On 24/02/2020 19:02, Jens Axboe wrote:
>> Usually doesn't work because of such possible "hackier assignments".
>> Ok, I have to go and experiment a bit. Anyway, it probably generates a lot of
>> useless stuff, e.g. for req->ctx
>
> Tried this, and it generates the same code...
Maybe it wasn't able to optimise in the first place
E.g. for the following code any compiler generates 2 reads (thanks godbolt).
extern void foo(int);
int bar(const int *v)
{
foo(*v);
return *v;
}
>
> diff --git a/fs/io_uring.c b/fs/io_uring.c
> index ba8d4e2d9f99..8de5863aa749 100644
> --- a/fs/io_uring.c
> +++ b/fs/io_uring.c
> @@ -598,7 +598,7 @@ struct io_kiocb {
>
> struct io_async_ctx *io;
> bool needs_fixed_file;
> - u8 opcode;
> + const u8 opcode;
>
> struct io_ring_ctx *ctx;
> struct list_head list;
> @@ -5427,6 +5427,8 @@ static bool io_get_sqring(struct io_ring_ctx *ctx, struct io_kiocb *req,
> */
> head = READ_ONCE(sq_array[ctx->cached_sq_head & ctx->sq_mask]);
> if (likely(head < ctx->sq_entries)) {
> + u8 *op;
> +
> /*
> * All io need record the previous position, if LINK vs DARIN,
> * it can be used to mark the position of the first IO in the
> @@ -5434,7 +5436,8 @@ static bool io_get_sqring(struct io_ring_ctx *ctx, struct io_kiocb *req,
> */
> req->sequence = ctx->cached_sq_head;
> *sqe_ptr = &ctx->sq_sqes[head];
> - req->opcode = READ_ONCE((*sqe_ptr)->opcode);
> + op = (void *) req + offsetof(struct io_kiocb, opcode);
> + *op = READ_ONCE((*sqe_ptr)->opcode);
> req->user_data = READ_ONCE((*sqe_ptr)->user_data);
> ctx->cached_sq_head++;
> return true;
>
--
Pavel Begunkov
Attachment:
signature.asc
Description: OpenPGP digital signature
