fs/eventfd.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  *  fs/eventfd.c
   4  *
   5  *  Copyright (C) 2007  Davide Libenzi <davidel@xmailserver.org>
   6  *
   7  */
   8
   9 #include <linux/file.h>
  10 #include <linux/poll.h>
  11 #include <linux/init.h>
  12 #include <linux/fs.h>
  13 #include <linux/sched/signal.h>
  14 #include <linux/kernel.h>
  15 #include <linux/slab.h>
  16 #include <linux/list.h>
  17 #include <linux/spinlock.h>
  18 #include <linux/anon_inodes.h>
  19 #include <linux/syscalls.h>
  20 #include <linux/export.h>
  21 #include <linux/kref.h>
  22 #include <linux/eventfd.h>
  23 #include <linux/proc_fs.h>
  24 #include <linux/seq_file.h>
  25 #include <linux/idr.h>
  26
  27 DEFINE_PER_CPU(int, eventfd_wake_count);
  28
  29 static DEFINE_IDA(eventfd_ida);
  30
  31 struct eventfd_ctx {
  32         struct kref kref;
  33         wait_queue_head_t wqh;
  34         /*
  35          * Every time that a write(2) is performed on an eventfd, the
  36          * value of the __u64 being written is added to "count" and a
  37          * wakeup is performed on "wqh". A read(2) will return the "count"
  38          * value to userspace, and will reset "count" to zero. The kernel
  39          * side eventfd_signal() also, adds to the "count" counter and
  40          * issue a wakeup.
  41          */
  42         __u64 count;
  43         unsigned int flags;
  44         int id;
  45 };
  46
  47 /**
  48  * eventfd_signal - Adds @n to the eventfd counter.
  49  * @ctx: [in] Pointer to the eventfd context.
  50  * @n: [in] Value of the counter to be added to the eventfd internal counter.
  51  *          The value cannot be negative.
  52  *
  53  * This function is supposed to be called by the kernel in paths that do not
  54  * allow sleeping. In this function we allow the counter to reach the ULLONG_MAX
  55  * value, and we signal this as overflow condition by returning a EPOLLERR
  56  * to poll(2).
  57  *
  58  * Returns the amount by which the counter was incremented.  This will be less
  59  * than @n if the counter has overflowed.
  60  */
  61 __u64 eventfd_signal(struct eventfd_ctx *ctx, __u64 n)
  62 {
  63         unsigned long flags;
  64
  65         /*
  66          * Deadlock or stack overflow issues can happen if we recurse here
  67          * through waitqueue wakeup handlers. If the caller users potentially
  68          * nested waitqueues with custom wakeup handlers, then it should
  69          * check eventfd_signal_count() before calling this function. If
  70          * it returns true, the eventfd_signal() call should be deferred to a
  71          * safe context.
  72          */
  73         if (WARN_ON_ONCE(this_cpu_read(eventfd_wake_count)))
  74                 return 0;
  75
  76         spin_lock_irqsave(&ctx->wqh.lock, flags);
  77         this_cpu_inc(eventfd_wake_count);
  78         if (ULLONG_MAX - ctx->count < n)
  79                 n = ULLONG_MAX - ctx->count;
  80         ctx->count += n;
  81         if (waitqueue_active(&ctx->wqh))
  82                 wake_up_locked_poll(&ctx->wqh, EPOLLIN);
  83         this_cpu_dec(eventfd_wake_count);
  84         spin_unlock_irqrestore(&ctx->wqh.lock, flags);
  85
  86         return n;
  87 }
  88 EXPORT_SYMBOL_GPL(eventfd_signal);
  89
  90 static void eventfd_free_ctx(struct eventfd_ctx *ctx)
  91 {
  92         if (ctx->id >= 0)
  93                 ida_simple_remove(&eventfd_ida, ctx->id);
  94         kfree(ctx);
  95 }
  96
  97 static void eventfd_free(struct kref *kref)
  98 {
  99         struct eventfd_ctx *ctx = container_of(kref, struct eventfd_ctx, kref);
 100
 101         eventfd_free_ctx(ctx);
 102 }
 103
 104 /**
 105  * eventfd_ctx_put - Releases a reference to the internal eventfd context.
 106  * @ctx: [in] Pointer to eventfd context.
 107  *
 108  * The eventfd context reference must have been previously acquired either
 109  * with eventfd_ctx_fdget() or eventfd_ctx_fileget().
