Subject: Re: [RFC 4/4] drm: add support for raw monotonic vblank timestamps Message-ID: 1349446447.17758.73.camel@thor.local Content-Type: text/plain; charset="ISO-8859-1"
On Fre, 2012-10-05 at 16:59 +0300, Imre Deak wrote:
On Fri, 2012-10-05 at 15:55 +0200, Michel D?nzer wrote:
On Fre, 2012-10-05 at 16:37 +0300, Imre Deak wrote:
In practice we never want the timestamps for vblank and page flip events to be affected by time adjustments, so in addition to the gettimeofday timestamps we used so far add support for raw monotonic timestamps.
For backward compatibility use flags to select between the old and new timestamp format.
Note that with this change we will save the timestamp in both formats, for cases where multiple clients are expecting an event notification in different time formats.
I wonder if all this trouble is really necessary. I honestly can't imagine any user of this API requiring non-monotonic timestamps and breaking with monotonic ones. I think it was simply a mistake that we didn't make them monotonic in the first place (or maybe it wasn't even possible when this API was first introduced).
Hi
The OML_sync_control spec requires monotonic timestamps, so yes, it was either a mistake or it wasn't possible at that time.
If an app developer follows the spec, he would consider the current behaviour broken -- or should be able to handle monotonic timestamps.
Yea, I'd rather simply switch over to monotonic timestamps too. But that would break apps that already compare against the wall time for whatever purpose (for example A/V sync).
Are there actually any such apps in the real world? Do they work when the wall time jumps?
Psychtoolbox as my example of a very timing sensitive app: Currently uses wall time as a reference for historical reasons. We tell/expect users not to do anything in the middle of a work session that could cause a sudden jump in wall time, but then ntp time adjustments are usually in very small steps, so not a problem, and big adjustments should only happen if users do something they'll certainly not do during a timing-sensitve session, like manually set the system clock, reboot/suspend/resume the machine etc. In practice, this works well enough over the typical duration of such sessions (minutes to a few hours) and is sometimes needed if multiple machines need to be time-synced over the local network.
But then Psychtoolbox checks each timestamp it gets from somewhere "outside" (OML_sync_control / INTEL_swap_events / ALSA audio timestamps, network receive timestamps, evdev, x11, ...) if it is in gettimeofday() aka CLOCK_REALTIME aka wall time or in CLOCK_MONOTONIC time and just remaps to whatever its reference clock is.
There's no way around this than to have no fixed expectations, but to remap stuff on the fly, because different parts of the Linux universe have decided on different time bases, or even switched somewhere from one kernel version to the next in the last years, e.g., ALSA, which at some time switched from wall clock to CLOCK_MONOTONIC. Sometimes clock_gettime() wasn't available at all in older setups, so there only was gettimeofday(). Or toolkits like GStreamer use different timebases dependent on OS and sometimes even on plugins.
I would expect that other timing sensitive apps have to have ways to handle this in similar ways.
Wrt. to the drm vblank/pageflip timestamps, the userspace extensions which expose these (INTEL_swap_events, OML_sync_control) don't allow apps to select which timebase to use, they define monotonic time as what is returned, so i don't know how a userspace app could actually ask the DRM for one or the other format? So i guess just switching to CLOCK_MONOTONIC shouldn't be that bad.
Kristian, i assume Wayland will also return presentation timestamps in the format and microsecond precision of the DRM, right?
On 05.10.12 18:22, intel-gfx-request@lists.freedesktop.org wrote:
Message: 7 Date: Fri, 5 Oct 2012 12:14:29 -0400 From: Kristian H?gsberg
...
I just had a quick look at driver/input/evdev.c, since evdev devices did a similar change recently to allow evdev timestamp from the monotonic clock. They're using a different time API:
time_mono = ktime_get(); time_real = ktime_sub(time_mono, ktime_get_monotonic_offset());
and
event->time = ktime_to_timeval(client->clkid == CLOCK_MONOTONIC ? mono : real);I'm not really up-to-date on kernel time APIs, but I wonder if that may be better? At least, I suspect we wouldn't need changes outside drm if we use this API.
Kristian
Userspace apps only have access to what gettimeofday() and clock_gettime() for CLOCK_REALTIME (== gettimeofday() afaik) and CLOCK_MONOTONIC return, so whatever is returned should be in CLOCK_MONOTONIC format, otherwise there will be lots of tears and dead kittens. I think what evdev does makes a lot of sense, but i'm also not up-to-date about the various layers of timing apis.
-mario
On Sat, 2012-10-06 at 03:41 +0200, Mario Kleiner wrote:
[...]
But then Psychtoolbox checks each timestamp it gets from somewhere "outside" (OML_sync_control / INTEL_swap_events / ALSA audio timestamps, network receive timestamps, evdev, x11, ...) if it is in gettimeofday() aka CLOCK_REALTIME aka wall time or in CLOCK_MONOTONIC time and just remaps to whatever its reference clock is.
There's no way around this than to have no fixed expectations, but to remap stuff on the fly, because different parts of the Linux universe have decided on different time bases, or even switched somewhere from one kernel version to the next in the last years, e.g., ALSA, which at some time switched from wall clock to CLOCK_MONOTONIC. Sometimes clock_gettime() wasn't available at all in older setups, so there only was gettimeofday(). Or toolkits like GStreamer use different timebases dependent on OS and sometimes even on plugins.
I would expect that other timing sensitive apps have to have ways to handle this in similar ways.
I think the question is, can we be sure? I don't think there is any guarantee that random application X will not be confused by an unconditional switch to monotonic timestamps.
