Re: [PATCH v7 01/31] iov_iter: Add ITER_XARRAY

From: Al Viro
Date: Sun Apr 25 2021 - 09:14:35 EST


On Fri, Apr 23, 2021 at 02:28:01PM +0100, David Howells wrote:

> diff --git a/include/linux/uio.h b/include/linux/uio.h
> index 27ff8eb786dc..5f5ffc45d4aa 100644
> --- a/include/linux/uio.h
> +++ b/include/linux/uio.h
> @@ -10,6 +10,7 @@
> #include <uapi/linux/uio.h>
>
> struct page;
> +struct address_space;
> struct pipe_inode_info;
>
> struct kvec {

What is that chunk for?

> +#define iterate_all_kinds(i, n, v, I, B, K, X) { \
> if (likely(n)) { \
> size_t skip = i->iov_offset; \
> if (unlikely(i->type & ITER_BVEC)) { \
> @@ -88,6 +125,9 @@
> struct kvec v; \
> iterate_kvec(i, n, v, kvec, skip, (K)) \
> } else if (unlikely(i->type & ITER_DISCARD)) { \
> + } else if (unlikely(i->type & ITER_XARRAY)) { \
> + struct bio_vec v; \
> + iterate_xarray(i, n, v, skip, (X)); \
> } else { \
> const struct iovec *iov; \
> struct iovec v; \
> @@ -96,7 +136,7 @@
> } \
> }

For the record - these forests of macros had been my mistake. I'm trying
to get rid of that crap right now, but your changes don't look likely to be
trouble in that respect.


> @@ -738,6 +783,16 @@ size_t _copy_mc_to_iter(const void *addr, size_t bytes, struct iov_iter *i)
> bytes = curr_addr - s_addr - rem;
> return bytes;
> }
> + }),
> + ({
> + rem = copy_mc_to_page(v.bv_page, v.bv_offset,
> + (from += v.bv_len) - v.bv_len, v.bv_len);
> + if (rem) {
> + curr_addr = (unsigned long) from;
> + bytes = curr_addr - s_addr - rem;
> + rcu_read_unlock();
> + return bytes;
> + }

That's broken, same way as kvec and bvec cases are in the same primitive.
Iterator not advanced on failure halfway through.

> @@ -1246,7 +1349,8 @@ unsigned long iov_iter_alignment(const struct iov_iter *i)
> iterate_all_kinds(i, size, v,
> (res |= (unsigned long)v.iov_base | v.iov_len, 0),
> res |= v.bv_offset | v.bv_len,
> - res |= (unsigned long)v.iov_base | v.iov_len
> + res |= (unsigned long)v.iov_base | v.iov_len,
> + res |= v.bv_offset | v.bv_len
> )
> return res;
> }

Hmm... That looks like a really bad overkill - do you need anything beyond count and
iov_offset in that case + perhaps "do we have the very last page"? IOW, do you need
to iterate anything at all here? What am I missing here?

> @@ -1268,7 +1372,9 @@ unsigned long iov_iter_gap_alignment(const struct iov_iter *i)
> (res |= (!res ? 0 : (unsigned long)v.bv_offset) |
> (size != v.bv_len ? size : 0)),
> (res |= (!res ? 0 : (unsigned long)v.iov_base) |
> - (size != v.iov_len ? size : 0))
> + (size != v.iov_len ? size : 0)),
> + (res |= (!res ? 0 : (unsigned long)v.bv_offset) |
> + (size != v.bv_len ? size : 0))
> );
> return res;
> }

Very limited use; it shouldn't be called for anything other than IOV_ITER case.

> @@ -1849,7 +2111,12 @@ int iov_iter_for_each_range(struct iov_iter *i, size_t bytes,
> kunmap(v.bv_page);
> err;}), ({
> w = v;
> - err = f(&w, context);})
> + err = f(&w, context);}), ({
> + w.iov_base = kmap(v.bv_page) + v.bv_offset;
> + w.iov_len = v.bv_len;
> + err = f(&w, context);
> + kunmap(v.bv_page);
> + err;})

Would be easier to have that sucker removed first...

Anyway, I can live with that; the only real bug is in sodding _copy_mc_to_iter(),
it's not anything new and it can be fixed at the same time we deal with kvec and
bvec cases there. Which, unfortunately, requires untangling the macro mess ;-/

What I've got in a local branch right now is
* removal of iov_iter_for_each_range() (yours, BTW)
* separation of flavour and direction (and the end of pseudo-bitmaps)
* untangling and optimizing iov_iter_advance(); iovec/kvec cases are
switched to the logics similar to bvec_iter_advance(), get considerably smaller
and should be faster
* fixing ITER_DISCARD iov_iter_advance() - move past the end should
quietly stop at the end.
* getting rid of iterate_all_kinds() in iov_iter_alignment(),
iov_iter_gap_alignment(), iov_iter_get_pages() and iov_iter_get_pages_alloc().

After that the only remaining irregular case of iterate_all_kinds() is in
iov_iter_npages(); that's what I'm trying to sort out right now. With that
done, all remaining uses will be for copying-style primitives, same as for
iterate_and_advance(). What I want to try after that is a separate "tracking"
argument, so that e.g. in _copy_to_iter() we'd have
iterate_and_advance(i, bytes, from, v,
copyout(v.iov_base, from, v.iov_len),
memcpy_to_page(v.bv_page, v.bv_offset, from, v.bv_len),
memcpy(v.iov_base, from, v.iov_len)
)
Next step will be to teach the damn thing about the possibility of short
copies in kvec/bvec cases. We'd get
#define iterate_and_advance(i, n, p, v, I, K, B) \
__iterate_and_advance(i, n, p, v, I, (K, 0), (B, 0))
and AFAICS it can be done in a way that won't screw code generation for
the normal ones. At that point _copy_mc_to_iter() mess gets cleared *AND*
we can merge K and B callbacks, handling B as kmap_atomic + K + kunmap_atomic
(_copy_mc_to_iter() is the main obstacle to that). Your callback (X) would
also fold into that.

After that I want to try and see how well iov_iter_advance() got optimized
and see if we can get e.g. _copy_to_iter() simply to

iterate_all_kinds(i, bytes, from, v,
copyout(v.iov_base, from, v.iov_len),
memcpy(v.iov_base, from, v.iov_len)
)
iov_iter_advance(i, from - addr);
return from - addr;
If iov_iter_advance() ends up being too much overhead - oh, well, we'll keep
iterate_and_advance() along with iterate_all_kinds(). Needs profiling,
obviously.