Message-ID: <20101001132300.GF10463@shisha.kicks-ass.net>
Date:	Fri, 1 Oct 2010 16:23:01 +0300
From:	Alexander Shishkin <virtuoso@...nd.org>
To:	Thomas Gleixner <tglx@...utronix.de>
Cc:	linux-kernel@...r.kernel.org, John Stultz <johnstul@...ibm.com>,
	Andrew Morton <akpm@...ux-foundation.org>,
	"H. Peter Anvin" <hpa@...or.com>,
	Kay Sievers <kay.sievers@...y.org>, Greg KH <gregkh@...e.de>,
	Chris Friesen <chris.friesen@...band.com>,
	Linus Torvalds <torvalds@...ux-foundation.org>,
	"Kirill A. Shutemov" <kirill@...temov.name>,
	Martin Schwidefsky <schwidefsky@...ibm.com>,
	Jon Hunter <jon-hunter@...com>, Ingo Molnar <mingo@...e.hu>,
	Peter Zijlstra <a.p.zijlstra@...llo.nl>,
	"Paul E. McKenney" <paulmck@...ux.vnet.ibm.com>,
	David Howells <dhowells@...hat.com>,
	Avi Kivity <avi@...hat.com>, John Kacur <jkacur@...hat.com>,
	Alexander Shishkin <virtuoso@...nd.org>
Subject: Re: [PATCH 1/7] notify userspace about time changes

On Fri, Sep 17, 2010 at 11:29:40 +0200, Thomas Gleixner wrote:
> On Fri, 17 Sep 2010, Alexander Shishkin wrote:
> 
> > Certain userspace applications (like "clock" desktop applets or cron) might
> > want to be notified when some other application changes the system time. It
> > might also be important for an application to be able to distinguish between
> > its own and somebody else's time changes.
> 
> That's overengineered. An application which fiddles with the time
> should be able to deal with that inside of the application.

Good point.

> > diff --git a/include/linux/syscalls.h b/include/linux/syscalls.h
> > index e6319d1..789f92e 100644
> > --- a/include/linux/syscalls.h
> > +++ b/include/linux/syscalls.h
> > @@ -819,6 +819,7 @@ asmlinkage long sys_fanotify_init(unsigned int flags, unsigned int event_f_flags
> >  asmlinkage long sys_fanotify_mark(int fanotify_fd, unsigned int flags,
> >  				  u64 mask, int fd,
> >  				  const char  __user *pathname);
> > +asmlinkage long sys_time_change_notify(int fd, unsigned int flags);
> 
> Please do not create a new syscall with a weird flag field. As I said
> above the selection of who changed the time is not necessary at all,
> so what you want is the selection of which change event and that
> should be an enum.

What if the application wants to be notified about events of both types?
enum { A, B, BOTH }; is more awkward than having them bitwise or'ed. Especially
if one day another event type turns up.
Userspace can, of course, have separate eventfds for different events, but
isn't that sort of interface limitation for no gain?

> >  int kernel_execve(const char *filename, const char *const argv[], const char *const envp[]);
> >  
> > diff --git a/include/linux/time.h b/include/linux/time.h
> > index 9f15ac7..d66045e 100644
> > --- a/include/linux/time.h
> > +++ b/include/linux/time.h
> > @@ -252,6 +252,26 @@ static __always_inline void timespec_add_ns(struct timespec *a, u64 ns)
> >  	a->tv_sec += __iter_div_u64_rem(a->tv_nsec + ns, NSEC_PER_SEC, &ns);
> >  	a->tv_nsec = ns;
> >  }
> > +
> > +/* time change events */
> > +#define TIME_EVENT_SET 0
> > +#define TIME_EVENT_ADJ 1
> 
> Enum please
> 
> > +#define TIME_CHANGE_NOTIFY_OTHERS	BIT(0)
> > +#define TIME_CHANGE_NOTIFY_OWN		BIT(1)
> > +#define TIME_CHANGE_NOTIFY_SET		BIT(2)
> > +#define TIME_CHANGE_NOTIFY_ADJUST	BIT(3)
> > +
> > +#define TIME_CHANGE_NOTIFY_MAX_USERS	1024
> 
> What's this random limit for ?

Well, I guess as long as some sensible rlimits are in effect, this is
indeed not needed.

> > diff --git a/init/Kconfig b/init/Kconfig
> > index 2de5b1c..504a51a 100644
> > --- a/init/Kconfig
> > +++ b/init/Kconfig
> > @@ -980,6 +980,13 @@ config PERF_USE_VMALLOC
> >  	help
> >  	  See tools/perf/design.txt for details
> >  
> > +config TIME_NOTIFY
> > +	bool "System time changes notification for userspace"
> > +	depends on EVENTFD
> > +	help
> > +	  Enable time change notification events to userspace via
> > +	  eventfd.
> > +
> 
> How is that config switch related to PERF ?   
> 
> >  menu "Kernel Performance Events And Counters"
> >  
> >  config PERF_EVENTS
> > diff --git a/kernel/Makefile b/kernel/Makefile
> > index 0b72d1a..ac53c67 100644
> > --- a/kernel/Makefile
> > +++ b/kernel/Makefile
> > @@ -104,6 +104,7 @@ obj-$(CONFIG_PERF_EVENTS) += perf_event.o
> >  obj-$(CONFIG_HAVE_HW_BREAKPOINT) += hw_breakpoint.o
> >  obj-$(CONFIG_USER_RETURN_NOTIFIER) += user-return-notifier.o
> >  obj-$(CONFIG_PADATA) += padata.o
> > +obj-$(CONFIG_TIME_NOTIFY) += time_notify.o
> 
> Please move this to kernel/time/

Ok.

