Am 06.02.20 um 17:42 schrieb Jens Axboe:
> On 2/6/20 9:37 AM, Stefan Metzmacher wrote:
>> Am 06.02.20 um 17:04 schrieb Jens Axboe:
>>> On 2/6/20 9:02 AM, Stefan Metzmacher wrote:
>>>> Signed-off-by: Stefan Metzmacher <metze@xxxxxxxxx>
>>>> ---
>>>> src/include/liburing.h | 2 +-
>>>> 1 file changed, 1 insertion(+), 1 deletion(-)
>>>>
>>>> diff --git a/src/include/liburing.h b/src/include/liburing.h
>>>> index faed2e7..44f18fd 100644
>>>> --- a/src/include/liburing.h
>>>> +++ b/src/include/liburing.h
>>>> @@ -179,7 +179,7 @@ static inline void io_uring_sqe_set_data(struct io_uring_sqe *sqe, void *data)
>>>> sqe->user_data = (unsigned long) data;
>>>> }
>>>>
>>>> -static inline void *io_uring_cqe_get_data(struct io_uring_cqe *cqe)
>>>> +static inline void *io_uring_cqe_get_data(const struct io_uring_cqe *cqe)
>>>> {
>>>> return (void *) (uintptr_t) cqe->user_data;
>>>> }
>>>
>>> Applied, thanks.
>>
>> Thanks!
>>
>>> Unrelated to this patch, but I'd like to release a 0.4 sooner rather
>>> than later.
>>
>> Funny, I thought about that today:-)
>> I prepared debian packaging for liburing-0.4 I'll send the patches soon.
>
> Great!
>
>> While doing that I found the following incompatible change against
>> liburing-0.3:
>>
>> static inline void io_uring_prep_files_update(struct io_uring_sqe *sqe,
>> - int *fds, unsigned nr_fds)
>> + int *fds, unsigned nr_fds,
>> + int offset)
>>
>> I'm not sure if we should do something about that.
>
> Hmm, that wasn't on purpose. But for this specific case, I think we can
> just pretend that never happened.
Ok.
>> It's also strange that in src/liburing.map LIBURING_0.3 doesn't
>> inherit LIBURING_0.2. There's not a single symbol with @LIBURING_0.3.
>> Also io_uring_{setup,enter,register} are still
>> listed under LIBURING_0.1, but they're not in the library anymore.
>
> That seems like a bug, I'd happily take a patch for that...
I'm not sure how what to change, so I'll just leave it for now.
>>> Let me know if you see any immediate work that needs doing
>>> before that happens.
>>
>> I had the idea to have a simple version of fd compounding.
>> We already have IORING_OP_FILES_UPDATE in order to update
>> specific indexes in the files array.
>> I'm wondering if we could have specify an index where
>> IORING_OP_ACCEPT, IORING_OP_OPENAT and IORING_OP_OPENAT2
>> could store the generated fd into the fixed array.
>> The index 0 is not valid, correct? So we can have it
>> without a flag similar to the personality, and for
>> all of these buf_index is not used.
>
> Just to make sure I'm undestanding your proposal, you want the result
> from those fd instantiating calls to be added to the array of registered
> files, instead of having the application do that?
Yes.
> If so, I think this is another case where the BPF driven links would be useful, as we could
> easily do it through that with an IORING_OP_FILES_UPDATE linked to
> either one of those commands.
Sure, it's just hard to imagine how to do BPF without it being there and
a real world examples...
> index 0 is valid, so we can't use that trick.
Ok :-)
>> While researching that I noticed that IOSQE_FIXED_FILE
>> seems to be ignored for some new commands, I think that
>> all commands with on input fd, should be able to use that flag.
>> Can this be fixed before 5.6 final?
>
> Do you have specifics? Generally the file grabbing happens as part of
> request prep, and the individual opcodes should not need to bother with
> it.
io_statx_prep():
io_openat_prep():
io_openat2_prep():
req->open.dfd = READ_ONCE(sqe->fd);
io_statx():
ret = filename_lookup(ctx->dfd, ctx->filename, lookup_flags...
io_openat2():
file = do_filp_open(req->open.dfd, req->open.filename, &op);
io_close_prep(): has this to make it clear that IOSQE_FIXED_FILE is not
supported, I guess because FILE_UPDATE with -1 needs to be used instead?
if (sqe->flags & IOSQE_FIXED_FILE)
return -EINVAL;
req->close.fd = READ_ONCE(sqe->fd);
I guess at least we need if (sqe->flags & IOSQE_FIXED_FILE) in all
cases, if we can't just fix it.
metze
Attachment:
signature.asc
Description: OpenPGP digital signature
