On 8/9/2019 9:52 AM, Bernard Metzler wrote:
-----"Tom Talpey" <tom@xxxxxxxxxx> wrote: -----
To: "Bernard Metzler" <BMT@xxxxxxxxxxxxxx>, "Krishnamraju Eraparaju"
<krishna2@xxxxxxxxxxx>
From: "Tom Talpey" <tom@xxxxxxxxxx>
Date: 08/09/2019 02:27PM
Cc: "jgg@xxxxxxxx" <jgg@xxxxxxxx>, "linux-rdma@xxxxxxxxxxxxxxx"
<linux-rdma@xxxxxxxxxxxxxxx>, "Potnuri Bharat Teja"
<bharat@xxxxxxxxxxx>, "Nirranjan Kirubaharan" <nirranjan@xxxxxxxxxxx>
Subject: [EXTERNAL] Re: [PATCH for-rc] siw: MPA Reply handler tries
to read beyond MPA message
On 8/9/2019 6:29 AM, Bernard Metzler wrote:
-----"Krishnamraju Eraparaju" <krishna2@xxxxxxxxxxx> wrote: -----
To: "Bernard Metzler" <BMT@xxxxxxxxxxxxxx>
From: "Krishnamraju Eraparaju" <krishna2@xxxxxxxxxxx>
Date: 08/08/2019 06:47PM
Cc: "Tom Talpey" <tom@xxxxxxxxxx>, "jgg@xxxxxxxx" <jgg@xxxxxxxx>,
"linux-rdma@xxxxxxxxxxxxxxx" <linux-rdma@xxxxxxxxxxxxxxx>,
"Potnuri
Bharat Teja" <bharat@xxxxxxxxxxx>, "Nirranjan Kirubaharan"
<nirranjan@xxxxxxxxxxx>
Subject: [EXTERNAL] Re: [PATCH for-rc] siw: MPA Reply handler
tries
to read beyond MPA message
On Thursday, August 08/08/19, 2019 at 15:05:12 +0000, Bernard
Metzler
wrote:
-----"Krishnamraju Eraparaju" <krishna2@xxxxxxxxxxx> wrote: -----
To: "Tom Talpey" <tom@xxxxxxxxxx>, <BMT@xxxxxxxxxxxxxx>
From: "Krishnamraju Eraparaju" <krishna2@xxxxxxxxxxx>
Date: 08/05/2019 07:26PM
Cc: "jgg@xxxxxxxx" <jgg@xxxxxxxx>, "linux-rdma@xxxxxxxxxxxxxxx"
<linux-rdma@xxxxxxxxxxxxxxx>, "Potnuri Bharat Teja"
<bharat@xxxxxxxxxxx>, "Nirranjan Kirubaharan"
<nirranjan@xxxxxxxxxxx>
Subject: [EXTERNAL] Re: [PATCH for-rc] siw: MPA Reply handler
tries
to read beyond MPA message
On Friday, August 08/02/19, 2019 at 18:17:37 +0530, Tom Talpey
wrote:
On 8/2/2019 7:18 AM, Bernard Metzler wrote:
-----"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).
Yeah, I guess so. Because nobody ever actually implemented
markers,
I think that they may more or less passively ignore this. But
you're
currect that it's invalid protocol behavior.
If your testing didn't uncover any issues with existing
implementations
failing to connect with your strict checking, I'm ok with it.
Tom & Bernard,
Thanks for the insight on MPA negotiation.
Could the below patch be considered as a proper fix?
diff --git a/drivers/infiniband/sw/siw/siw_cm.c
b/drivers/infiniband/sw/siw/siw_cm.c
index 9ce8a1b925d2..0aec1b5212f9 100644
--- a/drivers/infiniband/sw/siw/siw_cm.c
+++ b/drivers/infiniband/sw/siw/siw_cm.c
@@ -503,6 +503,7 @@ static int siw_recv_mpa_rr(struct siw_cep
*cep)
struct socket *s = cep->sock;
u16 pd_len;
int rcvd, to_rcv;
+ int extra_data_check = 4; /* 4Bytes, for MPA rules
violation
checking */
if (cep->mpa.bytes_rcvd < sizeof(struct mpa_rr)) {
rcvd = ksock_recv(s, (char *)hdr +
cep->mpa.bytes_rcvd,
@@ -553,23 +554,37 @@ static int siw_recv_mpa_rr(struct siw_cep
*cep)
return -EPROTO;
}
+ /*
+ * Peer must not send any extra data other than MPA
messages
until MPA
+ * negotiation is completed, an exception is MPA V2
client-server Mode,
+ * IE, in this mode the peer after sending MPA Reply can
immediately
+ * start sending data in RDMA mode.
+ */
This is unfortunately not true. The responder NEVER sends
an FPDU without having seen an FPDU from the initiator.
I just checked RFC 6581 again. The RTR (ready-to-receive)
indication from the initiator is still needed, but now
provided by the protocol and not the application: w/o
'enhanced MPA setup', the initiator sends the first
FPDU as an application message. With 'enhanced MPA setup',
the initiator protocol entity sends (w/o application
interaction) a zero length READ/WRITE/Send as a first FPDU,
as previously negotiated with the responder. Again: only
after the responder has seen the first FPDU, it can
start sending in FPDU mode.
