timerfd: support CLOCK_BOOTTIME clock
[linux-3.10.git] / fs / timerfd.c
index 5400524..0013142 100644 (file)
@@ -8,12 +8,14 @@
  *
  */
 
+#include <linux/alarmtimer.h>
 #include <linux/file.h>
 #include <linux/poll.h>
 #include <linux/init.h>
 #include <linux/fs.h>
 #include <linux/sched.h>
 #include <linux/kernel.h>
+#include <linux/slab.h>
 #include <linux/list.h>
 #include <linux/spinlock.h>
 #include <linux/time.h>
 #include <linux/anon_inodes.h>
 #include <linux/timerfd.h>
 #include <linux/syscalls.h>
+#include <linux/compat.h>
+#include <linux/rcupdate.h>
 
 struct timerfd_ctx {
-       struct hrtimer tmr;
+       union {
+               struct hrtimer tmr;
+               struct alarm alarm;
+       } t;
        ktime_t tintv;
+       ktime_t moffs;
        wait_queue_head_t wqh;
        u64 ticks;
        int expired;
        int clockid;
+       struct rcu_head rcu;
+       struct list_head clist;
+       bool might_cancel;
 };
 
+static LIST_HEAD(cancel_list);
+static DEFINE_SPINLOCK(cancel_lock);
+
+static inline bool isalarm(struct timerfd_ctx *ctx)
+{
+       return ctx->clockid == CLOCK_REALTIME_ALARM ||
+               ctx->clockid == CLOCK_BOOTTIME_ALARM;
+}
+
 /*
  * This gets called when the timer event triggers. We set the "expired"
  * flag, but we do not re-arm the timer (in case it's necessary,
  * tintv.tv64 != 0) until the timer is accessed.
  */
-static enum hrtimer_restart timerfd_tmrproc(struct hrtimer *htmr)
+static void timerfd_triggered(struct timerfd_ctx *ctx)
 {
-       struct timerfd_ctx *ctx = container_of(htmr, struct timerfd_ctx, tmr);
        unsigned long flags;
 
        spin_lock_irqsave(&ctx->wqh.lock, flags);
@@ -46,25 +65,104 @@ static enum hrtimer_restart timerfd_tmrproc(struct hrtimer *htmr)
        ctx->ticks++;
        wake_up_locked(&ctx->wqh);
        spin_unlock_irqrestore(&ctx->wqh.lock, flags);
+}
 
+static enum hrtimer_restart timerfd_tmrproc(struct hrtimer *htmr)
+{
+       struct timerfd_ctx *ctx = container_of(htmr, struct timerfd_ctx,
+                                              t.tmr);
+       timerfd_triggered(ctx);
        return HRTIMER_NORESTART;
 }
 
+static enum alarmtimer_restart timerfd_alarmproc(struct alarm *alarm,
+       ktime_t now)
+{
+       struct timerfd_ctx *ctx = container_of(alarm, struct timerfd_ctx,
+                                              t.alarm);
+       timerfd_triggered(ctx);
+       return ALARMTIMER_NORESTART;
+}
+
+/*
+ * Called when the clock was set to cancel the timers in the cancel
+ * list. This will wake up processes waiting on these timers. The
+ * wake-up requires ctx->ticks to be non zero, therefore we increment
+ * it before calling wake_up_locked().
+ */
+void timerfd_clock_was_set(void)
+{
+       ktime_t moffs = ktime_get_monotonic_offset();
+       struct timerfd_ctx *ctx;
+       unsigned long flags;
+
+       rcu_read_lock();
+       list_for_each_entry_rcu(ctx, &cancel_list, clist) {
+               if (!ctx->might_cancel)
+                       continue;
+               spin_lock_irqsave(&ctx->wqh.lock, flags);
+               if (ctx->moffs.tv64 != moffs.tv64) {
+                       ctx->moffs.tv64 = KTIME_MAX;
+                       ctx->ticks++;
+                       wake_up_locked(&ctx->wqh);
+               }
+               spin_unlock_irqrestore(&ctx->wqh.lock, flags);
+       }
+       rcu_read_unlock();
+}
+
+static void timerfd_remove_cancel(struct timerfd_ctx *ctx)
+{
+       if (ctx->might_cancel) {
+               ctx->might_cancel = false;
+               spin_lock(&cancel_lock);
+               list_del_rcu(&ctx->clist);
+               spin_unlock(&cancel_lock);
+       }
+}
+
+static bool timerfd_canceled(struct timerfd_ctx *ctx)
+{
+       if (!ctx->might_cancel || ctx->moffs.tv64 != KTIME_MAX)
+               return false;
+       ctx->moffs = ktime_get_monotonic_offset();
+       return true;
+}
+
+static void timerfd_setup_cancel(struct timerfd_ctx *ctx, int flags)
+{
+       if ((ctx->clockid == CLOCK_REALTIME ||
+            ctx->clockid == CLOCK_REALTIME_ALARM) &&
+           (flags & TFD_TIMER_ABSTIME) && (flags & TFD_TIMER_CANCEL_ON_SET)) {
+               if (!ctx->might_cancel) {
+                       ctx->might_cancel = true;
+                       spin_lock(&cancel_lock);
+                       list_add_rcu(&ctx->clist, &cancel_list);
+                       spin_unlock(&cancel_lock);
+               }
+       } else if (ctx->might_cancel) {
+               timerfd_remove_cancel(ctx);
+       }
+}
+
 static ktime_t timerfd_get_remaining(struct timerfd_ctx *ctx)
 {
-       ktime_t now, remaining;
+       ktime_t remaining;
 
