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 <linux/ratelimit.h>
17 #include <asm/ptrace.h>
18
19 #define CREATE_TRACE_POINTS
20 #include <trace/events/rseq.h>
21
22 /* The original rseq structure size (including padding) is 32 bytes. */
23 #define ORIG_RSEQ_SIZE 32
24
25 #define RSEQ_CS_NO_RESTART_FLAGS (RSEQ_CS_FLAG_NO_RESTART_ON_PREEMPT | \
26 RSEQ_CS_FLAG_NO_RESTART_ON_SIGNAL | \
27 RSEQ_CS_FLAG_NO_RESTART_ON_MIGRATE)
28
29 #ifdef CONFIG_DEBUG_RSEQ
rseq_kernel_fields(struct task_struct * t)30 static struct rseq *rseq_kernel_fields(struct task_struct *t)
31 {
32 return (struct rseq *) t->rseq_fields;
33 }
34
rseq_validate_ro_fields(struct task_struct * t)35 static int rseq_validate_ro_fields(struct task_struct *t)
36 {
37 static DEFINE_RATELIMIT_STATE(_rs,
38 DEFAULT_RATELIMIT_INTERVAL,
39 DEFAULT_RATELIMIT_BURST);
40 u32 cpu_id_start, cpu_id, node_id, mm_cid;
41 struct rseq __user *rseq = t->rseq;
42
43 /*
44 * Validate fields which are required to be read-only by
45 * user-space.
46 */
47 if (!user_read_access_begin(rseq, t->rseq_len))
48 goto efault;
49 unsafe_get_user(cpu_id_start, &rseq->cpu_id_start, efault_end);
50 unsafe_get_user(cpu_id, &rseq->cpu_id, efault_end);
51 unsafe_get_user(node_id, &rseq->node_id, efault_end);
52 unsafe_get_user(mm_cid, &rseq->mm_cid, efault_end);
53 user_read_access_end();
54
55 if ((cpu_id_start != rseq_kernel_fields(t)->cpu_id_start ||
56 cpu_id != rseq_kernel_fields(t)->cpu_id ||
57 node_id != rseq_kernel_fields(t)->node_id ||
58 mm_cid != rseq_kernel_fields(t)->mm_cid) && __ratelimit(&_rs)) {
59
60 pr_warn("Detected rseq corruption for pid: %d, name: %s\n"
61 "\tcpu_id_start: %u ?= %u\n"
62 "\tcpu_id: %u ?= %u\n"
63 "\tnode_id: %u ?= %u\n"
64 "\tmm_cid: %u ?= %u\n",
65 t->pid, t->comm,
66 cpu_id_start, rseq_kernel_fields(t)->cpu_id_start,
67 cpu_id, rseq_kernel_fields(t)->cpu_id,
68 node_id, rseq_kernel_fields(t)->node_id,
69 mm_cid, rseq_kernel_fields(t)->mm_cid);
70 }
71
72 /* For now, only print a console warning on mismatch. */
73 return 0;
74
75 efault_end:
76 user_read_access_end();
77 efault:
78 return -EFAULT;
79 }
80
81 /*
82 * Update an rseq field and its in-kernel copy in lock-step to keep a coherent
83 * state.
84 */
85 #define rseq_unsafe_put_user(t, value, field, error_label) \
86 do { \
87 unsafe_put_user(value, &t->rseq->field, error_label); \
88 rseq_kernel_fields(t)->field = value; \
89 } while (0)
90
91 #else
rseq_validate_ro_fields(struct task_struct * t)92 static int rseq_validate_ro_fields(struct task_struct *t)
93 {
94 return 0;
95 }
96
97 #define rseq_unsafe_put_user(t, value, field, error_label) \
98 unsafe_put_user(value, &t->rseq->field, error_label)
99 #endif
100
101 /*
102 *
103 * Restartable sequences are a lightweight interface that allows
104 * user-level code to be executed atomically relative to scheduler
105 * preemption and signal delivery. Typically used for implementing
106 * per-cpu operations.
107 *
108 * It allows user-space to perform update operations on per-cpu data
109 * without requiring heavy-weight atomic operations.
