Re: [Mpvdapi] MIF API Design Team

[Erik Kline <[e-mail]>]

Alright, I'm not gonna make it by tomorrow morning (too many work issues).

Dapeng's document explicitly describes a struct pvdinfo.  I like it,
but I fear it won't be extensible if we, for example, decide in the
future that a Foo Protocol Server is an essential piece of
configuration information.

Did you get a chance to think about the pvd_handle_t idea?  It means
we might end up with an ioctl-like call, were you would write:

    getpvdinfo(pvd_handle_t, PVD_INFO_FOOSERVERS, (void*)&foo_struct)

and portable code can test for the existence of PVD_INFO_FOOSERVERS.
Not the prettiest though. :(

Some of the things are not trivially represented, either.  Consider
how you would represent routes learned from various RIOs that make up
the pvd configuration.

On 6 July 2015 at 16:31, Dapeng Liu <[e-mail]> wrote:
> 在 2015年7月6日 星期一，下午3:04，Hui Deng 写道：
>
> Hi Dapeng
>
> At this moment, we would expect that Erik could lead this MIF MPVD API
> design tem, please try to contribute the design team work from your draft.
> thanks a lot for your cooperation
>
>
> Hi Hui,
>
> Sure.
> The draft is used used for the design team’s reference since it has some
> similarity with what Eric proposed in the list.
>
> Thanks,
> Dapeng Liu
>
>
> Best regards,
>
> DENG Hui
>
>
> ________________________________
> Date: Mon, 6 Jul 2015 14:25:12 +0800
> From: [e-mail]
> To: [e-mail]
> CC: [e-mail]; [e-mail]; [e-mail];
> [e-mail]; [e-mail]
> Subject: 回复： [Mpvdapi] MIF API Design Team
>
> The updated version: https://tools.ietf.org/html/draft-liu-mif-socket-api-00
>
> --
> Dapeng Liu
>
> 在 2015年7月5日 星期日，下午6:53，Dapeng Liu 写道：
>
> Hi Hui and Erik,
>
> Attachment is the draft I mentioned during the discussion. It was written
> before the Dallas meeting and it may not reflect the latest consensus we
> have.
>
> I am trying to update it.
>
>
>
> --
> Dapeng Liu
>
> 在 2015年7月5日 星期日，下午3:27，Hui Deng 写道：
>
> Hi Erik
>
> I guess that everbody is busy with deadline,
> could you kindly help to write down a draft about this proposal and then
> present it during coming MIF sesson,
> then we get more people to review on how Android implement this.
>
> Thanks a lot for your kind work
> Best regards,
>
> DENG Hui
>
>
>> From: [e-mail]
>> Date: Mon, 15 Jun 2015 19:55:15 +0900
>> Subject: Re: MIF API Design Team
>> To: [e-mail]
>> CC: [e-mail]; [e-mail]; [e-mail]
>>
>> Sorry about that: it's back to back 3GPP and now Android release issues.
>>
>> I spent the weekend reworking some notes on the things I described in
>> Dallas. That's not coming together quickly enough so let me lay
>> everything out here instead. In so doing I will try to describe more
>> about the approach Android took to see if people even think it makes
>> sense.
>>
>> Let's assume for the sake of argument that C API definitions are
>> nicely isolated, maybe in some #include <netpvd.h> header file.
>>
>> [1] Provisioning domains are represented by a "handle" kind of class,
>> android.net.Network:
>>
>> https://developer.android.com/reference/android/net/Network.html
>>
>> These are essentially handles to identify instances of attaches to
>> provisioning domains. The underlying identifier values are not
>> recycled (except after a really long time). Multiple temporally
>> disparate attaches to the same network each get a new instance
>> value--we have not yet found it of value to try to de-duplicate
>> attaches. Applications that want to know if you're attached to
>> "Starbucks" so they can startup and auto-sign-in use the relevant
>> handle value to query parameters (like SSID).
>>
>> I would propose that a C API would have something like
>>
>> typedef uint64_t pvd_handle_t
>> #define PVD_HANDLE_UNSPEC ((pvd_handle_t)0)
