| 1 | /* |
| 2 | * Restartable sequences system call |
| 3 | * |
| 4 | * Restartable sequences are a lightweight interface that allows |
| 5 | * user-level code to be executed atomically relative to scheduler |
| 6 | * preemption and signal delivery. Typically used for implementing |
| 7 | * per-cpu operations. |
| 8 | * |
| 9 | * It allows user-space to perform update operations on per-cpu data |
| 10 | * without requiring heavy-weight atomic operations. |
| 11 | * |
| 12 | * This program is free software; you can redistribute it and/or modify |
| 13 | * it under the terms of the GNU General Public License as published by |
| 14 | * the Free Software Foundation; either version 2 of the License, or |
| 15 | * (at your option) any later version. |
| 16 | * |
| 17 | * This program is distributed in the hope that it will be useful, |
| 18 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 19 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
| 20 | * GNU General Public License for more details. |
| 21 | * |
| 22 | * Copyright (C) 2015, Google, Inc., |
| 23 | * Paul Turner <pjt@google.com> and Andrew Hunter <ahh@google.com> |
| 24 | * Copyright (C) 2015-2016, EfficiOS Inc., |
| 25 | * Mathieu Desnoyers <mathieu.desnoyers@efficios.com> |
| 26 | */ |
| 27 | |
| 28 | #include <linux/sched.h> |
| 29 | #include <linux/uaccess.h> |
| 30 | #include <linux/syscalls.h> |
| 31 | #include <linux/rseq.h> |
| 32 | #include <linux/types.h> |
| 33 | #include <asm/ptrace.h> |
| 34 | |
| 35 | #define CREATE_TRACE_POINTS |
| 36 | #include <trace/events/rseq.h> |
| 37 | |
| 38 | /* |
| 39 | * The restartable sequences mechanism is the overlap of two distinct |
| 40 | * restart mechanisms: a sequence counter tracking preemption and signal |
| 41 | * delivery for high-level code, and an ip-fixup-based mechanism for the |
| 42 | * final assembly instruction sequence. |
| 43 | * |
| 44 | * A high-level summary of the algorithm to use rseq from user-space is |
| 45 | * as follows: |
| 46 | * |
| 47 | * The high-level code between rseq_start() and rseq_finish() loads the |
| 48 | * current value of the sequence counter in rseq_start(), and then it |
| 49 | * gets compared with the new current value within the rseq_finish() |
| 50 | * restartable instruction sequence. Between rseq_start() and |
| 51 | * rseq_finish(), the high-level code can perform operations that do not |
| 52 | * have side-effects, such as getting the current CPU number, and |
| 53 | * loading from variables. |
| 54 | * |
| 55 | * Stores are performed at the very end of the restartable sequence |
| 56 | * assembly block. Each assembly block within rseq_finish() defines a |
| 57 | * "struct rseq_cs" structure which describes the start_ip and |
| 58 | * post_commit_ip addresses, as well as the abort_ip address where the |
| 59 | * kernel should move the thread instruction pointer if a rseq critical |
| 60 | * section assembly block is preempted or if a signal is delivered on |
| 61 | * top of a rseq critical section assembly block. |
| 62 | * |
| 63 | * Detailed algorithm of rseq use: |
| 64 | * |
| 65 | * rseq_start() |
| 66 | * |
| 67 | * 0. Userspace loads the current event counter value from the |
| 68 | * event_counter field of the registered struct rseq TLS area, |
| 69 | * |
| 70 | * rseq_finish() |
| 71 | * |
| 72 | * Steps [1]-[3] (inclusive) need to be a sequence of instructions in |
| 73 | * userspace that can handle being moved to the abort_ip between any |
| 74 | * of those instructions. |
| 75 | * |
| 76 | * The abort_ip address needs to be less than start_ip, or |
| 77 | * greater-or-equal the post_commit_ip. Step [4] and the failure |
| 78 | * code step [F1] need to be at addresses lesser than start_ip, or |
| 79 | * greater-or-equal the post_commit_ip. |
| 80 | * |
| 81 | * [start_ip] |
| 82 | * 1. Userspace stores the address of the struct rseq_cs assembly |
| 83 | * block descriptor into the rseq_cs field of the registered |
