(5/8/12 3:36 AM), Pekka Enberg wrote:
On Tue, May 8, 2012 at 10:11 AM, KOSAKI Motohiro kosaki.motohiro@gmail.com wrote:
Ok, sane. Then I take my time a little and review current vmevent code briefly. (I read vmevent/core branch in pekka's tree. please let me know if there is newer repositry)
It's the latest one.
On Tue, May 8, 2012 at 10:11 AM, KOSAKI Motohiro kosaki.motohiro@gmail.com wrote:
- sample_period is brain damaged idea. If people ONLY need to
sampling stastics, they only need to read /proc/vmstat periodically. just remove it and implement push notification. _IF_ someone need unfrequent level trigger, just use "usleep(timeout); read(vmevent_fd)" on userland code.
That comes from a real-world requirement. See Leonid's email on the topic:
I know, many embedded guys prefer such timer interval. I also have an experience similar logic when I was TV box developer. but I must disagree. Someone hope timer housekeeping complexity into kernel. but I haven't seen any justification.
- VMEVENT_ATTR_STATE_ONE_SHOT is misleading name. That is effect as
edge trigger shot. not only once.
Would VMEVENT_ATTR_STATE_EDGE_TRIGGER be a better name?
maybe.
- vmevent_fd() seems sane interface. but it has name space unaware.
maybe we discuss how to harmonize name space feature. No hurry. but we have to think that issue since at beginning.
You mean VFS namespaces? Yeah, we need to take care of that.
If we keep current vmevent_fd() design, we may need to create new namespace concept likes ipc namespace. current vmevent_fd() is not VFS based.
- Currently, vmstat have per-cpu batch and vmstat updating makes 3
second delay at maximum. This is fine for usual case because almost userland watcher only read /proc/vmstat per second. But, for vmevent_fd() case, 3 seconds may be unacceptable delay. At worst, 128 batch x 4096 x 4k pagesize = 2G bytes inaccurate is there.
That's pretty awful. Anton, Leonid, comments?
- __VMEVENT_ATTR_STATE_VALUE_WAS_LT should be removed from userland
exporting files. When exporing kenrel internal, always silly gus used them and made unhappy.
Agreed. Anton, care to cook up a patch to do that?
- Also vmevent_event must hide from userland.
Why? That's part of the ABI.
Ahhh, if so, I missed something. as far as I look, vmevent_fd() only depend on vmevent_config. which syscall depend on vmevent_evennt?
- vmevent_config::size must be removed. In 20th century, M$ API
prefer to use this technique. But They dropped the way because a lot of application don't initialize size member and they can't use it for keeping upper compitibility.
It's there to support forward/backward ABI compatibility like perf does. I'm going to keep it for now but I'm open to dropping it when the ABI is more mature.
perf api is not intended to use from generic applications. then, I don't think it will make abi issue. tool/perf is sane, isn't it? but vmevent_fd() is generic api and we can't trust all userland guy have sane, unfortunately.
- memcg unaware
- numa unaware
- zone unaware
Yup.
And, we may need vm internal change if we really need lowmem notification. current kernel don't have such info. _And_ there is one more big problem. Currently the kernel maintain memory per zone. But almost all userland application aren't aware zone nor node. Thus raw notification aren't useful for userland. In the other hands, total memory and total free memory is useful? Definitely No! Even though total free memory are lots, system may start swap out and oom invokation. If we can't oom invocation, this feature has serious raison d'etre issue. (i.e. (4), (8), (9) and (19) are not ignorable issue. I think)
I'm guessing most of the existing solutions get away with approximations and soft limits because they're mostly used on UMA embedded machines.
But yes, we need to do better here.
Hm. If you want vmevent makes depend on CONFIG_EMBEDDED, I have no reason to complain this feature. At that world, almost all applications _know_ their system configuration. then I don't think api misuse issue is big matter.