>>
>> [2] Once an application has received a Network handle [let's save how
>> that happens for later] it can request that certain operations be
>> performed specifically within the associated PVD. Currently these
>> operations include:
>>
>> [bind a file descriptor to a PVD to force packets to be sent using
>> the PVD's routing table]
>> android.net.Network#bindSocket()
>>
>> https://developer.android.com/reference/android/net/Network.html#bindSocket(java.net.Socket)
>>
>> [perform DNS lookups using a PVD's DNS servers]
>> android.net.Network#getAllByName()
>>
>> https://developer.android.com/reference/android/net/Network.html#getAllByName(java.lang.String)
>>
>> [set the PVD for all file descriptors and DNS lookups for a
>> process, unless otherwise overridden]
>> android.net.ConnectivityManager#setProcessDefaultNetwork()
>>
>> https://developer.android.com/reference/android/net/ConnectivityManager.html#setProcessDefaultNetwork(android.net.Network)
>>
>> I would propose that a C API have at least the following:
>>
>> [a] get/setsockopt(int fd, SOL_SOCKET, SO_PVD_HANDLE, &pvd_handle_t);
>> [b] getaddrinfo/getnameinfo some taking a pvd_handle_t
>> [c] discussion of recvmsg/sendmsg/accept semantics
>>
>> Note that the last is particularly tricky. Operating systems need to
>> know how to "mark" incoming packets that would cause a new connection
>> such that they have the correct pvd_handle_t associated. This has
>> lots of impact on suitable address selection as well.
>>
>> The C API should also have at least the following:
>>
>> [d] system, process, and per-thread pvd_handle get/set calls
>>
>> // These values are used by PVD-aware function calls when a PVD index
>> // is not explicitly specified.
>> pvd_handle_t pvd_system_default();
>>
>> // Same as above, but operates at a per-process level. If no
>> // process-specific default has been set this MUST return the value
>> // of a call to pvd_system_default().
>> pvd_handle_t pvd_process_default();
>>
>> // Same as above, but operates at a per-thread level. If no
>> // thread-specific default has been set this MUST return the value
>> // of a call to pvd_current_process_default().
>> pvd_handle_t pvd_thread_default();
>>
>> int pvd_set_system_default(pvd_handle_t); // 0 or -1 && errno = EFOO
>> int pvd_set_process_default(pvd_handle_t); // 0 or -1 && errno = EFOO
>> int pvd_set_thread_default(pvd_handle_t); // 0 or -1 && errno = EFOO
>>
>> [e] good discussion of a variety of EFOO errno return values.
>>
>> When a PVD has gone away (e.g. we disconnected from WIFI), subsequent
>> operations on the associated pvd_handle should fail with errno =
>> ENONET.
>>
>> [f] good discussion of how the process and thread default values
>> behave across fork/exec
>>
>> FWIW, Android uses all of the above when NetworkMonitor is validating
>> a network--it does DNS lookups within the PVD, makes pvd-bound sockets
>> used for HTTP Direct or Proxy connections to check connectivity to a
>> URL:
>>
>>
>> https://android.googlesource.com/platform/frameworks/base/+/master/services/core/java/com/android/server/connectivity/NetworkMonitor.java
>>
>> [3] the device has to build up configuration data continuously over
>> time, as DHCP results come back, and RAs come in bringing new prefixes
>> or new DNS servers.
>>
>> Not that it's relevant, but Android accumulates these in an
>> android.net.LinkProperties object:
>>
>> https://developer.android.com/reference/android/net/LinkProperties.html
>>
>> as we currently only support a 1:1 mapping of PVDs to physical interfaces.
>>
>> The means by which the operating system allocates pvd_handles, gathers
>> this data and associates it with a pvd_handle (including marking
>> incoming packets that would causes a new connection) might best be
