-----"Tom Talpey" <tom@xxxxxxxxxx> wrote: -----
>To: "Bernard Metzler" <BMT@xxxxxxxxxxxxxx>, "Krishnamraju Eraparaju"
><krishna2@xxxxxxxxxxx>
>From: "Tom Talpey" <tom@xxxxxxxxxx>
>Date: 08/01/2019 08:54PM
>Cc: jgg@xxxxxxxx, linux-rdma@xxxxxxxxxxxxxxx, bharat@xxxxxxxxxxx,
>nirranjan@xxxxxxxxxxx, krishn2@xxxxxxxxxxx
>Subject: [EXTERNAL] Re: [PATCH for-rc] siw: MPA Reply handler tries
>to read beyond MPA message
>
>On 8/1/2019 6:56 AM, Bernard Metzler wrote:
>> -----"Krishnamraju Eraparaju" <krishna2@xxxxxxxxxxx> wrote: -----
>>
>>> To: jgg@xxxxxxxx, bmt@xxxxxxxxxxxxxx
>>> From: "Krishnamraju Eraparaju" <krishna2@xxxxxxxxxxx>
>>> Date: 07/31/2019 12:34PM
>>> Cc: linux-rdma@xxxxxxxxxxxxxxx, bharat@xxxxxxxxxxx,
>>> nirranjan@xxxxxxxxxxx, krishn2@xxxxxxxxxxx, "Krishnamraju
>Eraparaju"
>>> <krishna2@xxxxxxxxxxx>
>>> Subject: [EXTERNAL] [PATCH for-rc] siw: MPA Reply handler tries to
>>> read beyond MPA message
>>>
>>> while processing MPA Reply, SIW driver is trying to read extra 4
>>> bytes
>>> than what peer has advertised as private data length.
>>>
>>> If a FPDU data is received before even siw_recv_mpa_rr() completed
>>> reading MPA reply, then ksock_recv() in siw_recv_mpa_rr() could
>also
>>> read FPDU, if "size" is larger than advertised MPA reply length.
>>>
>>> 501 static int siw_recv_mpa_rr(struct siw_cep *cep)
>>> 502 {
>>> .............
>>> 572
>>> 573 if (rcvd > to_rcv)
>>> 574 return -EPROTO; <----- Failure here
>>>
>>> Looks like the intention here is to throw an ERROR if the received
>>> data
>>> is more than the total private data length advertised by the peer.
>>> But
>>> reading beyond MPA message causes siw_cm to generate
>>> RDMA_CM_EVENT_CONNECT_ERROR event when TCP socket recv buffer is
>>> already
>>> queued with FPDU messages.
>>>
>>> Hence, this function should only read upto private data length.
>>>
>>> Signed-off-by: Krishnamraju Eraparaju <krishna2@xxxxxxxxxxx>
>>> ---
>>> drivers/infiniband/sw/siw/siw_cm.c | 4 ++--
>>> 1 file changed, 2 insertions(+), 2 deletions(-)
>>>
>>> diff --git a/drivers/infiniband/sw/siw/siw_cm.c
>>> b/drivers/infiniband/sw/siw/siw_cm.c
>>> index a7cde98e73e8..8dc8cea2566c 100644
>>> --- a/drivers/infiniband/sw/siw/siw_cm.c
>>> +++ b/drivers/infiniband/sw/siw/siw_cm.c
>>> @@ -559,13 +559,13 @@ static int siw_recv_mpa_rr(struct siw_cep
>*cep)
>>> * A private data buffer gets allocated if hdr->params.pd_len !=
>0.
>>> */
>>> if (!cep->mpa.pdata) {
>>> - cep->mpa.pdata = kmalloc(pd_len + 4, GFP_KERNEL);
>>> + cep->mpa.pdata = kmalloc(pd_len, GFP_KERNEL);
>>> if (!cep->mpa.pdata)
>>> return -ENOMEM;
>>> }
>>> rcvd = ksock_recv(
>>> s, cep->mpa.pdata + cep->mpa.bytes_rcvd - sizeof(struct mpa_rr),
>>> - to_rcv + 4, MSG_DONTWAIT);
>>> + to_rcv, MSG_DONTWAIT);
>>>
>>> if (rcvd < 0)
>>> return rcvd;
>>> --
>>> 2.23.0.rc0
>>>
>>>
>>
>> The intention of this code is to make sure the
>> peer does not violates the MPA handshake rules.
>> The initiator MUST NOT send extra data after its
>> MPA request and before receiving the MPA reply.
>
>I think this is true only for MPA v2. With MPA v1, the
>initiator can begin sending immediately (before receiving
>the MPA reply), because there is no actual negotiation at
>the MPA layer.
>
>With MPA v2, the negotiation exchange is required because
>the type of the following message is predicated by the
>"Enhanced mode" a|b|c|d flags present in the first 32 bits
>of the private data buffer.
>
>So, it seems to me that additional logic is needed here to
>determine the MPA version, before sniffing additional octets
>from the incoming stream?
>
>Tom.
>
There still is the marker negotiation taking place.
RFC 5044 says in section 7.1.2:
"Note: Since the receiver's ability to deal with Markers is
unknown until the Request and Reply Frames have been
received, sending FPDUs before this occurs is not possible."
This section further discusses the responder's behavior,
where it MUST receive a first FPDU from the initiator
before sending its first FPDU:
"4. MPA Responder mode implementations MUST receive and validate at
least one FPDU before sending any FPDUs or Markers."
So it appears with MPA version 1, the current siw
code is correct. The initiator is entering FPDU mode
first, and only after receiving the MPA reply frame.
Only after the initiator sent a first FPDU, the responder
can start using the connection in FPDU mode.
Because of this somehow broken connection establishment
procedure (only the initiator can start sending data), a
later MPA version makes this first FPDU exchange explicit
and selectable (zero length READ/WRITE/Send).
Bernard.
>
>> So, for the MPA request case, this code is needed
>> to check for protocol correctness.
>> You are right for the MPA reply case - if we are
>> _not_ in peer2peer mode, the peer can immediately
>> start sending data in RDMA mode after its MPA Reply.
>> So we shall add appropriate code to be more specific
>> For an error, we are (1) processing an MPA Request,
>> OR (2) processing an MPA Reply AND we are not expected
>> to send an initial READ/WRITE/Send as negotiated with
>> the peer (peer2peer mode MPA handshake).
>>
>> Just removing this check would make siw more permissive,
>> but to a point where peer MPA protocol errors are
>> tolerated. I am not in favor of that level of
>> forgiveness.
>>
>> If possible, please provide an appropriate patch
>> or (if it causes current issues with another peer
>> iWarp implementation) just run in MPA peer2peer mode,
>> where the current check is appropriate.
>> Otherwise, I would provide an appropriate fix by Monday
>> (I am still out of office this week).
>>
>>
>> Many thanks and best regards,
>> Bernard.
>>
>>
>>
>
>
