Commit | Line | Data |
---|---|---|
d7822b1e MD |
1 | // SPDX-License-Identifier: GPL-2.0+ |
2 | /* | |
3 | * Restartable sequences system call | |
4 | * | |
5 | * Copyright (C) 2015, Google, Inc., | |
6 | * Paul Turner <pjt@google.com> and Andrew Hunter <ahh@google.com> | |
7 | * Copyright (C) 2015-2018, EfficiOS Inc., | |
8 | * Mathieu Desnoyers <mathieu.desnoyers@efficios.com> | |
9 | */ | |
10 | ||
11 | #include <linux/sched.h> | |
12 | #include <linux/uaccess.h> | |
13 | #include <linux/syscalls.h> | |
14 | #include <linux/rseq.h> | |
15 | #include <linux/types.h> | |
16 | #include <asm/ptrace.h> | |
17 | ||
18 | #define CREATE_TRACE_POINTS | |
19 | #include <trace/events/rseq.h> | |
20 | ||
0190e419 MD |
21 | #define RSEQ_CS_NO_RESTART_FLAGS (RSEQ_CS_FLAG_NO_RESTART_ON_PREEMPT | \ |
22 | RSEQ_CS_FLAG_NO_RESTART_ON_SIGNAL | \ | |
23 | RSEQ_CS_FLAG_NO_RESTART_ON_MIGRATE) | |
d7822b1e MD |
24 | |
25 | /* | |
26 | * | |
27 | * Restartable sequences are a lightweight interface that allows | |
28 | * user-level code to be executed atomically relative to scheduler | |
29 | * preemption and signal delivery. Typically used for implementing | |
30 | * per-cpu operations. | |
31 | * | |
32 | * It allows user-space to perform update operations on per-cpu data | |
33 | * without requiring heavy-weight atomic operations. | |
34 | * | |
35 | * Detailed algorithm of rseq user-space assembly sequences: | |
36 | * | |
37 | * init(rseq_cs) | |
38 | * cpu = TLS->rseq::cpu_id_start | |
39 | * [1] TLS->rseq::rseq_cs = rseq_cs | |
40 | * [start_ip] ---------------------------- | |
41 | * [2] if (cpu != TLS->rseq::cpu_id) | |
42 | * goto abort_ip; | |
43 | * [3] <last_instruction_in_cs> | |
44 | * [post_commit_ip] ---------------------------- | |
45 | * | |
46 | * The address of jump target abort_ip must be outside the critical | |
47 | * region, i.e.: | |
48 | * | |
49 | * [abort_ip] < [start_ip] || [abort_ip] >= [post_commit_ip] | |
50 | * | |
51 | * Steps [2]-[3] (inclusive) need to be a sequence of instructions in | |
52 | * userspace that can handle being interrupted between any of those | |
53 | * instructions, and then resumed to the abort_ip. | |
54 | * | |
55 | * 1. Userspace stores the address of the struct rseq_cs assembly | |
56 | * block descriptor into the rseq_cs field of the registered | |
57 | * struct rseq TLS area. This update is performed through a single | |
58 | * store within the inline assembly instruction sequence. | |
59 | * [start_ip] | |
60 | * | |
61 | * 2. Userspace tests to check whether the current cpu_id field match | |
62 | * the cpu number loaded before start_ip, branching to abort_ip | |
63 | * in case of a mismatch. | |
64 | * | |
65 | * If the sequence is preempted or interrupted by a signal | |
66 | * at or after start_ip and before post_commit_ip, then the kernel | |
67 | * clears TLS->__rseq_abi::rseq_cs, and sets the user-space return | |
68 | * ip to abort_ip before returning to user-space, so the preempted | |
69 | * execution resumes at abort_ip. | |
70 | * | |
71 | * 3. Userspace critical section final instruction before | |
72 | * post_commit_ip is the commit. The critical section is | |
73 | * self-terminating. | |
74 | * [post_commit_ip] | |
75 | * | |
76 | * 4. <success> | |
77 | * | |
78 | * On failure at [2], or if interrupted by preempt or signal delivery | |
79 | * between [1] and [3]: | |
80 | * | |
81 | * [abort_ip] | |
82 | * F1. <failure> | |
83 | */ | |
84 | ||
85 | static int rseq_update_cpu_id(struct task_struct *t) | |
