This make epoll use hrtimers for the timeout value which prevents
epoll_wait() from timing out up to a millisecond early.
This mirrors the behavior of select() and poll().
Signed-off-by: Shawn Bohrer <[email protected]>
---
fs/eventpoll.c | 35 +++++++++++++++++++----------------
fs/select.c | 2 +-
include/linux/poll.h | 2 ++
3 files changed, 22 insertions(+), 17 deletions(-)
diff --git a/fs/eventpoll.c b/fs/eventpoll.c
index 3817149..728f56c 100644
--- a/fs/eventpoll.c
+++ b/fs/eventpoll.c
@@ -77,9 +77,6 @@
/* Maximum number of nesting allowed inside epoll sets */
#define EP_MAX_NESTS 4
-/* Maximum msec timeout value storeable in a long int */
-#define EP_MAX_MSTIMEO min(1000ULL * MAX_SCHEDULE_TIMEOUT / HZ, (LONG_MAX - 999ULL) / HZ)
-
#define EP_MAX_EVENTS (INT_MAX / sizeof(struct epoll_event))
#define EP_UNACTIVE_PTR ((void *) -1L)
@@ -1116,18 +1113,22 @@ static int ep_send_events(struct eventpoll *ep,
static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
int maxevents, long timeout)
{
- int res, eavail;
+ int res, eavail, timed_out = 0;
unsigned long flags;
- long jtimeout;
+ long slack;
wait_queue_t wait;
-
- /*
- * Calculate the timeout by checking for the "infinite" value (-1)
- * and the overflow condition. The passed timeout is in milliseconds,
- * that why (t * HZ) / 1000.
- */
- jtimeout = (timeout < 0 || timeout >= EP_MAX_MSTIMEO) ?
- MAX_SCHEDULE_TIMEOUT : (timeout * HZ + 999) / 1000;
+ struct timespec end_time;
+ ktime_t expires, *to = NULL;
+
+ if (timeout > 0) {
+ ktime_get_ts(&end_time);
+ timespec_add_ns(&end_time, (u64)timeout * NSEC_PER_MSEC);
+ slack = estimate_accuracy(&end_time);
+ to = &expires;
+ *to = timespec_to_ktime(end_time);
+ } else if (timeout == 0) {
+ timed_out = 1;
+ }
retry:
spin_lock_irqsave(&ep->lock, flags);
@@ -1149,7 +1150,7 @@ retry:
* to TASK_INTERRUPTIBLE before doing the checks.
*/
set_current_state(TASK_INTERRUPTIBLE);
- if (!list_empty(&ep->rdllist) || !jtimeout)
+ if (!list_empty(&ep->rdllist) || timed_out)
break;
if (signal_pending(current)) {
res = -EINTR;
@@ -1157,7 +1158,9 @@ retry:
}
spin_unlock_irqrestore(&ep->lock, flags);
- jtimeout = schedule_timeout(jtimeout);
+ if (!schedule_hrtimeout_range(to, slack, HRTIMER_MODE_ABS))
+ timed_out = 1;
+
spin_lock_irqsave(&ep->lock, flags);
}
__remove_wait_queue(&ep->wq, &wait);
@@ -1175,7 +1178,7 @@ retry:
* more luck.
