On 9/9/2015 6:20 PM, Doug Ledford wrote:
On 09/05/2015 03:59 PM, Or Gerlitz wrote:
--> they must not put any
additional bits on the wire
Maybe. For base level interop, sure, but for enhanced service in a homogeneous
environment, not necessarily true.
--> RC isn't an option
Not following here Or. The IPoIB connected mode packet is 4 byte IPoIB
Header followed by whatever header the TCP stack put on the packet
(likely either IPv4 or IPv6). The same ipoib_hard_header function is
used for both CM and UD connections, the same ipoib_start_xmit() is used
for both CM and UD connections, and we just hand off to the appropriate
send routine when we have our neighbor lookup complete. Why would you
need new bits on the wire to do a checksum on an RC send and not for a UD send?
I didn't say that, I said (again) the offload engines will use just
plain UD along the IPoIB RFC, b/c each IPoIB implementation would
support that. It's true that an offload can use connected mode (CM) and
fall back to UD if the other end don't support CM.
To make things concrete, exactly nothing in the libibverbs patch is made
such it would work just for UD and RAW PACKET qps, except few TEXT
comments and man page line. In other words, if there is HW vendor that
supports checksum generation/validation on RC QPs too - the same API
would just work.
It's pretty easy to add one line of check for the vendor send function
such that it returns error if checksum generation isn't supported for
the given QP type, and in that respect, before going and saying this
submission is crappy, you could have just provided this reviewer comment
"remove the assumption on QP types, make sure in libmlx4 you return
error for non UD/RAW qps" and (see next)
, so for IPoIB we
just need an IPoIB UD QP in user space, and for Ethernet RAW PACKET
QP. This device capability is there ~10y for IB UD and we just
naturally extend it to Eth RAW. If/when a vendor comes up with
supporting csum for RC, we can add another dev cap and say the well
established API applies on them too, with just a slight modification
to man pages and such, makes sense?
So if we ever support RC, then any actual users of this API will have
hardcoded which types of QPs are supported into their apps and they will
*all* have to go modify their source code to re-hardcode the types into
their app and recompile.
Alternatively, we write a reasonable API. One where the types of QPs
are not set in stone, we tell users to query the API to determine if any
given QP supports IP CSUM offloads, and then if we add more QP types in
the future, the apps just automatically work even on the different QP
types (assuming they used those QP types at all) because it was written
to an API that allowed for it.
Face it Or, the API in this patchset is crap. I totally get why you are
fighting for it so hard. You already spelled it out: "This device
capability is there ~10y for IB UD and we just naturally extend it to
Eth RAW". I can read between the lines. You have users of the API,
probably some internal and some external, and if I go demanding a proper
API, all of these people have to recode their apps to the new API, and
you'd like to avoid that if possible. However, if I don't, and then RC
support is added, then they have to recode their apps anyway. Wouldn't
it be best to just do it right and get it over with?
(continuing) "show me a plan, how applications can determine the support
matrix for QPs types and checksum offloads" -- e.g one can enhance the
extended query device verbs to report that and such.
Indeed there are bunch of UD and RAW based applications out there which
make perfect use of this API and I don't think we need to gate the
libibverbs support with that extended query kernel patch.
Why not move forward with the vendor library flow to act along your
request, removal of assumption from man pages and a kernel patch, which
on top of, we can add another libibverbs patch that allows apps to query?
I find this approach practical on the one hand, and with clear plan to
make it fully robust on the other hand.
So, here's the API I'm proposing:
- Add ibv_query_qp_ex
- In new ibv_query_qp_ex struct, extend the ibv_qp_caps struct to add a
flags element
- Define a flag for IP_CSUM_SUPPORTED
- Define IP_CSUM flag for send operations
- Define the API so that IP_CSUM is ignored on all sends if the QP
doesn't support IP_CSUM and only checked on QPs that support it. This
way other QP types don't suffer a penalty on send to check this and
return EINVAL if its set
- Define the return flags in the wc struct so we can signal that a CSUM
was performed and succeeded
A user app would then basically follow this flow:
ibv_create_qp()
ibv_query_qp()
check for IP CSUM and cache result
ibv_post_send()
set IP_CSUM if QP supports it
ibv_poll_cq()
if qp supports IP_CSUM, check CSUM result in wc
That's what I would like to see for these changes.
basically OK, expect that the support matrix for QP types is a property
of the device, e.g in a similar manner we have for ODP, right? and hence
the plan I suggested above.
Or.
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