86 | { | |
87 | u32 cpu_id = raw_smp_processor_id(); | |
60af388d | 88 | struct rseq __user *rseq = t->rseq; |
d7822b1e | 89 | |
60af388d ED |
90 | if (!user_write_access_begin(rseq, sizeof(*rseq))) |
91 | goto efault; | |
92 | unsafe_put_user(cpu_id, &rseq->cpu_id_start, efault_end); | |
93 | unsafe_put_user(cpu_id, &rseq->cpu_id, efault_end); | |
94 | user_write_access_end(); | |
d7822b1e MD |
95 | trace_rseq_update(t); |
96 | return 0; | |
60af388d ED |
97 | |
98 | efault_end: | |
99 | user_write_access_end(); | |
100 | efault: | |
101 | return -EFAULT; | |
d7822b1e MD |
102 | } |
103 | ||
104 | static int rseq_reset_rseq_cpu_id(struct task_struct *t) | |
105 | { | |
106 | u32 cpu_id_start = 0, cpu_id = RSEQ_CPU_ID_UNINITIALIZED; | |
107 | ||
108 | /* | |
109 | * Reset cpu_id_start to its initial state (0). | |
110 | */ | |
8f281770 | 111 | if (put_user(cpu_id_start, &t->rseq->cpu_id_start)) |
d7822b1e MD |
112 | return -EFAULT; |
113 | /* | |
114 | * Reset cpu_id to RSEQ_CPU_ID_UNINITIALIZED, so any user coming | |
115 | * in after unregistration can figure out that rseq needs to be | |
116 | * registered again. | |
117 | */ | |
8f281770 | 118 | if (put_user(cpu_id, &t->rseq->cpu_id)) |
d7822b1e MD |
119 | return -EFAULT; |
120 | return 0; | |
121 | } | |
122 | ||
123 | static int rseq_get_rseq_cs(struct task_struct *t, struct rseq_cs *rseq_cs) | |
124 | { | |
125 | struct rseq_cs __user *urseq_cs; | |
ec9c82e0 | 126 | u64 ptr; |
d7822b1e MD |
127 | u32 __user *usig; |
128 | u32 sig; | |
129 | int ret; | |
130 | ||
5e0ccd4a | 131 | #ifdef CONFIG_64BIT |
bfdf4e62 | 132 | if (get_user(ptr, &t->rseq->rseq_cs)) |
5e0ccd4a ED |
133 | return -EFAULT; |
134 | #else | |
bfdf4e62 | 135 | if (copy_from_user(&ptr, &t->rseq->rseq_cs, sizeof(ptr))) |
ec9c82e0 | 136 | return -EFAULT; |
5e0ccd4a | 137 | #endif |
d7822b1e MD |
138 | if (!ptr) { |
139 | memset(rseq_cs, 0, sizeof(*rseq_cs)); | |
140 | return 0; | |
141 | } | |
ec9c82e0 MD |
142 | if (ptr >= TASK_SIZE) |
143 | return -EINVAL; | |
144 | urseq_cs = (struct rseq_cs __user *)(unsigned long)ptr; | |
d7822b1e MD |
145 | if (copy_from_user(rseq_cs, urseq_cs, sizeof(*rseq_cs))) |
146 | return -EFAULT; | |
d7822b1e | 147 | |
e96d7135 MD |
148 | if (rseq_cs->start_ip >= TASK_SIZE || |
149 | rseq_cs->start_ip + rseq_cs->post_commit_offset >= TASK_SIZE || | |
150 | rseq_cs->abort_ip >= TASK_SIZE || | |
151 | rseq_cs->version > 0) | |
152 | return -EINVAL; | |
153 | /* Check for overflow. */ | |
154 | if (rseq_cs->start_ip + rseq_cs->post_commit_offset < rseq_cs->start_ip) | |
155 | return -EINVAL; | |
d7822b1e MD |
156 | /* Ensure that abort_ip is not in the critical section. */ |
157 | if (rseq_cs->abort_ip - rseq_cs->start_ip < rseq_cs->post_commit_offset) | |
158 | return -EINVAL; | |
159 | ||
e96d7135 | 160 | usig = (u32 __user *)(unsigned long)(rseq_cs->abort_ip - sizeof(u32)); |
d7822b1e MD |
161 | ret = get_user(sig, usig); |
162 | if (ret) | |
163 | return ret; | |
164 | ||
165 | if (current->rseq_sig != sig) { | |
166 | printk_ratelimited(KERN_WARNING | |
167 | "Possible attack attempt. Unexpected rseq signature 0x%x, expecting 0x%x (pid=%d, addr=%p).\n", | |
168 | sig, current->rseq_sig, current->pid, usig); | |
e96d7135 | 169 | return -EINVAL; |
d7822b1e MD |
170 | } |
171 | return 0; | |
172 | } | |
173 | ||
174 | static int rseq_need_restart(struct task_struct *t, u32 cs_flags) | |
175 | { | |
176 | u32 flags, event_mask; | |
177 | int ret; | |
178 | ||
