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 | |
21 | /* The original rseq structure size (including padding) is 32 bytes. */ |
22 | #define ORIG_RSEQ_SIZE 32 |
23 | |
24 | #define RSEQ_CS_NO_RESTART_FLAGS (RSEQ_CS_FLAG_NO_RESTART_ON_PREEMPT | \ |
25 | RSEQ_CS_FLAG_NO_RESTART_ON_SIGNAL | \ |
26 | RSEQ_CS_FLAG_NO_RESTART_ON_MIGRATE) |
27 | |
28 | /* |
29 | * |
30 | * Restartable sequences are a lightweight interface that allows |
31 | * user-level code to be executed atomically relative to scheduler |
32 | * preemption and signal delivery. Typically used for implementing |
33 | * per-cpu operations. |
34 | * |
35 | * It allows user-space to perform update operations on per-cpu data |
36 | * without requiring heavy-weight atomic operations. |
37 | * |
38 | * Detailed algorithm of rseq user-space assembly sequences: |
39 | * |
40 | * init(rseq_cs) |
41 | * cpu = TLS->rseq::cpu_id_start |
42 | * [1] TLS->rseq::rseq_cs = rseq_cs |
43 | * [start_ip] ---------------------------- |
44 | * [2] if (cpu != TLS->rseq::cpu_id) |
45 | * goto abort_ip; |
46 | * [3] <last_instruction_in_cs> |
47 | * [post_commit_ip] ---------------------------- |
48 | * |
49 | * The address of jump target abort_ip must be outside the critical |
50 | * region, i.e.: |
51 | * |
52 | * [abort_ip] < [start_ip] || [abort_ip] >= [post_commit_ip] |
53 | * |
54 | * Steps [2]-[3] (inclusive) need to be a sequence of instructions in |
55 | * userspace that can handle being interrupted between any of those |
56 | * instructions, and then resumed to the abort_ip. |
57 | * |
58 | * 1. Userspace stores the address of the struct rseq_cs assembly |
59 | * block descriptor into the rseq_cs field of the registered |
60 | * struct rseq TLS area. This update is performed through a single |
61 | * store within the inline assembly instruction sequence. |
62 | * [start_ip] |
63 | * |
64 | * 2. Userspace tests to check whether the current cpu_id field match |
65 | * the cpu number loaded before start_ip, branching to abort_ip |
66 | * in case of a mismatch. |
67 | * |
68 | * If the sequence is preempted or interrupted by a signal |
69 | * at or after start_ip and before post_commit_ip, then the kernel |
70 | * clears TLS->__rseq_abi::rseq_cs, and sets the user-space return |
71 | * ip to abort_ip before returning to user-space, so the preempted |
72 | * execution resumes at abort_ip. |
73 | * |
74 | * 3. Userspace critical section final instruction before |
75 | * post_commit_ip is the commit. The critical section is |
76 | * self-terminating. |
77 | * [post_commit_ip] |
78 | * |
79 | * 4. <success> |
80 | * |
81 | * On failure at [2], or if interrupted by preempt or signal delivery |
82 | * between [1] and [3]: |
83 | * |
84 | * [abort_ip] |
85 | * F1. <failure> |
86 | */ |
87 | |
88 | static int rseq_update_cpu_node_id(struct task_struct *t) |
89 | { |
90 | struct rseq __user *rseq = t->rseq; |
91 | u32 cpu_id = raw_smp_processor_id(); |
92 | u32 node_id = cpu_to_node(cpu: cpu_id); |
93 | u32 mm_cid = task_mm_cid(t); |
94 | |
95 | WARN_ON_ONCE((int) mm_cid < 0); |
96 | if (!user_write_access_begin(rseq, t->rseq_len)) |
97 | goto efault; |
98 | unsafe_put_user(cpu_id, &rseq->cpu_id_start, efault_end); |
99 | unsafe_put_user(cpu_id, &rseq->cpu_id, efault_end); |
100 | unsafe_put_user(node_id, &rseq->node_id, efault_end); |
101 | unsafe_put_user(mm_cid, &rseq->mm_cid, efault_end); |
102 | /* |
103 | * Additional feature fields added after ORIG_RSEQ_SIZE |
104 | * need to be conditionally updated only if |
105 | * t->rseq_len != ORIG_RSEQ_SIZE. |
