On Fri, Jul 19, 2024 at 10:51:58AM +0000, Konstantin Taranov wrote:
> > > > > > > Yes, you are. If user asked for specific functionality
> > > > > > > (max_inline_data != 0) and your device doesn't support it, you
> > > > > > > should
> > > > return an error.
> > > > > > >
> > > > > > > pvrdma, mlx4 and rvt are not good examples, they should return
> > > > > > > an error as well, but because of being legacy code, we won't change
> > them.
> > > > > > >
> > > > > > > Thanks
> > > > > > >
> > > > > >
> > > > > > I see. So I guess we can return a larger value, but not smaller. Right?
> > > > > > I will send v2 that fails QP creation then.
> > > > > >
> > > > > > In this case, may I submit a patch to rdma-core that queries
> > > > > > device caps before trying to create a qp in rdma_client.c and
> > > > > > rdma_server.c? As that code violates what you described.
> > > > >
> > > > > Let's ask Jason, why is that? Do we allow to ignore max_inline_data?
> > > > >
> > > > > librdmacm/examples/rdma_client.c
> > > > > 63 memset(&attr, 0, sizeof attr);
> > > > > 64 attr.cap.max_send_wr = attr.cap.max_recv_wr = 1;
> > > > > 65 attr.cap.max_send_sge = attr.cap.max_recv_sge = 1;
> > > > > 66 attr.cap.max_inline_data = 16;
> > > > > 67 attr.qp_context = id;
> > > > > 68 attr.sq_sig_all = 1;
> > > > > 69 ret = rdma_create_ep(&id, res, NULL, &attr);
> > > > > 70 // Check to see if we got inline data allowed or not
> > > > > 71 if (attr.cap.max_inline_data >= 16)
> > > > > 72 send_flags = IBV_SEND_INLINE;
> > > > > 73 else
> > > > > 74 printf("rdma_client: device doesn't support
> > > > IBV_SEND_INLINE, "
> > > > > 75 "using sge sends\n");
> > > >
> > > > I think the idea expressed in this code is that if max_inline_data
> > > > requested too much it would be limited to the device capability.
> > > >
> > > > ie qp creation should limit the requests values to what the HW can
> > > > do, similar to how entries and other work.
> > > >
> > > > If the HW has no support it should return - for max_inline_data not
> > > > an error, I guess?
> > >
> > > Yes, this code implies that max_inline_data can be ignored at creation,
> > while the manual of ibv_create_qp says:
> > > "The function ibv_create_qp() will update the qp_init_attr->cap struct
> > > with the actual QP values of the QP that was created; the values will
> > > be **greater than or equal to** the values requested."
> >
> > Ah, well that seems to be some misunderstandings then, yes.
> >
> > > I see two options:
> > > 1) Remove code from rdma examples that rely on ignoring max_inline; add
> > a warning to libibverbs when drivers ignore that value.
> > > 2) Add to manual that max_inline_data might be ignored by drivers; and
> > allow my current patch that ignores max_inline_data in mana_ib.
> >
> > I don't know, what do the majority of drivers do? If enough are already doing
> > 1 then lets force everyone into 1, otherwise we have to document 2.
> >
> > And a pyverbs test should be added to cover this weirdness
>
> I quickly read create_qp code of all providers and it seems that max_inline_data is ignored by hw/pvrdma and sw/rvt.
> Other providers fail the creation when they cannot satisfy the inline_data cap.
> Some drivers ignore it for GSI, but I think it is reasonable.
>
> Then I guess the option 1 is better. Regarding pyverbs, should I add a test for the option 1?
> If yes, what should it test?
Probably, the test should check the max_inline_data value returned from device caps and try to create
QP with higher value. If the QP creation fails, the test should pass. For hw/pvrdma and sw/rvt, the QP
should be successfully created, despite the requested value.
Thanks
>
> >
> > Jason
