From: David Howells

Sent: Friday, August 18, 2023 12:43 PM

Linus Torvalds [off-list ref] wrote:

quoted

Well, that part is trivially fixable, and we should do that anyway for
other reasons.
..
 enum iter_type {
 	/* iter types */
-	ITER_IOVEC,
-	ITER_KVEC,
-	ITER_BVEC,
-	ITER_XARRAY,
-	ITER_DISCARD,
-	ITER_UBUF,
+	ITER_IOVEC = 1,
+	ITER_UBUF = 2,
+	ITER_KVEC = 4,
+	ITER_BVEC = 8,
+	ITER_XARRAY = 16,
+	ITER_DISCARD = 32,

That could be zero - no bits and default.

quoted

 };

It used to be this way, but Al switched it:

	8cd54c1c848031a87820e58d772166ffdf8c08c0
	iov_iter: separate direction from flavour

Except it also had the direction flag inside the enum.
That caused its own piles of grief.

IIRC Linus had type:6 - that doesn't leave any headroom
for additional types (even though they shouldn't proliferate).
It may be best to avoid bits 15+ (in a bitfield) due to
issues with large constants and sign extension.
On x86 (I think) 'and immediate' and 'bit test' are the same
size for bit 0 to 7, BIT wins for higher bits.

gcc generates strange code for some initialisers (see yesterday's
thread) and you definitely mustn't leave unused bits in a bitfield.
Might be better is the fields are assigned later!

(I also saw clang carefully preserving %eax on stabck!)

	David

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