Prev: [RFC][PATCH 0/6] xtime/wall_to_monotonic cleanups
Next: [PATCH] remove pointless debug printk in rt_ioctl_giwscan()
From: mark gross on 5 Jun 2010 00:10
On Fri, Jun 04, 2010 at 09:23:10AM -0500, James Bottomley wrote:
> On Thu, 2010-06-03 at 21:07 -0700, Arve Hj�nnev�g wrote:
> > On Thu, Jun 3, 2010 at 2:05 PM, Rafael J. Wysocki <rjw(a)sisk.pl> wrote:
> > > On Thursday 03 June 2010, James Bottomley wrote:
> > >> On Thu, 2010-06-03 at 00:10 -0700, Arve Hj�nnev�g wrote:
> > >> > On Wed, Jun 2, 2010 at 10:40 PM, mark gross <640e9920(a)gmail.com> wrote:
> > >> > > On Wed, Jun 02, 2010 at 09:54:15PM -0700, Brian Swetland wrote:
> > >> > >> On Wed, Jun 2, 2010 at 8:18 PM, mark gross <640e9920(a)gmail.com> wrote:
> > >> > >> > On Wed, Jun 02, 2010 at 02:58:30PM -0700, Arve Hj�nnev�g wrote:
> > >> > >> >>
> > >> > >> >> The list is not short. You have all the inactive and active
> > >> > >> >> constraints on the same list. If you change it to a two level list
> > >> > >> >> though, the list of unique values (which is the list you have to walk)
> > >> > >> >> may be short enough for a tree to be overkill.
> > >> > >> >
> > >> > >> > what have you seen in practice from the wake-lock stats?
> > >> > >> >
> > >> > >> > I'm having a hard time seeing where you could get more than just a
> > >> > >> > handfull. However; one could go to a dual list (like the scheduler) and
> > >> > >> > move inactive nodes from an active to inactive list, or we could simply
> > >> > >> > remove them from the list uppon inactivity. which would would well
> > >> > >> > after I change the api to have the client allocate the memory for the
> > >> > >> > nodes... BUT, if your moving things in and out of a list a lot, I'm not
> > >> > >> > sure the break even point where changing the structure helps.
> > >> > >> >
> > >> > >> > We'll need to try it.
> > >> > >> >
> > >> > >> > I think we will almost never see more than 10 list elements.
> > >> > >> >
> > >> > >> > --mgross
> > >> > >> >
> > >> > >> >
> > >> > >>
> > >> > >> I see about 80 (based on the batteryinfo dump) on my Nexus One
> > >> > >> (QSD8250, Android Froyo):
> > >> > >
> > >> > > shucks.
> > >> > >
> > >> > > well I think for a pm_qos class that has boolean dynamic range we can
> > >> > > get away with not walking the list on every request update. we can use
> > >> > > a counter, and the list will be for mostly for stats.
> > >> > >
> > >> >
> > >> > Did you give any thought to my suggestion to only use one entry per
> > >> > unique value on the first level list and then use secondary lists of
> > >> > identical values. That way if you only have two constraints values the
> > >> > list you have to walk when updating a request will never have more
> > >> > than two entries regardless of how many total request you have.
> > >> >
> > >> > A request update then becomes something like this:
> > >> > if on primary list {
> > >> > unlink from primary list
> > >> > if secondary list is not empty
> > >> > get next secondary entry and add in same spot on primary list
> > >> > }
> > >> > unlink from secondary list
> > >> > find new spot on primary list
> > >> > if already there
> > >> > add to secondary list
> > >> > else
> > >> > add to primary list
> > >>
> > >> This is just reinventing hash bucketed lists. To get the benefits, all
> > >> we do is implement an N state constraint as backed by an N bucketed hash
> > >> list, which the kernel already has all the internal mechanics for.
> > >
> > > Agreed.
> > >
> >
> > No, a hash is used for quick lookup of a specific value, not to find
> > an extreme value.
>
> If you only have N possible values an N bucket hash list is rather
> efficient (provided N is small). But I would agree that knowing what N
> is represents an API change, and since plists can do this without
> changing the API, they're better.
>
> > It is however extremely similar to plists. The only
> > difference is that plists link all the secondary lists together.
>
> Right, so they would solve the *current* problem exactly.
>
> > If we
> > want to have constraints that autoexpire, then keeping the secondary
> > lists separate allows the same optimization as I did for
> > wakelock/suspend_blocker timeouts where no timer is active if an
> > (equal or stricter) non-expiring constraint is active.
>
> But this is a future discussion and not part of the patch. The way open
> source works is that we sort out the best implementation for the current
> conditions. If the implementation has to change because of future
> stuff, then we change it when the future stuff comes along. Changing
> implementations is easy (they don't have any externally visible impact).
> Changing the in-kernel API is slightly harder, but easily doable. It's
> only changing the user visible ABI that we worry about and try not to
> do.
>