110 *
111 * Detailed algorithm of rseq user-space assembly sequences:
112 *
113 * init(rseq_cs)
114 * cpu = TLS->rseq::cpu_id_start
115 * [1] TLS->rseq::rseq_cs = rseq_cs
116 * [start_ip] ----------------------------
117 * [2] if (cpu != TLS->rseq::cpu_id)
118 * goto abort_ip;
119 * [3] <last_instruction_in_cs>
120 * [post_commit_ip] ----------------------------
121 *
122 * The address of jump target abort_ip must be outside the critical
123 * region, i.e.:
124 *
125 * [abort_ip] < [start_ip] || [abort_ip] >= [post_commit_ip]
126 *
127 * Steps [2]-[3] (inclusive) need to be a sequence of instructions in
128 * userspace that can handle being interrupted between any of those
129 * instructions, and then resumed to the abort_ip.
130 *
131 * 1. Userspace stores the address of the struct rseq_cs assembly
132 * block descriptor into the rseq_cs field of the registered
133 * struct rseq TLS area. This update is performed through a single
134 * store within the inline assembly instruction sequence.
135 * [start_ip]
136 *
137 * 2. Userspace tests to check whether the current cpu_id field match
138 * the cpu number loaded before start_ip, branching to abort_ip
139 * in case of a mismatch.
140 *
141 * If the sequence is preempted or interrupted by a signal
142 * at or after start_ip and before post_commit_ip, then the kernel
143 * clears TLS->__rseq_abi::rseq_cs, and sets the user-space return
144 * ip to abort_ip before returning to user-space, so the preempted
145 * execution resumes at abort_ip.
146 *
147 * 3. Userspace critical section final instruction before
148 * post_commit_ip is the commit. The critical section is
149 * self-terminating.
150 * [post_commit_ip]
151 *
152 * 4. <success>
153 *
154 * On failure at [2], or if interrupted by preempt or signal delivery
155 * between [1] and [3]:
156 *
157 * [abort_ip]
158 * F1. <failure>
159 */
160
rseq_update_cpu_node_id(struct task_struct * t)161 static int rseq_update_cpu_node_id(struct task_struct *t)
162 {
163 struct rseq __user *rseq = t->rseq;
164 u32 cpu_id = raw_smp_processor_id();
165 u32 node_id = cpu_to_node(cpu_id);
166 u32 mm_cid = task_mm_cid(t);
167
168 /*
169 * Validate read-only rseq fields.
170 */
171 if (rseq_validate_ro_fields(t))
172 goto efault;
173 WARN_ON_ONCE((int) mm_cid < 0);
174 if (!user_write_access_begin(rseq, t->rseq_len))
175 goto efault;
176
177 rseq_unsafe_put_user(t, cpu_id, cpu_id_start, efault_end);
178 rseq_unsafe_put_user(t, cpu_id, cpu_id, efault_end);
179 rseq_unsafe_put_user(t, node_id, node_id, efault_end);
180 rseq_unsafe_put_user(t, mm_cid, mm_cid, efault_end);
181
182 /*
183 * Additional feature fields added after ORIG_RSEQ_SIZE
184 * need to be conditionally updated only if
185 * t->rseq_len != ORIG_RSEQ_SIZE.
186 */
187 user_write_access_end();
188 trace_rseq_update(t);
189 return 0;
190
191 efault_end:
192 user_write_access_end();
193 efault:
194 return -EFAULT;
195 }
196
rseq_reset_rseq_cpu_node_id(struct task_struct * t)197 static int rseq_reset_rseq_cpu_node_id(struct task_struct *t)
198 {
199 struct rseq __user *rseq = t->rseq;
200 u32 cpu_id_start = 0, cpu_id = RSEQ_CPU_ID_UNINITIALIZED, node_id = 0,
201 mm_cid = 0;
202
203 /*
204 * Validate read-only rseq fields.
205 */
206 if (rseq_validate_ro_fields(t))
207 goto efault;
208
209 if (!user_write_access_begin(rseq, t->rseq_len))
210 goto efault;
211
212 /*
213 * Reset all fields to their initial state.
214 *
215 * All fields have an initial state of 0 except cpu_id which is set to
216 * RSEQ_CPU_ID_UNINITIALIZED, so that any user coming in after
217 * unregistration can figure out that rseq needs to be registered
218 * again.
219 */
220 rseq_unsafe_put_user(t, cpu_id_start, cpu_id_start, efault_end);
221 rseq_unsafe_put_user(t, cpu_id, cpu_id, efault_end);
222 rseq_unsafe_put_user(t, node_id, node_id, efault_end);
223 rseq_unsafe_put_user(t, mm_cid, mm_cid, efault_end);
224
225 /*
226 * Additional feature fields added after ORIG_RSEQ_SIZE
227 * need to be conditionally reset only if
228 * t->rseq_len != ORIG_RSEQ_SIZE.
