> On Jun 20, 2024, at 8:28 PM, Mike Snitzer <snitzer@xxxxxxxxxx> wrote:
>
> On Thu, Jun 20, 2024 at 07:28:14PM -0400, Chuck Lever wrote:
>> On Thu, Jun 20, 2024 at 06:35:38PM -0400, Mike Snitzer wrote:
>>> On Fri, Jun 21, 2024 at 08:12:56AM +1000, NeilBrown wrote:
>>>> On Thu, 20 Jun 2024, Chuck Lever wrote:
>>>>> On Wed, Jun 19, 2024 at 04:40:31PM -0400, Mike Snitzer wrote:
>>>>>> This document gives an overview of the LOCALIO auxiliary RPC protocol
>>>>>> added to the Linux NFS client and server (both v3 and v4) to allow a
>>>>>> client and server to reliably handshake to determine if they are on the
>>>>>> same host. The LOCALIO auxiliary protocol's implementation, which uses
>>>>>> the same connection as NFS traffic, follows the pattern established by
>>>>>> the NFS ACL protocol extension.
>>>>>>
>>>>>> The robust handshake between local client and server is just the
>>>>>> beginning, the ultimate usecase this locality makes possible is the
>>>>>> client is able to issue reads, writes and commits directly to the server
>>>>>> without having to go over the network. This is particularly useful for
>>>>>> container usecases (e.g. kubernetes) where it is possible to run an IO
>>>>>> job local to the server.
>>>>>>
>>>>>> Signed-off-by: Mike Snitzer <snitzer@xxxxxxxxxx>
>>>>>> ---
>>>>>> Documentation/filesystems/nfs/localio.rst | 148 ++++++++++++++++++++++
>>>>>> include/linux/nfslocalio.h | 2 +
>>>>>> 2 files changed, 150 insertions(+)
>>>>>> create mode 100644 Documentation/filesystems/nfs/localio.rst
>>>>>>
>>>>>> diff --git a/Documentation/filesystems/nfs/localio.rst b/Documentation/filesystems/nfs/localio.rst
>>>>>> new file mode 100644
>>>>>> index 000000000000..a43c3dab2cab
>>>>>> --- /dev/null
>>>>>> +++ b/Documentation/filesystems/nfs/localio.rst
>>>>>> @@ -0,0 +1,148 @@
>>>>>> +===========
>>>>>> +NFS localio
>>>>>> +===========
>>>>>> +
>>>>>> +This document gives an overview of the LOCALIO auxiliary RPC protocol
>>>>>> +added to the Linux NFS client and server (both v3 and v4) to allow a
>>>>>> +client and server to reliably handshake to determine if they are on the
>>>>>> +same host. The LOCALIO auxiliary protocol's implementation, which uses
>>>>>> +the same connection as NFS traffic, follows the pattern established by
>>>>>> +the NFS ACL protocol extension.
>>>>>> +
>>>>>> +The LOCALIO auxiliary protocol is needed to allow robust discovery of
>>>>>> +clients local to their servers. Prior to this LOCALIO protocol a
>>>>>> +fragile sockaddr network address based match against all local network
>>>>>> +interfaces was attempted. But unlike the LOCALIO protocol, the
>>>>>> +sockaddr-based matching didn't handle use of iptables or containers.
>>>>>> +
>>>>>> +The robust handshake between local client and server is just the
>>>>>> +beginning, the ultimate usecase this locality makes possible is the
>>>>>> +client is able to issue reads, writes and commits directly to the server
>>>>>> +without having to go over the network. This is particularly useful for
>>>>>> +container usecases (e.g. kubernetes) where it is possible to run an IO
>>>>>> +job local to the server.
>>>>>> +
>>>>>> +The performance advantage realized from localio's ability to bypass
>>>>>> +using XDR and RPC for reads, writes and commits can be extreme, e.g.:
>>>>>> +fio for 20 secs with 24 libaio threads, 64k directio reads, qd of 8,
>>>>>> +- With localio:
>>>>>> + read: IOPS=691k, BW=42.2GiB/s (45.3GB/s)(843GiB/20002msec)
>>>>>> +- Without localio:
>>>>>> + read: IOPS=15.7k, BW=984MiB/s (1032MB/s)(19.2GiB/20013msec)
>>>>>> +
>>>>>> +RPC
>>>>>> +---
>>>>>> +
>>>>>> +The LOCALIO auxiliary RPC protocol consists of a single "GETUUID" RPC
>>>>>> +method that allows the Linux nfs client to retrieve a Linux nfs server's
>>>>>> +uuid. This protocol isn't part of an IETF standard, nor does it need to
>>>>>> +be considering it is Linux-to-Linux auxiliary RPC protocol that amounts
>>>>>> +to an implementation detail.