>> left for a separate document, since sections 1 and 2 and background
>> theory will require a fair amount of text.
>>
>> [4] Android lets processes register callbacks to be called when
>> certain network properties change, or when a new network shows up, or
>> one goes away. This is how applications receive the handles (Network
>> objects).
>>
>> The UNIX way of doing these sorts of things...well, frequently leaves
>> much to be desired...but we should support multiple event-driven
>> models, I think.
>>
>> [5] we'll want an API for an application to pass in a handle and get
>> an ever expanding list of parameters (expanding as we think to add
>> them and write documents about them).
>>
>> The Android ConnectivityManager mitigates all this stuff and passes
>> around NetworkInfo objects which encapsulate extra information like
>> SSIDs, whether the Network is Ethernet, Mobile, Wi-Fi, or whether it's
>> metered, and so on.
>>
>> The closest UNIX-style analogy here would be some interface that is
>> getsockopt()-like, I think. Something like:
>>
>> int pvd_get_attribute(pvd_handle_t, PVD_ATTR_(HWTYPE|METERED|...),
>> void *returned_blob);
>>
>> And then it will be possible for code to be written like
>>
>> #ifdef PVD_ATTR_HTTP_PROXY_URL
>>
>> char proxyString[PVD_ATTR_HTTP_PROXY_URL_MAXLEN];
>> memset(proxyString, 0, sizeof(proxyString));
>> if (!pvg_get_attribute(pvd_handle, PVD_ATTR_HTTP_PROXY_URL, proxyString))
>> {
>> perror("Sadness!");
>> return -1;
>> }
>>
>> #endif
>>
>>
>> ---
>>
>> That's a massive brain dump, and I apologize for it and also that it's so
>> late.
>>
>> Let me know what y'all think.
>> -Erik
>>
>> On Mon, Jun 8, 2015 at 11:18 PM, Hui Deng <[e-mail]> wrote:
>> > Hi Erik
>> >
>> > We are waiting for your lead to start this work asap. Feel free to us
>> > know
>> > when you are ready.
>> >
>> > Hi Ian, thanks a lot for letting us know that you are going to work with
>> > a
>> > university on the implementations.
>> > Erik was quite busy for 3GPP SA2 work recently, we are waiting for his
>> > return to lead the work.
>> >
>> > Best regards,
>> >
>> > DENG Hui
>> >
>> >
>> >> From: [e-mail]
>> >> Subject: MIF API Design Team
>> >> Date: Mon, 8 Jun 2015 14:07:02 +0200
>> >> CC: [e-mail]
>> >> To: [e-mail]
>> >>
>> >> Hi Hui,
>> >>
>> >> My apologies for not being in contact sooner. Things don’t always move
>> >> as
>> >> quickly as hoped around here!
>> >>
>> >> The current status is that we have provisionally obtained the necessary
>> >> resources to start working on a proof of concept implementation of the
>> >> MIF
>> >> architecture. We will be doing this in partnership with a University.
>> >>
>> >> Tomorrow (9th July) we will start the scoping work with the
>> >> implementor.
>> >> What I want to come out of this work is a proof of concept
>> >> implementation of
>> >> the MIF architecture on a ‘generic’ Linux based platform. As Eric is
>> >> working
>> >> on an Android implementation for the mobile use case, I think that this
>> >> should complement his work quite well and serve to provide some good
>> >> implementation experience which can be fed back into the update of the
>> >> MIF
>> >> API document.
>> >>
>> >> I’ve sent a subscription request to the MIF API design team mailing
>> >> list.
>> >> Once we’ve got some firmer details agreed with the implementors, I’ll
>> >> make
>> >> an announcement on the m/l of what we’re doing and we can discuss what
>> >> the
>> >> best way of aligning the two implementation efforts.
>> >>
>> >> Best regards,
>> >> Ian
> _______________________________________________
> Mpvdapi mailing list
> [e-mail]
> https://www.ietf.org/mailman/listinfo/mpvdapi
>
>
>
> 附件：
> - draft-liu-mif-socket-api-00.txt
>
>
>