fs/signalfd.c

   1 /*
   2  *  fs/signalfd.c
   3  *
   4  *  Copyright (C) 2003  Linus Torvalds
   5  *
   6  *  Mon Mar 5, 2007: Davide Libenzi <davidel@xmailserver.org>
   7  *      Changed ->read() to return a siginfo strcture instead of signal number.
   8  *      Fixed locking in ->poll().
   9  *      Added sighand-detach notification.
  10  *      Added fd re-use in sys_signalfd() syscall.
  11  *      Now using anonymous inode source.
  12  *      Thanks to Oleg Nesterov for useful code review and suggestions.
  13  *      More comments and suggestions from Arnd Bergmann.
  14  * Sat May 19, 2007: Davi E. M. Arnaut <davi@haxent.com.br>
  15  *      Retrieve multiple signals with one read() call
  16  */
  17
  18 #include <linux/file.h>
  19 #include <linux/poll.h>
  20 #include <linux/init.h>
  21 #include <linux/fs.h>
  22 #include <linux/sched.h>
  23 #include <linux/kernel.h>
  24 #include <linux/signal.h>
  25 #include <linux/list.h>
  26 #include <linux/anon_inodes.h>
  27 #include <linux/signalfd.h>
  28
  29 struct signalfd_ctx {
  30         struct list_head lnk;
  31         wait_queue_head_t wqh;
  32         sigset_t sigmask;
  33         struct task_struct *tsk;
  34 };
  35
  36 struct signalfd_lockctx {
  37         struct task_struct *tsk;
  38         unsigned long flags;
  39 };
  40
  41 /*
  42  * Tries to acquire the sighand lock. We do not increment the sighand
  43  * use count, and we do not even pin the task struct, so we need to
  44  * do it inside an RCU read lock, and we must be prepared for the
  45  * ctx->tsk going to NULL (in signalfd_deliver()), and for the sighand
  46  * being detached. We return 0 if the sighand has been detached, or
  47  * 1 if we were able to pin the sighand lock.
  48  */
  49 static int signalfd_lock(struct signalfd_ctx *ctx, struct signalfd_lockctx *lk)
  50 {
  51         struct sighand_struct *sighand = NULL;
  52
  53         rcu_read_lock();
  54         lk->tsk = rcu_dereference(ctx->tsk);
  55         if (likely(lk->tsk != NULL))
  56                 sighand = lock_task_sighand(lk->tsk, &lk->flags);
  57         rcu_read_unlock();
  58
  59         if (sighand && !ctx->tsk) {
  60                 unlock_task_sighand(lk->tsk, &lk->flags);
  61                 sighand = NULL;
  62         }
  63
  64         return sighand != NULL;
  65 }
  66
  67 static void signalfd_unlock(struct signalfd_lockctx *lk)
  68 {
  69         unlock_task_sighand(lk->tsk, &lk->flags);
  70 }
  71
  72 /*
  73  * This must be called with the sighand lock held.
  74  */
  75 void signalfd_deliver(struct task_struct *tsk, int sig)
  76 {
  77         struct sighand_struct *sighand = tsk->sighand;
  78         struct signalfd_ctx *ctx, *tmp;
  79
  80         BUG_ON(!sig);
  81         list_for_each_entry_safe(ctx, tmp, &sighand->signalfd_list, lnk) {
  82                 /*
  83                  * We use a negative signal value as a way to broadcast that the
  84                  * sighand has been orphaned, so that we can notify all the
  85                  * listeners about this. Remember the ctx->sigmask is inverted,
  86                  * so if the user is interested in a signal, that corresponding
  87                  * bit will be zero.
  88                  */
  89                 if (sig < 0) {
  90                         if (ctx->tsk == tsk) {
  91                                 ctx->tsk = NULL;
  92                                 list_del_init(&ctx->lnk);
  93                                 wake_up(&ctx->wqh);
  94                         }
  95                 } else {
  96                         if (!sigismember(&ctx->sigmask, sig))
  97                                 wake_up(&ctx->wqh);
  98                 }
  99         }
 100 }
 101
 102 static void signalfd_cleanup(struct signalfd_ctx *ctx)
 103 {
 104         struct signalfd_lockctx lk;
 105
 106         /*
 107          * This is tricky. If the sighand is gone, we do not need to remove
 108          * context from the list, the list itself won't be there anymore.
 109          */
 110         if (signalfd_lock(ctx, &lk)) {
 111                 list_del(&ctx->lnk);
 112                 signalfd_unlock(&lk);
 113         }
 114         kfree(ctx);
 115 }
 116
 117 static int signalfd_release(struct inode *inode, struct file *file)
 118 {
 119         signalfd_cleanup(file->private_data);
 120         return 0;
 121 }
 122
 123 static unsigned int signalfd_poll(struct file *file, poll_table *wait)
 124 {
 125         struct signalfd_ctx *ctx = file->private_data;
 126         unsigned int events = 0;
 127         struct signalfd_lockctx lk;
 128
 129         poll_wait(file, &ctx->wqh, wait);
 130
 131         /*
 132          * Let the caller get a POLLIN in this case, ala socket recv() when
 133          * the peer disconnects.