229 */
230 user_write_access_end();
231 return 0;
232
233 efault_end:
234 user_write_access_end();
235 efault:
236 return -EFAULT;
237 }
238
239 /*
240 * Get the user-space pointer value stored in the 'rseq_cs' field.
241 */
rseq_get_rseq_cs_ptr_val(struct rseq __user * rseq,u64 * rseq_cs)242 static int rseq_get_rseq_cs_ptr_val(struct rseq __user *rseq, u64 *rseq_cs)
243 {
244 if (!rseq_cs)
245 return -EFAULT;
246
247 #ifdef CONFIG_64BIT
248 if (get_user(*rseq_cs, &rseq->rseq_cs))
249 return -EFAULT;
250 #else
251 if (copy_from_user(rseq_cs, &rseq->rseq_cs, sizeof(*rseq_cs)))
252 return -EFAULT;
253 #endif
254
255 return 0;
256 }
257
258 /*
259 * If the rseq_cs field of 'struct rseq' contains a valid pointer to
260 * user-space, copy 'struct rseq_cs' from user-space and validate its fields.
261 */
rseq_get_rseq_cs(struct task_struct * t,struct rseq_cs * rseq_cs)262 static int rseq_get_rseq_cs(struct task_struct *t, struct rseq_cs *rseq_cs)
263 {
264 struct rseq_cs __user *urseq_cs;
265 u64 ptr;
266 u32 __user *usig;
267 u32 sig;
268 int ret;
269
270 ret = rseq_get_rseq_cs_ptr_val(t->rseq, &ptr);
271 if (ret)
272 return ret;
273
274 /* If the rseq_cs pointer is NULL, return a cleared struct rseq_cs. */
275 if (!ptr) {
276 memset(rseq_cs, 0, sizeof(*rseq_cs));
277 return 0;
278 }
279 /* Check that the pointer value fits in the user-space process space. */
280 if (ptr >= TASK_SIZE)
281 return -EINVAL;
282 urseq_cs = (struct rseq_cs __user *)(unsigned long)ptr;
283 if (copy_from_user(rseq_cs, urseq_cs, sizeof(*rseq_cs)))
284 return -EFAULT;
285
286 if (rseq_cs->start_ip >= TASK_SIZE ||
287 rseq_cs->start_ip + rseq_cs->post_commit_offset >= TASK_SIZE ||
288 rseq_cs->abort_ip >= TASK_SIZE ||
289 rseq_cs->version > 0)
290 return -EINVAL;
291 /* Check for overflow. */
292 if (rseq_cs->start_ip + rseq_cs->post_commit_offset < rseq_cs->start_ip)
293 return -EINVAL;
294 /* Ensure that abort_ip is not in the critical section. */
295 if (rseq_cs->abort_ip - rseq_cs->start_ip < rseq_cs->post_commit_offset)
296 return -EINVAL;
297
298 usig = (u32 __user *)(unsigned long)(rseq_cs->abort_ip - sizeof(u32));
299 ret = get_user(sig, usig);
300 if (ret)
301 return ret;
302
303 if (current->rseq_sig != sig) {
304 printk_ratelimited(KERN_WARNING
305 "Possible attack attempt. Unexpected rseq signature 0x%x, expecting 0x%x (pid=%d, addr=%p).\n",
306 sig, current->rseq_sig, current->pid, usig);
307 return -EINVAL;
308 }
309 return 0;
310 }
311
rseq_warn_flags(const char * str,u32 flags)312 static bool rseq_warn_flags(const char *str, u32 flags)
313 {
314 u32 test_flags;
315
316 if (!flags)
317 return false;
318 test_flags = flags & RSEQ_CS_NO_RESTART_FLAGS;
319 if (test_flags)
320 pr_warn_once("Deprecated flags (%u) in %s ABI structure", test_flags, str);
321 test_flags = flags & ~RSEQ_CS_NO_RESTART_FLAGS;
322 if (test_flags)
323 pr_warn_once("Unknown flags (%u) in %s ABI structure", test_flags, str);
324 return true;
325 }
326
rseq_need_restart(struct task_struct * t,u32 cs_flags)327 static int rseq_need_restart(struct task_struct *t, u32 cs_flags)
328 {
329 u32 flags, event_mask;
330 int ret;
331
332 if (rseq_warn_flags("rseq_cs", cs_flags))
333 return -EINVAL;
334
335 /* Get thread flags. */
336 ret = get_user(flags, &t->rseq->flags);
337 if (ret)
338 return ret;
339
340 if (rseq_warn_flags("rseq", flags))
341 return -EINVAL;
342
343 /*
344 * Load and clear event mask atomically with respect to
345 * scheduler preemption.
