Re: [PATCH] cifs: Fix TCP_Server_Info::credits to be signed

[Shyam Prasad N]

On Mon, Oct 20, 2025 at 10:28 PM Steve French <smfrench@xxxxxxxxx> wrote:
>
> On Mon, Oct 20, 2025 at 9:08 AM David Howells <dhowells@xxxxxxxxxx> wrote:
> >
> > Enzo Matsumiya <ematsumiya@xxxxxxx> wrote:
> >
> > > Both semantically and technically, credits shouldn't go negative.
> > > Shouldn't those other fields/functions become unsigned instead?
> >
> > That's really a question for Steve, but it makes it easier to handle
> > underflow, and I'm guessing that the maximum credits isn't likely to exceed
> > 2G.
> >
> > David
>
> Interesting question - I do like the idea of keeping signed if it
> makes it easier to check
> for underflows but IIRC that hasn't been a problem in a long time (adding Pavel
> and Ronnie in case they remember) but more important than the signed
> vs. unsigned
> in my opinion is at least keeping the field consistent.
>
> I have seen a few stress related xfstests that often generate
> reconnects, so we may want
> to run with the tracepoint enabled
> (smb3_reconnect_with_invalid_credits) to see if that
> is actually happening (the underflow of credits)
>
> I also was thinking that we should doublecheck that lease break acks
> will never run out credits
> (since that can deadlock servers for more than 30 seconds as they wait
> for timeouts), even if
> "lease break storms" are rare.   Maybe we should allow e.g. lease
> break acks to borrow echo
> credits e.g. as minor improvement?
>
> --
> Thanks,
>
> Steve

I agree with Steve.
I don't mind credits being a signed int. If credits go negative, it's
a clear indication of a bug and jumps out at you more than a large
integer would.

-- 
Regards,
Shyam