Re: [PATCH for-next 05/23] RDMA/core: Add writev to uverbs file descriptor

[Leon Romanovsky <leon@kernel.org>]

On Tue, Jul 01, 2025 at 01:17:48PM -0400, Dennis Dalessandro wrote:
> On 7/1/25 12:04 PM, Leon Romanovsky wrote:
> > On Tue, Jul 01, 2025 at 09:57:23AM -0400, Dennis Dalessandro wrote:
> >> On 7/1/25 8:32 AM, Leon Romanovsky wrote:
> >>> On Mon, Jun 30, 2025 at 11:30:12AM -0400, Dennis Dalessandro wrote:
> >>>> From: Dean Luick <dean.luick@cornelisnetworks.com>
> >>>>
> >>>> Add a writev pass-through between the uverbs file descriptor and
> >>>> infiniband devices.  Interested devices may subscribe to this
> >>>> functionality.
> >>>
> >>> Aren't we use IOCTL and not write interface now?
> >>> Why do we need this?
> >>
> >> We wanted to keep all the semantics of the user interface the same so it's an
> >> easy migration to the uverbs cdev from the private cdev. The idea is that all
> >> the command and control is still ioctl, but the "data path" is still using the
> >> writev() to pass in the iovecs.
> > 
> > ok, just add this to the commit message.
> > 
> 
> Will do.
> 
> >> By the way I'll get the rdma_core (user space) changes posted soon so you can
> >> see both sides.
> > 
> > BTW, I looked whole series and upto MAD/verbs interface everything looks ok.
> > The latter simply broke me with >3K LOC in single patch.
> 
> Yeah sorry about all that. The advantage to the new chip is we don't need to do
> all of the MAD stuff in the kernel because it's now done in FW. So most of it is
> just carried over from hfi1. The driver just needs to find out about some
> things. Would it help if I point out the major differences? I could take a crack
> and trying to break that up more even.

It will help if you can separate machine generated code vs. real logic.
MAD patches are not the worst part. For example, this one:
https://lore.kernel.org/all/175129743814.1859400.4253022820082459886.stgit@awdrv-04.cornelisnetworks.com/
 3 files changed, 17722 insertions(+)

With a lot of functions like this, with random casting, multiple
parameters without use e.t.c

+static u64 access_sw_pio_drain(const struct cntr_entry *entry,
+			       void *context, int vl, int mode, u64 data)
+{
+	struct hfi2_devdata *dd = (struct hfi2_devdata *)context;
+
+	return dd->verbs_dev.n_piodrain;
+}

+static u64 access_sw_ctx0_seq_drop(const struct cntr_entry *entry,
+                                  void *context, int vl, int mode, u64 data)
+{
+       struct hfi2_devdata *dd = context;
+
+       return dd->ctx0_seq_drop;
+}

> 
> -Denny
>