346 */
347 preempt_disable();
348 event_mask = t->rseq_event_mask;
349 t->rseq_event_mask = 0;
350 preempt_enable();
351
352 return !!event_mask;
353 }
354
clear_rseq_cs(struct rseq __user * rseq)355 static int clear_rseq_cs(struct rseq __user *rseq)
356 {
357 /*
358 * The rseq_cs field is set to NULL on preemption or signal
359 * delivery on top of rseq assembly block, as well as on top
360 * of code outside of the rseq assembly block. This performs
361 * a lazy clear of the rseq_cs field.
362 *
363 * Set rseq_cs to NULL.
364 */
365 #ifdef CONFIG_64BIT
366 return put_user(0UL, &rseq->rseq_cs);
367 #else
368 if (clear_user(&rseq->rseq_cs, sizeof(rseq->rseq_cs)))
369 return -EFAULT;
370 return 0;
371 #endif
372 }
373
374 /*
375 * Unsigned comparison will be true when ip >= start_ip, and when
376 * ip < start_ip + post_commit_offset.
377 */
in_rseq_cs(unsigned long ip,struct rseq_cs * rseq_cs)378 static bool in_rseq_cs(unsigned long ip, struct rseq_cs *rseq_cs)
379 {
380 return ip - rseq_cs->start_ip < rseq_cs->post_commit_offset;
381 }
382
rseq_ip_fixup(struct pt_regs * regs)383 static int rseq_ip_fixup(struct pt_regs *regs)
384 {
385 unsigned long ip = instruction_pointer(regs);
386 struct task_struct *t = current;
387 struct rseq_cs rseq_cs;
388 int ret;
389
390 ret = rseq_get_rseq_cs(t, &rseq_cs);
391 if (ret)
392 return ret;
393
394 /*
395 * Handle potentially not being within a critical section.
396 * If not nested over a rseq critical section, restart is useless.
397 * Clear the rseq_cs pointer and return.
398 */
399 if (!in_rseq_cs(ip, &rseq_cs))
400 return clear_rseq_cs(t->rseq);
401 ret = rseq_need_restart(t, rseq_cs.flags);
402 if (ret <= 0)
403 return ret;
404 ret = clear_rseq_cs(t->rseq);
405 if (ret)
406 return ret;
407 trace_rseq_ip_fixup(ip, rseq_cs.start_ip, rseq_cs.post_commit_offset,
408 rseq_cs.abort_ip);
409 instruction_pointer_set(regs, (unsigned long)rseq_cs.abort_ip);
410 return 0;
411 }
412
413 /*
414 * This resume handler must always be executed between any of:
415 * - preemption,
416 * - signal delivery,
417 * and return to user-space.
418 *
419 * This is how we can ensure that the entire rseq critical section
420 * will issue the commit instruction only if executed atomically with
421 * respect to other threads scheduled on the same CPU, and with respect
422 * to signal handlers.
423 */
__rseq_handle_notify_resume(struct ksignal * ksig,struct pt_regs * regs)424 void __rseq_handle_notify_resume(struct ksignal *ksig, struct pt_regs *regs)
425 {
426 struct task_struct *t = current;
427 int ret, sig;
428
429 if (unlikely(t->flags & PF_EXITING))
430 return;
431
432 /*
433 * regs is NULL if and only if the caller is in a syscall path. Skip
434 * fixup and leave rseq_cs as is so that rseq_sycall() will detect and
435 * kill a misbehaving userspace on debug kernels.
436 */
437 if (regs) {
438 ret = rseq_ip_fixup(regs);
439 if (unlikely(ret < 0))
440 goto error;
441 }
442 if (unlikely(rseq_update_cpu_node_id(t)))
443 goto error;
444 return;
445
446 error:
447 sig = ksig ? ksig->sig : 0;
448 force_sigsegv(sig);
449 }
450
451 #ifdef CONFIG_DEBUG_RSEQ
452
453 /*
454 * Terminate the process if a syscall is issued within a restartable
455 * sequence.