 110  */
 111 void eventfd_ctx_put(struct eventfd_ctx *ctx)
 112 {
 113         kref_put(&ctx->kref, eventfd_free);
 114 }
 115 EXPORT_SYMBOL_GPL(eventfd_ctx_put);
 116
 117 static int eventfd_release(struct inode *inode, struct file *file)
 118 {
 119         struct eventfd_ctx *ctx = file->private_data;
 120
 121         wake_up_poll(&ctx->wqh, EPOLLHUP);
 122         eventfd_ctx_put(ctx);
 123         return 0;
 124 }
 125
 126 static __poll_t eventfd_poll(struct file *file, poll_table *wait)
 127 {
 128         struct eventfd_ctx *ctx = file->private_data;
 129         __poll_t events = 0;
 130         u64 count;
 131
 132         poll_wait(file, &ctx->wqh, wait);
 133
 134         /*
 135          * All writes to ctx->count occur within ctx->wqh.lock.  This read
 136          * can be done outside ctx->wqh.lock because we know that poll_wait
 137          * takes that lock (through add_wait_queue) if our caller will sleep.
 138          *
 139          * The read _can_ therefore seep into add_wait_queue's critical
 140          * section, but cannot move above it!  add_wait_queue's spin_lock acts
 141          * as an acquire barrier and ensures that the read be ordered properly
 142          * against the writes.  The following CAN happen and is safe:
 143          *
 144          *     poll                               write
 145          *     -----------------                  ------------
 146          *     lock ctx->wqh.lock (in poll_wait)
 147          *     count = ctx->count
 148          *     __add_wait_queue
 149          *     unlock ctx->wqh.lock
 150          *                                        lock ctx->qwh.lock
 151          *                                        ctx->count += n
 152          *                                        if (waitqueue_active)
 153          *                                          wake_up_locked_poll
 154          *                                        unlock ctx->qwh.lock
 155          *     eventfd_poll returns 0
 156          *
 157          * but the following, which would miss a wakeup, cannot happen:
 158          *
 159          *     poll                               write
 160          *     -----------------                  ------------
 161          *     count = ctx->count (INVALID!)
 162          *                                        lock ctx->qwh.lock
 163          *                                        ctx->count += n
 164          *                                        **waitqueue_active is false**
 165          *                                        **no wake_up_locked_poll!**
 166          *                                        unlock ctx->qwh.lock
 167          *     lock ctx->wqh.lock (in poll_wait)
 168          *     __add_wait_queue
 169          *     unlock ctx->wqh.lock
 170          *     eventfd_poll returns 0
 171          */
 172         count = READ_ONCE(ctx->count);
 173
 174         if (count > 0)
 175                 events |= EPOLLIN;
 176         if (count == ULLONG_MAX)
 177                 events |= EPOLLERR;
 178         if (ULLONG_MAX - 1 > count)
 179                 events |= EPOLLOUT;
 180
 181         return events;
 182 }
 183
 184 static void eventfd_ctx_do_read(struct eventfd_ctx *ctx, __u64 *cnt)
 185 {
 186         *cnt = (ctx->flags & EFD_SEMAPHORE) ? 1 : ctx->count;
 187         ctx->count -= *cnt;
 188 }
 189
 190 /**
 191  * eventfd_ctx_remove_wait_queue - Read the current counter and removes wait queue.
 192  * @ctx: [in] Pointer to eventfd context.
 193  * @wait: [in] Wait queue to be removed.
 194  * @cnt: [out] Pointer to the 64-bit counter value.
 195  *
 196  * Returns %0 if successful, or the following error codes:
 197  *
 198  * -EAGAIN      : The operation would have blocked.
 199  *
 200  * This is used to atomically remove a wait queue entry from the eventfd wait
 201  * queue head, and read/reset the counter value.