c17a6ff9 | 179 | if (WARN_ON_ONCE(cs_flags & RSEQ_CS_NO_RESTART_FLAGS) || cs_flags) |
0190e419 MD |
180 | return -EINVAL; |
181 | ||
d7822b1e | 182 | /* Get thread flags. */ |
8f281770 | 183 | ret = get_user(flags, &t->rseq->flags); |
d7822b1e MD |
184 | if (ret) |
185 | return ret; | |
186 | ||
c17a6ff9 | 187 | if (WARN_ON_ONCE(flags & RSEQ_CS_NO_RESTART_FLAGS) || flags) |
d7822b1e MD |
188 | return -EINVAL; |
189 | ||
190 | /* | |
191 | * Load and clear event mask atomically with respect to | |
192 | * scheduler preemption. | |
193 | */ | |
194 | preempt_disable(); | |
195 | event_mask = t->rseq_event_mask; | |
196 | t->rseq_event_mask = 0; | |
197 | preempt_enable(); | |
198 | ||
0190e419 | 199 | return !!event_mask; |
d7822b1e MD |
200 | } |
201 | ||
202 | static int clear_rseq_cs(struct task_struct *t) | |
203 | { | |
204 | /* | |
205 | * The rseq_cs field is set to NULL on preemption or signal | |
206 | * delivery on top of rseq assembly block, as well as on top | |
207 | * of code outside of the rseq assembly block. This performs | |
208 | * a lazy clear of the rseq_cs field. | |
209 | * | |
0fb9a1ab | 210 | * Set rseq_cs to NULL. |
d7822b1e | 211 | */ |
5e0ccd4a | 212 | #ifdef CONFIG_64BIT |
bfdf4e62 | 213 | return put_user(0UL, &t->rseq->rseq_cs); |
5e0ccd4a | 214 | #else |
bfdf4e62 | 215 | if (clear_user(&t->rseq->rseq_cs, sizeof(t->rseq->rseq_cs))) |
ec9c82e0 MD |
216 | return -EFAULT; |
217 | return 0; | |
5e0ccd4a | 218 | #endif |
d7822b1e MD |
219 | } |
220 | ||
221 | /* | |
222 | * Unsigned comparison will be true when ip >= start_ip, and when | |
223 | * ip < start_ip + post_commit_offset. | |
224 | */ | |
225 | static bool in_rseq_cs(unsigned long ip, struct rseq_cs *rseq_cs) | |
226 | { | |
227 | return ip - rseq_cs->start_ip < rseq_cs->post_commit_offset; | |
228 | } | |
229 | ||
230 | static int rseq_ip_fixup(struct pt_regs *regs) | |
231 | { | |
232 | unsigned long ip = instruction_pointer(regs); | |
233 | struct task_struct *t = current; | |
234 | struct rseq_cs rseq_cs; | |
235 | int ret; | |
236 | ||
237 | ret = rseq_get_rseq_cs(t, &rseq_cs); | |
238 | if (ret) | |
239 | return ret; | |
240 | ||
241 | /* | |
242 | * Handle potentially not being within a critical section. | |
243 | * If not nested over a rseq critical section, restart is useless. | |
244 | * Clear the rseq_cs pointer and return. | |
245 | */ | |
246 | if (!in_rseq_cs(ip, &rseq_cs)) | |
247 | return clear_rseq_cs(t); | |
248 | ret = rseq_need_restart(t, rseq_cs.flags); | |
249 | if (ret <= 0) | |
250 | return ret; | |
251 | ret = clear_rseq_cs(t); | |
252 | if (ret) | |
253 | return ret; | |
254 | trace_rseq_ip_fixup(ip, rseq_cs.start_ip, rseq_cs.post_commit_offset, | |
255 | rseq_cs.abort_ip); | |
256 | instruction_pointer_set(regs, (unsigned long)rseq_cs.abort_ip); | |
257 | return 0; | |
258 | } | |
259 | ||
260 | /* | |
261 | * This resume handler must always be executed between any of: | |
262 | * - preemption, | |
263 | * - signal delivery, | |
264 | * and return to user-space. | |
265 | * | |
bff9504b | 266 | * This is how we can ensure that the entire rseq critical section |
d7822b1e MD |
267 | * will issue the commit instruction only if executed atomically with |
268 | * respect to other threads scheduled on the same CPU, and with respect | |
269 | * to signal handlers. | |
270 | */ | |
784e0300 | 271 | void __rseq_handle_notify_resume(struct ksignal *ksig, struct pt_regs *regs) |
d7822b1e MD |
272 | { |
273 | struct task_struct *t = current; | |
