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