On 12/5/24 18:04, Keith Busch wrote:
...
diff --git a/include/uapi/linux/io_uring.h b/include/uapi/linux/io_uring.h
index aac9a4f8fa9a..38f0d6b10eaf 100644
--- a/include/uapi/linux/io_uring.h
+++ b/include/uapi/linux/io_uring.h
@@ -98,6 +98,10 @@ struct io_uring_sqe {
__u64 addr3;
__u64 __pad2[1];
};
+ struct {
+ __u64 attr_ptr; /* pointer to attribute information */
+ __u64 attr_type_mask; /* bit mask of attributes */
+ };
I can't say I'm a fan of how this turned out. I'm merging up the write
hint stuff, and these new fields occupy where that 16-bit value was
initially going. Okay, so I guess I need to just add a new attribute
flag? That might work if I am only appending exactly one extra attribute
per SQE, but what if I need both PI and Write Hint? Do the attributes
need to appear in a strict order?
Martin put it well, this version requires attributes to be placed
in their id order, but FWIW without any holes, i.e. the following
is fine:
ATTR_PI = 1
ATTR_WRITE_STREAM = 2
strcut attr_stream attr = { ... };
sqe->attr_mask = ATTR_WRITE_HINT;
sqe->attr_ptr = &attr;
In case of multiple attributes:
struct compound_attr {
struct attr_pi a;
struct attr_stream b;
} attr = { ... };
sqe->attr_mask = ATTR_WRITE_HINT | ATTR_PI;
sqe->attr_ptr = &attr;
The other option for the uapi, and Martin mentioned it as well, is
to add a type field to all attributes.
FWIW, I think the space aliased with sqe->splice_fd_in is not
used by read/write, but I haven't looked into streams to get
an opinion on which way is more appropriate.
--
Pavel Begunkov
