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