If I understand your statements correctly: in MPA v2 clint-server
mode,
the responder application should have some logic to wait(after
ESTABLISH
event) until the first FPDU message is received from the
initiator?
Thanks,
Krishna.
The responder may either delay the established event
until it sees the first peer FPDU, or it delays sending
the first (now already posted) FPDU, until it sees the
first initiator FPDU. This, since (theoretically), the
responder does not know yet the result of CRC and
Marker negotiation. siw delays the established event,
which seems the more clear path to me.
This MPA handshake is cumbersome. siw is permissive
for violations of these rules to the extend, that the
responder shall not send back to back the first FPDU
with the MPA reply. If a (Chelsio ?) iWarp adapter
would send the first responder FPDU only after it
got TCP acked its MPA reply, the chances are high the
(siw) iWarp initiator has already transitioned to RTS
and it would accept that FPU, even if not having sent
it's first FPDU.
siw sticks to this rule to that extend, since all receive
processing is triggered by socket callbacks (from
sk->sk_data_ready()). As long as the QP is not
in RTS, all TCP bytestream processing is done by the
MPA protocol (code in siw_cm.c). If the QP reaches
RTS, RX processing is done bt the full iWarp
protocol (code in siw_qp_rx.c for RX).
Now, if the higher layer (application) protocol has
a semantic where the responder sends the first message,
we could end up in an infinite wait for that packet
at initiator application side.
This, since the complete first FPDU was already
received at the TCP socket, while the QP was not in
RTS. It will not trigger any additional sk->sk_data_ready()
and we are stuck with a FPDU in the socket rx buffer.
I implemented that 4 bytes extra data testing only to
keep the siw protocol state machine as simple as
possible (it is already unfortunately complex, if
you count the lines in siw_cm.c), while following
the protocol rules.
I suggest to correctly implement peer2peer mode and
delay the established event at the responder side until
it got the negotiated first zero length FPDU. Out
of the possible options, siw supports both zero length
WRITE and READ, but no SEND, since this would consume
an extra receive WQE.
As a last resort, I might consider extending the
siw state machine to handle those - non-complaint -
cases gracefully. But, that would result in different
code than just avoiding checking for peer protocol
violation.
Bernard, everything you describe in siw seems perfectly valid
to me, and in fact that's why the MPA enhanced connection mode
was designed the way it is. I disagree that it's "cumbersome",
but that's a nit.
The issue is that the responder reaches RTS at a different time
than the initiator reaches RTR. The original iWARP connection
model did not make any requirement, and races were possible.
MPAv2 fixes those.
Tom, thanks, exactly. MPAv2 fixes it. And an implementation
of that fix MUST adhere to the defined rules. If one
negotiates an extra handshake to synchronize, it
MUST adhere to that handshake rules. There is no point in
negotiating an extra handshake, and right away ignoring it.
If the responder wants a zero length WRITE, it MUST use
its reception to synchronize its transition to RTS.
Shall we extend the siw state machine to support silly
peer behavior? As said, if I read there is no way of
Of course the right answer to this is "no", but I'd reserve
drawing that conclusion until I am certain it won't cause
trouble. If existing iWARP implementations don't enforce
this, and SIW becomes the odd duck failing connections, then
I'm afraid I would have to say "yes" instead.
making the peer implementation RFC compliant (it would
be unexpected), I'll look into the applicability of
Postel's Rule. But let's make those things explicit.
Explicit is best, agreed.
Tom.
SIW is certainly within its rights to prevent peers from
sending prior to responder RTR. I'd suggest that if possible,
you try to detect the deadlock rather than flatly rejecting any
incoming bytes. The Internet Principle (be liberal in what you
accept...), after all.
This same dance happens in IB/RoCE, btw. Just via different
messages.
Interesting!
Thanks
Bernard.
Tom.
Thanks,
Bernard.
Sorry for the confusion. But the current
siw code appears to be just correct.
Thanks
Bernard
+ if ((__mpa_rr_revision(cep->mpa.hdr.params.bits) ==
MPA_REVISION_2) &&
+ (cep->state == SIW_EPSTATE_AWAIT_MPAREP)) {
+ int mpa_p2p_mode = cep->mpa.v2_ctrl_req.ord &
+ (MPA_V2_RDMA_WRITE_RTR |
MPA_V2_RDMA_READ_RTR);
+ if (!mpa_p2p_mode)
+ extra_data_check = 0;
+ }
+
/*
* At this point, we must have hdr->params.pd_len != 0.
* 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 +
extra_data_check,
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 + extra_data_check, MSG_DONTWAIT);
if (rcvd < 0)
return rcvd;
- if (rcvd > to_rcv)
+ if (extra_data_check && (rcvd > to_rcv))
return -EPROTO;
cep->mpa.bytes_rcvd += rcvd;
-Krishna.
Tom.
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.
