> Il giorno 18 apr 2018, alle ore 00:57, Jens Axboe <axboe@xxxxxxxxx> ha scritto:
>
> On 4/17/18 3:48 PM, Jens Axboe wrote:
>> On 4/17/18 3:47 PM, Kees Cook wrote:
>>> On Tue, Apr 17, 2018 at 2:39 PM, Jens Axboe <axboe@xxxxxxxxx> wrote:
>>>> On 4/17/18 3:25 PM, Kees Cook wrote:
>>>>> On Tue, Apr 17, 2018 at 1:46 PM, Kees Cook <keescook@xxxxxxxxxxxx> wrote:
>>>>>> I see elv.priv[1] assignments made in a few places -- is it possible
>>>>>> there is some kind of uninitialized-but-not-NULL state that can leak
>>>>>> in there?
>>>>>
>>>>> Got it. This fixes it for me:
>>>>>
>>>>> diff --git a/block/blk-mq.c b/block/blk-mq.c
>>>>> index 0dc9e341c2a7..859df3160303 100644
>>>>> --- a/block/blk-mq.c
>>>>> +++ b/block/blk-mq.c
>>>>> @@ -363,7 +363,7 @@ static struct request *blk_mq_get_request(struct
>>>>> request_queue *q,
>>>>>
>>>>> rq = blk_mq_rq_ctx_init(data, tag, op);
>>>>> if (!op_is_flush(op)) {
>>>>> - rq->elv.icq = NULL;
>>>>> + memset(&rq->elv, 0, sizeof(rq->elv));
>>>>> if (e && e->type->ops.mq.prepare_request) {
>>>>> if (e->type->icq_cache && rq_ioc(bio))
>>>>> blk_mq_sched_assign_ioc(rq, bio);
>>>>> @@ -461,7 +461,7 @@ void blk_mq_free_request(struct request *rq)
>>>>> e->type->ops.mq.finish_request(rq);
>>>>> if (rq->elv.icq) {
>>>>> put_io_context(rq->elv.icq->ioc);
>>>>> - rq->elv.icq = NULL;
>>>>> + memset(&rq->elv, 0, sizeof(rq->elv));
>>>>> }
>>>>> }
>>>>
>>>> This looks like a BFQ problem, this should not be necessary. Paolo,
>>>> you're calling your own prepare request handler from the insert
>>>> as well, and your prepare request does nothing if rq->elv.icq == NULL.
>>>
>>> I sent the patch anyway, since it's kind of a robustness improvement,
>>> I'd hope. If you fix BFQ also, please add:
>>
>> It's also a memset() in the hot path, would prefer to avoid that...
>> The issue here is really the convoluted bfq usage of insert/prepare,
>> I'm sure Paolo can take it from here.
>
Hi,
I'm very sorry for tuning in very late, but, at the same time, very
glad to find the problem probably already solved ;) (in this respect, I swear,
my delay was not intentional)
> Does this fix it?
>
> diff --git a/block/bfq-iosched.c b/block/bfq-iosched.c
> index f0ecd98509d8..d883469a1582 100644
> --- a/block/bfq-iosched.c
> +++ b/block/bfq-iosched.c
> @@ -4934,8 +4934,11 @@ static void bfq_prepare_request(struct request *rq, struct bio *bio)
> bool new_queue = false;
> bool bfqq_already_existing = false, split = false;
>
> - if (!rq->elv.icq)
> + if (!rq->elv.icq) {
> + rq->elv.priv[0] = rq->elv.priv[1] = NULL;
> return;
> + }
> +
This does solve the problem at hand. But it also arouses a question,
related to a possible subtle bug.
For BFQ, !rq->elv.icq basically means "this request is not for me, as
I am an icq-based scheduler". But, IIUC the main points in this
thread, then this assumption is false. If it is actually false, then
I hope that all requests with !rq->elv.icq that are sent to BFQ do
verify the condition (at_head || blk_rq_is_passthrough(rq)). In fact,
requests that do not verify that condition are those that BFQ must put
in a bfq_queue. So, even if this patch makes the crash disappear, we
drive BFQ completely crazy (and we may expect other strange failures)
if we send BFQ a request with !((at_head || blk_rq_is_passthrough(rq))
and !rq->elv.icq. BFQ has to put that rq into a bfq_queue, but simply
cannot.
Jens, or any other, could you please shed a light on this, and explain
how things are exactly?
Thanks,
Paolo
> bic = icq_to_bic(rq->elv.icq);
>
> spin_lock_irq(&bfqd->lock);
>
> --
> Jens Axboe