106 | */ |
107 | user_write_access_end(); |
108 | trace_rseq_update(t); |
109 | return 0; |
110 | |
111 | efault_end: |
112 | user_write_access_end(); |
113 | efault: |
114 | return -EFAULT; |
115 | } |
116 | |
117 | static int rseq_reset_rseq_cpu_node_id(struct task_struct *t) |
118 | { |
119 | u32 cpu_id_start = 0, cpu_id = RSEQ_CPU_ID_UNINITIALIZED, node_id = 0, |
120 | mm_cid = 0; |
121 | |
122 | /* |
123 | * Reset cpu_id_start to its initial state (0). |
124 | */ |
125 | if (put_user(cpu_id_start, &t->rseq->cpu_id_start)) |
126 | return -EFAULT; |
127 | /* |
128 | * Reset cpu_id to RSEQ_CPU_ID_UNINITIALIZED, so any user coming |
129 | * in after unregistration can figure out that rseq needs to be |
130 | * registered again. |
131 | */ |
132 | if (put_user(cpu_id, &t->rseq->cpu_id)) |
133 | return -EFAULT; |
134 | /* |
135 | * Reset node_id to its initial state (0). |
136 | */ |
137 | if (put_user(node_id, &t->rseq->node_id)) |
138 | return -EFAULT; |
139 | /* |
140 | * Reset mm_cid to its initial state (0). |
141 | */ |
142 | if (put_user(mm_cid, &t->rseq->mm_cid)) |
143 | return -EFAULT; |
144 | /* |
145 | * Additional feature fields added after ORIG_RSEQ_SIZE |
146 | * need to be conditionally reset only if |
147 | * t->rseq_len != ORIG_RSEQ_SIZE. |
148 | */ |
149 | return 0; |
150 | } |
151 | |
152 | static int rseq_get_rseq_cs(struct task_struct *t, struct rseq_cs *rseq_cs) |
153 | { |
154 | struct rseq_cs __user *urseq_cs; |
155 | u64 ptr; |
156 | u32 __user *usig; |
157 | u32 sig; |
158 | int ret; |
159 | |
160 | #ifdef CONFIG_64BIT |
161 | if (get_user(ptr, &t->rseq->rseq_cs)) |
162 | return -EFAULT; |
163 | #else |
164 | if (copy_from_user(&ptr, &t->rseq->rseq_cs, sizeof(ptr))) |
165 | return -EFAULT; |
166 | #endif |
167 | if (!ptr) { |
168 | memset(rseq_cs, 0, sizeof(*rseq_cs)); |
169 | return 0; |
170 | } |
171 | if (ptr >= TASK_SIZE) |
172 | return -EINVAL; |
173 | urseq_cs = (struct rseq_cs __user *)(unsigned long)ptr; |
174 | if (copy_from_user(to: rseq_cs, from: urseq_cs, n: sizeof(*rseq_cs))) |
175 | return -EFAULT; |
176 | |
177 | if (rseq_cs->start_ip >= TASK_SIZE || |
178 | rseq_cs->start_ip + rseq_cs->post_commit_offset >= TASK_SIZE || |
179 | rseq_cs->abort_ip >= TASK_SIZE || |
180 | rseq_cs->version > 0) |
181 | return -EINVAL; |
182 | /* Check for overflow. */ |
183 | if (rseq_cs->start_ip + rseq_cs->post_commit_offset < rseq_cs->start_ip) |
184 | return -EINVAL; |
185 | /* Ensure that abort_ip is not in the critical section. */ |
186 | if (rseq_cs->abort_ip - rseq_cs->start_ip < rseq_cs->post_commit_offset) |
187 | return -EINVAL; |
188 | |
189 | usig = (u32 __user *)(unsigned long)(rseq_cs->abort_ip - sizeof(u32)); |
190 | ret = get_user(sig, usig); |
191 | if (ret) |
192 | return ret; |
193 | |
194 | if (current->rseq_sig != sig) { |
195 | printk_ratelimited(KERN_WARNING |
196 | "Possible attack attempt. Unexpected rseq signature 0x%x, expecting 0x%x (pid=%d, addr=%p).\n" , |
197 | sig, current->rseq_sig, current->pid, usig); |
198 | return -EINVAL; |
199 | } |
200 | return 0; |
201 | } |
202 | |
203 | static bool rseq_warn_flags(const char *str, u32 flags) |
204 | { |
205 | u32 test_flags; |
206 | |
207 | if (!flags) |
208 | return false; |
209 | test_flags = flags & RSEQ_CS_NO_RESTART_FLAGS; |
210 | if (test_flags) |
211 | pr_warn_once("Deprecated flags (%u) in %s ABI structure" , test_flags, str); |
212 | test_flags = flags & ~RSEQ_CS_NO_RESTART_FLAGS; |
213 | if (test_flags) |
214 | pr_warn_once("Unknown flags (%u) in %s ABI structure" , test_flags, str); |