*/
if (!res && eavail &&
- !(res = ep_send_events(ep, events, maxevents)) && jtimeout)
+ !(res = ep_send_events(ep, events, maxevents)) && !timed_out)
goto retry;
return res;
diff --git a/fs/select.c b/fs/select.c
index 500a669..003cb77 100644
--- a/fs/select.c
+++ b/fs/select.c
@@ -67,7 +67,7 @@ static long __estimate_accuracy(struct timespec *tv)
return slack;
}
-static long estimate_accuracy(struct timespec *tv)
+long estimate_accuracy(struct timespec *tv)
{
unsigned long ret;
struct timespec now;
diff --git a/include/linux/poll.h b/include/linux/poll.h
index 600cc1f..52be81f 100644
--- a/include/linux/poll.h
+++ b/include/linux/poll.h
@@ -73,6 +73,8 @@ extern void poll_initwait(struct poll_wqueues *pwq);
extern void poll_freewait(struct poll_wqueues *pwq);
extern int poll_schedule_timeout(struct poll_wqueues *pwq, int state,
ktime_t *expires, unsigned long slack);
+extern long estimate_accuracy(struct timespec *tv);
+
static inline int poll_schedule(struct poll_wqueues *pwq, int state)
{
--
1.7.2.1
I'd say this is an epoll patch, not an hrtimer patch. So I renamed it
to "epoll: make epoll_wait() use the hrtimer range feature".
Davide looks after epoll, so let's cc him.
On Sun, 8 Aug 2010 17:45:32 -0500
Shawn Bohrer <[email protected]> wrote:
> This make epoll use hrtimers for the timeout value which prevents
> epoll_wait() from timing out up to a millisecond early.
>
> This mirrors the behavior of select() and poll().
>
> ...
>
> --- a/fs/eventpoll.c
> +++ b/fs/eventpoll.c
Davide stuff ;)
> --- a/fs/select.c
> +++ b/fs/select.c
> @@ -67,7 +67,7 @@ static long __estimate_accuracy(struct timespec *tv)
> return slack;
> }
>
> -static long estimate_accuracy(struct timespec *tv)
> +long estimate_accuracy(struct timespec *tv)
"estimate_accuracy" is a rotten name for a global symbol. I queued a
preparatory patch which renames this to "select_estimate_accuracy".
On Sun, 8 Aug 2010, Shawn Bohrer wrote:
> This make epoll use hrtimers for the timeout value which prevents
> epoll_wait() from timing out up to a millisecond early.
>
> This mirrors the behavior of select() and poll().
I saw this now, since I got notifications from Andrew's patch machine.
Is this really needed?
Is that really worth the extra code in the fast path?
Aren't we breaking existing behavior?
I have some doubts.
- Davide
On Thu, 26 Aug 2010, Davide Libenzi wrote:
> On Sun, 8 Aug 2010, Shawn Bohrer wrote:
>
> > This make epoll use hrtimers for the timeout value which prevents
> > epoll_wait() from timing out up to a millisecond early.
> >
> > This mirrors the behavior of select() and poll().
>
> I saw this now, since I got notifications from Andrew's patch machine.
> Is this really needed?
Yes, as it is the last user space interface which is jiffies based.
> Is that really worth the extra code in the fast path?
It's not that much overhead and we had no complaints when we converted
everything else
> Aren't we breaking existing behavior?
We had no problems when we moved all the other interfaces over
> I have some doubts.
/me not - though I need to look at the patch itself
Thanks,
tglx
On Fri, 27 Aug 2010, Thomas Gleixner wrote:
> On Thu, 26 Aug 2010, Davide Libenzi wrote:
> > On Sun, 8 Aug 2010, Shawn Bohrer wrote:
> >
> > > This make epoll use hrtimers for the timeout value which prevents
> > > epoll_wait() from timing out up to a millisecond early.
> > >
> > > This mirrors the behavior of select() and poll().
> >
> > I saw this now, since I got notifications from Andrew's patch machine.
> > Is this really needed?
>
> Yes, as it is the last user space interface which is jiffies based.
What? You killed jiffies?! :)
Alright then, the patch itself, besides the different bahavior on timeout
greater/equal to EP_MAX_MSTIMEO (which is not documented, hence does not
exist), looks OK to me.
- Davide
On Sun, Aug 8, 2010 at 18:45, Shawn Bohrer wrote:
> @@ -1116,18 +1113,22 @@ static int ep_send_events(struct eventpoll *ep,
> static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
> int maxevents, long timeout)
> {
> - int res, eavail;
> + int res, eavail, timed_out = 0;
> unsigned long flags;
> - long jtimeout;
> + long slack;
> wait_queue_t wait;
> -
> - /*
> - * Calculate the timeout by checking for the "infinite" value (-1)
> - * and the overflow condition. The passed timeout is in milliseconds,
> - * that why (t * HZ) / 1000.