-       now = ctx->tmr.base->get_time();
-       remaining = ktime_sub(ctx->tmr.expires, now);
+       if (isalarm(ctx))
+               remaining = alarm_expires_remaining(&ctx->t.alarm);
+       else
+               remaining = hrtimer_expires_remaining(&ctx->t.tmr);
 
        return remaining.tv64 < 0 ? ktime_set(0, 0): remaining;
 }
 
-static void timerfd_setup(struct timerfd_ctx *ctx, int flags,
-                         const struct itimerspec *ktmr)
+static int timerfd_setup(struct timerfd_ctx *ctx, int flags,
+                        const struct itimerspec *ktmr)
 {
        enum hrtimer_mode htmode;
        ktime_t texp;
+       int clockid = ctx->clockid;
 
        htmode = (flags & TFD_TIMER_ABSTIME) ?
                HRTIMER_MODE_ABS: HRTIMER_MODE_REL;
@@ -73,19 +171,45 @@ static void timerfd_setup(struct timerfd_ctx *ctx, int flags,
        ctx->expired = 0;
        ctx->ticks = 0;
        ctx->tintv = timespec_to_ktime(ktmr->it_interval);
-       hrtimer_init(&ctx->tmr, ctx->clockid, htmode);
-       ctx->tmr.expires = texp;
-       ctx->tmr.function = timerfd_tmrproc;
-       if (texp.tv64 != 0)
-               hrtimer_start(&ctx->tmr, texp, htmode);
+
+       if (isalarm(ctx)) {
+               alarm_init(&ctx->t.alarm,
+                          ctx->clockid == CLOCK_REALTIME_ALARM ?
+                          ALARM_REALTIME : ALARM_BOOTTIME,
+                          timerfd_alarmproc);
+       } else {
+               hrtimer_init(&ctx->t.tmr, clockid, htmode);
+               hrtimer_set_expires(&ctx->t.tmr, texp);
+               ctx->t.tmr.function = timerfd_tmrproc;
+       }
+
+       if (texp.tv64 != 0) {
+               if (isalarm(ctx)) {
+                       if (flags & TFD_TIMER_ABSTIME)
+                               alarm_start(&ctx->t.alarm, texp);
+                       else
+                               alarm_start_relative(&ctx->t.alarm, texp);
+               } else {
+                       hrtimer_start(&ctx->t.tmr, texp, htmode);
+               }
+
+               if (timerfd_canceled(ctx))
+                       return -ECANCELED;
+       }
+       return 0;
 }
 
 static int timerfd_release(struct inode *inode, struct file *file)
 {
        struct timerfd_ctx *ctx = file->private_data;
 
-       hrtimer_cancel(&ctx->tmr);
-       kfree(ctx);
+       timerfd_remove_cancel(ctx);
+
+       if (isalarm(ctx))
+               alarm_cancel(&ctx->t.alarm);
+       else
+               hrtimer_cancel(&ctx->t.tmr);
+       kfree_rcu(ctx, rcu);
        return 0;
 }
 
