xref: /linux/kernel/rseq.c (revision ed843d6ed7310a27cf7c8ee0a82a482eed0cb4a6)
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 #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)
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();
88 	struct rseq __user *rseq = t->rseq;
89 
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();
95 	trace_rseq_update(t);
96 	return 0;
97 
98 efault_end:
99 	user_write_access_end();
100 efault:
101 	return -EFAULT;
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 	 */
111 	if (put_user(cpu_id_start, &t->rseq->cpu_id_start))
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 	 */
118 	if (put_user(cpu_id, &t->rseq->cpu_id))
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;
126 	u64 ptr;
127 	u32 __user *usig;
128 	u32 sig;
129 	int ret;
130 
131 #ifdef CONFIG_64BIT
132 	if (get_user(ptr, &t->rseq->rseq_cs))
133 		return -EFAULT;
134 #else
135 	if (copy_from_user(&ptr, &t->rseq->rseq_cs, sizeof(ptr)))
136 		return -EFAULT;
137 #endif
138 	if (!ptr) {
139 		memset(rseq_cs, 0, sizeof(*rseq_cs));
140 		return 0;
141 	}
142 	if (ptr >= TASK_SIZE)
143 		return -EINVAL;
144 	urseq_cs = (struct rseq_cs __user *)(unsigned long)ptr;
145 	if (copy_from_user(rseq_cs, urseq_cs, sizeof(*rseq_cs)))
146 		return -EFAULT;
147 
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;
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 
160 	usig = (u32 __user *)(unsigned long)(rseq_cs->abort_ip - sizeof(u32));
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);
169 		return -EINVAL;
170 	}
171 	return 0;
172 }
173 
174 static bool rseq_warn_flags(const char *str, u32 flags)
175 {
176 	u32 test_flags;
177 
178 	if (!flags)
179 		return false;
180 	test_flags = flags & RSEQ_CS_NO_RESTART_FLAGS;
181 	if (test_flags)
182 		pr_warn_once("Deprecated flags (%u) in %s ABI structure", test_flags, str);
183 	test_flags = flags & ~RSEQ_CS_NO_RESTART_FLAGS;
184 	if (test_flags)
185 		pr_warn_once("Unknown flags (%u) in %s ABI structure", test_flags, str);
186 	return true;
187 }
188 
189 static int rseq_need_restart(struct task_struct *t, u32 cs_flags)
190 {
191 	u32 flags, event_mask;
192 	int ret;
193 
194 	if (rseq_warn_flags("rseq_cs", cs_flags))
195 		return -EINVAL;
196 
197 	/* Get thread flags. */
198 	ret = get_user(flags, &t->rseq->flags);
199 	if (ret)
200 		return ret;
201 
202 	if (rseq_warn_flags("rseq", flags))
203 		return -EINVAL;
204 
205 	/*
206 	 * Load and clear event mask atomically with respect to
207 	 * scheduler preemption.
208 	 */
209 	preempt_disable();
210 	event_mask = t->rseq_event_mask;
211 	t->rseq_event_mask = 0;
212 	preempt_enable();
213 
214 	return !!event_mask;
215 }
216 
217 static int clear_rseq_cs(struct task_struct *t)
218 {
219 	/*
220 	 * The rseq_cs field is set to NULL on preemption or signal
221 	 * delivery on top of rseq assembly block, as well as on top
222 	 * of code outside of the rseq assembly block. This performs
223 	 * a lazy clear of the rseq_cs field.
224 	 *
225 	 * Set rseq_cs to NULL.
226 	 */
227 #ifdef CONFIG_64BIT
228 	return put_user(0UL, &t->rseq->rseq_cs);
229 #else
230 	if (clear_user(&t->rseq->rseq_cs, sizeof(t->rseq->rseq_cs)))
231 		return -EFAULT;
232 	return 0;
233 #endif
234 }
235 
236 /*
237  * Unsigned comparison will be true when ip >= start_ip, and when
238  * ip < start_ip + post_commit_offset.
239  */
240 static bool in_rseq_cs(unsigned long ip, struct rseq_cs *rseq_cs)
241 {
242 	return ip - rseq_cs->start_ip < rseq_cs->post_commit_offset;
243 }
244 
245 static int rseq_ip_fixup(struct pt_regs *regs)
246 {
247 	unsigned long ip = instruction_pointer(regs);
248 	struct task_struct *t = current;
249 	struct rseq_cs rseq_cs;
250 	int ret;
251 
252 	ret = rseq_get_rseq_cs(t, &rseq_cs);
253 	if (ret)
254 		return ret;
255 
256 	/*
257 	 * Handle potentially not being within a critical section.
258 	 * If not nested over a rseq critical section, restart is useless.
259 	 * Clear the rseq_cs pointer and return.
