1 // SPDX-License-Identifier: GPL-2.0+
3 * Restartable sequences system call
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>
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>
18 #define CREATE_TRACE_POINTS
19 #include <trace/events/rseq.h>
21 #define RSEQ_CS_PREEMPT_MIGRATE_FLAGS (RSEQ_CS_FLAG_NO_RESTART_ON_MIGRATE | \
22 RSEQ_CS_FLAG_NO_RESTART_ON_PREEMPT)
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
31 * It allows user-space to perform update operations on per-cpu data
32 * without requiring heavy-weight atomic operations.
34 * Detailed algorithm of rseq user-space assembly sequences:
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)
42 * [3] <last_instruction_in_cs>
43 * [post_commit_ip] ----------------------------
45 * The address of jump target abort_ip must be outside the critical
48 * [abort_ip] < [start_ip] || [abort_ip] >= [post_commit_ip]
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.
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.
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.
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.
70 * 3. Userspace critical section final instruction before
71 * post_commit_ip is the commit. The critical section is
77 * On failure at [2], or if interrupted by preempt or signal delivery
78 * between [1] and [3]:
84 static int rseq_update_cpu_id(struct task_struct *t)
86 u32 cpu_id = raw_smp_processor_id();
88 if (put_user(cpu_id, &t->rseq->cpu_id_start))
90 if (put_user(cpu_id, &t->rseq->cpu_id))
96 static int rseq_reset_rseq_cpu_id(struct task_struct *t)
98 u32 cpu_id_start = 0, cpu_id = RSEQ_CPU_ID_UNINITIALIZED;
101 * Reset cpu_id_start to its initial state (0).
103 if (put_user(cpu_id_start, &t->rseq->cpu_id_start))
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
110 if (put_user(cpu_id, &t->rseq->cpu_id))
115 static int rseq_get_rseq_cs(struct task_struct *t, struct rseq_cs *rseq_cs)
117 struct rseq_cs __user *urseq_cs;
123 ret = get_user(ptr, &t->rseq->rseq_cs);
127 memset(rseq_cs, 0, sizeof(*rseq_cs));
130 urseq_cs = (struct rseq_cs __user *)ptr;
131 if (copy_from_user(rseq_cs, urseq_cs, sizeof(*rseq_cs)))
134 if (rseq_cs->start_ip >= TASK_SIZE ||
135 rseq_cs->start_ip + rseq_cs->post_commit_offset >= TASK_SIZE ||
136 rseq_cs->abort_ip >= TASK_SIZE ||
137 rseq_cs->version > 0)
139 /* Check for overflow. */
140 if (rseq_cs->start_ip + rseq_cs->post_commit_offset < rseq_cs->start_ip)
142 /* Ensure that abort_ip is not in the critical section. */
143 if (rseq_cs->abort_ip - rseq_cs->start_ip < rseq_cs->post_commit_offset)
146 usig = (u32 __user *)(unsigned long)(rseq_cs->abort_ip - sizeof(u32));
147 ret = get_user(sig, usig);
151 if (current->rseq_sig != sig) {
152 printk_ratelimited(KERN_WARNING
153 "Possible attack attempt. Unexpected rseq signature 0x%x, expecting 0x%x (pid=%d, addr=%p).\n",
154 sig, current->rseq_sig, current->pid, usig);
160 static int rseq_need_restart(struct task_struct *t, u32 cs_flags)
162 u32 flags, event_mask;
165 /* Get thread flags. */
166 ret = get_user(flags, &t->rseq->flags);
170 /* Take critical section flags into account. */
174 * Restart on signal can only be inhibited when restart on
175 * preempt and restart on migrate are inhibited too. Otherwise,
176 * a preempted signal handler could fail to restart the prior
177 * execution context on sigreturn.
179 if (unlikely((flags & RSEQ_CS_FLAG_NO_RESTART_ON_SIGNAL) &&
180 (flags & RSEQ_CS_PREEMPT_MIGRATE_FLAGS) !=
181 RSEQ_CS_PREEMPT_MIGRATE_FLAGS))
185 * Load and clear event mask atomically with respect to
186 * scheduler preemption.
189 event_mask = t->rseq_event_mask;
190 t->rseq_event_mask = 0;
193 return !!(event_mask & ~flags);
196 static int clear_rseq_cs(struct task_struct *t)
199 * The rseq_cs field is set to NULL on preemption or signal
200 * delivery on top of rseq assembly block, as well as on top
201 * of code outside of the rseq assembly block. This performs
202 * a lazy clear of the rseq_cs field.