>>>>>> +
>>>>>> +The GETUUID method encodes the server's uuid_t in terms of the fixed
>>>>>> +UUID_SIZE (16 bytes). The fixed size opaque encode and decode XDR
>>>>>> +methods are used instead of the less efficient variable sized methods.
>>>>>> +
>>>>>> +The RPC program number for the NFS_LOCALIO_PROGRAM is currently defined
>>>>>> +as 0x20000002 (but a request for a unique RPC program number assignment
>>>>>> +has been submitted to IANA.org).
>>>>>> +
>>>>>> +The following approximately describes the LOCALIO in a pseudo rpcgen .x
>>>>>> +syntax:
>>>>>> +
>>>>>> +#define UUID_SIZE 16
>>>>>> +typedef u8 uuid_t<UUID_SIZE>;
>>>>>> +
>>>>>> +program NFS_LOCALIO_PROGRAM {
>>>>>> + version NULLVERS {
>>>>>> + void NULL(void) = 0;
>>>>>> + } = 1;
>>>>>> + version GETUUIDVERS {
>>>>>> + uuid_t GETUUID(void) = 1;
>>>>>> + } = 1;
>>>>>> +} = 0x20000002;
>>>>>> +
>>>>>> +The above is the skeleton for the LOCALIO protocol, it doesn't account
>>>>>> +for NFS v3 and v4 RPC boilerplate (which also marshalls RPC status) that
>>>>>> +is used to implement GETUUID.
>>>>>> +
>>>>>> +Here are the respective XDR results for nfsd and nfs:
>>>>>
>>>>> Hi Mike!
>>>>>
>>>>> A protocol spec describes the on-the-wire data formats, not the
>>>>> in-memory structure layouts. The below C structures are not
>>>>> relevant to this specification. This should be all you need here,
>>>>> if I understand your protocol correctly:
>>>>>
>>>>> /* raw RFC 9562 UUID */
>>>>> #define UUID_SIZE 16
>>>>> typedef u8 uuid_t<UUID_SIZE>;
>>>>>
>>>>> union GETUUID1res switch (uint32 status) {
>>>>
>>>> I don't think we need a status in the protocol. GETUUID always returns
>>>> a UUID. There is no possible error condition.
>>>
>>> By having localio use NFS's XDR we're able to piggyback on a status
>>> being returned by standard NFS RPC handling.
>>>
>>> See:
>>> nfs3svc_encode_getuuidres and nfs4svc_encode_getuuidres.
>>> nfs3_xdr_dec_getuuidres and nfs4_xdr_dec_getuuidres (and note the
>>> FIXME comment about abusing nfs_opnum4).
>>
>> No, let's not piggyback like that. Please make it a separate
>> XDR implementation just like NFSACL is. Again, LOCALIO is not
>> an extension of the NFS protocol. Making that claim confuses
>> people for whom the term "extension" has a very precise meaning.
>> If we were extending NFS, then yes, adding the new procedures
>> to the NFS XDR implementation is appropriate, but that's not
>> what you are doing: you are adding a new side-band protocol.
>
> I reworded yesterday, when you seized on my saying "extension" in the
> Documentation before and I changed it to "auxilliary" (I used
> "extension" because that's how NFS ACL was framed, I had no idea it
> was a loaded or nuanced term). localio the protocol amounts to a
> single RPC, no idea why I need to implement/duplicate a bunch of
> boilerplate code to add a single GETUUID RPC. Localio as submitted
> isn't concerned with some weird hijacking/extension of the NFS
> protocol, it is enabling the NFS client and server to be more
> efficient with how it fulfills IO. Localio has no tie to NFS other
> than it serves as connective glue between the NFS client and server
> to optimize reads, writes, and commits.
What I'm saying is please don't implement this as
nfs3_xdr_dec_getuuidres and nfs4_xdr_dec_getuuidres
GETUUID is not an NFS procedure, as you say. Just
use:
localio1_xdr_dec_getuuidres
Program name and version, XDR decoder, result data
type name. This is consistent with every other
modern (ie, recently-written) XDR function.
>> I have a long-term goal to ensure we can generate the source
>> code of the XDR layer of all the kernel RPC protocols via an
>> rpcgen like tool. A code generator can ensure that the
>> marshalling and unmarshalling code is memory-safe.