@@ -111,33 +235,29 @@ static ssize_t timerfd_read(struct file *file, char __user *buf, size_t count,
        struct timerfd_ctx *ctx = file->private_data;
        ssize_t res;
        u64 ticks = 0;
-       DECLARE_WAITQUEUE(wait, current);
 
        if (count < sizeof(ticks))
                return -EINVAL;
        spin_lock_irq(&ctx->wqh.lock);
-       res = -EAGAIN;
-       if (!ctx->ticks && !(file->f_flags & O_NONBLOCK)) {
-               __add_wait_queue(&ctx->wqh, &wait);
-               for (res = 0;;) {
-                       set_current_state(TASK_INTERRUPTIBLE);
-                       if (ctx->ticks) {
-                               res = 0;
-                               break;
-                       }
-                       if (signal_pending(current)) {
-                               res = -ERESTARTSYS;
-                               break;
-                       }
-                       spin_unlock_irq(&ctx->wqh.lock);
-                       schedule();
-                       spin_lock_irq(&ctx->wqh.lock);
-               }
-               __remove_wait_queue(&ctx->wqh, &wait);
-               __set_current_state(TASK_RUNNING);
+       if (file->f_flags & O_NONBLOCK)
+               res = -EAGAIN;
+       else
+               res = wait_event_interruptible_locked_irq(ctx->wqh, ctx->ticks);
+
+       /*
+        * If clock has changed, we do not care about the
+        * ticks and we do not rearm the timer. Userspace must
+        * reevaluate anyway.
+        */
+       if (timerfd_canceled(ctx)) {
+               ctx->ticks = 0;
+               ctx->expired = 0;
+               res = -ECANCELED;
        }
+
        if (ctx->ticks) {
                ticks = ctx->ticks;
+
                if (ctx->expired && ctx->tintv.tv64) {
                        /*
                         * If tintv.tv64 != 0, this is a periodic timer that
@@ -145,9 +265,15 @@ static ssize_t timerfd_read(struct file *file, char __user *buf, size_t count,
                         * callback to avoid DoS attacks specifying a very
                         * short timer period.
                         */
-                       ticks += hrtimer_forward_now(&ctx->tmr,
-                                                    ctx->tintv) - 1;
-                       hrtimer_restart(&ctx->tmr);
+                       if (isalarm(ctx)) {
+                               ticks += alarm_forward_now(
+                                       &ctx->t.alarm, ctx->tintv) - 1;
+                               alarm_restart(&ctx->t.alarm);
+                       } else {
+                               ticks += hrtimer_forward_now(&ctx->t.tmr,
+                                                            ctx->tintv) - 1;
+                               hrtimer_restart(&ctx->t.tmr);
+                       }
                }
                ctx->expired = 0;
                ctx->ticks = 0;
@@ -162,34 +288,37 @@ static const struct file_operations timerfd_fops = {
        .release        = timerfd_release,
        .poll           = timerfd_poll,
        .read           = timerfd_read,
+       .llseek         = noop_llseek,
 };
 
-static struct file *timerfd_fget(int fd)
+static int timerfd_fget(int fd, struct fd *p)
 {
-       struct file *file;
-
-       file = fget(fd);
-       if (!file)
-               return ERR_PTR(-EBADF);
-       if (file->f_op != &timerfd_fops) {
-               fput(file);
-               return ERR_PTR(-EINVAL);
+       struct fd f = fdget(fd);
+       if (!f.file)
+               return -EBADF;
+       if (f.file->f_op != &timerfd_fops) {
+               fdput(f);
+               return -EINVAL;
        }
-
-       return file;
+       *p = f;
+       return 0;
 }
 
-asmlinkage long sys_timerfd_create(int clockid, int flags)
+SYSCALL_DEFINE2(timerfd_create, int, clockid, int, flags)
 {
-       int error, ufd;
+       int ufd;
        struct timerfd_ctx *ctx;
-       struct file *file;
-       struct inode *inode;
 
-       if (flags)
-               return -EINVAL;
-       if (clockid != CLOCK_MONOTONIC &&
-           clockid != CLOCK_REALTIME)
+       /* Check the TFD_* constants for consistency.  */
+       BUILD_BUG_ON(TFD_CLOEXEC != O_CLOEXEC);
+       BUILD_BUG_ON(TFD_NONBLOCK != O_NONBLOCK);
+
+       if ((flags & ~TFD_CREATE_FLAGS) ||
+           (clockid != CLOCK_MONOTONIC &&
+            clockid != CLOCK_REALTIME &&
+            clockid != CLOCK_REALTIME_ALARM &&
+            clockid != CLOCK_BOOTTIME &&
+            clockid != CLOCK_BOOTTIME_ALARM))
                return -EINVAL;
 
        ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
@@ -198,37 +327,44 @@ asmlinkage long sys_timerfd_create(int clockid, int flags)
 
        init_waitqueue_head(&ctx->wqh);
        ctx->clockid = clockid;
-       hrtimer_init(&ctx->tmr, clockid, HRTIMER_MODE_ABS);
 
-       error = anon_inode_getfd(&ufd, &inode, &file, "[timerfd]",
-                                &timerfd_fops, ctx);
-       if (error) {
+       if (isalarm(ctx))
+               alarm_init(&ctx->t.alarm,
+                          ctx->clockid == CLOCK_REALTIME_ALARM ?
+                          ALARM_REALTIME : ALARM_BOOTTIME,
+                          timerfd_alarmproc);
+       else
+               hrtimer_init(&ctx->t.tmr, clockid, HRTIMER_MODE_ABS);
+
+       ctx->moffs = ktime_get_monotonic_offset();
+
+       ufd = anon_inode_getfd("[timerfd]", &timerfd_fops, ctx,
+                              O_RDWR | (flags & TFD_SHARED_FCNTL_FLAGS));
+       if (ufd < 0)
                kfree(ctx);
-               return error;
-       }
 
        return ufd;
 }
 
-asmlinkage long sys_timerfd_settime(int ufd, int flags,
-                                   const struct itimerspec __user *utmr,
-                                   struct itimerspec __user *otmr)
+static int do_timerfd_settime(int ufd, int flags, 
+               const struct itimerspec *new,
+               struct itimerspec *old)
 {
-       struct file *file;
+       struct fd f;
        struct timerfd_ctx *ctx;
-       struct itimerspec ktmr, kotmr;
+       int ret;
 
-       if (copy_from_user(&ktmr, utmr, sizeof(ktmr)))
-               return -EFAULT;
-
-       if (!timespec_valid(&ktmr.it_value) ||
-           !timespec_valid(&ktmr.it_interval))
+       if ((flags & ~TFD_SETTIME_FLAGS) ||
+           !timespec_valid(&new->it_value) ||
+           !timespec_valid(&new->it_interval))
                return -EINVAL;
 
-       file = timerfd_fget(ufd);
-       if (IS_ERR(file))
-               return PTR_ERR(file);
-       ctx = file->private_data;
+       ret = timerfd_fget(ufd, &f);
+       if (ret)
+               return ret;
+       ctx = f.file->private_data;
+
+       timerfd_setup_cancel(ctx, flags);
 
        /*
         * We need to stop the existing timer before reprogramming
@@ -236,8 +372,14 @@ asmlinkage long sys_timerfd_settime(int ufd, int flags,
         */
        for (;;) {
                spin_lock_irq(&ctx->wqh.lock);
-               if (hrtimer_try_to_cancel(&ctx->tmr) >= 0)
-                       break;
+
+               if (isalarm(ctx)) {
+                       if (alarm_try_to_cancel(&ctx->t.alarm) >= 0)
+                               break;
+               } else {
+                       if (hrtimer_try_to_cancel(&ctx->t.tmr) >= 0)
+                               break;
+               }
                spin_unlock_irq(&ctx->wqh.lock);
                cpu_relax();
        }
@@ -248,48 +390,110 @@ asmlinkage long sys_timerfd_settime(int ufd, int flags,
         * We do not update "ticks" and "expired" since the timer will be
         * re-programmed again in the following timerfd_setup() call.
         */
-       if (ctx->expired && ctx->tintv.tv64)
-               hrtimer_forward_now(&ctx->tmr, ctx->tintv);
+       if (ctx->expired && ctx->tintv.tv64) {
+               if (isalarm(ctx))
+                       alarm_forward_now(&ctx->t.alarm, ctx->tintv);
+               else
+                       hrtimer_forward_now(&ctx->t.tmr, ctx->tintv);
+       }
 
-       kotmr.it_value = ktime_to_timespec(timerfd_get_remaining(ctx));
-       kotmr.it_interval = ktime_to_timespec(ctx->tintv);
+       old->it_value = ktime_to_timespec(timerfd_get_remaining(ctx));
+       old->it_interval = ktime_to_timespec(ctx->tintv);
 
