCPU_CPU_ID = 0,
89 RSEQ_PERCPU_MM_CID = 1,
90 };
91
rseq_get_abi(void)92 static inline struct rseq_abi *rseq_get_abi(void)
93 {
94 return (struct rseq_abi *) ((uintptr_t) rseq_thread_pointer() + rseq_offset);
95 }
96
97 #define rseq_likely(x) __builtin_expect(!!(x), 1)
98 #define rseq_unlikely(x) __builtin_expect(!!(x), 0)
99 #define rseq_barrier() __asm__ __volatile__("" : : : "memory")
100
101 #define RSEQ_ACCESS_ONCE(x) (*(__volatile__ __typeof__(x) *)&(x))
102 #define RSEQ_WRITE_ONCE(x, v) __extension__ ({ RSEQ_ACCESS_ONCE(x) = (v); })
103 #define RSEQ_READ_ONCE(x) RSEQ_ACCESS_ONCE(x)
104
105 #define __rseq_str_1(x) #x
106 #define __rseq_str(x) __rseq_str_1(x)
107
108 #define rseq_log(fmt, args...) \
109 fprintf(stderr, fmt "(in %s() at " __FILE__ ":" __rseq_str(__LINE__)"\n", \
110 ## args, __func__)
111
112 #define rseq_bug(fmt, args...) \
113 do { \
114 rseq_log(fmt, ##args); \
115 abort(); \
116 } while (0)
117
118 #if defined(__x86_64__) || defined(__i386__)
119 #include <rseq-x86.h>
120 #elif defined(__ARMEL__)
121 #include <rseq-arm.h>
122 #elif defined (__AARCH64EL__)
123 #include <rseq-arm64.h>
124 #elif defined(__PPC__)
125 #include <rseq-ppc.h>
126 #elif defined(__mips__)
127 #include <rseq-mips.h>
128 #elif defined(__s390__)
129 #include <rseq-s390.h>
130 #elif defined(__riscv)
131 #include <rseq-riscv.h>
132 #elif defined(__or1k__)
133 #include <rseq-or1k.h>
134 #else
135 #error unsupported target
136 #endif
137
138 /*
139 * Register rseq for the current thread. This needs to be called once
140 * by any thread which uses restartable sequences, before they start
141 * using restartable sequences, to ensure restartable sequences
142 * succeed. A restartable sequence executed from a non-registered
143 * thread will always fail.
144 */
145 int rseq_register_current_thread(void);
146
147 /*
148 * Unregister rseq for current thread.
149 */
150 int rseq_unregister_current_thread(void);
151
152 /*
153 * Restartable sequence fallback for reading the current CPU number.
154 */
155 int32_t rseq_fallback_current_cpu(void);
156
157 /*
158 * Restartable sequence fallback for reading the current node number.
159 */
160 int32_t rseq_fallback_current_node(void);
161
162 /*
163 * Returns true if rseq is supported.
164 */
165 bool rseq_available(void);
166
167 /*
168 * Values returned can be either the current CPU number, -1 (rseq is
169 * uninitialized), or -2 (rseq initialization has failed).
170 */
rseq_current_cpu_raw(void)171 static inline int32_t rseq_current_cpu_raw(void)
172 {
173 return RSEQ_ACCESS_ONCE(rseq_get_abi()->cpu_id);
174 }
175
176 /*
177 * Returns a possible CPU number, which is typically the current CPU.
178 * The returned CPU number can be used to prepare for an rseq critical
179 * section, which will confirm whether the cpu number is indeed the
180 * current one, and whether rseq is initialized.
181 *
182 * The CPU number returned by rseq_cpu_start should always be validated
183 * by passing it to a rseq asm sequence, or by comparing it to the
184 * return value of rseq_current_cpu_raw() if the rseq asm sequence
185 * does not need to be invoked.
186 */
rseq_cpu_start(void)187 static inline uint32_t rseq_cpu_start(void)
188 {
189 return RSEQ_ACCESS_ONCE(rseq_get_abi()->cpu_id_start);
190 }
191
rseq_current_cpu(void)192 static inline uint32_t rseq_current_cpu(void)
193 {
194 int32_t cpu;
195
196 cpu = rseq_current_cpu_raw();
197 if (rseq_unlikely(cpu < 0))
198 cpu = rseq_fallback_current_cpu();
199 return cpu;
200 }
201
rseq_node_id_available(void)202 static inline bool rseq_node_id_available(void)
203 {
204 return (int) rseq_size >= rseq_offsetofend(struct rseq_abi, node_id);
205 }
206
207 /*
208 * Current NUMA node number.
