kernel/rseq.c

   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
  84 static int rseq_update_cpu_id(struct task_struct *t)
  85 {
  86         u32 cpu_id = raw_smp_processor_id();
  87         struct rseq __user *rseq = t->rseq;
  88
  89         if (!user_write_access_begin(rseq, sizeof(*rseq)))
  90                 goto efault;
  91         unsafe_put_user(cpu_id, &rseq->cpu_id_start, efault_end);
  92         unsafe_put_user(cpu_id, &rseq->cpu_id, efault_end);
  93         user_write_access_end();
  94         trace_rseq_update(t);
  95         return 0;
  96
  97 efault_end:
  98         user_write_access_end();
  99 efault:
 100         return -EFAULT;
 101 }
 102
 103 static int rseq_reset_rseq_cpu_id(struct task_struct *t)
 104 {
 105         u32 cpu_id_start = 0, cpu_id = RSEQ_CPU_ID_UNINITIALIZED;
 106
 107         /*
 108          * Reset cpu_id_start to its initial state (0).
 109          */
 110         if (put_user(cpu_id_start, &t->rseq->cpu_id_start))
 111                 return -EFAULT;
 112         /*
 113          * Reset cpu_id to RSEQ_CPU_ID_UNINITIALIZED, so any user coming
 114          * in after unregistration can figure out that rseq needs to be
 115          * registered again.
 116          */
 117         if (put_user(cpu_id, &t->rseq->cpu_id))
 118                 return -EFAULT;
 119         return 0;
 120 }
 121
 122 static int rseq_get_rseq_cs(struct task_struct *t, struct rseq_cs *rseq_cs)
 123 {
 124         struct rseq_cs __user *urseq_cs;
 125         u64 ptr;
 126         u32 __user *usig;
 127         u32 sig;
 128         int ret;
 129
 130 #ifdef CONFIG_64BIT
 131         if (get_user(ptr, &t->rseq->rseq_cs.ptr64))
 132                 return -EFAULT;
 133 #else
 134         if (copy_from_user(&ptr, &t->rseq->rseq_cs.ptr64, sizeof(ptr)))
 135                 return -EFAULT;
 136 #endif
 137         if (!ptr) {
 138                 memset(rseq_cs, 0, sizeof(*rseq_cs));
 139                 return 0;
 140         }
 141         if (ptr >= TASK_SIZE)
 142                 return -EINVAL;
 143         urseq_cs = (struct rseq_cs __user *)(unsigned long)ptr;
 144         if (copy_from_user(rseq_cs, urseq_cs, sizeof(*rseq_cs)))
 145                 return -EFAULT;
 146
 147         if (rseq_cs->start_ip >= TASK_SIZE ||
 148             rseq_cs->start_ip + rseq_cs->post_commit_offset >= TASK_SIZE ||
 149             rseq_cs->abort_ip >= TASK_SIZE ||
 150             rseq_cs->version > 0)
 151                 return -EINVAL;
 152         /* Check for overflow. */
 153         if (rseq_cs->start_ip + rseq_cs->post_commit_offset < rseq_cs->start_ip)
 154                 return -EINVAL;
 155         /* Ensure that abort_ip is not in the critical section. */
 156         if (rseq_cs->abort_ip - rseq_cs->start_ip < rseq_cs->post_commit_offset)
 157                 return -EINVAL;
 158
 159         usig = (u32 __user *)(unsigned long)(rseq_cs->abort_ip - sizeof(u32));
 160         ret = get_user(sig, usig);
 161         if (ret)
 162                 return ret;
 163
 164         if (current->rseq_sig != sig) {
 165                 printk_ratelimited(KERN_WARNING
 166                         "Possible attack attempt. Unexpected rseq signature 0x%x, expecting 0x%x (pid=%d, addr=%p).\n",
 167                         sig, current->rseq_sig, current->pid, usig);
 168                 return -EINVAL;
 169         }
 170         return 0;
 171 }
 172
 173 static int rseq_need_restart(struct task_struct *t, u32 cs_flags)
 174 {
 175         u32 flags, event_mask;
 176         int ret;
 177
 178         /* Get thread flags. */
 179         ret = get_user(flags, &t->rseq->flags);
 180         if (ret)
 181                 return ret;
 182
 183         /* Take critical section flags into account. */
 184         flags |= cs_flags;
 185
 186         /*
 187          * Restart on signal can only be inhibited when restart on
 188          * preempt and restart on migrate are inhibited too. Otherwise,
 189          * a preempted signal handler could fail to restart the prior
 190          * execution context on sigreturn.
 191          */
 192         if (unlikely((flags & RSEQ_CS_FLAG_NO_RESTART_ON_SIGNAL) &&
 193                      (flags & RSEQ_CS_PREEMPT_MIGRATE_FLAGS) !=
 194                      RSEQ_CS_PREEMPT_MIGRATE_FLAGS))
 195                 return -EINVAL;
 196
 197         /*
 198          * Load and clear event mask atomically with respect to
 199          * scheduler preemption.
 200          */
 201         preempt_disable();
 202         event_mask = t->rseq_event_mask;
 203         t->rseq_event_mask = 0;
 204         preempt_enable();
 205
 206         return !!(event_mask & ~flags);
 207 }
 208
 209 static int clear_rseq_cs(struct task_struct *t)
 210 {
 211         /*
 212          * The rseq_cs field is set to NULL on preemption or signal
 213          * delivery on top of rseq assembly block, as well as on top
 214          * of code outside of the rseq assembly block. This performs
 215          * a lazy clear of the rseq_cs field.
 216          *
 217          * Set rseq_cs to NULL.