        /*
         * Re-program the timer to the new value ...
         */
-       timerfd_setup(ctx, flags, &ktmr);
+       ret = timerfd_setup(ctx, flags, new);
 
        spin_unlock_irq(&ctx->wqh.lock);
-       fput(file);
-       if (otmr && copy_to_user(otmr, &kotmr, sizeof(kotmr)))
-               return -EFAULT;
-
-       return 0;
+       fdput(f);
+       return ret;
 }
 
-asmlinkage long sys_timerfd_gettime(int ufd, struct itimerspec __user *otmr)
+static int do_timerfd_gettime(int ufd, struct itimerspec *t)
 {
-       struct file *file;
+       struct fd f;
        struct timerfd_ctx *ctx;
-       struct itimerspec kotmr;
-
-       file = timerfd_fget(ufd);
-       if (IS_ERR(file))
-               return PTR_ERR(file);
-       ctx = file->private_data;
+       int ret = timerfd_fget(ufd, &f);
+       if (ret)
+               return ret;
+       ctx = f.file->private_data;
 
        spin_lock_irq(&ctx->wqh.lock);
        if (ctx->expired && ctx->tintv.tv64) {
                ctx->expired = 0;
-               ctx->ticks +=
-                       hrtimer_forward_now(&ctx->tmr, ctx->tintv) - 1;
-               hrtimer_restart(&ctx->tmr);
+
+               if (isalarm(ctx)) {
+                       ctx->ticks +=
+                               alarm_forward_now(
+                                       &ctx->t.alarm, ctx->tintv) - 1;
+                       alarm_restart(&ctx->t.alarm);
+               } else {
+                       ctx->ticks +=
+                               hrtimer_forward_now(&ctx->t.tmr, ctx->tintv)
+                               - 1;
+                       hrtimer_restart(&ctx->t.tmr);
+               }
        }
-       kotmr.it_value = ktime_to_timespec(timerfd_get_remaining(ctx));
-       kotmr.it_interval = ktime_to_timespec(ctx->tintv);
+       t->it_value = ktime_to_timespec(timerfd_get_remaining(ctx));
+       t->it_interval = ktime_to_timespec(ctx->tintv);
        spin_unlock_irq(&ctx->wqh.lock);
-       fput(file);
+       fdput(f);
+       return 0;
+}
+
+SYSCALL_DEFINE4(timerfd_settime, int, ufd, int, flags,
+               const struct itimerspec __user *, utmr,
+               struct itimerspec __user *, otmr)
+{
+       struct itimerspec new, old;
+       int ret;
+
+       if (copy_from_user(&new, utmr, sizeof(new)))
+               return -EFAULT;
+       ret = do_timerfd_settime(ufd, flags, &new, &old);
+       if (ret)
+               return ret;
+       if (otmr && copy_to_user(otmr, &old, sizeof(old)))
+               return -EFAULT;
+
+       return ret;
+}
 
+SYSCALL_DEFINE2(timerfd_gettime, int, ufd, struct itimerspec __user *, otmr)
+{
+       struct itimerspec kotmr;
+       int ret = do_timerfd_gettime(ufd, &kotmr);
+       if (ret)
+               return ret;
        return copy_to_user(otmr, &kotmr, sizeof(kotmr)) ? -EFAULT: 0;
 }
 
+#ifdef CONFIG_COMPAT
+COMPAT_SYSCALL_DEFINE4(timerfd_settime, int, ufd, int, flags,
+               const struct compat_itimerspec __user *, utmr,
+               struct compat_itimerspec __user *, otmr)
+{
+       struct itimerspec new, old;
+       int ret;
+
+       if (get_compat_itimerspec(&new, utmr))
+               return -EFAULT;
+       ret = do_timerfd_settime(ufd, flags, &new, &old);
+       if (ret)
+               return ret;
+       if (otmr && put_compat_itimerspec(otmr, &old))
+               return -EFAULT;
+       return ret;
+}
+
+COMPAT_SYSCALL_DEFINE2(timerfd_gettime, int, ufd,
+               struct compat_itimerspec __user *, otmr)
+{
+       struct itimerspec kotmr;
+       int ret = do_timerfd_gettime(ufd, &kotmr);
+       if (ret)
+               return ret;
+       return put_compat_itimerspec(otmr, &kotmr) ? -EFAULT: 0;
+}
+#endif