> - */
> - jtimeout = (timeout < 0 || timeout >= EP_MAX_MSTIMEO) ?
> - MAX_SCHEDULE_TIMEOUT : (timeout * HZ + 999) / 1000;
> + struct timespec end_time;
> + ktime_t expires, *to = NULL;
> +
> + if (timeout > 0) {
> + ktime_get_ts(&end_time);
> + timespec_add_ns(&end_time, (u64)timeout * NSEC_PER_MSEC);
> + slack = estimate_accuracy(&end_time);
> + to = &expires;
> + *to = timespec_to_ktime(end_time);
> + } else if (timeout == 0) {
> + timed_out = 1;
> + }
>
> retry:
> spin_lock_irqsave(&ep->lock, flags);
> @@ -1149,7 +1150,7 @@ retry:
> * to TASK_INTERRUPTIBLE before doing the checks.
> */
> set_current_state(TASK_INTERRUPTIBLE);
> - if (!list_empty(&ep->rdllist) || !jtimeout)
> + if (!list_empty(&ep->rdllist) || timed_out)
> break;
> if (signal_pending(current)) {
> res = -EINTR;
> @@ -1157,7 +1158,9 @@ retry:
> }
>
> spin_unlock_irqrestore(&ep->lock, flags);
> - jtimeout = schedule_timeout(jtimeout);
> + if (!schedule_hrtimeout_range(to, slack, HRTIMER_MODE_ABS))
> + timed_out = 1;
> +
> spin_lock_irqsave(&ep->lock, flags);
> }
> __remove_wait_queue(&ep->wq, &wait);
this code introduces a warning:
fs/eventpoll.c: In function ‘ep_poll’:
fs/eventpoll.c:1119: warning: ‘slack’ may be used uninitialized in this function
looks to me like you arent properly handling negative timeouts.
certainly epoll_wait() passes the timeout value straight from
userspace to ep_poll() without any error checking, so if userspace
passes a negative timeout value, it looks like "slack" will be used
uninitialized.
-mike
????{.n?+???????+%?????ݶ??w??{.n?+????{??G?????{ay? ʇڙ?,j??f???h?????????z_??(?階?ݢj"???m??????G????????????&???~???iO???z??v?^?m???? ????????I?
On Wed, Nov 24, 2010 at 03:33:02AM -0500, Mike Frysinger wrote:
> On Sun, Aug 8, 2010 at 18:45, Shawn Bohrer wrote:
> > @@ -1116,18 +1113,22 @@ static int ep_send_events(struct eventpoll *ep,
> > static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
> > int maxevents, long timeout)
> > {
> > - int res, eavail;
> > + int res, eavail, timed_out = 0;
> > unsigned long flags;
> > - long jtimeout;
> > + long slack;
> > wait_queue_t wait;
> > -
> > - /*
> > - * Calculate the timeout by checking for the "infinite" value (-1)
> > - * and the overflow condition. The passed timeout is in milliseconds,
> > - * that why (t * HZ) / 1000.
> > - */
> > - jtimeout = (timeout < 0 || timeout >= EP_MAX_MSTIMEO) ?
> > - MAX_SCHEDULE_TIMEOUT : (timeout * HZ + 999) / 1000;
> > + struct timespec end_time;
> > + ktime_t expires, *to = NULL;
> > +
> > + if (timeout > 0) {
> > + ktime_get_ts(&end_time);
> > + timespec_add_ns(&end_time, (u64)timeout * NSEC_PER_MSEC);
> > + slack = estimate_accuracy(&end_time);
> > + to = &expires;
> > + *to = timespec_to_ktime(end_time);
> > + } else if (timeout == 0) {
> > + timed_out = 1;
> > + }
> >
> > retry:
> > spin_lock_irqsave(&ep->lock, flags);
> > @@ -1149,7 +1150,7 @@ retry:
> > * to TASK_INTERRUPTIBLE before doing the checks.