 202  */
 203 int eventfd_ctx_remove_wait_queue(struct eventfd_ctx *ctx, wait_queue_entry_t *wait,
 204                                   __u64 *cnt)
 205 {
 206         unsigned long flags;
 207
 208         spin_lock_irqsave(&ctx->wqh.lock, flags);
 209         eventfd_ctx_do_read(ctx, cnt);
 210         __remove_wait_queue(&ctx->wqh, wait);
 211         if (*cnt != 0 && waitqueue_active(&ctx->wqh))
 212                 wake_up_locked_poll(&ctx->wqh, EPOLLOUT);
 213         spin_unlock_irqrestore(&ctx->wqh.lock, flags);
 214
 215         return *cnt != 0 ? 0 : -EAGAIN;
 216 }
 217 EXPORT_SYMBOL_GPL(eventfd_ctx_remove_wait_queue);
 218
 219 static ssize_t eventfd_read(struct file *file, char __user *buf, size_t count,
 220                             loff_t *ppos)
 221 {
 222         struct eventfd_ctx *ctx = file->private_data;
 223         ssize_t res;
 224         __u64 ucnt = 0;
 225         DECLARE_WAITQUEUE(wait, current);
 226
 227         if (count < sizeof(ucnt))
 228                 return -EINVAL;
 229
 230         spin_lock_irq(&ctx->wqh.lock);
 231         res = -EAGAIN;
 232         if (ctx->count > 0)
 233                 res = sizeof(ucnt);
 234         else if (!(file->f_flags & O_NONBLOCK)) {
 235                 __add_wait_queue(&ctx->wqh, &wait);
 236                 for (;;) {
 237                         set_current_state(TASK_INTERRUPTIBLE);
 238                         if (ctx->count > 0) {
 239                                 res = sizeof(ucnt);
 240                                 break;
 241                         }
 242                         if (signal_pending(current)) {
 243                                 res = -ERESTARTSYS;
 244                                 break;
 245                         }
 246                         spin_unlock_irq(&ctx->wqh.lock);
 247                         schedule();
 248                         spin_lock_irq(&ctx->wqh.lock);
 249                 }
 250                 __remove_wait_queue(&ctx->wqh, &wait);
 251                 __set_current_state(TASK_RUNNING);
 252         }
 253         if (likely(res > 0)) {
 254                 eventfd_ctx_do_read(ctx, &ucnt);
 255                 if (waitqueue_active(&ctx->wqh))
 256                         wake_up_locked_poll(&ctx->wqh, EPOLLOUT);
 257         }
 258         spin_unlock_irq(&ctx->wqh.lock);
 259
 260         if (res > 0 && put_user(ucnt, (__u64 __user *)buf))
 261                 return -EFAULT;
 262
 263         return res;
 264 }
 265
 266 static ssize_t eventfd_write(struct file *file, const char __user *buf, size_t count,
 267                              loff_t *ppos)
 268 {
 269         struct eventfd_ctx *ctx = file->private_data;
 270         ssize_t res;
 271         __u64 ucnt;
 272         DECLARE_WAITQUEUE(wait, current);
 273
 274         if (count < sizeof(ucnt))
 275                 return -EINVAL;
 276         if (copy_from_user(&ucnt, buf, sizeof(ucnt)))
 277                 return -EFAULT;
 278         if (ucnt == ULLONG_MAX)
 279                 return -EINVAL;
 280         spin_lock_irq(&ctx->wqh.lock);
 281         res = -EAGAIN;
 282         if (ULLONG_MAX - ctx->count > ucnt)
 283                 res = sizeof(ucnt);
 284         else if (!(file->f_flags & O_NONBLOCK)) {
 285                 __add_wait_queue(&ctx->wqh, &wait);
 286                 for (res = 0;;) {
 287                         set_current_state(TASK_INTERRUPTIBLE);
 288                         if (ULLONG_MAX - ctx->count > ucnt) {
 289                                 res = sizeof(ucnt);
 290                                 break;
 291                         }
 292                         if (signal_pending(current)) {
 293                                 res = -ERESTARTSYS;
 294                                 break;
 295                         }
 296                         spin_unlock_irq(&ctx->wqh.lock);
 297                         schedule();
 298                         spin_lock_irq(&ctx->wqh.lock);
 299                 }
 300                 __remove_wait_queue(&ctx->wqh, &wait);
 301                 __set_current_state(TASK_RUNNING);
 302         }
 303         if (likely(res > 0)) {
 304                 ctx->count += ucnt;
 305                 if (waitqueue_active(&ctx->wqh))
 306                         wake_up_locked_poll(&ctx->wqh, EPOLLIN);
 307         }
 308         spin_unlock_irq(&ctx->wqh.lock);
 309
 310         return res;
 311 }
 312
 313 #ifdef CONFIG_PROC_FS
 314 static void eventfd_show_fdinfo(struct seq_file *m, struct file *f)
 315 {
 316         struct eventfd_ctx *ctx = f->private_data;
 317
 318         spin_lock_irq(&ctx->wqh.lock);
 319         seq_printf(m, "eventfd-count: %16llx\n",
 320                    (unsigned long long)ctx->count);
 321         spin_unlock_irq(&ctx->wqh.lock);
 322         seq_printf(m, "eventfd-id: %d\n", ctx->id);
 323 }
 324 #endif
 325
 326 static const struct file_operations eventfd_fops = {
 327 #ifdef CONFIG_PROC_FS
 328         .show_fdinfo    = eventfd_show_fdinfo,
 329 #endif
 330         .release        = eventfd_release,
 331         .poll           = eventfd_poll,
 332         .read           = eventfd_read,
 333         .write          = eventfd_write,
 334         .llseek         = noop_llseek,
 335 };
 336
 337 /**
 338  * eventfd_fget - Acquire a reference of an eventfd file descriptor.
