RFR: 8339538: Wrong timeout computations in DnsClient [v2]

[Daniel Jeliński]

On Mon, 9 Sep 2024 11:42:25 GMT, Aleksei Efimov <aefimov at openjdk.org> wrote:

>> This PR proposes the following changes to address wrong timeout computations in the `com.sun.jndi.dns.DnsClient`:
>> - The `DnsClient` has been updated to use a monotonic high-resolution (nano) clock. The existing `Timeout` test has also been updated to use the nano clock to measure observed timeout value.
>> 
>> - The left timeout computation has been fixed to decrease the timeout value during each retry attempt. A new test, `TimeoutWithEmptyDatagrams`, has been added to test it.
>> 
>> - The `DnsClient.blockingReceive` has been updated:
>>     - to detect if any data is received
>>     - to avoid contention with `Selector.close()` that could be called by a cleaner thread
>> 
>> - The expected timeout calculation in the `Timeout` test has been updated to take into account the minimum retry timeout (50ms). Additionally, the max allowed difference between the observed timeout and the expected one has been increased from 50% to 67%. Taking into account 50 ms retry timeout decrease the maximum allowed difference is effectively set to 61%. This change is expected to improve the stability of the `Timeout` test which has been seen to fail [intermittentlly](https://bugs.openjdk.org/browse/JDK-8220213). If no objections, I'm planning to close [JDK-8220213](https://bugs.openjdk.org/browse/JDK-8220213) as duplicate of this one.
>> 
>> JNDI/DNS jtreg tests has been executed multiple times (500+) to check if the new and the modified tests are stable. No failures been observed (so far?).
>
> Aleksei Efimov has updated the pull request incrementally with two additional commits since the last revision:
> 
>  - guard against possible integer value overflows
>  - make startTime a local variable

src/jdk.naming.dns/share/classes/com/sun/jndi/dns/DnsClient.java line 477:

> 475:                     long elapsedMillis = TimeUnit.NANOSECONDS.toMillis(end - start);
> 476:                     // Setting the Math.clamp min to 1 ensures that the timeout is decreased
> 477:                     timeoutLeft = timeoutLeft - Math.clamp(elapsedMillis, 1, Integer.MAX_VALUE);

I think I'd prefer to calculate the remaining timeout based on the total elapsed time in this loop, as opposed to the time spent in blockingReceive.

-------------

PR Review Comment: https://git.openjdk.org/jdk/pull/20892#discussion_r1750323475