True.

The following is what I think I'll work on this weekend. (hopefully
have some sort of patch that at least compiles...)


Changes to pm_qos to enable "wakelock" or "suspend blocker" support:

Requirements:
1) atomic context support.
2) make updating request fast enough for hot path users.
3) Add a request class for "interactive_suspend", this particular
request has a dynamic range of 1 (its 0 or 1). If zero then ok to auto
suspend, if 1 then only user driven suspend. Whenever the aggregate
request changes value the registered notifiers are called.


Implementation:
* change api to have caller allocated the qos request structures. (solve
the kalloc issue)
* add plist use for constraint classes with hash-able dynamic ranges
* consider auto hashing / plist-ing everything. (most users of pm_qos
don't do any fancy stuff and tend to request the only 2 different
values, ever) but for now only request with dynamic range of 0,1 will
get the fancy lists.

I think the rest are implementation details.

--mgross




--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo(a)vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
From: mark gross on 6 Jun 2010 17:50
On Sat, Jun 05, 2010 at 09:07:32PM +0200, Rafael J. Wysocki wrote:
> On Saturday 05 June 2010, mark gross wrote:
> > On Fri, Jun 04, 2010 at 09:23:10AM -0500, James Bottomley wrote:
> > > On Thu, 2010-06-03 at 21:07 -0700, Arve Hj�nnev�g wrote:
> > > > On Thu, Jun 3, 2010 at 2:05 PM, Rafael J. Wysocki <rjw(a)sisk.pl> wrote:
> > > > > On Thursday 03 June 2010, James Bottomley wrote:
> > > > >> On Thu, 2010-06-03 at 00:10 -0700, Arve Hj�nnev�g wrote:
> > > > >> > On Wed, Jun 2, 2010 at 10:40 PM, mark gross <640e9920(a)gmail.com> wrote:
> > > > >> > > On Wed, Jun 02, 2010 at 09:54:15PM -0700, Brian Swetland wrote:
> > > > >> > >> On Wed, Jun 2, 2010 at 8:18 PM, mark gross <640e9920(a)gmail.com> wrote:
> > > > >> > >> > On Wed, Jun 02, 2010 at 02:58:30PM -0700, Arve Hj�nnev�g wrote:
> > > > >> > >> >>
> > > > >> > >> >> The list is not short. You have all the inactive and active
> > > > >> > >> >> constraints on the same list. If you change it to a two level list
> > > > >> > >> >> though, the list of unique values (which is the list you have to walk)
> > > > >> > >> >> may be short enough for a tree to be overkill.
> > > > >> > >> >
> > > > >> > >> > what have you seen in practice from the wake-lock stats?
> > > > >> > >> >
> > > > >> > >> > I'm having a hard time seeing where you could get more than just a
> > > > >> > >> > handfull. However; one could go to a dual list (like the scheduler) and
> > > > >> > >> > move inactive nodes from an active to inactive list, or we could simply
> > > > >> > >> > remove them from the list uppon inactivity. which would would well
> > > > >> > >> > after I change the api to have the client allocate the memory for the
> > > > >> > >> > nodes... BUT, if your moving things in and out of a list a lot, I'm not
> > > > >> > >> > sure the break even point where changing the structure helps.
> > > > >> > >> >
> > > > >> > >> > We'll need to try it.
> > > > >> > >> >
> > > > >> > >> > I think we will almost never see more than 10 list elements.
> > > > >> > >> >
> > > > >> > >> > --mgross
> > > > >> > >> >
> > > > >> > >> >
> > > > >> > >>
> > > > >> > >> I see about 80 (based on the batteryinfo dump) on my Nexus One
> > > > >> > >> (QSD8250, Android Froyo):
> > > > >> > >
> > > > >> > > shucks.
> > > > >> > >
> > > > >> > > well I think for a pm_qos class that has boolean dynamic range we can
> > > > >> > > get away with not walking the list on every request update. we can use
> > > > >> > > a counter, and the list will be for mostly for stats.
> > > > >> > >
> > > > >> >
> > > > >> > Did you give any thought to my suggestion to only use one entry per
