在 2024/9/25 17:26, Leon Romanovsky 写道:
On Wed, Sep 25, 2024 at 10:04:21AM +0800, Zhu Yanjun wrote:
在 2024/9/24 23:16, Haakon Bugge 写道:
On 24 Sep 2024, at 15:59, Zhu Yanjun <yanjun.zhu@xxxxxxxxx> wrote:
在 2024/9/24 14:54, Christoph Hellwig 写道:
On Tue, Sep 24, 2024 at 09:58:24AM +0800, Zhu Yanjun wrote:
The users that I mentioned is not in the kernel tree.
And why do you think that would matter the slightest?
I noticed that the same cq functions are used. And I also made tests with this patch series. Without this patch series, dim mechanism will not be invoked.
Christoph alluded to say: Do not modify the old cq_create_cq() code in order to support DIM, it is better to change the ULP to use ib_alloc_cq(), and get DIM enabled that way.
Hi, Haakon
To be honest, I like your original commit that enable DIM for legacy ULPs
because this can fix this problem once for all and improve the old
ib_create_cq function.
The idea from Christoph will cause a lot of changes in ULPs. I am not very
sure if these changes cause risks or not.
Thus, I prefer to your original commit. But I will follow the advice from
Leon and Jason.
Christoph was very clear and he summarized our position very well. We
said similar thing to SMC folks in 2022 [1] and RDS is no different here.
Thanks, Leon. I will read this link carefully.
So no, "old ib_create_cq" shouldn't be used by ULPs.
Got it. I have replaced ib_create_cq with ib cq pool APIs. Perhaps
drivers/infiniband/ulp/srpt/ib_srpt.c is a good example to use ib cq
pool APIs.
Best Regards,
Zhu Yanjun
Thanks
[1] https://lore.kernel.org/netdev/YePesYRnrKCh1vFy@unreal/
Zhu Yanjun
Thxs, Håkon
