On Fri, Aug 9, 2024 at 8:29 PM Al Viro [off-list ref] wrote:

On Thu, Aug 08, 2024 at 09:51:34AM -0700, Alexei Starovoitov wrote:

quoted

The bpf changes look ok and Andrii's approach is easier to grasp.
It's better to route bpf conversion to CLASS(fd,..) via bpf-next,
so it goes through bpf CI and our other testing.

bpf patches don't seem to depend on newly added CLASS(fd_pos, ...
and fderr, so pretty much independent from other patches.

Representation change and switch to accessors do matter, though.
OTOH, I can put just those into never-rebased branch (basically,
"introduce fd_file(), convert all accessors to it" +
"struct fd representation change" + possibly "add struct fd constructors,
get rid of __to_fd()", for completeness sake), so you could pull it.
Otherwise you'll get textual conflicts on all those f.file vs. fd_file(f)...

Yep, makes sense. Let's do that, we can merge that branch into
bpf-next/master and I will follow up with my changes on top of that.

Let's just drop the do_one_ldimm64() extraction, and keep fdput(f)
logic, plus add fd_file() accessor changes. I'll then add a switch to
CLASS(fd) after a bit more BPF-specific clean ups. This code is pretty
sensitive, so I'd rather have all the non-trivial refactoring done
separately. Thanks!