> > diff --git a/kernel/sysctl.c b/kernel/sysctl.c
> > index f88552c..9e7d7a9 100644
> > --- a/kernel/sysctl.c
> > +++ b/kernel/sysctl.c
> > @@ -1441,6 +1441,17 @@ static struct ctl_table fs_table[] = {
> >  		.proc_handler	= proc_doulongvec_minmax,
> >  	},
> >  #endif /* CONFIG_AIO */
> > +#ifdef CONFIG_TIME_NOTIFY
> > +	{
> > +		.procname	= "time-change-notify-max-users",
> > +		.data		= &time_change_notify_max_users,
> > +		.maxlen		= sizeof(unsigned int),
> > +		.mode		= 0644,
> > +		.proc_handler	= proc_dointvec_minmax,
> > +		.extra1		= &zero,
> > +		.extra2		= &ten_thousand,
> > +	},
> 
> Do we really need another sysctl which nobody ever will use ?  

If we're sure that we don't want to limit the number of users, we
don't need this as well.

> > +#endif
> >  #ifdef CONFIG_INOTIFY_USER
> >  	{
> >  		.procname	= "inotify",
> > diff --git a/kernel/time.c b/kernel/time.c
> > index ba9b338..b4155b8 100644
> > --- a/kernel/time.c
> > +++ b/kernel/time.c
> > @@ -92,7 +92,9 @@ SYSCALL_DEFINE1(stime, time_t __user *, tptr)
> >  	if (err)
> >  		return err;
> >  
> > -	do_settimeofday(&tv);
> > +	err = do_settimeofday(&tv);
> > +	if (!err)
> > +		time_notify_all(TIME_EVENT_SET);
> 
> Crap. We really do _NOT_ want to sprinkle this into all syscalls which
> can be used to set the time. You already missed the case in

I do hope that there aren't any new ones coming.

> do_sys_settimeofday() when only a tz change happens. Can't we simply
> put that into do_settimeofday() and cover all cases ?

Yes, that's a better idea.

> >  }
> > @@ -215,6 +220,8 @@ SYSCALL_DEFINE1(adjtimex, struct timex __user *, txc_p)
> >  	if(copy_from_user(&txc, txc_p, sizeof(struct timex)))
> >  		return -EFAULT;
> >  	ret = do_adjtimex(&txc);
> > +	if (!ret)
> > +		time_notify_all(TIME_EVENT_ADJ);
> 
> That's nonsense. adjtimex is not necessarily setting the time. adjtimex
> can just read the current values w/o doing any modification. Also if
> adjtimex merily adjusts the clock drift parameters then we do NOT set
> the clock. So this is rather pointless AFAICT.

Well, I have an application here that would very much like to be notified
if something tries to adjust the clock drift. But yes, I should check for
txc->modes here.

> > +/*
> > + * A process can "subscribe" to receive a notification via eventfd that
> > + * some other process has called stime/settimeofday/adjtimex.
> > + */
> > +struct time_event {
> > +	struct eventfd_ctx	*eventfd;
> > +	struct task_struct	*watcher;
> > +	unsigned int		want_others:1;
> > +	unsigned int		want_own:1;
> > +	unsigned int		want_set:1;
> > +	unsigned int		want_adj:1;
> > +	struct work_struct	remove;
> > +	wait_queue_t		wq;
> > +	wait_queue_head_t	*wqh;
> > +	poll_table		pt;
> > +	struct list_head	list;
> > +};
> > +
> > +static LIST_HEAD(event_list);
> > +static int nevents;
> > +static DEFINE_SPINLOCK(event_lock);
> > +
> > +/*
> > + * Do the necessary cleanup when the eventfd is being closed
> > + */
> > +static void time_event_remove(struct work_struct *work)
> > +{
> > +	struct time_event *evt = container_of(work, struct time_event, remove);
> > +
> > +	BUG_ON(nevents <= 0);
> 
>   Errm. The accounting belongs to the place where we remove the event
>   from the event list. This counter can go away anyway.
> 
>   You leak the refcount on the eventfd_ctx here.
> 
> > +
> > +	kfree(evt);
> > +	nevents--;
> > +}
> > +
> > +static int time_event_wakeup(wait_queue_t *wq, unsigned int mode, int sync,
> > +			     void *key)
> > +{
> > +	struct time_event *evt = container_of(wq, struct time_event, wq);
> > +	unsigned long flags = (unsigned long)key;
> > +
> > +	if (flags & POLLHUP) {
> > +		__remove_wait_queue(evt->wqh, &evt->wq);
> 
>   Please don't do that. Merily delete it from the list and schedule it
>   for removal. Then in the remove function use
>   eventfd_ctx_remove_wait_queue() to cleanup the eventfd.

