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_PREEMPT_MIGRATE_FLAGS (RSEQ_CS_FLAG_NO_RESTART_ON_MIGRATE | \
22 RSEQ_CS_FLAG_NO_RESTART_ON_PREEMPT)
23
24 /*
25 *
26 * Restartable sequences are a lightweight interface that allows
27 * user-level code to be executed atomically relative to scheduler
28 * preemption and signal delivery. Typically used for implementing
29 * per-cpu operations.
30 *
31 * It allows user-space to perform update operations on per-cpu data
32 * without requiring heavy-weight atomic operations.
33 *
34 * Detailed algorithm of rseq user-space assembly sequences:
35 *
36 * init(rseq_cs)
37 * cpu = TLS->rseq::cpu_id_start
38 * [1] TLS->rseq::rseq_cs = rseq_cs
39 * [start_ip] ----------------------------
40 * [2] if (cpu != TLS->rseq::cpu_id)
41 * goto abort_ip;
42 * [3] <last_instruction_in_cs>
43 * [post_commit_ip] ----------------------------
44 *
45 * The address of jump target abort_ip must be outside the critical
46 * region, i.e.:
47 *
48 * [abort_ip] < [start_ip] || [abort_ip] >= [post_commit_ip]
49 *
50 * Steps [2]-[3] (inclusive) need to be a sequence of instructions in
51 * userspace that can handle being interrupted between any of those
52 * instructions, and then resumed to the abort_ip.
53 *
54 * 1. Userspace stores the address of the struct rseq_cs assembly
55 * block descriptor into the rseq_cs field of the registered
56 * struct rseq TLS area. This update is performed through a single
57 * store within the inline assembly instruction sequence.
58 * [start_ip]
59 *
60 * 2. Userspace tests to check whether the current cpu_id field match
61 * the cpu number loaded before start_ip, branching to abort_ip
62 * in case of a mismatch.
63 *
64 * If the sequence is preempted or interrupted by a signal
65 * at or after start_ip and before post_commit_ip, then the kernel
66 * clears TLS->__rseq_abi::rseq_cs, and sets the user-space return
67 * ip to abort_ip before returning to user-space, so the preempted
68 * execution resumes at abort_ip.
69 *
70 * 3. Userspace critical section final instruction before
71 * post_commit_ip is the commit. The critical section is
72 * self-terminating.
73 * [post_commit_ip]
74 *
75 * 4. <success>
76 *
77 * On failure at [2], or if interrupted by preempt or signal delivery
78 * between [1] and [3]:
79 *
80 * [abort_ip]
81 * F1. <failure>
82 */
83
rseq_update_cpu_id(struct task_struct * t)84 static int rseq_update_cpu_id(struct task_struct *t)
85 {
86 u32 cpu_id = raw_smp_processor_id();
87
88 if (put_user(cpu_id, &t->rseq->cpu_id_start))
89 return -EFAULT;
90 if (put_user(cpu_id, &t->rseq->cpu_id))
91 return -EFAULT;
92 trace_rseq_update(t);
93 return 0;
94 }
95
rseq_reset_rseq_cpu_id(struct task_struct * t)96 static int rseq_reset_rseq_cpu_id(struct task_struct *t)
97 {
98 u32 cpu_id_start = 0, cpu_id = RSEQ_CPU_ID_UNINITIALIZED;
99
100 /*
101 * Reset cpu_id_start to its initial state (0).
102 */
103 if (put_user(cpu_id_start, &t->rseq->cpu_id_start))
104 return -EFAULT;
105 /*
106 * Reset cpu_id to RSEQ_CPU_ID_UNINITIALIZED, so any user coming
107 * in after unregistration can figure out that rseq needs to be
108 * registered again.
109 */
110 if (put_user(cpu_id, &t->rseq->cpu_id))
111 return -EFAULT;
112 return 0;
113 }
114
rseq_get_rseq_cs(struct task_struct * t,struct rseq_cs * rseq_cs)115 static int rseq_get_rseq_cs(struct task_struct *t, struct rseq_cs *rseq_cs)
116 {
117 struct rseq_cs __user *urseq_cs;
118 u64 ptr;
119 u32 __user *usig;
120 u32 sig;
121 int ret;
122
123 #ifdef CONFIG_64BIT
124 if (get_user(ptr, &t->rseq->rseq_cs))
125 return -EFAULT;
126 #else
127 if (copy_from_user(&ptr, &t->rseq->rseq_cs, sizeof(ptr)))
128 return -EFAULT;
129 #endif
130 if (!ptr) {
131 memset(rseq_cs, 0, sizeof(*rseq_cs));
132 return 0;
133 }
134 if (ptr >= TASK_SIZE)
135 return -EINVAL;
136 urseq_cs = (struct rseq_cs __user *)(unsigned long)ptr;
137 if (copy_from_user(rseq_cs, urseq_cs, sizeof(*rseq_cs)))
138 return -EFAULT;
139
140 if (rseq_cs->start_ip >= TASK_SIZE ||
141 rseq_cs->start_ip + rseq_cs->post_commit_offset >= TASK_SIZE ||
142 rseq_cs->abort_ip >= TASK_SIZE ||
143 rseq_cs->version > 0)
144 return -EINVAL;
145 /* Check for overflow. */
146 if (rseq_cs->start_ip + rseq_cs->post_commit_offset < rseq_cs->start_ip)
147 return -EINVAL;
148 /* Ensure that abort_ip is not in the critical section. */
149 if (rseq_cs->abort_ip - rseq_cs->start_ip < rseq_cs->post_commit_offset)
150 return -EINVAL;
151
152 usig = (u32 __user *)(unsigned long)(rseq_cs->abort_ip - sizeof(u32));
153 ret = get_user(sig, usig);
154 if (ret)
155 return ret;
156
157 if (current->rseq_sig != sig) {
158 printk_ratelimited(KERN_WARNING
159 "Possible attack attempt. Unexpected rseq signature 0x%x, expecting 0x%x (pid=%d, addr=%p).\n",
160 sig, current->rseq_sig, current->pid, usig);
161 return -EINVAL;
162 }
163 return 0;
164 }
165
rseq_need_restart(struct task_struct * t,u32 cs_flags)166 static int rseq_need_restart(struct task_struct *t, u32 cs_flags)
167 {
168 u32 flags, event_mask;
169 int ret;
170
171 /* Get thread flags. */
172 ret = get_user(flags, &t->rseq->flags);
173 if (ret)
174 return ret;
175
176 /* Take critical section flags into account. */
177 flags |= cs_flags;
178
179 /*
180 * Restart on signal can only be inhibited when restart on
181 * preempt and restart on migrate are inhibited too. Otherwise,
182 * a preempted signal handler could fail to restart the prior
183 * execution context on sigreturn.
184 */
185 if (unlikely((flags & RSEQ_CS_FLAG_NO_RESTART_ON_SIGNAL) &&
186 (flags & RSEQ_CS_PREEMPT_MIGRATE_FLAGS) !=
187 RSEQ_CS_PREEMPT_MIGRATE_FLAGS))
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
199 return !!(event_mask & ~flags);
200 }
201
clear_rseq_cs(struct task_struct * t)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 *
210 * Set rseq_cs to NULL.
211 */
212 #ifdef CONFIG_64BIT
213 return put_user(0UL, &t->rseq->rseq_cs);
214 #else
215 if (clear_user(&t->rseq->rseq_cs, sizeof(t->rseq->rseq_cs)))
216 return -EFAULT;
217 return 0;
218 #endif
219 }
220
221 /*
222 * Unsigned comparison will be true when ip >= start_ip, and when
223 * ip < start_ip + post_commit_offset.