 339  * @fd: [in] Eventfd file descriptor.
 340  *
 341  * Returns a pointer to the eventfd file structure in case of success, or the
 342  * following error pointer:
 343  *
 344  * -EBADF    : Invalid @fd file descriptor.
 345  * -EINVAL   : The @fd file descriptor is not an eventfd file.
 346  */
 347 struct file *eventfd_fget(int fd)
 348 {
 349         struct file *file;
 350
 351         file = fget(fd);
 352         if (!file)
 353                 return ERR_PTR(-EBADF);
 354         if (file->f_op != &eventfd_fops) {
 355                 fput(file);
 356                 return ERR_PTR(-EINVAL);
 357         }
 358
 359         return file;
 360 }
 361 EXPORT_SYMBOL_GPL(eventfd_fget);
 362
 363 /**
 364  * eventfd_ctx_fdget - Acquires a reference to the internal eventfd context.
 365  * @fd: [in] Eventfd file descriptor.
 366  *
 367  * Returns a pointer to the internal eventfd context, otherwise the error
 368  * pointers returned by the following functions:
 369  *
 370  * eventfd_fget
 371  */
 372 struct eventfd_ctx *eventfd_ctx_fdget(int fd)
 373 {
 374         struct eventfd_ctx *ctx;
 375         struct fd f = fdget(fd);
 376         if (!f.file)
 377                 return ERR_PTR(-EBADF);
 378         ctx = eventfd_ctx_fileget(f.file);
 379         fdput(f);
 380         return ctx;
 381 }
 382 EXPORT_SYMBOL_GPL(eventfd_ctx_fdget);
 383
 384 /**
 385  * eventfd_ctx_fileget - Acquires a reference to the internal eventfd context.
 386  * @file: [in] Eventfd file pointer.
 387  *
 388  * Returns a pointer to the internal eventfd context, otherwise the error
 389  * pointer:
 390  *
 391  * -EINVAL   : The @fd file descriptor is not an eventfd file.
 392  */
 393 struct eventfd_ctx *eventfd_ctx_fileget(struct file *file)
 394 {
 395         struct eventfd_ctx *ctx;
 396
 397         if (file->f_op != &eventfd_fops)
 398                 return ERR_PTR(-EINVAL);
 399
 400         ctx = file->private_data;
 401         kref_get(&ctx->kref);
 402         return ctx;
 403 }
 404 EXPORT_SYMBOL_GPL(eventfd_ctx_fileget);
 405
 406 static int do_eventfd(unsigned int count, int flags)
 407 {
 408         struct eventfd_ctx *ctx;
 409         int fd;
 410
 411         /* Check the EFD_* constants for consistency.  */
 412         BUILD_BUG_ON(EFD_CLOEXEC != O_CLOEXEC);
 413         BUILD_BUG_ON(EFD_NONBLOCK != O_NONBLOCK);
 414
 415         if (flags & ~EFD_FLAGS_SET)
 416                 return -EINVAL;
 417
 418         ctx = kmalloc(sizeof(*ctx), GFP_KERNEL);
 419         if (!ctx)
 420                 return -ENOMEM;
 421
 422         kref_init(&ctx->kref);
 423         init_waitqueue_head(&ctx->wqh);
 424         ctx->count = count;
 425         ctx->flags = flags;
 426         ctx->id = ida_simple_get(&eventfd_ida, 0, 0, GFP_KERNEL);
 427
 428         fd = anon_inode_getfd("[eventfd]", &eventfd_fops, ctx,
 429                               O_RDWR | (flags & EFD_SHARED_FCNTL_FLAGS));
 430         if (fd < 0)
 431                 eventfd_free_ctx(ctx);
 432
 433         return fd;
 434 }
 435
 436 SYSCALL_DEFINE2(eventfd2, unsigned int, count, int, flags)
 437 {
 438         return do_eventfd(count, flags);
 439 }
 440
 441 SYSCALL_DEFINE1(eventfd, unsigned int, count)
 442 {
 443         return do_eventfd(count, 0);
 444 }
 445