456 */
rseq_syscall(struct pt_regs * regs)457 void rseq_syscall(struct pt_regs *regs)
458 {
459 unsigned long ip = instruction_pointer(regs);
460 struct task_struct *t = current;
461 struct rseq_cs rseq_cs;
462
463 if (!t->rseq)
464 return;
465 if (rseq_get_rseq_cs(t, &rseq_cs) || in_rseq_cs(ip, &rseq_cs))
466 force_sig(SIGSEGV);
467 }
468
469 #endif
470
471 /*
472 * sys_rseq - setup restartable sequences for caller thread.
473 */
SYSCALL_DEFINE4(rseq,struct rseq __user *,rseq,u32,rseq_len,int,flags,u32,sig)474 SYSCALL_DEFINE4(rseq, struct rseq __user *, rseq, u32, rseq_len,
475 int, flags, u32, sig)
476 {
477 int ret;
478 u64 rseq_cs;
479
480 if (flags & RSEQ_FLAG_UNREGISTER) {
481 if (flags & ~RSEQ_FLAG_UNREGISTER)
482 return -EINVAL;
483 /* Unregister rseq for current thread. */
484 if (current->rseq != rseq || !current->rseq)
485 return -EINVAL;
486 if (rseq_len != current->rseq_len)
487 return -EINVAL;
488 if (current->rseq_sig != sig)
489 return -EPERM;
490 ret = rseq_reset_rseq_cpu_node_id(current);
491 if (ret)
492 return ret;
493 current->rseq = NULL;
494 current->rseq_sig = 0;
495 current->rseq_len = 0;
496 return 0;
497 }
498
499 if (unlikely(flags))
500 return -EINVAL;
501
502 if (current->rseq) {
503 /*
504 * If rseq is already registered, check whether
505 * the provided address differs from the prior
506 * one.
507 */
508 if (current->rseq != rseq || rseq_len != current->rseq_len)
509 return -EINVAL;
510 if (current->rseq_sig != sig)
511 return -EPERM;
512 /* Already registered. */
513 return -EBUSY;
514 }
515
516 /*
517 * If there was no rseq previously registered, ensure the provided rseq
518 * is properly aligned, as communcated to user-space through the ELF
519 * auxiliary vector AT_RSEQ_ALIGN. If rseq_len is the original rseq
520 * size, the required alignment is the original struct rseq alignment.
521 *
522 * In order to be valid, rseq_len is either the original rseq size, or
523 * large enough to contain all supported fields, as communicated to
524 * user-space through the ELF auxiliary vector AT_RSEQ_FEATURE_SIZE.
525 */
526 if (rseq_len < ORIG_RSEQ_SIZE ||
527 (rseq_len == ORIG_RSEQ_SIZE && !IS_ALIGNED((unsigned long)rseq, ORIG_RSEQ_SIZE)) ||
528 (rseq_len != ORIG_RSEQ_SIZE && (!IS_ALIGNED((unsigned long)rseq, __alignof__(*rseq)) ||
529 rseq_len < offsetof(struct rseq, end))))
530 return -EINVAL;
531 if (!access_ok(rseq, rseq_len))
532 return -EFAULT;
533
534 /*
535 * If the rseq_cs pointer is non-NULL on registration, clear it to
536 * avoid a potential segfault on return to user-space. The proper thing
537 * to do would have been to fail the registration but this would break
538 * older libcs that reuse the rseq area for new threads without
539 * clearing the fields.
540 */
541 if (rseq_get_rseq_cs_ptr_val(rseq, &rseq_cs))
542 return -EFAULT;
543 if (rseq_cs && clear_rseq_cs(rseq))
544 return -EFAULT;
545
546 #ifdef CONFIG_DEBUG_RSEQ
547 /*
548 * Initialize the in-kernel rseq fields copy for validation of
549 * read-only fields.
550 */
551 if (get_user(rseq_kernel_fields(current)->cpu_id_start, &rseq->cpu_id_start) ||
552 get_user(rseq_kernel_fields(current)->cpu_id, &rseq->cpu_id) ||
553 get_user(rseq_kernel_fields(current)->node_id, &rseq->node_id) ||
554 get_user(rseq_kernel_fields(current)->mm_cid, &rseq->mm_cid))
555 return -EFAULT;
556 #endif
557 /*
558 * Activate the registration by setting the rseq area address, length
559 * and signature in the task struct.
560 */
561 current->rseq = rseq;
562 current->rseq_len = rseq_len;
563 current->rseq_sig = sig;
564
565 /*
566 * If rseq was previously inactive, and has just been
567 * registered, ensure the cpu_id_start and cpu_id fields
568 * are updated before returning to user-space.
569 */
570 rseq_set_notify_resume(current);
571
572 return 0;
573 }
574