netdev - Re: [RFC mptcp-next] tcp: mptcp: use mptcp receive buffer space to select rcv window

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [day] [month] [year] [list]

Date:   Wed, 18 Mar 2020 21:59:20 +0100
From:   Florian Westphal <fw@...len.de>
To:     Eric Dumazet <eric.dumazet@...il.com>
Cc:     Florian Westphal <fw@...len.de>, netdev@...r.kernel.org,
        mptcp@...ts.01.org, Eric Dumazet <edumazet@...gle.com>
Subject: Re: [RFC mptcp-next] tcp: mptcp: use mptcp receive buffer space to
 select rcv window

Eric Dumazet <eric.dumazet@...il.com> wrote:
> >>> +/* If ssk has an mptcp parent socket, use the mptcp rcvbuf occupancy,
> >>> + * not the ssk one.
> >>> + *
> >>> + * In mptcp, rwin is about the mptcp-level connection data.
> >>> + *
> >>> + * Data that is still on the ssk rx queue can thus be ignored,
> >>> + * as far as mptcp peer is concerened that data is still inflight.
> >>> + */
> >>> +void mptcp_space(const struct sock *ssk, int *space, int *full_space)
> >>> +{
> >>> +	const struct mptcp_subflow_context *subflow = mptcp_subflow_ctx(ssk);
> >>> +	const struct sock *sk = READ_ONCE(subflow->conn);
> >>
> >> What are the rules protecting subflow->conn lifetime ?
> >>
> >> Why dereferencing sk after this line is safe ?
> > 
> > Subflow sockets hold a reference on the master/parent mptcp-socket.
> > 
> 
> Presence of READ_ONCE() tells something might happen on
> this pointer after you read it.

Right, sorry about this. The READ_ONCE() isn't needed anymore after
recent improvement from Paolo.

> Can this pointer be set while this thread is owning the socket lock ?

Only by the one holding the sk lock, so no race.

> If not, then you do not need READ_ONCE(), this is confusing.

Yes.

> If yes, then it means that whatever changes the pointer might also release the reference
> on the old object.

The reference is released only after aquiring the socket lock.