On Tue, 2021-12-07 at 09:19 +1000, Erwin van Londen wrote:

Hello Martin, Muneendra.

As I kicked this discussion off in the beginning of the year and

seeing the Muneendra and the broadcom people have come up with the

first iteration I can only applaud the efforts. On behalf of all

storage and linux administrators I would say "Thank you".

As for your remark Martin my view would be to try and create a

modular approach where the transport layer drivers can hook into and

inform multipathd of any event. The module in multipathd would then

decide based on configured characteristics what the actions should

be. (Take it offline, suspend for X amount of time, introduce X us

delay etc...) That way when more transport methods are used these can

then dynamically be linked into the configuration without having any

impact on other parts of the transport stack. I can imagine that

Infiniband. ethernet, SAS and others utilise different transport

characteristics and as such may need to inform the attached hosts of

one or more events. On FC this is FPIN but a similar module may be

written for other transports.

Interesting idea. Are you aware of a technology for non-FC transports

that could take the role of FPIN? I have to admit I'm not, but that

doesn't mean they don't exist or won't exist in the future.

In the first place we'd need to "hook in" an event listener. Like with

Muneendra's patch, we're adding a new class of events that we're

listening to. The events would then than collected and processed by

separate worker thread (which unlike the listener would take the

multipath lock), setting paths states to marginal or back to normal.

I don't think we want to add plug-ins that spawn their own independent

threads, though. That sounds very difficult to handle properly, and we

already have more than enough complexity.

If we want to modularize this, we need a *generic* event listener

thread. A module would basically provide an fd for that thread to poll

on, and a callback to be called when an event occurs. This idea appeals

to me a lot, in particular because we already have an event listener

(the uevent listener thread) which is sitting idle most of the time.

So Muneendra, instead of creating a new receiver thread, you would

extend the existing uevent listener to handle the FPIN events as well.

The thread would now add uevents to the uevent list and FPIN events to

the FPIN events list.

Next, we'd also need a generic event consumer, with callbacks for

different types of marginal state handlers. Perhaps this could even be

the uevent trigger thread? The uevent trigger has more work to do than

the uevent listener. But any handler thread that wants to modify path

state would need to take the lock anyway, effectively serializing all

operations. So I guess we might as well use both uevent threads for

"transport event notification" reception and processing, respectively.

We also need to think about whether the currently existing marginal

path handler could fit into this framework. Not so well probably,

because it's not event driven and hooks into check_path(). OTOH, maybe

possible future mechanisms might hook into check_path(), too, so we'd

need a generic callback there?

Moreover, the existing marginal paths handler has two different modes

of operation, the "classical" one that disables reinstate, and the

more modern one that uses marginal pathgroups. I am wondering whether

we need the first mode in the long run. In particular if we want to

generalize this feature, we may want to get rind of the "classical"

mode altogether. I'm not aware of any distinct advantages of that

algorithm compared to marginal path groups.

@Ben, Muneendra, what do you think?

One word of caution here: we must be careful not to over-engineer.

As long as no other mechanism like FPIN for other transports is

conceivable, generalizing the concept makes only so much sense.

Therefore we shouldn't hold back the FPIN patches until we have

conceived of a generic mechanism, which may take a lot of time to

develop. If another mechanism becomes available, we could still try to

generalize the concept, if we keep the current additions clean and

well-separated from the core multipathd code.

However I am really thrilled by the prospect of generalizing event

handling and reusing the uevent threads for FPIN. That would reduce

complexity a lot, which is a good thing IMO.

@Ben, Muneendra, again, your opinions?

Best

Martin