| 84 | * struct rseq TLS area. This update is performed through a single |
| 85 | * store, followed by a compiler barrier which prevents the |
| 86 | * compiler from moving following loads or stores before this |
| 87 | * store. |
| 88 | * |
| 89 | * 2. Userspace tests to see whether the current event counter value |
| 90 | * match the value loaded at [0]. Manually jumping to [F1] in case |
| 91 | * of a mismatch. |
| 92 | * |
| 93 | * Note that if we are preempted or interrupted by a signal |
| 94 | * after [1] and before post_commit_ip, then the kernel also |
| 95 | * performs the comparison performed in [2], and conditionally |
| 96 | * clears the rseq_cs field of struct rseq, then jumps us to |
| 97 | * abort_ip. |
| 98 | * |
| 99 | * 3. Userspace critical section final instruction before |
| 100 | * post_commit_ip is the commit. The critical section is |
| 101 | * self-terminating. |
| 102 | * [post_commit_ip] |
| 103 | * |
| 104 | * 4. Userspace clears the rseq_cs field of the struct rseq |
| 105 | * TLS area. |
| 106 | * |
| 107 | * 5. Return true. |
| 108 | * |
| 109 | * On failure at [2]: |
| 110 | * |
| 111 | * F1. Userspace clears the rseq_cs field of the struct rseq |
| 112 | * TLS area. Followed by step [F2]. |
| 113 | * |
| 114 | * [abort_ip] |
| 115 | * F2. Return false. |
| 116 | */ |
| 117 | |
| 118 | /* |
| 119 | * The rseq_event_counter allow user-space to detect preemption and |
| 120 | * signal delivery. It increments at least once before returning to |
| 121 | * user-space if a thread is preempted or has a signal delivered. It is |
| 122 | * not meant to be an exact counter of such events. |
| 123 | * |
| 124 | * Overflow of the event counter is not a problem in practice. It |
| 125 | * increments at most once between each user-space thread instruction |
| 126 | * executed, so we would need a thread to execute 2^32 instructions or |
| 127 | * more between rseq_start() and rseq_finish(), while single-stepping, |
| 128 | * for this to be an issue. |
| 129 | * |
| 130 | * On 64-bit architectures, both cpu_id and event_counter can be updated |
| 131 | * with a single 64-bit store. On 32-bit architectures, __put_user() is |
| 132 | * expected to perform two 32-bit single-copy stores to guarantee |
| 133 | * single-copy atomicity semantics for other threads. |
| 134 | */ |
| 135 | static bool rseq_update_cpu_id_event_counter(struct task_struct *t) |
| 136 | { |
| 137 | union rseq_cpu_event u; |
| 138 | |
| 139 | u.e.cpu_id = raw_smp_processor_id(); |
| 140 | u.e.event_counter = ++t->rseq_event_counter; |
| 141 | if (__put_user(u.v, &t->rseq->u.v)) |
| 142 | return false; |
| 143 | trace_rseq_update(t); |
| 144 | return true; |
| 145 | } |
| 146 | |
| 147 | static bool rseq_get_rseq_cs(struct task_struct *t, |
| 148 | void __user **start_ip, |
| 149 | void __user **post_commit_ip, |
| 150 | void __user **abort_ip) |
| 151 | { |
| 152 | unsigned long ptr; |
| 153 | struct rseq_cs __user *urseq_cs; |
| 154 | struct rseq_cs rseq_cs; |
| 155 | |
| 156 | if (__get_user(ptr, &t->rseq->rseq_cs)) |
| 157 | return false; |
| 158 | if (!ptr) |
| 159 | return true; |
| 160 | urseq_cs = (struct rseq_cs __user *)ptr; |
| 161 | if (copy_from_user(&rseq_cs, urseq_cs, sizeof(rseq_cs))) |
| 162 | return false; |
| 163 | *start_ip = (void __user *)rseq_cs.start_ip; |
| 164 | *post_commit_ip = (void __user *)rseq_cs.post_commit_ip; |
| 165 | *abort_ip = (void __user *)rseq_cs.abort_ip; |
| 166 | return true; |
| 167 | } |
| 168 | |
| 169 | static bool rseq_ip_fixup(struct pt_regs *regs) |
| 170 | { |
| 171 | struct task_struct *t = current; |
| 172 | void __user *start_ip = NULL; |
| 173 | void __user *post_commit_ip = NULL; |
| 174 | void __user *abort_ip = NULL; |
| 175 | bool ret; |
| 176 | |
| 177 | ret = rseq_get_rseq_cs(t, &start_ip, &post_commit_ip, &abort_ip); |
| 178 | trace_rseq_ip_fixup((void __user *)instruction_pointer(regs), |
| 179 | start_ip, post_commit_ip, abort_ip, t->rseq_event_counter, |
| 180 | ret); |