> > > > >> > unique value on the first level list and then use secondary lists of
> > > > >> > identical values. That way if you only have two constraints values the
> > > > >> > list you have to walk when updating a request will never have more
> > > > >> > than two entries regardless of how many total request you have.
> > > > >> >
> > > > >> > A request update then becomes something like this:
> > > > >> > if on primary list {
> > > > >> > unlink from primary list
> > > > >> > if secondary list is not empty
> > > > >> > get next secondary entry and add in same spot on primary list
> > > > >> > }
> > > > >> > unlink from secondary list
> > > > >> > find new spot on primary list
> > > > >> > if already there
> > > > >> > add to secondary list
> > > > >> > else
> > > > >> > add to primary list
> > > > >>
> > > > >> This is just reinventing hash bucketed lists. To get the benefits, all
> > > > >> we do is implement an N state constraint as backed by an N bucketed hash
> > > > >> list, which the kernel already has all the internal mechanics for.
> > > > >
> > > > > Agreed.
> > > > >
> > > >
> > > > No, a hash is used for quick lookup of a specific value, not to find
> > > > an extreme value.
> > >
> > > If you only have N possible values an N bucket hash list is rather
> > > efficient (provided N is small). But I would agree that knowing what N
> > > is represents an API change, and since plists can do this without
> > > changing the API, they're better.
> > >
> > > > It is however extremely similar to plists. The only
> > > > difference is that plists link all the secondary lists together.
> > >
> > > Right, so they would solve the *current* problem exactly.
> > >
> > > > If we
> > > > want to have constraints that autoexpire, then keeping the secondary
> > > > lists separate allows the same optimization as I did for
> > > > wakelock/suspend_blocker timeouts where no timer is active if an
> > > > (equal or stricter) non-expiring constraint is active.
> > >
> > > But this is a future discussion and not part of the patch. The way open
> > > source works is that we sort out the best implementation for the current
> > > conditions. If the implementation has to change because of future
> > > stuff, then we change it when the future stuff comes along. Changing
> > > implementations is easy (they don't have any externally visible impact).
> > > Changing the in-kernel API is slightly harder, but easily doable. It's
> > > only changing the user visible ABI that we worry about and try not to
> > > do.
> > >
> >
> > True.
> >
> > The following is what I think I'll work on this weekend. (hopefully
> > have some sort of patch that at least compiles...)
> >
> >
> > Changes to pm_qos to enable "wakelock" or "suspend blocker" support:
> >
> > Requirements:
> > 1) atomic context support.
> > 2) make updating request fast enough for hot path users.
> > 3) Add a request class for "interactive_suspend", this particular
> > request has a dynamic range of 1 (its 0 or 1). If zero then ok to auto
> > suspend, if 1 then only user driven suspend. Whenever the aggregate
> > request changes value the registered notifiers are called.
> >
> >
> > Implementation:
> > * change api to have caller allocated the qos request structures. (solve
> > the kalloc issue)
> > * add plist use for constraint classes with hash-able dynamic ranges
>
> Please have a look at the patch James posted earlier today on linux-pm,
> you were CCed.
>

I like James' post. I'm sorry I missed it before I sent this note, but
I'm really glad he's posted it. Now I can do some testing.

--mgross


--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo(a)vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
 | 
Pages: 1
Prev: [RFC][PATCH 0/6] xtime/wall_to_monotonic cleanups
Next: [PATCH] remove pointless debug printk in rt_ioctl_giwscan()