Re: [MERGE] Send DNS A and AAAA queries in parallel (v2)

On 28.02.2012 14:26, Alex Rousskov wrote:
> On 01/16/2012 10:32 AM, Henrik Nordström wrote:
>
>> This implements sending DNS A & AAAA queries in parallel by
>> creating "slave" idns_query requests. Current implementation
>> uses the A lookup as "master" and AAAA as "slave" query.
>
> As implemented, the half-baked slave query cannot exist without the
> master because it does not have callback information (at least).
> However, the error and timeout handling code still thinks that each
> query is fully capable of standing on its own:
>

The disable-A lookup directive (dns_v4_fallback) was removed. There is
always an A lookup now in the pair now.

Over the last few days I've been thinking it might be worth moving to a
model where we have a master object without specific RR type. Just a
data object to hold the query template, callback and statistics info.
And a pointer to a linked list of actual RR lookups being performed.
That way we are not locked down to a specific order of RR lookups like
now, and can extend it trivially to other types like PTR and SRV.

Ideally long-term the dns caches would also take the controller object
and store it as the cached entry. No mapping to an array of addresses.
But that is just speed polishing.

>
>> if (q->rcode != 0)
>> idnsCallback(q, NULL, -q->rcode,
>> rfc1035ErrorMessage(q->rcode)!
>> else
>> idnsCallback(q, NULL, -16, "Timeout");
>>
>> cbdataFree(q);
>
> Squid dumps core in idnsCallback() when the slave query times out
> because q->callback is nil and the callback_data validity check is
> useless.
>

Is this actually happening? in what query events?

> There may be other similar leftovers from the good old days where we
> did
> not have half-baked queries; I have not checked.
>
>
> One solution would be to change the slave query class type so that it
> does not have the currently unset fields (such as callback). This
> would
> expose bugs like the one above at compile time.
>

Long term yes I think we should. see above.

Short-term a dummy handler can scream "bug" in the log or something.

Amos
Received on Tue Feb 28 2012 - 02:45:53 MST