 134          */
 135         if (signalfd_lock(ctx, &lk)) {
 136                 if ((lk.tsk == current &&
 137                      next_signal(&lk.tsk->pending, &ctx->sigmask) > 0) ||
 138                     next_signal(&lk.tsk->signal->shared_pending,
 139                                 &ctx->sigmask) > 0)
 140                         events |= POLLIN;
 141                 signalfd_unlock(&lk);
 142         } else
 143                 events |= POLLIN;
 144
 145         return events;
 146 }
 147
 148 /*
 149  * Copied from copy_siginfo_to_user() in kernel/signal.c
 150  */
 151 static int signalfd_copyinfo(struct signalfd_siginfo __user *uinfo,
 152                              siginfo_t const *kinfo)
 153 {
 154         long err;
 155
 156         BUILD_BUG_ON(sizeof(struct signalfd_siginfo) != 128);
 157
 158         /*
 159          * Unused memebers should be zero ...
 160          */
 161         err = __clear_user(uinfo, sizeof(*uinfo));
 162
 163         /*
 164          * If you change siginfo_t structure, please be sure
 165          * this code is fixed accordingly.
 166          */
 167         err |= __put_user(kinfo->si_signo, &uinfo->signo);
 168         err |= __put_user(kinfo->si_errno, &uinfo->err);
 169         err |= __put_user((short)kinfo->si_code, &uinfo->code);
 170         switch (kinfo->si_code & __SI_MASK) {
 171         case __SI_KILL:
 172                 err |= __put_user(kinfo->si_pid, &uinfo->pid);
 173                 err |= __put_user(kinfo->si_uid, &uinfo->uid);
 174                 break;
 175         case __SI_TIMER:
 176                  err |= __put_user(kinfo->si_tid, &uinfo->tid);
 177                  err |= __put_user(kinfo->si_overrun, &uinfo->overrun);
 178                  err |= __put_user((long)kinfo->si_ptr, &uinfo->svptr);
 179                 break;
 180         case __SI_POLL:
 181                 err |= __put_user(kinfo->si_band, &uinfo->band);
 182                 err |= __put_user(kinfo->si_fd, &uinfo->fd);
 183                 break;
 184         case __SI_FAULT:
 185                 err |= __put_user((long)kinfo->si_addr, &uinfo->addr);
 186 #ifdef __ARCH_SI_TRAPNO
 187                 err |= __put_user(kinfo->si_trapno, &uinfo->trapno);
 188 #endif
 189                 break;
 190         case __SI_CHLD:
 191                 err |= __put_user(kinfo->si_pid, &uinfo->pid);
 192                 err |= __put_user(kinfo->si_uid, &uinfo->uid);
 193                 err |= __put_user(kinfo->si_status, &uinfo->status);
 194                 err |= __put_user(kinfo->si_utime, &uinfo->utime);
 195                 err |= __put_user(kinfo->si_stime, &uinfo->stime);
 196                 break;
 197         case __SI_RT: /* This is not generated by the kernel as of now. */
 198         case __SI_MESGQ: /* But this is */
 199                 err |= __put_user(kinfo->si_pid, &uinfo->pid);
 200                 err |= __put_user(kinfo->si_uid, &uinfo->uid);
 201                 err |= __put_user((long)kinfo->si_ptr, &uinfo->svptr);
 202                 break;
 203         default: /* this is just in case for now ... */
 204                 err |= __put_user(kinfo->si_pid, &uinfo->pid);
 205                 err |= __put_user(kinfo->si_uid, &uinfo->uid);
 206                 break;
 207         }
 208
 209         return err ? -EFAULT: sizeof(*uinfo);
 210 }
 211
 212 static ssize_t signalfd_dequeue(struct signalfd_ctx *ctx, siginfo_t *info,
 213                                 int nonblock)
 214 {
 215         ssize_t ret;
 216         struct signalfd_lockctx lk;
 217         DECLARE_WAITQUEUE(wait, current);
 218
 219         if (!signalfd_lock(ctx, &lk))
 220                 return 0;
 221
 222         ret = dequeue_signal(lk.tsk, &ctx->sigmask, info);
 223         switch (ret) {
 224         case 0:
 225                 if (!nonblock)
 226                         break;
 227                 ret = -EAGAIN;
 228         default:
 229                 signalfd_unlock(&lk);
 230                 return ret;
 231         }
 232
 233         add_wait_queue(&ctx->wqh, &wait);
 234         for (;;) {
 235                 set_current_state(TASK_INTERRUPTIBLE);
 236                 ret = dequeue_signal(lk.tsk, &ctx->sigmask, info);
 237                 signalfd_unlock(&lk);
 238                 if (ret != 0)
 239                         break;
 240                 if (signal_pending(current)) {
 241                         ret = -ERESTARTSYS;
 242                         break;
 243                 }
 244                 schedule();
 245                 ret = signalfd_lock(ctx, &lk);
 246                 if (unlikely(!ret)) {
 247                         /*
 248                          * Let the caller read zero byte, ala socket
 249                          * recv() when the peer disconnect. This test
 250                          * must be done before doing a dequeue_signal(),
 251                          * because if the sighand has been orphaned,
 252                          * the dequeue_signal() call is going to crash
 253                          * because ->sighand will be long gone.
