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