 218          */
 219 #ifdef CONFIG_64BIT
 220         return put_user(0UL, &t->rseq->rseq_cs.ptr64);
 221 #else
 222         if (clear_user(&t->rseq->rseq_cs.ptr64, sizeof(t->rseq->rseq_cs.ptr64)))
 223                 return -EFAULT;
 224         return 0;
 225 #endif
 226 }
 227
 228 /*
 229  * Unsigned comparison will be true when ip >= start_ip, and when
 230  * ip < start_ip + post_commit_offset.
 231  */
 232 static bool in_rseq_cs(unsigned long ip, struct rseq_cs *rseq_cs)
 233 {
 234         return ip - rseq_cs->start_ip < rseq_cs->post_commit_offset;
 235 }
 236
 237 static int rseq_ip_fixup(struct pt_regs *regs)
 238 {
 239         unsigned long ip = instruction_pointer(regs);
 240         struct task_struct *t = current;
 241         struct rseq_cs rseq_cs;
 242         int ret;
 243
 244         ret = rseq_get_rseq_cs(t, &rseq_cs);
 245         if (ret)
 246                 return ret;
 247
 248         /*
 249          * Handle potentially not being within a critical section.
 250          * If not nested over a rseq critical section, restart is useless.
 251          * Clear the rseq_cs pointer and return.
 252          */
 253         if (!in_rseq_cs(ip, &rseq_cs))
 254                 return clear_rseq_cs(t);
 255         ret = rseq_need_restart(t, rseq_cs.flags);
 256         if (ret <= 0)
 257                 return ret;
 258         ret = clear_rseq_cs(t);
 259         if (ret)
 260                 return ret;
 261         trace_rseq_ip_fixup(ip, rseq_cs.start_ip, rseq_cs.post_commit_offset,
 262                             rseq_cs.abort_ip);
 263         instruction_pointer_set(regs, (unsigned long)rseq_cs.abort_ip);
 264         return 0;
 265 }
 266
 267 /*
 268  * This resume handler must always be executed between any of:
 269  * - preemption,
 270  * - signal delivery,
 271  * and return to user-space.
 272  *
 273  * This is how we can ensure that the entire rseq critical section
 274  * will issue the commit instruction only if executed atomically with
 275  * respect to other threads scheduled on the same CPU, and with respect
 276  * to signal handlers.
 277  */
 278 void __rseq_handle_notify_resume(struct ksignal *ksig, struct pt_regs *regs)
 279 {
 280         struct task_struct *t = current;
 281         int ret, sig;
 282
 283         if (unlikely(t->flags & PF_EXITING))
 284                 return;
 285         ret = rseq_ip_fixup(regs);
 286         if (unlikely(ret < 0))
 287                 goto error;
 288         if (unlikely(rseq_update_cpu_id(t)))
 289                 goto error;
 290         return;
 291
 292 error:
 293         sig = ksig ? ksig->sig : 0;
 294         force_sigsegv(sig);
 295 }
 296
 297 #ifdef CONFIG_DEBUG_RSEQ
 298
 299 /*
 300  * Terminate the process if a syscall is issued within a restartable
 301  * sequence.
 302  */
 303 void rseq_syscall(struct pt_regs *regs)
 304 {
 305         unsigned long ip = instruction_pointer(regs);
 306         struct task_struct *t = current;
 307         struct rseq_cs rseq_cs;
 308
 309         if (!t->rseq)
 310                 return;
 311         if (rseq_get_rseq_cs(t, &rseq_cs) || in_rseq_cs(ip, &rseq_cs))
 312                 force_sig(SIGSEGV);
 313 }
 314
 315 #endif
 316
 317 /*
 318  * sys_rseq - setup restartable sequences for caller thread.
 319  */
 320 SYSCALL_DEFINE4(rseq, struct rseq __user *, rseq, u32, rseq_len,
 321                 int, flags, u32, sig)
 322 {
 323         int ret;
 324
 325         if (flags & RSEQ_FLAG_UNREGISTER) {
 326                 if (flags & ~RSEQ_FLAG_UNREGISTER)
 327                         return -EINVAL;
 328                 /* Unregister rseq for current thread. */
 329                 if (current->rseq != rseq || !current->rseq)
 330                         return -EINVAL;
 331                 if (rseq_len != sizeof(*rseq))
 332                         return -EINVAL;
 333                 if (current->rseq_sig != sig)
 334                         return -EPERM;
 335                 ret = rseq_reset_rseq_cpu_id(current);
 336                 if (ret)
 337                         return ret;
 338                 current->rseq = NULL;
 339                 current->rseq_sig = 0;
 340                 return 0;
 341         }
 342
 343         if (unlikely(flags))
 344                 return -EINVAL;
 345
 346         if (current->rseq) {
 347                 /*
 348                  * If rseq is already registered, check whether
 349                  * the provided address differs from the prior
 350                  * one.
 351                  */
 352                 if (current->rseq != rseq || rseq_len != sizeof(*rseq))
 353                         return -EINVAL;
 354                 if (current->rseq_sig != sig)
 355                         return -EPERM;
 356                 /* Already registered. */
 357                 return -EBUSY;
 358         }
 359
 360         /*
 361          * If there was no rseq previously registered,
 362          * ensure the provided rseq is properly aligned and valid.
 363          */
 364         if (!IS_ALIGNED((unsigned long)rseq, __alignof__(*rseq)) ||
 365             rseq_len != sizeof(*rseq))
 366                 return -EINVAL;
 367         if (!access_ok(rseq, rseq_len))
 368                 return -EFAULT;
 369         current->rseq = rseq;
 370         current->rseq_sig = sig;
 371         /*
 372          * If rseq was previously inactive, and has just been
 373          * registered, ensure the cpu_id_start and cpu_id fields
 374          * are updated before returning to user-space.
 375          */
 376         rseq_set_notify_resume(current);
 377
 378         return 0;
 379 }