 254                          */
 255                          break;
 256                 }
 257         }
 258
 259         remove_wait_queue(&ctx->wqh, &wait);
 260         __set_current_state(TASK_RUNNING);
 261
 262         return ret;
 263 }
 264
 265 /*
 266  * Returns either the size of a "struct signalfd_siginfo", or zero if the
 267  * sighand we are attached to, has been orphaned. The "count" parameter
 268  * must be at least the size of a "struct signalfd_siginfo".
 269  */
 270 static ssize_t signalfd_read(struct file *file, char __user *buf, size_t count,
 271                              loff_t *ppos)
 272 {
 273         struct signalfd_ctx *ctx = file->private_data;
 274         struct signalfd_siginfo __user *siginfo;
 275         int nonblock = file->f_flags & O_NONBLOCK;
 276         ssize_t ret, total = 0;
 277         siginfo_t info;
 278
 279         count /= sizeof(struct signalfd_siginfo);
 280         if (!count)
 281                 return -EINVAL;
 282
 283         siginfo = (struct signalfd_siginfo __user *) buf;
 284
 285         do {
 286                 ret = signalfd_dequeue(ctx, &info, nonblock);
 287                 if (unlikely(ret <= 0))
 288                         break;
 289                 ret = signalfd_copyinfo(siginfo, &info);
 290                 if (ret < 0)
 291                         break;
 292                 siginfo++;
 293                 total += ret;
 294                 nonblock = 1;
 295         } while (--count);
 296
 297         return total ? total : ret;
 298 }
 299
 300 static const struct file_operations signalfd_fops = {
 301         .release        = signalfd_release,
 302         .poll           = signalfd_poll,
 303         .read           = signalfd_read,
 304 };
 305
 306 /*
 307  * Create a file descriptor that is associated with our signal
 308  * state. We can pass it around to others if we want to, but
 309  * it will always be _our_ signal state.
 310  */
 311 asmlinkage long sys_signalfd(int ufd, sigset_t __user *user_mask, size_t sizemask)
 312 {
 313         int error;
 314         sigset_t sigmask;
 315         struct signalfd_ctx *ctx;
 316         struct sighand_struct *sighand;
 317         struct file *file;
 318         struct inode *inode;
 319         struct signalfd_lockctx lk;
 320
 321         if (sizemask != sizeof(sigset_t) ||
 322             copy_from_user(&sigmask, user_mask, sizeof(sigmask)))
 323                 return -EINVAL;
 324         sigdelsetmask(&sigmask, sigmask(SIGKILL) | sigmask(SIGSTOP));
 325         signotset(&sigmask);
 326
 327         if (ufd == -1) {
 328                 ctx = kmalloc(sizeof(*ctx), GFP_KERNEL);
 329                 if (!ctx)
 330                         return -ENOMEM;
 331
 332                 init_waitqueue_head(&ctx->wqh);
 333                 ctx->sigmask = sigmask;
 334                 ctx->tsk = current;
 335
 336                 sighand = current->sighand;
 337                 /*
 338                  * Add this fd to the list of signal listeners.
 339                  */
 340                 spin_lock_irq(&sighand->siglock);
 341                 list_add_tail(&ctx->lnk, &sighand->signalfd_list);
 342                 spin_unlock_irq(&sighand->siglock);
 343
 344                 /*
 345                  * When we call this, the initialization must be complete, since
 346                  * anon_inode_getfd() will install the fd.
 347                  */
 348                 error = anon_inode_getfd(&ufd, &inode, &file, "[signalfd]",
 349                                          &signalfd_fops, ctx);
 350                 if (error)
 351                         goto err_fdalloc;
 352         } else {
 353                 file = fget(ufd);
 354                 if (!file)
 355                         return -EBADF;
 356                 ctx = file->private_data;
 357                 if (file->f_op != &signalfd_fops) {
 358                         fput(file);
 359                         return -EINVAL;
 360                 }
 361                 /*
 362                  * We need to be prepared of the fact that the sighand this fd
 363                  * is attached to, has been detched. In that case signalfd_lock()
 364                  * will return 0, and we'll just skip setting the new mask.
 365                  */
 366                 if (signalfd_lock(ctx, &lk)) {
 367                         ctx->sigmask = sigmask;
 368                         signalfd_unlock(&lk);
 369                 }
 370                 wake_up(&ctx->wqh);
 371                 fput(file);
 372         }
 373
 374         return ufd;
 375
 376 err_fdalloc:
 377         signalfd_cleanup(ctx);
 378         return error;
 379 }
 380