260 	 */
261 	if (!in_rseq_cs(ip, &rseq_cs))
262 		return clear_rseq_cs(t);
263 	ret = rseq_need_restart(t, rseq_cs.flags);
264 	if (ret <= 0)
265 		return ret;
266 	ret = clear_rseq_cs(t);
267 	if (ret)
268 		return ret;
269 	trace_rseq_ip_fixup(ip, rseq_cs.start_ip, rseq_cs.post_commit_offset,
270 			    rseq_cs.abort_ip);
271 	instruction_pointer_set(regs, (unsigned long)rseq_cs.abort_ip);
272 	return 0;
273 }
274 
275 /*
276  * This resume handler must always be executed between any of:
277  * - preemption,
278  * - signal delivery,
279  * and return to user-space.
280  *
281  * This is how we can ensure that the entire rseq critical section
282  * will issue the commit instruction only if executed atomically with
283  * respect to other threads scheduled on the same CPU, and with respect
284  * to signal handlers.
285  */
286 void __rseq_handle_notify_resume(struct ksignal *ksig, struct pt_regs *regs)
287 {
288 	struct task_struct *t = current;
289 	int ret, sig;
290 
291 	if (unlikely(t->flags & PF_EXITING))
292 		return;
293 
294 	/*
295 	 * regs is NULL if and only if the caller is in a syscall path.  Skip
296 	 * fixup and leave rseq_cs as is so that rseq_sycall() will detect and
297 	 * kill a misbehaving userspace on debug kernels.
298 	 */
299 	if (regs) {
300 		ret = rseq_ip_fixup(regs);
301 		if (unlikely(ret < 0))
302 			goto error;
303 	}
304 	if (unlikely(rseq_update_cpu_id(t)))
305 		goto error;
306 	return;
307 
308 error:
309 	sig = ksig ? ksig->sig : 0;
310 	force_sigsegv(sig);
311 }
312 
313 #ifdef CONFIG_DEBUG_RSEQ
314 
315 /*
316  * Terminate the process if a syscall is issued within a restartable
317  * sequence.
318  */
319 void rseq_syscall(struct pt_regs *regs)
320 {
321 	unsigned long ip = instruction_pointer(regs);
322 	struct task_struct *t = current;
323 	struct rseq_cs rseq_cs;
324 
325 	if (!t->rseq)
326 		return;
327 	if (rseq_get_rseq_cs(t, &rseq_cs) || in_rseq_cs(ip, &rseq_cs))
328 		force_sig(SIGSEGV);
329 }
330 
331 #endif
332 
333 /*
334  * sys_rseq - setup restartable sequences for caller thread.
335  */
336 SYSCALL_DEFINE4(rseq, struct rseq __user *, rseq, u32, rseq_len,
337 		int, flags, u32, sig)
338 {
339 	int ret;
340 
341 	if (flags & RSEQ_FLAG_UNREGISTER) {
342 		if (flags & ~RSEQ_FLAG_UNREGISTER)
343 			return -EINVAL;
344 		/* Unregister rseq for current thread. */
345 		if (current->rseq != rseq || !current->rseq)
346 			return -EINVAL;
347 		if (rseq_len != sizeof(*rseq))
348 			return -EINVAL;
349 		if (current->rseq_sig != sig)
350 			return -EPERM;
351 		ret = rseq_reset_rseq_cpu_id(current);
352 		if (ret)
353 			return ret;
354 		current->rseq = NULL;
355 		current->rseq_sig = 0;
356 		return 0;
357 	}
358 
359 	if (unlikely(flags))
360 		return -EINVAL;
361 
362 	if (current->rseq) {
363 		/*
364 		 * If rseq is already registered, check whether
365 		 * the provided address differs from the prior
366 		 * one.
367 		 */
368 		if (current->rseq != rseq || rseq_len != sizeof(*rseq))
369 			return -EINVAL;
370 		if (current->rseq_sig != sig)
371 			return -EPERM;
372 		/* Already registered. */
373 		return -EBUSY;
374 	}
375 
376 	/*
377 	 * If there was no rseq previously registered,
378 	 * ensure the provided rseq is properly aligned and valid.
379 	 */
380 	if (!IS_ALIGNED((unsigned long)rseq, __alignof__(*rseq)) ||
381 	    rseq_len != sizeof(*rseq))
382 		return -EINVAL;
383 	if (!access_ok(rseq, rseq_len))
384 		return -EFAULT;
385 	current->rseq = rseq;
386 	current->rseq_sig = sig;
387 	/*
388 	 * If rseq was previously inactive, and has just been
389 	 * registered, ensure the cpu_id_start and cpu_id fields
390 	 * are updated before returning to user-space.
391 	 */
392 	rseq_set_notify_resume(current);
393 
394 	return 0;
395 }
396