RE: [RFC PATCH v2] iov_iter: Convert iterate*() to inline funcs

[David Laight]

From: David Howells
> Sent: Wednesday, August 16, 2023 10:50 AM
> 
> David Laight <David.Laight@xxxxxxxxxx> wrote:
> 
> > It is harder to compare because of some of the random name changes.
> 
> I wouldn't say 'random' exactly, but if you prefer, some of the name changing
> can be split out into a separate patch.  The macros are kind of the worst
> since they picked up variable names from the callers.
> 
> > The version of the source I found seems to pass priv2 to functions
> > that don't use it?
> 
> That can't be avoided if I convert everything to inline functions and function
> pointers - but the optimiser can get rid of it where it can inline the step
> function.

AFAICT the IOVEC one was only called directly.

> I tried passing the iterator to the step functions instead, but that just made
> things bigger.  memcpy_from_iter_mc() is interesting to deal with.  I would
> prefer to deal with it in the caller so we only do the check once, but that
> might mean duplicating the caller.

You could try something slightly horrid that the compiler
might optimise for you.
Instead of passing in a function pointer pass a number.
Then do something like:
#define call_iter(id, ...) \
	(id == x ? fn_x(__VA_ARGS__) : id == y ? fn_y(__VA_ARGS) ...)
constant folding on the inline should kill the function pointer.
You might get away with putting the args on the end.

...
> > I rather hope the should_fail_usercopy() and instrument_copy_xxx()
> > calls are usually either absent or, at most, nops.
> 
> Okay - it's probably worth marking those too, then.

Thinking I'm sure they are KASAN annotations.
The are few enough calls that I suspect that replicating them
won't affect KASAN (etc) builds.

> > This all seems to have a lot fewer options than last time I looked.
> 
> I'm not sure what you mean by 'a lot fewer options'?

It might just be ITER_PIPE that has gone.

> > Is it worth optimising the KVEC case with a single buffer?
> 
> You mean an equivalent of UBUF?  Maybe.  There are probably a whole bunch of
> netfs places that do single-kvec writes, though I'm trying to convert these
> over to bvec arrays, combining them with their data, and MSG_SPLICE_PAGES.

I'm thinking of what happens with kernel callers of things
like the socket code - especially for address/option buffers.
Probably io_uring and bpf (and my out of tree drivers!).

Could be the equivalent of UBUF, but checking for KVEC with
a count of 1 wouldn't really add any more cmp/jmp pairs.

I've also noticed in the past that some of this code seems
to be optimised for zero length buffers/fragments.
Surely they just need to work?

	David

-
Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
Registration No: 1397386 (Wales)