215 | return true; |
216 | } |
217 | |
218 | static int rseq_need_restart(struct task_struct *t, u32 cs_flags) |
219 | { |
220 | u32 flags, event_mask; |
221 | int ret; |
222 | |
223 | if (rseq_warn_flags(str: "rseq_cs" , flags: cs_flags)) |
224 | return -EINVAL; |
225 | |
226 | /* Get thread flags. */ |
227 | ret = get_user(flags, &t->rseq->flags); |
228 | if (ret) |
229 | return ret; |
230 | |
231 | if (rseq_warn_flags(str: "rseq" , flags)) |
232 | return -EINVAL; |
233 | |
234 | /* |
235 | * Load and clear event mask atomically with respect to |
236 | * scheduler preemption. |
237 | */ |
238 | preempt_disable(); |
239 | event_mask = t->rseq_event_mask; |
240 | t->rseq_event_mask = 0; |
241 | preempt_enable(); |
242 | |
243 | return !!event_mask; |
244 | } |
245 | |
246 | static int clear_rseq_cs(struct task_struct *t) |
247 | { |
248 | /* |
249 | * The rseq_cs field is set to NULL on preemption or signal |
250 | * delivery on top of rseq assembly block, as well as on top |
251 | * of code outside of the rseq assembly block. This performs |
252 | * a lazy clear of the rseq_cs field. |
253 | * |
254 | * Set rseq_cs to NULL. |
255 | */ |
256 | #ifdef CONFIG_64BIT |
257 | return put_user(0UL, &t->rseq->rseq_cs); |
258 | #else |
259 | if (clear_user(&t->rseq->rseq_cs, sizeof(t->rseq->rseq_cs))) |
260 | return -EFAULT; |
261 | return 0; |
262 | #endif |
263 | } |
264 | |
265 | /* |
266 | * Unsigned comparison will be true when ip >= start_ip, and when |
267 | * ip < start_ip + post_commit_offset. |
268 | */ |
269 | static bool in_rseq_cs(unsigned long ip, struct rseq_cs *rseq_cs) |
270 | { |
271 | return ip - rseq_cs->start_ip < rseq_cs->post_commit_offset; |
272 | } |
273 | |
274 | static int rseq_ip_fixup(struct pt_regs *regs) |
275 | { |
276 | unsigned long ip = instruction_pointer(regs); |
277 | struct task_struct *t = current; |
278 | struct rseq_cs rseq_cs; |
279 | int ret; |
280 | |
281 | ret = rseq_get_rseq_cs(t, rseq_cs: &rseq_cs); |
282 | if (ret) |
283 | return ret; |
284 | |
285 | /* |
286 | * Handle potentially not being within a critical section. |
287 | * If not nested over a rseq critical section, restart is useless. |
288 | * Clear the rseq_cs pointer and return. |
289 | */ |
290 | if (!in_rseq_cs(ip, rseq_cs: &rseq_cs)) |
291 | return clear_rseq_cs(t); |
292 | ret = rseq_need_restart(t, cs_flags: rseq_cs.flags); |
293 | if (ret <= 0) |
294 | return ret; |
295 | ret = clear_rseq_cs(t); |
296 | if (ret) |
297 | return ret; |
298 | trace_rseq_ip_fixup(regs_ip: ip, start_ip: rseq_cs.start_ip, post_commit_offset: rseq_cs.post_commit_offset, |
299 | abort_ip: rseq_cs.abort_ip); |
300 | instruction_pointer_set(regs, val: (unsigned long)rseq_cs.abort_ip); |
301 | return 0; |
302 | } |
303 | |
304 | /* |
305 | * This resume handler must always be executed between any of: |
306 | * - preemption, |
307 | * - signal delivery, |
308 | * and return to user-space. |
309 | * |
310 | * This is how we can ensure that the entire rseq critical section |
311 | * will issue the commit instruction only if executed atomically with |
312 | * respect to other threads scheduled on the same CPU, and with respect |
313 | * to signal handlers. |
314 | */ |
315 | void __rseq_handle_notify_resume(struct ksignal *ksig, struct pt_regs *regs) |
316 | { |
317 | struct task_struct *t = current; |
318 | int ret, sig; |
319 | |
320 | if (unlikely(t->flags & PF_EXITING)) |
321 | return; |
322 | |
323 | /* |
324 | * regs is NULL if and only if the caller is in a syscall path. Skip |