> > */
> > set_current_state(TASK_INTERRUPTIBLE);
> > - if (!list_empty(&ep->rdllist) || !jtimeout)
> > + if (!list_empty(&ep->rdllist) || timed_out)
> > break;
> > if (signal_pending(current)) {
> > res = -EINTR;
> > @@ -1157,7 +1158,9 @@ retry:
> > }
> >
> > spin_unlock_irqrestore(&ep->lock, flags);
> > - jtimeout = schedule_timeout(jtimeout);
> > + if (!schedule_hrtimeout_range(to, slack, HRTIMER_MODE_ABS))
> > + timed_out = 1;
> > +
> > spin_lock_irqsave(&ep->lock, flags);
> > }
> > __remove_wait_queue(&ep->wq, &wait);
>
> this code introduces a warning:
> fs/eventpoll.c: In function ‘ep_poll’:
> fs/eventpoll.c:1119: warning: ‘slack’ may be used uninitialized in this function
>
> looks to me like you arent properly handling negative timeouts.
> certainly epoll_wait() passes the timeout value straight from
> userspace to ep_poll() without any error checking, so if userspace
> passes a negative timeout value, it looks like "slack" will be used
> uninitialized.
If a negative timeout is passed in then 'to' remains NULL. When 'to
is NULL schedule_hrtimeout_range() has an infinite timeout and the
'slack' parameter is never used. So technically everything should be
fine here.
Of course it would be safest and best to simply initialize slack to 0.
I can send a patch this evening with the fix.
--
Shawn
On Wed, Nov 24, 2010 at 09:52, Shawn Bohrer wrote:
> On Wed, Nov 24, 2010 at 03:33:02AM -0500, Mike Frysinger wrote:
>> On Sun, Aug 8, 2010 at 18:45, Shawn Bohrer wrote:
>> > @@ -1116,18 +1113,22 @@ static int ep_send_events(struct eventpoll *ep,
>> > static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
>> > int maxevents, long timeout)
>> > {
>> > - int res, eavail;
>> > + int res, eavail, timed_out = 0;
>> > unsigned long flags;
>> > - long jtimeout;
>> > + long slack;
>> > wait_queue_t wait;
>> > -
>> > - /*
>> > - * Calculate the timeout by checking for the "infinite" value (-1)
>> > - * and the overflow condition. The passed timeout is in milliseconds,
>> > - * that why (t * HZ) / 1000.
>> > - */
>> > - jtimeout = (timeout < 0 || timeout >= EP_MAX_MSTIMEO) ?
>> > - MAX_SCHEDULE_TIMEOUT : (timeout * HZ + 999) / 1000;
>> > + struct timespec end_time;
>> > + ktime_t expires, *to = NULL;
>> > +
>> > + if (timeout > 0) {
>> > + ktime_get_ts(&end_time);
>> > + timespec_add_ns(&end_time, (u64)timeout * NSEC_PER_MSEC);
>> > + slack = estimate_accuracy(&end_time);
>> > + to = &expires;
>> > + *to = timespec_to_ktime(end_time);
>> > + } else if (timeout == 0) {
>> > + timed_out = 1;
>> > + }
>> >
>> > retry:
>> > spin_lock_irqsave(&ep->lock, flags);
>> > @@ -1149,7 +1150,7 @@ retry:
>> > * to TASK_INTERRUPTIBLE before doing the checks.