209 */
rseq_current_node_id(void)210 static inline uint32_t rseq_current_node_id(void)
211 {
212 assert(rseq_node_id_available());
213 return RSEQ_ACCESS_ONCE(rseq_get_abi()->node_id);
214 }
215
rseq_mm_cid_available(void)216 static inline bool rseq_mm_cid_available(void)
217 {
218 return (int) rseq_size >= rseq_offsetofend(struct rseq_abi, mm_cid);
219 }
220
rseq_current_mm_cid(void)221 static inline uint32_t rseq_current_mm_cid(void)
222 {
223 return RSEQ_ACCESS_ONCE(rseq_get_abi()->mm_cid);
224 }
225
rseq_clear_rseq_cs(void)226 static inline void rseq_clear_rseq_cs(void)
227 {
228 RSEQ_WRITE_ONCE(rseq_get_abi()->rseq_cs.arch.ptr, 0);
229 }
230
231 /*
232 * rseq_prepare_unload() should be invoked by each thread executing a rseq
233 * critical section at least once between their last critical section and
234 * library unload of the library defining the rseq critical section (struct
235 * rseq_cs) or the code referred to by the struct rseq_cs start_ip and
236 * post_commit_offset fields. This also applies to use of rseq in code
237 * generated by JIT: rseq_prepare_unload() should be invoked at least once by
238 * each thread executing a rseq critical section before reclaim of the memory
239 * holding the struct rseq_cs or reclaim of the code pointed to by struct
240 * rseq_cs start_ip and post_commit_offset fields.
241 */
rseq_prepare_unload(void)242 static inline void rseq_prepare_unload(void)
243 {
244 rseq_clear_rseq_cs();
245 }
246
247 static inline __attribute__((always_inline))
rseq_cmpeqv_storev(enum rseq_mo rseq_mo,enum rseq_percpu_mode percpu_mode,intptr_t * v,intptr_t expect,intptr_t newv,int cpu)248 int rseq_cmpeqv_storev(enum rseq_mo rseq_mo, enum rseq_percpu_mode percpu_mode,
249 intptr_t *v, intptr_t expect,
250 intptr_t newv, int cpu)
251 {
252 if (rseq_mo != RSEQ_MO_RELAXED)
253 return -1;
254 switch (percpu_mode) {
255 case RSEQ_PERCPU_CPU_ID:
256 return rseq_cmpeqv_storev_relaxed_cpu_id(v, expect, newv, cpu);
257 case RSEQ_PERCPU_MM_CID:
258 return rseq_cmpeqv_storev_relaxed_mm_cid(v, expect, newv, cpu);
259 }
260 return -1;
261 }
262
263 /*
264 * Compare @v against @expectnot. When it does _not_ match, load @v
265 * into @load, and store the content of *@v + voffp into @v.
266 */
267 static inline __attribute__((always_inline))
rseq_cmpnev_storeoffp_load(enum rseq_mo rseq_mo,enum rseq_percpu_mode percpu_mode,intptr_t * v,intptr_t expectnot,long voffp,intptr_t * load,int cpu)268 int rseq_cmpnev_storeoffp_load(enum rseq_mo rseq_mo, enum rseq_percpu_mode percpu_mode,
269 intptr_t *v, intptr_t expectnot, long voffp, intptr_t *load,
270 int cpu)
271 {
272 if (rseq_mo != RSEQ_MO_RELAXED)
273 return -1;
274 switch (percpu_mode) {
275 case RSEQ_PERCPU_CPU_ID:
276 return rseq_cmpnev_storeoffp_load_relaxed_cpu_id(v, expectnot, voffp, load, cpu);
277 case RSEQ_PERCPU_MM_CID:
278 return rseq_cmpnev_storeoffp_load_relaxed_mm_cid(v, expectnot, voffp, load, cpu);
279 }
280 return -1;
281 }
282
283 static inline __attribute__((always_inline))
rseq_addv(enum rseq_mo rseq_mo,enum rseq_percpu_mode percpu_mode,intptr_t * v,intptr_t count,int cpu)284 int rseq_addv(enum rseq_mo rseq_mo, enum rseq_percpu_mode percpu_mode,
285 intptr_t *v, intptr_t count, int cpu)
286 {
287 if (rseq_mo != RSEQ_MO_RELAXED)
288 return -1;
289 switch (percpu_mode) {
290 case RSEQ_PERCPU_CPU_ID:
291 return rseq_addv_relaxed_cpu_id(v, count, cpu);
292 case RSEQ_PERCPU_MM_CID:
293 return rseq_addv_relaxed_mm_cid(v, count, cpu);
294 }
295 return -1;
296 }
297
298 #ifdef RSEQ_ARCH_HAS_OFFSET_DEREF_ADDV
299 /*
300 * pval = *(ptr+off)
301 * *pval += inc;
302 */