325 | * fixup and leave rseq_cs as is so that rseq_sycall() will detect and |
326 | * kill a misbehaving userspace on debug kernels. |
327 | */ |
328 | if (regs) { |
329 | ret = rseq_ip_fixup(regs); |
330 | if (unlikely(ret < 0)) |
331 | goto error; |
332 | } |
333 | if (unlikely(rseq_update_cpu_node_id(t))) |
334 | goto error; |
335 | return; |
336 | |
337 | error: |
338 | sig = ksig ? ksig->sig : 0; |
339 | force_sigsegv(sig); |
340 | } |
341 | |
342 | #ifdef CONFIG_DEBUG_RSEQ |
343 | |
344 | /* |
345 | * Terminate the process if a syscall is issued within a restartable |
346 | * sequence. |
347 | */ |
348 | void rseq_syscall(struct pt_regs *regs) |
349 | { |
350 | unsigned long ip = instruction_pointer(regs); |
351 | struct task_struct *t = current; |
352 | struct rseq_cs rseq_cs; |
353 | |
354 | if (!t->rseq) |
355 | return; |
356 | if (rseq_get_rseq_cs(t, rseq_cs: &rseq_cs) || in_rseq_cs(ip, rseq_cs: &rseq_cs)) |
357 | force_sig(SIGSEGV); |
358 | } |
359 | |
360 | #endif |
361 | |
362 | /* |
363 | * sys_rseq - setup restartable sequences for caller thread. |
364 | */ |
365 | SYSCALL_DEFINE4(rseq, struct rseq __user *, rseq, u32, rseq_len, |
366 | int, flags, u32, sig) |
367 | { |
368 | int ret; |
369 | |
370 | if (flags & RSEQ_FLAG_UNREGISTER) { |
371 | if (flags & ~RSEQ_FLAG_UNREGISTER) |
372 | return -EINVAL; |
373 | /* Unregister rseq for current thread. */ |
374 | if (current->rseq != rseq || !current->rseq) |
375 | return -EINVAL; |
376 | if (rseq_len != current->rseq_len) |
377 | return -EINVAL; |
378 | if (current->rseq_sig != sig) |
379 | return -EPERM; |
380 | ret = rseq_reset_rseq_cpu_node_id(current); |
381 | if (ret) |
382 | return ret; |
383 | current->rseq = NULL; |
384 | current->rseq_sig = 0; |
385 | current->rseq_len = 0; |
386 | return 0; |
387 | } |
388 | |
389 | if (unlikely(flags)) |
390 | return -EINVAL; |
391 | |
392 | if (current->rseq) { |
393 | /* |
394 | * If rseq is already registered, check whether |
395 | * the provided address differs from the prior |
396 | * one. |
397 | */ |
398 | if (current->rseq != rseq || rseq_len != current->rseq_len) |
399 | return -EINVAL; |
400 | if (current->rseq_sig != sig) |
401 | return -EPERM; |
402 | /* Already registered. */ |
403 | return -EBUSY; |
404 | } |
405 | |
406 | /* |
407 | * If there was no rseq previously registered, ensure the provided rseq |
408 | * is properly aligned, as communcated to user-space through the ELF |
409 | * auxiliary vector AT_RSEQ_ALIGN. If rseq_len is the original rseq |
410 | * size, the required alignment is the original struct rseq alignment. |
411 | * |
412 | * In order to be valid, rseq_len is either the original rseq size, or |
413 | * large enough to contain all supported fields, as communicated to |
414 | * user-space through the ELF auxiliary vector AT_RSEQ_FEATURE_SIZE. |
415 | */ |
416 | if (rseq_len < ORIG_RSEQ_SIZE || |
417 | (rseq_len == ORIG_RSEQ_SIZE && !IS_ALIGNED((unsigned long)rseq, ORIG_RSEQ_SIZE)) || |
418 | (rseq_len != ORIG_RSEQ_SIZE && (!IS_ALIGNED((unsigned long)rseq, __alignof__(*rseq)) || |
419 | rseq_len < offsetof(struct rseq, end)))) |
420 | return -EINVAL; |
421 | if (!access_ok(rseq, rseq_len)) |
422 | return -EFAULT; |
423 | current->rseq = rseq; |
424 | current->rseq_len = rseq_len; |
425 | current->rseq_sig = sig; |
426 | /* |
427 | * If rseq was previously inactive, and has just been |
428 | * registered, ensure the cpu_id_start and cpu_id fields |
429 | * are updated before returning to user-space. |
430 | */ |
431 | rseq_set_notify_resume(current); |
432 | |
433 | return 0; |
434 | } |
435 | |