>> > */
>> > set_current_state(TASK_INTERRUPTIBLE);
>> > - if (!list_empty(&ep->rdllist) || !jtimeout)
>> > + if (!list_empty(&ep->rdllist) || timed_out)
>> > break;
>> > if (signal_pending(current)) {
>> > res = -EINTR;
>> > @@ -1157,7 +1158,9 @@ retry:
>> > }
>> >
>> > spin_unlock_irqrestore(&ep->lock, flags);
>> > - jtimeout = schedule_timeout(jtimeout);
>> > + if (!schedule_hrtimeout_range(to, slack, HRTIMER_MODE_ABS))
>> > + timed_out = 1;
>> > +
>> > spin_lock_irqsave(&ep->lock, flags);
>> > }
>> > __remove_wait_queue(&ep->wq, &wait);
>>
>> this code introduces a warning:
>> fs/eventpoll.c: In function ‘ep_poll’:
>> fs/eventpoll.c:1119: warning: ‘slack’ may be used uninitialized in this function
>>
>> looks to me like you arent properly handling negative timeouts.
>> certainly epoll_wait() passes the timeout value straight from
>> userspace to ep_poll() without any error checking, so if userspace
>> passes a negative timeout value, it looks like "slack" will be used
>> uninitialized.
>
> If a negative timeout is passed in then 'to' remains NULL. When 'to
> is NULL schedule_hrtimeout_range() has an infinite timeout and the
> 'slack' parameter is never used. So technically everything should be
> fine here.
ok, but that depends on an external function never changing behavior
and makes changing the API pretty hard since all callers must be
closely analyzed
> Of course it would be safest and best to simply initialize slack to 0.
> I can send a patch this evening with the fix.
thanks !
-mike
When a negative timeout value is passed to epoll the 'slack' variable is
currently uninitialized:
fs/eventpoll.c: In function ‘ep_poll’:
fs/eventpoll.c:1119: warning: ‘slack’ may be used uninitialized in this function
In this case a NULL pointer is passed to schedule_hrtimeout_range()
specifying an infinite timeout. The current implementation of
schedule_hrtimeout_range() does not use slack in this case, but we
should still initialize slack to 0 in case future implementations use it.
Signed-off-by: Shawn Bohrer <[email protected]>
---
fs/eventpoll.c | 2 +-
1 files changed, 1 insertions(+), 1 deletions(-)
diff --git a/fs/eventpoll.c b/fs/eventpoll.c
index 8cf0724..c24a032 100644
--- a/fs/eventpoll.c
+++ b/fs/eventpoll.c
@@ -1116,7 +1116,7 @@ static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
{
int res, eavail, timed_out = 0;
unsigned long flags;
- long slack;
+ long slack = 0;
wait_queue_t wait;
struct timespec end_time;
ktime_t expires, *to = NULL;
--
1.7.3.2
On Wed, 24 Nov 2010, Shawn Bohrer wrote:
> When a negative timeout value is passed to epoll the 'slack' variable is
> currently uninitialized:
>
> fs/eventpoll.c: In function ‘ep_poll’:
> fs/eventpoll.c:1119: warning: ‘slack’ may be used uninitialized in this function
>
> In this case a NULL pointer is passed to schedule_hrtimeout_range()
> specifying an infinite timeout. The current implementation of
> schedule_hrtimeout_range() does not use slack in this case, but we
> should still initialize slack to 0 in case future implementations use it.
Thanks.
> Signed-off-by: Shawn Bohrer <[email protected]>
Acked-by: Davide Libenzi <[email protected]>
> ---
> fs/eventpoll.c | 2 +-
> 1 files changed, 1 insertions(+), 1 deletions(-)
>
> diff --git a/fs/eventpoll.c b/fs/eventpoll.c
> index 8cf0724..c24a032 100644
> --- a/fs/eventpoll.c
> +++ b/fs/eventpoll.c
> @@ -1116,7 +1116,7 @@ static int ep_poll(struct eventpoll *ep, struct epoll_event __user *events,
> {
> int res, eavail, timed_out = 0;
> unsigned long flags;
> - long slack;
> + long slack = 0;
> wait_queue_t wait;
> struct timespec end_time;
> ktime_t expires, *to = NULL;
- Davide