| 181 | if (!ret) |
| 182 | return false; |
| 183 | |
| 184 | /* Handle potentially not being within a critical section. */ |
| 185 | if ((void __user *)instruction_pointer(regs) >= post_commit_ip || |
| 186 | (void __user *)instruction_pointer(regs) < start_ip) |
| 187 | return true; |
| 188 | |
| 189 | /* |
| 190 | * We need to clear rseq_cs upon entry into a signal |
| 191 | * handler nested on top of a rseq assembly block, so |
| 192 | * the signal handler will not be fixed up if itself |
| 193 | * interrupted by a nested signal handler or preempted. |
| 194 | */ |
| 195 | if (clear_user(&t->rseq->rseq_cs, sizeof(t->rseq->rseq_cs))) |
| 196 | return false; |
| 197 | |
| 198 | /* |
| 199 | * We set this after potentially failing in |
| 200 | * clear_user so that the signal arrives at the |
| 201 | * faulting rip. |
| 202 | */ |
| 203 | instruction_pointer_set(regs, (unsigned long)abort_ip); |
| 204 | return true; |
| 205 | } |
| 206 | |
| 207 | /* |
| 208 | * This resume handler should always be executed between any of: |
| 209 | * - preemption, |
| 210 | * - signal delivery, |
| 211 | * and return to user-space. |
| 212 | * |
| 213 | * This is how we can ensure that the entire rseq critical section, |
| 214 | * consisting of both the C part and the assembly instruction sequence, |
| 215 | * will issue the commit instruction only if executed atomically with |
| 216 | * respect to other threads scheduled on the same CPU, and with respect |
| 217 | * to signal handlers. |
| 218 | */ |
| 219 | void __rseq_handle_notify_resume(struct pt_regs *regs) |
| 220 | { |
| 221 | struct task_struct *t = current; |
| 222 | |
| 223 | if (unlikely(t->flags & PF_EXITING)) |
| 224 | return; |
| 225 | if (!access_ok(VERIFY_WRITE, t->rseq, sizeof(*t->rseq))) |
| 226 | goto error; |
| 227 | if (!rseq_update_cpu_id_event_counter(t)) |
| 228 | goto error; |
| 229 | if (!rseq_ip_fixup(regs)) |
| 230 | goto error; |
| 231 | return; |
| 232 | |
| 233 | error: |
| 234 | force_sig(SIGSEGV, t); |
| 235 | } |
| 236 | |
| 237 | /* |
| 238 | * sys_rseq - setup restartable sequences for caller thread. |
| 239 | */ |
| 240 | SYSCALL_DEFINE2(rseq, struct rseq __user *, rseq, int, flags) |
| 241 | { |
| 242 | if (!rseq) { |
| 243 | /* Unregister rseq for current thread. */ |
| 244 | if (unlikely(flags & ~RSEQ_FORCE_UNREGISTER)) |
| 245 | return -EINVAL; |
| 246 | if (flags & RSEQ_FORCE_UNREGISTER) { |
| 247 | current->rseq = NULL; |
| 248 | current->rseq_refcount = 0; |
| 249 | return 0; |
| 250 | } |
| 251 | if (!current->rseq_refcount) |
| 252 | return -ENOENT; |
| 253 | if (!--current->rseq_refcount) |
| 254 | current->rseq = NULL; |
| 255 | return 0; |
| 256 | } |
| 257 | |
| 258 | if (unlikely(flags)) |
| 259 | return -EINVAL; |
| 260 | |
| 261 | if (current->rseq) { |
| 262 | /* |
| 263 | * If rseq is already registered, check whether |
| 264 | * the provided address differs from the prior |
| 265 | * one. |
| 266 | */ |
| 267 | BUG_ON(!current->rseq_refcount); |
| 268 | if (current->rseq != rseq) |
| 269 | return -EBUSY; |
| 270 | if (current->rseq_refcount == UINT_MAX) |
| 271 | return -EOVERFLOW; |
| 272 | current->rseq_refcount++; |
| 273 | } else { |
| 274 | /* |
| 275 | * If there was no rseq previously registered, |
| 276 | * we need to ensure the provided rseq is |
| 277 | * properly aligned and valid. |
| 278 | */ |
| 279 | BUG_ON(current->rseq_refcount); |
| 280 | if (!IS_ALIGNED((unsigned long)rseq, __alignof__(*rseq))) |
| 281 | return -EINVAL; |
| 282 | if (!access_ok(VERIFY_WRITE, rseq, sizeof(*rseq))) |
| 283 | return -EFAULT; |
| 284 | current->rseq = rseq; |
| 285 | current->rseq_refcount = 1; |
| 286 | /* |
| 287 | * If rseq was previously inactive, and has just |
| 288 | * been registered, ensure the cpu_id and |
| 289 | * event_counter fields are updated before |
| 290 | * returning to user-space. |
| 291 | */ |
| 292 | rseq_set_notify_resume(current); |
| 293 | } |
| 294 | |
| 295 | return 0; |
| 296 | } |