Message-ID: <00a01000-92ea-f5d9-3275-c4cc730f9f24@oracle.com>
Date:   Fri, 29 Jan 2021 12:01:54 -0800
From:   Shoaib Rao <rao.shoaib@...cle.com>
To:     Matthew Wilcox <willy@...radead.org>
Cc:     Jakub Kicinski <kuba@...nel.org>, linux-kernel@...r.kernel.org,
        linux-api@...r.kernel.org, netdev@...r.kernel.org,
        "David S. Miller" <davem@...emloft.net>, andy.rudoff@...el.com
Subject: Re: [PATCH] af_unix: Allow Unix sockets to raise SIGURG


On 1/29/21 11:54 AM, Shoaib Rao wrote:
>
> On 1/29/21 11:19 AM, Matthew Wilcox wrote:
>> On Fri, Jan 29, 2021 at 09:56:48AM -0800, Shoaib Rao wrote:
>>> On 1/25/21 3:36 PM, Jakub Kicinski wrote:
>>>> On Fri, 22 Jan 2021 15:06:37 +0000 Matthew Wilcox (Oracle) wrote:
>>>>> From: Rao Shoaib <rao.shoaib@...cle.com>
>>>>>
>>>>> TCP sockets allow SIGURG to be sent to the process holding the other
>>>>> end of the socket.  Extend Unix sockets to have the same ability.
>>>>>
>>>>> The API is the same in that the sender uses sendmsg() with MSG_OOB to
>>>>> raise SIGURG.  Unix sockets behave in the same way as TCP sockets 
>>>>> with
>>>>> SO_OOBINLINE set.
>>>> Noob question, if we only want to support the inline mode, why 
>>>> don't we
>>>> require SO_OOBINLINE to have been called on @other? Wouldn't that
>>>> provide more consistent behavior across address families?
>>>>
>>>> With the current implementation the receiver will also not see MSG_OOB
>>>> set in msg->msg_flags, right?
>>> SO_OOBINLINE does not control the delivery of signal, It controls how
>>> OOB Byte is delivered. It may not be obvious but this change does not
>>> deliver any Byte, just a signal. So, as long as sendmsg flag contains
>>> MSG_OOB, signal will be delivered just like it happens for TCP.
>> I don't think that's the question Jakub is asking.  As I understand it,
>> if you send a MSG_OOB on a TCP socket and the receiver calls recvmsg(),
>> it will see MSG_OOB set, even if SO_OOBINLINE is set.
> No it wont. Application just gets a signal.
>>    That wouldn't
>> happen with Unix sockets.  I'm OK with that difference in behaviour,
>> because MSG_OOB on Unix sockets _is not_ for sending out of band data.
>> It's just for sending an urgent signal.
> That is what I just explained in my email
>>
>> As you say, MSG_OOB does not require data to be sent for unix sockets
>> (unlike TCP which always requires at least one byte), but one can
>> choose to send data as part of a message which has MSG_OOB set. It
>> won't be tagged in any special way.
>>
>> To Jakub's other question, we could require SO_OOBINLINE to be set.
>> That'd provide another layer of insurance against applications being
>> surprised by a SIGURG they weren't expecting.  I don't know that it's
>> really worth it though.
>
> SO_OOBINLINE has a meaning, that the urgent byte is part of the stream 
> and using SO_OOBLINE to allow signalling would be wrong/confusing. We 
> could add a socket option on the receiver to indicate if it supports 
> or wants the signal.
>
>>
>> One thing I wasn't clear about, and maybe you know, if we send a 
>> MSG_OOB,
>> does this patch cause this part of the tcp(7) manpage to be true for
>> unix sockets too?
>>
>>         When out-of-band data is present, select(2) indicates the 
>> file descrip‐
>>         tor as having an exceptional condition and poll (2) indicates 
>> a POLLPRI
>>         event.
>
> No because there is no data involved. Poll is associated with data not 
> signals.
>
> Shoaib

SO_OOBINLINE implies there is urgent data inline -- This patch will send 
a signal even if there is no data.

Shoaib

>
>>

Powered by blists - more mailing lists