Ok, but I can't quite see the difference in this particular case, when we
only care about POLLHUP (apart from using eventfd interface for waitqueue
removal). There's a spinlock in that function, but I'm not quite sure what
can race when we're in POLLHUP situation. What am I missing?

> > +		spin_lock(&event_lock);
> > +		list_del(&evt->list);
> > +		spin_unlock(&event_lock);
> > +
> > +		schedule_work(&evt->remove);
> > +	}
> > +
> > +	return 0;
> > +}
> > +
> > +static void time_event_ptable_queue_proc(struct file *file,
> > +					 wait_queue_head_t *wqh, poll_table *pt)
> > +{
> > +	struct time_event *evt = container_of(pt, struct time_event, pt);
> > +
> > +	evt->wqh = wqh;
> > +	add_wait_queue(wqh, &evt->wq);
> > +}
> > +
> > +/*
> > + * time_change_notify() registers a given eventfd to receive time change
> > + * notifications
> > + */
> > +SYSCALL_DEFINE2(time_change_notify, int, fd, unsigned int, flags)
> > +{
> > +	int ret;
> > +	struct file *file;
> > +	struct time_event *evt;
> 
>   Reverting the order of these makes is 5 times easier to read.

Very true.

> > +
> > +	evt = kmalloc(sizeof(*evt), GFP_KERNEL);
> > +	if (!evt)
> > +		return -ENOMEM;
> > +
> > +	evt->want_others = !!(flags & TIME_CHANGE_NOTIFY_OTHERS);
> > +	evt->want_own = !!(flags & TIME_CHANGE_NOTIFY_OWN);
> > +	evt->want_set = !!(flags & TIME_CHANGE_NOTIFY_SET);
> > +	evt->want_adj = !!(flags & TIME_CHANGE_NOTIFY_ADJUST);
> 
>   Shudder. Please get rid of this.

Ok.

> > +	file = eventfd_fget(fd);
> > +	if (IS_ERR(file)) {
> > +		ret = -EINVAL;
> 
>   Grr. PTR_ERR(file) perhaps ?

Indeed.

> > +		goto out_free;
> > +	}
> > +
> > +	evt->eventfd = eventfd_ctx_fileget(file);
> > +	if (IS_ERR(evt->eventfd)) {
> > +		ret = PTR_ERR(evt->eventfd);
> > +		goto out_fput;
> > +	}
> > +
> > +	INIT_LIST_HEAD(&evt->list);
> > +	INIT_WORK(&evt->remove, time_event_remove);
> > +
> > +	init_waitqueue_func_entry(&evt->wq, time_event_wakeup);
> > +	init_poll_funcptr(&evt->pt, time_event_ptable_queue_proc);
> > +
> > +	evt->watcher = current;
> 
>   No need for this
> 
> > +	spin_lock(&event_lock);
> > +	if (nevents == time_change_notify_max_users) {
> > +		spin_unlock(&event_lock);
> > +		ret = -EBUSY;
> 
> 
>   Gah. You leak the refcount to eventfd.
> 
> > +		goto out_fput;
> > +	}
> > +
> > +	nevents++;
> > +	list_add(&evt->list, &event_list);
> > +	spin_unlock(&event_lock);
> > +
> > +	if (file->f_op->poll(file, &evt->pt) & POLLHUP) {
> > +		ret = 0;
> 
>   Ditto.
> 
> > +		goto out_fput;
> > +	}
> > +
> > +	fput(file);
> > +
> > +	return 0;
> > +
> > +out_fput:
> > +	fput(file);
> > +
> > +out_free:
> > +	kfree(evt);
> > +
> > +	return ret;
> > +}
> > +
> > +void time_notify_all(int type)
> > +{
> > +	struct list_head *tmp;
> > +
> > +	spin_lock(&event_lock);
> > +	list_for_each(tmp, &event_list) {
> > +		struct time_event *e = container_of(tmp, struct time_event,
> > +						    list);
> > +
> > +		if (type == TIME_EVENT_SET && !e->want_set)
> > +			continue;
> > +		else if (type == TIME_EVENT_ADJ && !e->want_adj)
> > +			continue;
> > +
> > +		if (e->watcher == current && !e->want_own)
> > +			continue;
> > +		else if (e->watcher != current && !e->want_others)
> > +			continue;
> 
>   This needs a major cleanup.
> 
> > +		eventfd_signal(e->eventfd, 1);
> > +	}
> > +	spin_unlock(&event_lock);
> > +}
> > +
> > +static int time_notify_init(void)
> > +{
> > +	return 0;
> > +}
> > +
> > +core_initcall(time_notify_init);
> 
> What's the exact purpose of this initcall ?

This should have been removed, yes.

Regards,
--
Alex
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists