RE: [PATCH-for-9.1 v2 2/3] migration: Remove RDMA protocol handling
> For rdma programming, the current mainstream implementation is
to use
rdma_cm to establish a connection, and then use verbs to transmit data.
>
> rdma_cm and ibverbs create two FDs respectively. The two FDs have
> different responsibilities. rdma_cm fd is used to notify connection
> establishment events, and verbs fd is used to notify new CQEs. When
poll/epoll monitoring is directly performed on the rdma_cm fd, only a pollin
event can be monitored, which means that an rdma_cm event occurs. When
the verbs fd is directly polled/epolled, only the pollin event can be listened,
which indicates that a new CQE is generated.
>
> Rsocket is a sub-module attached to the rdma_cm library and provides
> rdma calls that are completely similar to socket interfaces. However,
> this library returns only the rdma_cm fd for listening to link setup-related
events and does not expose the verbs fd (readable and writable events for
listening to data). Only the rpoll interface provided by the RSocket can be used
to listen to related events. However, QEMU uses the ppoll interface to listen to
the rdma_cm fd (gotten by raccept API).
> And cannot listen to the verbs fd event. Only some hacking methods can be
used to address this problem.
>
> Do you guys have any ideas? Thanks.
The current rsocket code allows calling rpoll() with non-rsocket fd's, so an app can
use rpoll() directly in place of poll(). It may be easiest to add an rppoll() call to
rsockets and call that when using RDMA.
In case the easy path isn't feasible:
An extension could allow extracting the actual fd's under an rsocket, in order to
allow a user to call poll()/ppoll() directly. But it would be non-trivial.
The 'fd' that represents an rsocket happens to be the fd related to the RDMA CM.
That's because an rsocket needs a unique integer value to report as an 'fd'
value which will not conflict with any other fd value that the app may have. I would
consider the fd value an implementation detail, rather than something which an app should
depend upon. (For example, the 'fd' value returned for a datagram rsocket is
actually a UDP socket fd).
Once an rsocket is in the connected state, it's possible an extended rgetsockopt() or
rfcntl() call could return the fd related to the CQ. But if an app tried to call poll()
on that fd, the results would not be as expected. For example, it's possible for data
to be available to receive on the rsocket without the CQ fd being signaled. Calling
poll() on the CQ fd in this state could leave the app hanging. This is a natural? result
of races in the RDMA CQ signaling. If you look at the rsocket rpoll() implementation,
you'll see that it checks for data prior to sleeping.
For an app to safely wait in poll/ppoll on the CQ fd, it would need to invoke some sort of
'pre-poll' routine, which would perform the same checks done in rpoll() prior to
blocking. As a reference to a similar pre-poll routine, see the fi_trywait() call from
this man page:
https://ofiwg.github.io/libfabric/v1.21.0/man/fi_poll.3.html
This is for a different library but deals with the same underlying problem. Obviously
adding an rtrywait() to rsockets is possible but wouldn't align with any socket API
equivalent.
- Sean