204 * Set rseq_cs to NULL with single-copy atomicity.
206 return put_user(0UL, &t->rseq->rseq_cs);
210 * Unsigned comparison will be true when ip >= start_ip, and when
211 * ip < start_ip + post_commit_offset.
213 static bool in_rseq_cs(unsigned long ip, struct rseq_cs *rseq_cs)
215 return ip - rseq_cs->start_ip < rseq_cs->post_commit_offset;
218 static int rseq_ip_fixup(struct pt_regs *regs)
220 unsigned long ip = instruction_pointer(regs);
221 struct task_struct *t = current;
222 struct rseq_cs rseq_cs;
225 ret = rseq_get_rseq_cs(t, &rseq_cs);
230 * Handle potentially not being within a critical section.
231 * If not nested over a rseq critical section, restart is useless.
232 * Clear the rseq_cs pointer and return.
234 if (!in_rseq_cs(ip, &rseq_cs))
235 return clear_rseq_cs(t);
236 ret = rseq_need_restart(t, rseq_cs.flags);
239 ret = clear_rseq_cs(t);
242 trace_rseq_ip_fixup(ip, rseq_cs.start_ip, rseq_cs.post_commit_offset,
244 instruction_pointer_set(regs, (unsigned long)rseq_cs.abort_ip);
249 * This resume handler must always be executed between any of:
252 * and return to user-space.
254 * This is how we can ensure that the entire rseq critical section,
255 * consisting of both the C part and the assembly instruction sequence,
256 * will issue the commit instruction only if executed atomically with
257 * respect to other threads scheduled on the same CPU, and with respect
258 * to signal handlers.
260 void __rseq_handle_notify_resume(struct ksignal *ksig, struct pt_regs *regs)
262 struct task_struct *t = current;
265 if (unlikely(t->flags & PF_EXITING))
267 if (unlikely(!access_ok(VERIFY_WRITE, t->rseq, sizeof(*t->rseq))))
269 ret = rseq_ip_fixup(regs);
270 if (unlikely(ret < 0))
272 if (unlikely(rseq_update_cpu_id(t)))
277 sig = ksig ? ksig->sig : 0;
278 force_sigsegv(sig, t);
281 #ifdef CONFIG_DEBUG_RSEQ
284 * Terminate the process if a syscall is issued within a restartable
287 void rseq_syscall(struct pt_regs *regs)
289 unsigned long ip = instruction_pointer(regs);
290 struct task_struct *t = current;
291 struct rseq_cs rseq_cs;
295 if (!access_ok(VERIFY_READ, t->rseq, sizeof(*t->rseq)) ||
296 rseq_get_rseq_cs(t, &rseq_cs) || in_rseq_cs(ip, &rseq_cs))
297 force_sig(SIGSEGV, t);
303 * sys_rseq - setup restartable sequences for caller thread.
305 SYSCALL_DEFINE4(rseq, struct rseq __user *, rseq, u32, rseq_len,
306 int, flags, u32, sig)
310 if (flags & RSEQ_FLAG_UNREGISTER) {
311 /* Unregister rseq for current thread. */
312 if (current->rseq != rseq || !current->rseq)
314 if (current->rseq_len != rseq_len)
316 if (current->rseq_sig != sig)
318 ret = rseq_reset_rseq_cpu_id(current);
321 current->rseq = NULL;
322 current->rseq_len = 0;
323 current->rseq_sig = 0;
332 * If rseq is already registered, check whether
333 * the provided address differs from the prior
336 if (current->rseq != rseq || current->rseq_len != rseq_len)
338 if (current->rseq_sig != sig)
340 /* Already registered. */
345 * If there was no rseq previously registered,
346 * ensure the provided rseq is properly aligned and valid.
348 if (!IS_ALIGNED((unsigned long)rseq, __alignof__(*rseq)) ||
349 rseq_len != sizeof(*rseq))
351 if (!access_ok(VERIFY_WRITE, rseq, rseq_len))
353 current->rseq = rseq;
354 current->rseq_len = rseq_len;
355 current->rseq_sig = sig;
357 * If rseq was previously inactive, and has just been
358 * registered, ensure the cpu_id_start and cpu_id fields
359 * are updated before returning to user-space.
361 rseq_set_notify_resume(current);