Em Wed, 23 Sep 2015 00:56:05 +0300
Sakari Ailus <sakari.ailus@xxxxxxxxxxxxxxx> escreveu:
> Hi Mauro,
>
> Mauro Carvalho Chehab wrote:
> > Em Sun, 20 Sep 2015 10:26:09 +0300
> > Sakari Ailus <sakari.ailus@xxxxxxxxxxxxxxx> escreveu:
> >
> >> Hi Mauro,
> >>
> >> Mauro Carvalho Chehab wrote:
> >>> Hi Sakari,
> >>>
> >>> On Fri, 11 Sep 2015 13:09:03 +0300
> >>> Sakari Ailus <sakari.ailus@xxxxxxxxxxxxxxx> wrote:
> >>>
> >>>> Hi all,
> >>>>
> >>>> This patchset adds an API for managing entity enumerations, i.e.
> >>>> storing a bit of information per entity. The entity ID is no longer
> >>>> limited to small integers (e.g. 31 or 63), but
> >>>> MEDIA_ENTITY_MAX_LOW_ID. The drivers are also converted to use that
> >>>> instead of a fixed number.
> >>>>
> >>>> If the number of entities in a real use case grows beyond the
> >>>> capabilities of the approach, the algorithm may be changed. But most
> >>>> importantly, the API is used to perform the same operation everywhere
> >>>> it's done.
> >>>>
> >>>> I'm sending this for review only, the code itself is untested.
> >>>>
> >>>> I haven't entirely made up my mind on the fourth patch. It could be
> >>>> dropped, as it uses the API for a somewhat different purpose.
> >>>>
> >>>
> >>> Sorry for not reviewing this earlier, but I'm traveling this week to
> >>> China, and I was having some troubles with the Internet. Btw, I don't
> >>> have my notebook here (just a chromebook without the media tree).
> >>> So, please consider this as just a preliminary review.
> >>>
> >>> I won't be comment this series patch by patch, because it is really
> >>> painful to do it while here with this Internet.
> >>>
> >>> Also, I want to discuss the patch series as a hole.
> >>>
> >>> From what we've agreed last week, we won't be using a separate ID
> >>> range for the entity ID, but this patch series is actually adding
> >>> it, and, IMHO, using a confusing nomenclature: instead of calling the
> >>> entity ID range as "entity_id" at the media_device struct, you're
> >>> now calling it "low_id". That sounds confusing to me. If you think
> >>> we should keep a separate range for entities, calling it as
> >>> "entity_id" is clearer.
> >>
> >> It's *not* the entity ID. It's a number used internally to keep track of
> >> the entities, and only used for this purpose, nothing else. If you look
> >> at the patch, the number of places where low_id is used is very limited.
> >> Individual drivers do not and must not access the low_id field.
> >>
> >> The underlying algorithm for keeping track of entities does not change,
> >> but that could be changed later on without affecting the users of the
> >> alrogithm. --- See patch 3.
> >>
> >> There are two main reasons for this:
> >>
> >> 1. No need to implement a new algorithm which would be less efficient
> >> but could cope in cases we do not have yet; this can be done later on,
> >> as patch 3 adds an API to access the information without making
> >> assumptions on the implementation.
> >
> > If this is an internal number used only by the graph traversal
> > algorithm, then the implementation doesn't seem correct. I mean:
> > such number should be generated internally when starting the
> > graph traversal algorithm, and it would be better to store such
> > graph-traversal internal algorithm data on a separate struct.
>
> In that case the number would become enumeration specific. I have no
> problem in that, but if a single entity specific number works for the
> purpose, I think that could be used as well. Using something else than a
> constant value per entity quickly will require memory allocation, and
> graph traversal is performance critical in many cases.
I guess a single loop at graph traversal start would solve it. Yet,
it means an extra O(n) to fill in the numbers, but the algorithm
that fills such numbers will be at least O(n) either here or at the
entity creation.
Perhaps one solution would be to have a "dirty" flag that would be
raised if a new entity would be created (and removed?). If dirty
flag is raised, it would do the count at the beginning of the graph
start.
> >
> >> 2. It does solve the problem. The graph object IDs of the entities can
> >> be large without adversely affecting the functionality of existing drivers.
> >
> > Right now, with just those 9 patches, it doesn't ;)
> >
> > I mean: if I apply your patches after the MC next gen ones and try to use the
> > graph traversal, it will do the wrong thing for hybrid TV cards, as the number
> > of entities there are more than 64. Ok, easy to fix after this series by
> > just changing the value of MEDIA_ENTITY_MAX_LOW_ID, but this will only
> > work while we don't implement dynamic support.
>
> Fair enough. Still, it didn't work before either: the graph traversal
> algorithm depends on entity ID not exceeding MEDIA_ENTITY_ENUM_MAX_ID
> (which this set renames).
Well, for existing drivers that use graph traversal, this number is
not exceeded.
> I still consider that a separate problem.
>
> >
> >>>
> >>> Also, you said at the low_id comment that if an entity is
> >>> unregistered and then re-registered, it would preserve the same
> >>
> >> I never claimed that, and the patchset does not implement that either.
> >
> > That's what I understood from this comment:
> >
> > Rename the macro as it no longer is a maximum ID that an entity can have,
> > but a low ID which is used for internal purposes of enumeration. This is
> > the maximum number of concurrent entities that may exist in a media
> > device.