224 */
in_rseq_cs(unsigned long ip,struct rseq_cs * rseq_cs)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
rseq_ip_fixup(struct pt_regs * regs)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 *
266 * This is how we can ensure that the entire rseq critical section
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 */
__rseq_handle_notify_resume(struct ksignal * ksig,struct pt_regs * regs)271 void __rseq_handle_notify_resume(struct ksignal *ksig, struct pt_regs *regs)
272 {
273 struct task_struct *t = current;
274 int ret, sig;
275
276 if (unlikely(t->flags & PF_EXITING))
277 return;
278 if (unlikely(!access_ok(t->rseq, sizeof(*t->rseq))))
279 goto error;
280 /*
281 * regs is NULL if and only if the caller is in a syscall path. Skip
282 * fixup and leave rseq_cs as is so that rseq_sycall() will detect and
283 * kill a misbehaving userspace on debug kernels.
284 */
285 if (regs) {
286 ret = rseq_ip_fixup(regs);
287 if (unlikely(ret < 0))
288 goto error;
289 }
290 if (unlikely(rseq_update_cpu_id(t)))
291 goto error;
292 return;
293
294 error:
295 sig = ksig ? ksig->sig : 0;
296 force_sigsegv(sig);
297 }
298
299 #ifdef CONFIG_DEBUG_RSEQ
300
301 /*
302 * Terminate the process if a syscall is issued within a restartable
303 * sequence.
304 */
rseq_syscall(struct pt_regs * regs)305 void rseq_syscall(struct pt_regs *regs)
306 {
307 unsigned long ip = instruction_pointer(regs);
308 struct task_struct *t = current;
309 struct rseq_cs rseq_cs;
310
311 if (!t->rseq)
312 return;
313 if (!access_ok(t->rseq, sizeof(*t->rseq)) ||
314 rseq_get_rseq_cs(t, &rseq_cs) || in_rseq_cs(ip, &rseq_cs))
315 force_sig(SIGSEGV);
316 }
317
318 #endif
319
320 /*
321 * sys_rseq - setup restartable sequences for caller thread.
322 */
SYSCALL_DEFINE4(rseq,struct rseq __user *,rseq,u32,rseq_len,int,flags,u32,sig)323 SYSCALL_DEFINE4(rseq, struct rseq __user *, rseq, u32, rseq_len,
324 int, flags, u32, sig)
325 {
326 int ret;
327
328 if (flags & RSEQ_FLAG_UNREGISTER) {
329 if (flags & ~RSEQ_FLAG_UNREGISTER)
330 return -EINVAL;
331 /* Unregister rseq for current thread. */
332 if (current->rseq != rseq || !current->rseq)
333 return -EINVAL;
334 if (rseq_len != sizeof(*rseq))
335 return -EINVAL;
336 if (current->rseq_sig != sig)
337 return -EPERM;
338 ret = rseq_reset_rseq_cpu_id(current);
339 if (ret)
340 return ret;
341 current->rseq = NULL;
342 current->rseq_sig = 0;
343 return 0;
344 }
345
346 if (unlikely(flags))
347 return -EINVAL;
348
349 if (current->rseq) {
350 /*
351 * If rseq is already registered, check whether
352 * the provided address differs from the prior
353 * one.
354 */
355 if (current->rseq != rseq || rseq_len != sizeof(*rseq))
356 return -EINVAL;
357 if (current->rseq_sig != sig)
358 return -EPERM;
359 /* Already registered. */
360 return -EBUSY;
361 }
362
363 /*
364 * If there was no rseq previously registered,
365 * ensure the provided rseq is properly aligned and valid.
366 */
367 if (!IS_ALIGNED((unsigned long)rseq, __alignof__(*rseq)) ||
368 rseq_len != sizeof(*rseq))
369 return -EINVAL;
370 if (!access_ok(rseq, rseq_len))
371 return -EFAULT;
372 current->rseq = rseq;
373 current->rseq_sig = sig;
374 /*
375 * If rseq was previously inactive, and has just been
376 * registered, ensure the cpu_id_start and cpu_id fields
377 * are updated before returning to user-space.
378 */
379 rseq_set_notify_resume(current);
380
381 return 0;
382 }
383