>
> The code doesn't show itself to be developed with such code generation
> at all. If anything it all feels very open-coded.
That's correct, it was open-coded in the 90's because
back then that was the way to build code that was
fast.
These days, compilers and CPUs optimize the shit out
of the code and instruction sequence so it matters
much less what we do in our source files. Since this
is a community-maintained effort, it makes sense to
spend our effort to use the source code as
documentation for the humans who have to read it
and update it, anywhere the utmost speed is not a
requirement.
In particular, right now there's no easy way to audit
our XDR implementation and prove/verify that we have
implemented exactly what the spec says. And, the open
coded nature of this implementation means it is
subject to memory bugs, which is bad for a widely
deployed network protocol.
In user space, RPC developers almost always start with
a .x file and use rpcgen to generate C (or Python)
code to handle marshaling of arguments and results.
That way, the XDR implementation starts with C code
that comes directly from the specification, untouched
by human hands.
We really need to take that approach in the kernel
too. The only reason not to do that is because there
is decades of technical debt baked into the current
open coded implementation. But I've spent a lot of
effort to move the server over to consistent and
properly layered XDR functions so that converting
it to use machine-generated XDR will be a burp
rather than a heart attack. Lots of work yet to do.
>> By piggybacking, you are building LOCALIO into another
>> protocol's XDR implementation, which makes it a special case,
>> and thus harder to implement via code that is generated
>> automatically from unmodified XDR language specs.
>
> GETUUID really couldn't be simpler. I'll do whatever you think best,
> but I just want to be clear: when told to implement a side-band
> protocol to coordinate I distilled it down to "I need a unique UUID
> from the server" (hence GETUUID). When implementing it I set out to
> work with what I saw in the existing code -- not reinvent boilerplate
> code. NFS ACL did that. If I went further by (ab)using NFS xdr
> helpers it was because I didn't want to write that code. If asked to
> decouple, I'm just going to factor out cleaner code that LOCALIO and
> NFS share to do what I already have it doing.
That "clean shared code" is the RPC XDR layer. There's
really no need to share any NFS-specific code between
NFS and LOCALIO. Please have a look at
include/linux/sunrpc/xdr.h for utility functions to
use for XDR; I suspect you will find everything you
need.
And I recommend having a look at
https://www.rfc-editor.org/rfc/rfc4506
for more detail about what XDR is and why we use it.
>> Maybe the client side maintainers don't care about that, but
>> please don't piggyback LOCALIO onto the NFSD's NFS XDR
>> implementation.
>>
>> Then, if it's a separate implementation, you can remove the status
>> code. I was wondering why the server would reply with an error. If
>> LOCALIO/GETUUID is not supported, then an RPC level error occurs
>> anyway.
>
> In my extensive discovery by trial and error I found that the NFS
> status codes and infrastructure necessary. Otherwise basic/advanced
> connectivity issues with sunrpc et al aren't accounted for properly.
That is quite the opposite of my experience in
this code base. Can you share an example of the
connectivity issues you encountered?
> NFS is pretty tightly coupled to sunrpc, and vice-versa. Not
> following why things have all of a sudden gotten "heightened", guess I
> chose the wrong word ("piggybacking") when being very clear about what
> the LOCALIO protocol does: piggyback on NFS XDR methods. I'm really
> caught out for that being viewed as so bad (but my FIXME in
> nfs4_xdr_dec_getuuidres really does need fixing).
Thanks for being clear about it, that is quite
helpful. But I'd rather this was implemented the way
every other RPC protocol is already done in the
kernel. There's more than one reason to do it that
way.
>> If you think you need an error like "Yes, I recognize that program
>> and procedure, but this file system doesn't allow local access
>> in any case" then that needs to be added to the protocol XDR
>> specification.
>
> The NFS status baked into NFS RPCs offer utility.
I'm agnostic about whether GETUUID1res has a status
field or not... When Neil pointed it out, I was just
requesting that we have a little more rationale for
it than "the XDR layer seems to need this" because,
IME, it doesn't.
Note that RPC allows a little extensibility: if
you send an unrecognized procedure number, the
RPC server will bark at you. So we can add new
procedures without much difficulty; the client can
tell if the server doesn't support them.
And if the argument or result data items have to
change in existing procedures, then we create a
new version of the RPC program.
So there isn't a strong need to add fields we think
might be needed some day but don't have a use for
now, even in a private non-standard protocol.
--
Chuck Lever