Wrt. to the drm vblank/pageflip timestamps, the userspace extensions which expose these (INTEL_swap_events, OML_sync_control) don't allow apps to select which timebase to use, they define monotonic time as what is returned, so i don't know how a userspace app could actually ask the DRM for one or the other format? So i guess just switching to CLOCK_MONOTONIC shouldn't be that bad.
An application could just use the kernel DRM interface directly. I admit this is not very likely but this is what should determine the rules by which we change the ABI.
Kristian, i assume Wayland will also return presentation timestamps in the format and microsecond precision of the DRM, right?
On 05.10.12 18:22, intel-gfx-request@lists.freedesktop.org wrote:
Message: 7 Date: Fri, 5 Oct 2012 12:14:29 -0400 From: Kristian H?gsberg
...
I just had a quick look at driver/input/evdev.c, since evdev devices did a similar change recently to allow evdev timestamp from the monotonic clock. They're using a different time API:
time_mono = ktime_get(); time_real = ktime_sub(time_mono, ktime_get_monotonic_offset());
and
event->time = ktime_to_timeval(client->clkid == CLOCK_MONOTONIC ? mono : real);I'm not really up-to-date on kernel time APIs, but I wonder if that may be better? At least, I suspect we wouldn't need changes outside drm if we use this API.
Kristian
Userspace apps only have access to what gettimeofday() and clock_gettime() for CLOCK_REALTIME (== gettimeofday() afaik) and CLOCK_MONOTONIC return, so whatever is returned should be in CLOCK_MONOTONIC format, otherwise there will be lots of tears and dead kittens. I think what evdev does makes a lot of sense, but i'm also not up-to-date about the various layers of timing apis.
Yes, this should be the case, regardless of which kernel interface we decide to use.
--Imre
On 08.10.12 13:35, Imre Deak wrote:
On Sat, 2012-10-06 at 03:41 +0200, Mario Kleiner wrote:
[...]
But then Psychtoolbox checks each timestamp it gets from somewhere "outside" (OML_sync_control / INTEL_swap_events / ALSA audio timestamps, network receive timestamps, evdev, x11, ...) if it is in gettimeofday() aka CLOCK_REALTIME aka wall time or in CLOCK_MONOTONIC time and just remaps to whatever its reference clock is.
There's no way around this than to have no fixed expectations, but to remap stuff on the fly, because different parts of the Linux universe have decided on different time bases, or even switched somewhere from one kernel version to the next in the last years, e.g., ALSA, which at some time switched from wall clock to CLOCK_MONOTONIC. Sometimes clock_gettime() wasn't available at all in older setups, so there only was gettimeofday(). Or toolkits like GStreamer use different timebases dependent on OS and sometimes even on plugins.
I would expect that other timing sensitive apps have to have ways to handle this in similar ways.
I think the question is, can we be sure? I don't think there is any guarantee that random application X will not be confused by an unconditional switch to monotonic timestamps.
Wrt. to the drm vblank/pageflip timestamps, the userspace extensions which expose these (INTEL_swap_events, OML_sync_control) don't allow apps to select which timebase to use, they define monotonic time as what is returned, so i don't know how a userspace app could actually ask the DRM for one or the other format? So i guess just switching to CLOCK_MONOTONIC shouldn't be that bad.
An application could just use the kernel DRM interface directly. I admit this is not very likely but this is what should determine the rules by which we change the ABI.
Ok, that's a good point, agreed. -mario
On Tue, 2012-10-09 at 01:00 +0200, Mario Kleiner wrote:
On 08.10.12 13:35, Imre Deak wrote:
On Sat, 2012-10-06 at 03:41 +0200, Mario Kleiner wrote:
[...]
But then Psychtoolbox checks each timestamp it gets from somewhere "outside" (OML_sync_control / INTEL_swap_events / ALSA audio timestamps, network receive timestamps, evdev, x11, ...) if it is in gettimeofday() aka CLOCK_REALTIME aka wall time or in CLOCK_MONOTONIC time and just remaps to whatever its reference clock is.
There's no way around this than to have no fixed expectations, but to remap stuff on the fly, because different parts of the Linux universe have decided on different time bases, or even switched somewhere from one kernel version to the next in the last years, e.g., ALSA, which at some time switched from wall clock to CLOCK_MONOTONIC. Sometimes clock_gettime() wasn't available at all in older setups, so there only was gettimeofday(). Or toolkits like GStreamer use different timebases dependent on OS and sometimes even on plugins.
I would expect that other timing sensitive apps have to have ways to handle this in similar ways.
I think the question is, can we be sure? I don't think there is any guarantee that random application X will not be confused by an unconditional switch to monotonic timestamps.
Wrt. to the drm vblank/pageflip timestamps, the userspace extensions which expose these (INTEL_swap_events, OML_sync_control) don't allow apps to select which timebase to use, they define monotonic time as what is returned, so i don't know how a userspace app could actually ask the DRM for one or the other format? So i guess just switching to CLOCK_MONOTONIC shouldn't be that bad.
An application could just use the kernel DRM interface directly. I admit this is not very likely but this is what should determine the rules by which we change the ABI.
Ok, that's a good point, agreed.
Still pondering about this, since I wouldn't want to complicate things unnecessarily by storing both timestamps.
Based on the feedback you guys gave, I now also think we could simply switch over to monotonic timestamps and have a module flag to fall back to the old behavior. This is what ALSA does and what V4L2 people are planning to do as well. As you pointed out the mainstream users of this interface already expect a monotonic timestamp and I couldn't find either a good reason to support both formats at the same time.
--Imre
dri-devel@lists.freedesktop.org
-
Imre Deak -
Mario Kleiner