Am 29.01.20 um 11:17 schrieb Pavel Begunkov:
> On 29/01/2020 03:54, Jens Axboe wrote:
>> On 1/28/20 5:24 PM, Jens Axboe wrote:
>>> On 1/28/20 5:21 PM, Pavel Begunkov wrote:
>>>> On 29/01/2020 03:20, Jens Axboe wrote:
>>>>> On 1/28/20 5:10 PM, Pavel Begunkov wrote:
>>>>>>>>> Checked out ("don't use static creds/mm assignments")
>>>>>>>>>
>>>>>>>>> 1. do we miscount cred refs? We grab one in get_current_cred() for each async
>>>>>>>>> request, but if (worker->creds != work->creds) it will never be put.
>>>>>>>>
>>>>>>>> Yeah I think you're right, that needs a bit of fixing up.
>>>>>>>
>>>>>>
>>>>>> Hmm, it seems it leaks it unconditionally, as it grabs in a ref in
>>>>>> override_creds().
>>>>>>
>>>>>
>>>>> We grab one there, and an extra one. Then we drop one of them inline,
>>>>> and the other in __io_req_aux_free().
>>>>>
>>>> Yeah, with the last patch it should make it even
>>>
>>> OK good we agree on that. I should probably pull back that bit to the
>>> original patch to avoid having a hole in there...
>>
>> Done
>>
> 
> ("io_uring/io-wq: don't use static creds/mm assignments") and ("io_uring:
> support using a registered personality for commands") looks good now.
> 
> Reviewed-by: Pavel Begunkov <asml.silence@gmail.com>


I'm very happy with the design, thanks!
That exactly what I had in mind:-)

It would also work with IORING_SETUP_SQPOLL, correct?

However I think there're a few things to improve/simplify.

> https://git.kernel.dk/cgit/linux-block/commit/?h=for-5.6/io_uring-vfs&id=a26d26412e1e1783473f9dc8f030c3af3d54b1a6

In fs/io_uring.c mmgrab() and get_current_cred() are used together in
two places, why is put_cred() called in __io_req_aux_free while
mmdrop() is called from io_put_work(). I think both should be called
in io_put_work(), that makes the code much easier to understand.

My guess is that you choose __io_req_aux_free() for put_cred() because
of the following patches, but I'll explain on the other commit
why it's not needed.

> https://git.kernel.dk/cgit/linux-block/commit/?h=for-5.6/io_uring-vfs&id=d9db233adf034bd7855ba06190525e10a05868be

A minor one would be starting with 1 instead of 0 and using
idr_alloc_cyclic() in order to avoid immediate reuse of ids.
That way we could include the id in the tracing message and
0 would mean the current creds were used.

> +static int io_remove_personalities(int id, void *p, void *data)
> +{
> +	struct io_ring_ctx *ctx = data;
> +
> +	idr_remove(&ctx->personality_idr, id);

Here we need something like:
put_creds((const struct cred *)p);

> +	return 0;
> +}


The io_uring_register() calles would look like this, correct?

 id = io_uring_register(ring_fd, IORING_REGISTER_PERSONALITY, NULL, 0);
 io_uring_register(ring_fd, IORING_UNREGISTER_PERSONALITY, NULL, id);

> https://git.kernel.dk/cgit/linux-block/commit/?h=for-5.6/io_uring-vfs&id=eec9e69e0ad9ad364e1b6a5dfc52ad576afee235
> +
> +	if (sqe_flags & IOSQE_PERSONALITY) {
> +		int id = READ_ONCE(sqe->personality);
> +
> +		req->work.creds = idr_find(&ctx->personality_idr, id);
> +		if (unlikely(!req->work.creds)) {
> +			ret = -EINVAL;
> +			goto err_req;
> +		}
> +		get_cred(req->work.creds);> +		old_creds = override_creds(req->work.creds);
> +	}
> +

Here we could use a helper variable
const struct cred *personality_creds;
and leave req->work.creds as NULL.
It means we can avoid the explicit get_cred() call
and can skip the following hunk too:

> @@ -3977,7 +3977,8 @@ static int io_req_defer_prep(struct io_kiocb *req,
>  		mmgrab(current->mm);
>  		req->work.mm = current->mm;
>  	}
> -	req->work.creds = get_current_cred();
> +	if (!req->work.creds)
> +		req->work.creds = get_current_cred();
>  
>  	switch (req->opcode) {
>  	case IORING_OP_NOP:

The override_creds(personality_creds) has changed current->cred
and get_current_cred() will just pick it up as in the default case.

This would make the patch much simpler and allows put_cred() to be
in io_put_work() instead of __io_req_aux_free() as explained above.

metze