Re: [RFC net-next 11/17] net/dibs: Move struct device to dibs_dev

From
Alexandra Winter <wintera@linux.ibm.com>
Date
2025-08-15 12:00:03
Also in
linux-rdma, linux-s390
Reply to
Dust Li, 2025-08-15

On 15.08.25 03:56, Dust Li wrote:
On 2025-08-14 10:51:27, Alexandra Winter wrote:
quoted
On 06.08.25 17:41, Alexandra Winter wrote:
[...]
quoted
Replace smcd->ops->get_dev(smcd) by dibs_get_dev().
Looking at the resulting code, I don't really like this concept of a *_get_dev() function,
that does not call get_device().
I plan to replace that by using dibs->dev directly in the next version.
May I ask why? Because of the function name ? If so, maybe we can change the name.
Yes the name. Especially, as it is often used as argument for get_device() or put_device().
Eventually I would like to provide dibs_get_dev()/dibs_put_dev() that actually
do refcounting.
And then I thought defining dibs_read_dev() is not helping readability.
While I don't have a strong preference either way, I personally favor
hiding the members of the dibs_dev structure from the upper layer. In my
opinion, it would be better to avoid direct access to dibs members from
upper layers and instead provide dedicated interface functions.

For example, I even think we should not expose dibs->ops->xxx directly
to the SMC layer. Encapsulating such details would improve modularity
and maintainability. Just like what IB subsystem has done before :)

For example:
# git grep dibs net/smc
[...]
net/smc/smc_ism.c:      return dibs->ops->query_remote_gid(dibs, &ism_rgid, vlan_id ? 1 : 0,
net/smc/smc_ism.c:      return smcd->dibs->ops->get_fabric_id(smcd->dibs);
net/smc/smc_ism.c:      if (!smcd->dibs->ops->add_vlan_id)
net/smc/smc_ism.c:      if (smcd->dibs->ops->add_vlan_id(smcd->dibs, vlanid)) {
net/smc/smc_ism.c:      if (!smcd->dibs->ops->del_vlan_id)
net/smc/smc_ism.c:      if (smcd->dibs->ops->del_vlan_id(smcd->dibs, vlanid))
[...]

Best regards,
Dust

I see your point and I remember you brought that up in your review of
[RFC net-next 0/7] Provide an ism layer
already.

I tried to keep this series to a meaningful minimum, which is a tradeoff.
If possible, I just wanted to move code around and add the dibs layer
in-between. There are several areas where I would like to see even more
de-coupling. eg.:
- handle_irq(): Clients should not run in interrupt context,
  a receive_data() callback function would be better.
- The device drivers should not loop through the client array
- dibs_dev_op.*_dmb() functions reveal unnecessary details of the
  internal dmb struct to the clients
- ...

So instead of adding a set of 1:1 caller functions / interface functions
for dibs_dev_ops and dibs_client_ops now, I would like to propose to work
on further decoupling devices and clients by adding more abstractions that
bring benefit. And then replace the remaining calls to ops by 1:1 caller
functions. Does that make sense? Or does anybody feel strongly that I need
to provide interface functions now?

BTW, there are some client-only functions and some device-driver-only functions
in dibs.h already. So that is the direction.




Thread (48 messages) 48 messages, 4 authors, last reply 2025-09-01
Keyboard shortcuts
hback out one level
jnext message in thread / next series-diff position
kprevious message in thread / previous series-diff position
ldrill in
Escclose help / fold thread tree
?toggle this help