RFR: 8231640: (prop) Canonical property storage [v7]

[Roger Riggs]

On Thu, 9 Sep 2021 09:56:30 GMT, Jaikiran Pai <jpai at openjdk.org> wrote:

>> The commit in this PR implements the proposal for enhancement that was discussed in the core-libs-dev mailing list recently[1], for https://bugs.openjdk.java.net/browse/JDK-8231640
>> 
>> At a high level - the `store()` APIs in `Properties` have been modified to now look for the `SOURCE_DATE_EPOCH` environment variable[2]. If that env variable is set, then instead of writing out the current date time as a date comment, the `store()` APIs instead will use the value set for this env variable to parse it to a `Date` and write out the string form of such a date. The implementation here uses the `d MMM yyyy HH:mm:ss 'GMT'` date format and `Locale.ROOT` to format and write out such a date. This should provide reproducibility whenever the `SOURCE_DATE_EPOCH` is set. Furthermore, intentionally, no changes in the date format of the "current date" have been done.
>> 
>> These  modified `store()` APIs work in the presence of the `SecurityManager` too. The caller is expected to have a read permission on the `SOURCE_DATE_EPOCH` environment variable. If the caller doesn't have that permission, then the implementation of these `store()` APIs will write out the "current date" and will ignore any value that has been set for the `SOURCE_DATE_EPOCH` env variable. This should allow for backward compatibility of existing applications, where, when they run under a `SecurityManager` and perhaps with an existing restrictive policy file, the presence of `SOURCE_DATE_EPOCH` shouldn't impact their calls to the `store()` APIs.
>> 
>> The modified `store()` APIs will also ignore any value for `SOURCE_DATE_EPOCH` that  cannot be parsed to an `long` value. In such cases, the `store()` APIs will write out the "current date" and ignore the value set for this environment variable. No exceptions will be thrown for such invalid values. This is an additional backward compatibility precaution to prevent any rogue value for `SOURCE_DATE_EPOCH` from breaking applications.
>> 
>> An additional change in the implementation of these `store()` APIs and unrelated to the date comment, is that these APIs will now write out the property keys in a deterministic order. The keys will be written out in the natural ordering as specified by `java.lang.String#compareTo()` API.
>> 
>> The combination of the ordering of the property keys when written out and the usage of `SOURCE_DATE_EPOCH` environment value to determine the date comment should together allow for reproducibility of the output generated by these `store()` APIs.
>> 
>> New jtreg test classes have been introduced to verify these changes. The primary focus of `PropertiesStoreTest` is the ordering aspects of the property keys that are written out. On the other hand `StoreReproducibilityTest` focuses on the reproducibility of the output generated by these APIs.  The `StoreReproducibilityTest` runs these tests both in the presence and absence of `SecurityManager`. Plus, in the presence of SecurityManager, it tests both the scenarios where the caller is granted the requisite permission and in other case not granted that permission.
>> 
>> These new tests and existing tests under `test/jdk/java/util/Properties/` pass with these changes.
>> 
>> [1] https://mail.openjdk.java.net/pipermail/core-libs-dev/2021-August/080758.html
>> [2] https://reproducible-builds.org/specs/source-date-epoch/
>
> Jaikiran Pai has updated the pull request incrementally with one additional commit since the last revision:
> 
>   dummy commit to trigger GitHub actions job to try and reproduce an unexplained failure with the new tests that happened around 24 hours back, this time yesterday (rule out any time/date/timezone specific issues)

Overall looks good, though a system property would be more in keeping with existing JDK pattern. See comments.

src/java.base/share/classes/java/util/Properties.java line 187:

> 185:                                         ? System.getenv("SOURCE_DATE_EPOCH")
> 186:                                         : AccessController.doPrivileged((PrivilegedAction<String>)
> 187:                                                 () -> System.getenv("SOURCE_DATE_EPOCH"));

The format of SOURCE_DATE_EPOCH is fine, but referring to an external document as part of the OpenJDK spec is undesirable in the long run. Previous comments gave support to using the environment variable directly but
it very unlike similar cases that use system properties when the behavior of an API needs to be overridden or modified.  

As a system property, specified on the command line, it would be visible when the program is invoked,
explicitly intended to change behavior and not having a context sensitive effect.  Named perhaps, java.util.Properties.storeDate.

It can be documented as part of the spec with the javadoc tag "{@systemProperty  ...}"
There is a cache of properties in jdk.internal.util.StaticProperty that captures the state when the runtime is initialized. For specific properties, they are cached and available without regard to the setting of the SecurityManager or not.  (BTW, there is no need to special case doPriv calls, they are pretty well optimized when the security manager is not set and it keeps the code simpler.)

Given the low frequency of calls to store(), even caching the parsed date doesn't save much.
And the cost is the cost of the size and loading an extra class.

src/java.base/share/classes/java/util/Properties.java line 887:

> 885:      * If the value set for {@code SOURCE_DATE_EPOCH} cannot be parsed to a {@code long},
> 886:      * then the current date and time will be written.
> 887:      *

Update to refer to the property and add the {@systemProperty ...} javadoc tag so the property is listed in the system property reference.

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

Changes requested by rriggs (Reviewer).

PR: https://git.openjdk.java.net/jdk/pull/5372