> >
> > If this is the "max number of concurrent entities", you need to reassign
> > those numbers when entities are removed.
> >
> > If this is not what you're meaning, then fix the patch description
> > to let it clearer.
> >
> > I guess then that "low ID" is actually an upper bound for that
> > graph-traversal only control number. We should really not use the word "ID"
> > here, as this is not an ID. it is just some index/control number that will
> > be granted to be below some upper bound.
>
> Index sounds good to me, it's indeed clearer. I didn't like "low"
> either, it just happens to be that way right now.
OK.
> With a different
> algorithm, I think we can get rid of the entire field --- but it will
> have to be replaced by something that requires memory allocation.
It is hard to tell how a new algorithm would be without writing the
code for it and do some experimentation ;)
>
> ...
>
> >>>
> >>> - some other mechanism would be available for drivers that
> >>> would support dynamic entity creation.
> >>>
> >>> So, I don't see how this would solve the problems that we
> >>> discussed at the last week IRC chats.
> >>>
> >>> Am I missing something?
> >>
> >> This set indeed solves a problem, and that problem was unrestricting the
> >> graph object ID of the entities. There are other problems remaining
> >> before entities can be e.g. unregistered one by one, but they are
> >> separate problems.
> >
> > That's not the way I see it ;) I mean, for the current MC code, this patch
> > series is not needed, as all drivers right now have less than 64 entities.
>
> In the last IRC meeting we agreed to have a single ID range for graph
> object IDs. Now that ID will be given to all graph objects, 64 will be
> reached much sooner than it used to, and it probably happens on at least
> some existing drivers as well.
Yes, using a single range will make this an issue. So, yes we need to
apply the patches that fixes the upper bound problem before applying
the patch that moves from 4 ranges into one. So, the order would be:
- the 82 patches
- the graph traversal change to use an "index" for the entity count
(the equivalent of this series)
- the patch that changes from 4 to 1 range.
>
> >
> > This need only starts after/during the MC next gen patches, in order
> > to address two changes that come on this series:
> >
> > - having an arbitrary entity ID number that will be a way bigger
> > than the current upper bound (the number of entities);
>
> It's indeed *the number of entities*, not entity ID.
>
> >
> > - support more than 64 entities.
> >
> > So, I would be expecting a patch series that would either:
> >
> > 1) change the graph traversal algorithms to not depend on some
> > upper bound limit;
> >
> > or:
> >
> > 2) dynamically create whatever internal index number and to
> > dynamically determine the upper bound limit that would be needed for the
> > graph traversal algorithm to work, allocating those data when the graph
> > traversal algorithm starts and freeing them when the graph traversal
> > stops, and/or when the media device is unregistered.
>
> Even with 4 k entities, the memory required by that algorithm to track
> the entities would be just 512 bytes + 16 * depth. I wonder if a more
> sophisticated algorithm would be able to operate with less or as
> efficiently.
Hard to tell without seeing the code ;)
The current algorithm is not bad, but I never really needed to look at
other alternatives for it.
Yet, with dynamic entity creation/removal, we'll need to add some extra
code to re-numerate the new entities and re-calculate the upper bound,
as entities are created/removed. So, it might justify the need or a better
algorithm, in order to avoid adding too much penalty to the performance.
>
> The memory would need to come from elsewhere than from the stack obviously.
Yes.
>
> >
> >>
> >>>
> >>> Regards,
> >>> Mauro
> >>>
> >>> PS.: sparse already complains on two places at the media-entity where
> >>> bitmaps are declared at the stack. With max entities equal to 64,
> >>> that's not an issue, but if we change to a higher number, those will
> >>> need to be dynamically allocated, in order to avoid stack overflows.
> >>> I didn't see any patches touching that.
> >>
> >> I agree. The aim of the set was not to increase the number of concurrent
> >> entities. That is a separate problem which can be solved later on once
> >> we have a use case for it.
> >>
> >
> > I reviewed that code. The problem there is actually unrelated to what
> > this patch series is trying to solve, as the problem there is due to the
> > number of PADs (and not on the number of entities). I mean about this
> > code at media_entity_pipeline_start():
> >
> > while ((entity = media_entity_graph_walk_next(&graph))) {
> > DECLARE_BITMAP(active, entity->num_pads);
> > DECLARE_BITMAP(has_no_links, entity->num_pads);
> >
> > That limits the max number of pads, as a big number would cause a
> > Linux stack overflow.
> >
> > Those bitmaps should likely be moved to struct media_pipeline or
> > struct media_device and be either dynamically created/removed or
> > having them relying on a max number of pads.
>
> How many pads are we looking forwards to have?
>
> I don't think we have a fixed limit as such, but AFAIR there's no driver
> using more than around five at the moment, so there was really no need
> to worry about it.
>
> I guess it'd be good to have a maximum though just to be on the safe side.
DVB drivers use currently 256 pads. At the patch I wrote earlier today:
https://patchwork.linuxtv.org/patch/31512/
(not sure why it was not at patchwork yet... Had to re-send it now)
I used the next power of two as the max limit, so 512 pads. I opted to
just add the two bitmaps used at the graph traversal logic as part of
the private fields of struct media_device. I don't expect any performance
penalty with that.
One thing I forgot to add there was a test if a driver is using more than
the maximum number of PADs. We should add such test on a version 2 of the
patch.
Regards,
Mauro
>
--
To unsubscribe from this list: send the line "unsubscribe linux-media" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
