在 2024/8/27 3:04, Jan Karcher 写道:
On 24/08/2024 12:04, liujian (CE) wrote:
在 2024/8/21 9:03, Dust Li 写道:
On 2024-08-20 15:16:57, Jan Karcher wrote:
On 09/08/2024 10:31, Liu Jian wrote:
Make SMC-R can work with rxe devices. This allows us to easily test
and
learn the SMC-R protocol without relying on a physical RoCE NIC.
Hi Liu,
sorry for taking quite some time to answer.
Looking into this i cannot accept this series at the given point of
time.
FWIU, RXE is mainly for testing and development and i agree that it
would be
a nice thing to have for SMC-R.
The problem is that there is no clean layer for different RoCE devices
currently. Adding RXE to it works but isn't clean.
Hi jan,
Also we have no way to do a "test" build which would have such a device
supported and a "prod" build which would not support it.
> I don't quite understand what you mean here, Maybe I missed
something ?
IIUC, we can control whether to use RXE by simpling insmod or rmmod
rdma_rxe.ko
Hi,
Yes that enables RXE in general, but not the use of RXE in SMC.
>>>
Yes, in the "prod" environment, we can completely turn off
CONFIG_RDMA_RXE.
Same as above + this is a compile time switch that is enabled for
distros like rh. Simply disabling it won't work here.
Got it, I misunderstood what you meant above. Thank you.
I believe having RXE support is beneficial for testing, especially in
simple physical networking setups where many corner cases are unlikely
to occur. By using RXE, we can easily configure unusual scenarios with
the existing iptables/netfilter infrastructure to simulate real-world
situations, such as packet dropping or network retransmission. This
approach can be advantageous for finding hidden bugs.
Yes, one of my main original intentions was to make testing smc-r
easier. This change is relatively simple, mainly patch2 and patch4,
and there are no logical changes.
I agree with you. It would be beneficial for testing.
This is not a never, this is a not right now.
If you want to push this forward as something you need now, feel free to
encapsulate it and introduce a vendor specific experimental option as
defined in the v2.1 protocol version [1] for it. This would be
compromise for me at the current time.
Got it, thanks.
Thanks
- Jan
[1]
https://www.ibm.com/support/pages/system/files/inline-files/IBM%20Shared%20Memory%20Communications%20Version%202.1%20Emulated-ISM_0.pdf
>>> Best regards,
Dust
Please give us time to investigate how to solve this in a neat way
without
building up to much technical debt.
Thanks for your contribution and making us aware of this area of
improvment.
- Jan
Liu Jian (4):
rdma/device: export ib_device_get_netdev()
net/smc: use ib_device_get_netdev() helper to get netdev info
net/smc: fix one NULL pointer dereference in
smc_ib_is_sg_need_sync()
RDMA/rxe: Set queue pair cur_qp_state when being queried
drivers/infiniband/core/core_priv.h | 3 ---
drivers/infiniband/core/device.c | 1 +
drivers/infiniband/sw/rxe/rxe_verbs.c | 2 ++
include/rdma/ib_verbs.h | 2 ++
net/smc/smc_ib.c | 10 +++++-----
net/smc/smc_pnet.c | 6 +-----
6 files changed, 11 insertions(+), 13 deletions(-)