303 static inline __attribute__((always_inline))
rseq_offset_deref_addv(enum rseq_mo rseq_mo,enum rseq_percpu_mode percpu_mode,intptr_t * ptr,long off,intptr_t inc,int cpu)304 int rseq_offset_deref_addv(enum rseq_mo rseq_mo, enum rseq_percpu_mode percpu_mode,
305 intptr_t *ptr, long off, intptr_t inc, int cpu)
306 {
307 if (rseq_mo != RSEQ_MO_RELAXED)
308 return -1;
309 switch (percpu_mode) {
310 case RSEQ_PERCPU_CPU_ID:
311 return rseq_offset_deref_addv_relaxed_cpu_id(ptr, off, inc, cpu);
312 case RSEQ_PERCPU_MM_CID:
313 return rseq_offset_deref_addv_relaxed_mm_cid(ptr, off, inc, cpu);
314 }
315 return -1;
316 }
317 #endif
318
319 static inline __attribute__((always_inline))
rseq_cmpeqv_trystorev_storev(enum rseq_mo rseq_mo,enum rseq_percpu_mode percpu_mode,intptr_t * v,intptr_t expect,intptr_t * v2,intptr_t newv2,intptr_t newv,int cpu)320 int rseq_cmpeqv_trystorev_storev(enum rseq_mo rseq_mo, enum rseq_percpu_mode percpu_mode,
321 intptr_t *v, intptr_t expect,
322 intptr_t *v2, intptr_t newv2,
323 intptr_t newv, int cpu)
324 {
325 switch (rseq_mo) {
326 case RSEQ_MO_RELAXED:
327 switch (percpu_mode) {
328 case RSEQ_PERCPU_CPU_ID:
329 return rseq_cmpeqv_trystorev_storev_relaxed_cpu_id(v, expect, v2, newv2, newv, cpu);
330 case RSEQ_PERCPU_MM_CID:
331 return rseq_cmpeqv_trystorev_storev_relaxed_mm_cid(v, expect, v2, newv2, newv, cpu);
332 }
333 return -1;
334 case RSEQ_MO_RELEASE:
335 switch (percpu_mode) {
336 case RSEQ_PERCPU_CPU_ID:
337 return rseq_cmpeqv_trystorev_storev_release_cpu_id(v, expect, v2, newv2, newv, cpu);
338 case RSEQ_PERCPU_MM_CID:
339 return rseq_cmpeqv_trystorev_storev_release_mm_cid(v, expect, v2, newv2, newv, cpu);
340 }
341 return -1;
342 default:
343 return -1;
344 }
345 }
346
347 static inline __attribute__((always_inline))
rseq_cmpeqv_cmpeqv_storev(enum rseq_mo rseq_mo,enum rseq_percpu_mode percpu_mode,intptr_t * v,intptr_t expect,intptr_t * v2,intptr_t expect2,intptr_t newv,int cpu)348 int rseq_cmpeqv_cmpeqv_storev(enum rseq_mo rseq_mo, enum rseq_percpu_mode percpu_mode,
349 intptr_t *v, intptr_t expect,
350 intptr_t *v2, intptr_t expect2,
351 intptr_t newv, int cpu)
352 {
353 if (rseq_mo != RSEQ_MO_RELAXED)
354 return -1;
355 switch (percpu_mode) {
356 case RSEQ_PERCPU_CPU_ID:
357 return rseq_cmpeqv_cmpeqv_storev_relaxed_cpu_id(v, expect, v2, expect2, newv, cpu);
358 case RSEQ_PERCPU_MM_CID:
359 return rseq_cmpeqv_cmpeqv_storev_relaxed_mm_cid(v, expect, v2, expect2, newv, cpu);
360 }
361 return -1;
362 }
363
364 static inline __attribute__((always_inline))
rseq_cmpeqv_trymemcpy_storev(enum rseq_mo rseq_mo,enum rseq_percpu_mode percpu_mode,intptr_t * v,intptr_t expect,void * dst,void * src,size_t len,intptr_t newv,int cpu)365 int rseq_cmpeqv_trymemcpy_storev(enum rseq_mo rseq_mo, enum rseq_percpu_mode percpu_mode,
366 intptr_t *v, intptr_t expect,
367 void *dst, void *src, size_t len,
368 intptr_t newv, int cpu)
369 {
370 switch (rseq_mo) {
371 case RSEQ_MO_RELAXED:
372 switch (percpu_mode) {
373 case RSEQ_PERCPU_CPU_ID:
374 return rseq_cmpeqv_trymemcpy_storev_relaxed_cpu_id(v, expect, dst, src, len, newv, cpu);
375 case RSEQ_PERCPU_MM_CID:
376 return rseq_cmpeqv_trymemcpy_storev_relaxed_mm_cid(v, expect, dst, src, len, newv, cpu);
377 }
378 return -1;
379 case RSEQ_MO_RELEASE:
380 switch (percpu_mode) {
381 case RSEQ_PERCPU_CPU_ID:
382 return rseq_cmpeqv_trymemcpy_storev_release_cpu_id(v, expect, dst, src, len, newv, cpu);
383 case RSEQ_PERCPU_MM_CID:
384 return rseq_cmpeqv_trymemcpy_storev_release_mm_cid(v, expect, dst, src, len, newv, cpu);
385 }
386 return -1;
387 default:
388 return -1;
389 }
390 }
391
392 #endif /* RSEQ_H_ */
393