784e0300 | 274 | int ret, sig; |
d7822b1e MD |
275 | |
276 | if (unlikely(t->flags & PF_EXITING)) | |
277 | return; | |
8646e536 SC |
278 | |
279 | /* | |
280 | * regs is NULL if and only if the caller is in a syscall path. Skip | |
281 | * fixup and leave rseq_cs as is so that rseq_sycall() will detect and | |
282 | * kill a misbehaving userspace on debug kernels. | |
283 | */ | |
284 | if (regs) { | |
285 | ret = rseq_ip_fixup(regs); | |
286 | if (unlikely(ret < 0)) | |
287 | goto error; | |
288 | } | |
d7822b1e MD |
289 | if (unlikely(rseq_update_cpu_id(t))) |
290 | goto error; | |
291 | return; | |
292 | ||
293 | error: | |
784e0300 | 294 | sig = ksig ? ksig->sig : 0; |
cb44c9a0 | 295 | force_sigsegv(sig); |
d7822b1e MD |
296 | } |
297 | ||
298 | #ifdef CONFIG_DEBUG_RSEQ | |
299 | ||
300 | /* | |
301 | * Terminate the process if a syscall is issued within a restartable | |
302 | * sequence. | |
303 | */ | |
304 | void rseq_syscall(struct pt_regs *regs) | |
305 | { | |
306 | unsigned long ip = instruction_pointer(regs); | |
307 | struct task_struct *t = current; | |
308 | struct rseq_cs rseq_cs; | |
309 | ||
310 | if (!t->rseq) | |
311 | return; | |
0ed96051 | 312 | if (rseq_get_rseq_cs(t, &rseq_cs) || in_rseq_cs(ip, &rseq_cs)) |
3cf5d076 | 313 | force_sig(SIGSEGV); |
d7822b1e MD |
314 | } |
315 | ||
316 | #endif | |
317 | ||
318 | /* | |
319 | * sys_rseq - setup restartable sequences for caller thread. | |
320 | */ | |
321 | SYSCALL_DEFINE4(rseq, struct rseq __user *, rseq, u32, rseq_len, | |
322 | int, flags, u32, sig) | |
323 | { | |
324 | int ret; | |
325 | ||
326 | if (flags & RSEQ_FLAG_UNREGISTER) { | |
66528a45 MD |
327 | if (flags & ~RSEQ_FLAG_UNREGISTER) |
328 | return -EINVAL; | |
d7822b1e MD |
329 | /* Unregister rseq for current thread. */ |
330 | if (current->rseq != rseq || !current->rseq) | |
331 | return -EINVAL; | |
83b0b15b | 332 | if (rseq_len != sizeof(*rseq)) |
d7822b1e MD |
333 | return -EINVAL; |
334 | if (current->rseq_sig != sig) | |
335 | return -EPERM; | |
336 | ret = rseq_reset_rseq_cpu_id(current); | |
337 | if (ret) | |
338 | return ret; | |
339 | current->rseq = NULL; | |
d7822b1e MD |
340 | current->rseq_sig = 0; |
341 | return 0; | |
342 | } | |
343 | ||
344 | if (unlikely(flags)) | |
345 | return -EINVAL; | |
346 | ||
347 | if (current->rseq) { | |
348 | /* | |
349 | * If rseq is already registered, check whether | |
350 | * the provided address differs from the prior | |
351 | * one. | |
352 | */ | |
83b0b15b | 353 | if (current->rseq != rseq || rseq_len != sizeof(*rseq)) |
d7822b1e MD |
354 | return -EINVAL; |
355 | if (current->rseq_sig != sig) | |
356 | return -EPERM; | |
357 | /* Already registered. */ | |
358 | return -EBUSY; | |
359 | } | |
360 | ||
361 | /* | |
362 | * If there was no rseq previously registered, | |
363 | * ensure the provided rseq is properly aligned and valid. | |
364 | */ | |
365 | if (!IS_ALIGNED((unsigned long)rseq, __alignof__(*rseq)) || | |
366 | rseq_len != sizeof(*rseq)) | |
367 | return -EINVAL; | |
96d4f267 | 368 | if (!access_ok(rseq, rseq_len)) |
d7822b1e MD |
369 | return -EFAULT; |
370 | current->rseq = rseq; | |
d7822b1e MD |
371 | current->rseq_sig = sig; |
372 | /* | |
373 | * If rseq was previously inactive, and has just been | |
374 | * registered, ensure the cpu_id_start and cpu_id fields | |
375 | * are updated before returning to user-space. | |
376 | */ | |
377 | rseq_set_notify_resume(current); | |
378 | ||
379 | return 0; | |
380 | } |