RFR: 8015666: test/tools/pack200/TimeStamp.java failing

[Xueming Shen]

On 6/21/13 6:33 AM, Alan Bateman wrote:
> On 20/06/2013 19:48, Xueming Shen wrote:
>> :
>>
>> I'm proposing the following approach to add the functionality of 
>> supporting
>> the "utc-date/time-with-1-second granularity" and keep the old behavior
>> of the get/setTime() of the ZipEntry.
>>
>> (1) keep the time/setTime()/getTime() for the MS-DOS standard date/time.
>>      To set via the old setTime() will only store the time into zip's 
>> standard
>>      date/time field, which is in MS-DOS date/time. And getTime() 
>> only returns
>>      the date/time from that field, when read from the zip file/stream.
>> (2) add mtime/set/getLastModifiedTime() to work on the UTC time fields,
>>      and the last modified time set via the new method will also set 
>> the "time",
>>      and the getLastModifiedTime() also returns the "time", if the 
>> UTC time
>>      stamp fields are not set in the zip file header. The idea is 
>> that for the new
>>      application, the recommendation is to use 
>> ZipEntry.set/getLastModifiedTime()
>>      for better/correct time stamp, but the existing apps keep the 
>> same behavior.
>> (3) jar and ZipOutputStream are updated to use the 
>> set/getLastModifiedTime().
>> (4) Pack/unpack continues to use the set/getTime(), so the current 
>> workaround
>>      continues work. I will leave this to Kuma to decide how it 
>> should be handled
>>      going forward. (there are two facts need to be considered here, 
>> a) the
>>      existing jar file might not have the utc time instored, and b) 
>> all "extra" data
>>      are wiped out during the pack/unpacking process)
>> (5) additionally add another pair of atime/get/setLastAccessTime and
>>      ctime/get/setCreationTime().
>> (6) The newly added 3 pairs of the m/a/ctime get/set methods use the 
>> "new"
>>      nio FileTime, instead of the "long". This may add some 
>> additional cost of
>>      conversion when working with them, but may also help improve the
>>      performance if the time stamps are directly from nio  file 
>> system when
>>      get/set XYZTime. Good/bad?
>>
>> http://cr.openjdk.java.net/~sherman/8015666/webrev/
>>
>> Comment, option and suggestion are appreciated.
> At a high-level this looks a reasonable approach. An alternative would 
> be to define an API over the extra block but I don't think we want to 
> go there.
>
> For consistency you might consider setTime setting mtime to null, 
> otherwise the ordering when both setXXX methods are used is 
> significant. Another thing to consider is whether the "parsing" of the 
> extra block could be changed to being lazy now as it's only 
> interesting when someone asks for it or the higher precision time 
> stamps. Thinking back to the original changes then I wonder if we 
> discussed the UNIX extra field, maybe you said that Windows tools 
> can't handle such zip files?
>
> A possible concern is that the footprint of ZipEntry increases. We 
> could eliminate this increase if the new methods operated on the extra 
> block. There's a clearly a trade-off here.

I do have a version that set/getXYZTime() on extra data...but given the 
expectation is
the new methods will be recommended in new code shift away from the dos 
time, so
the extra will take some footprint of ZE as well...but yes, it may have 
the benefit of
saving at least the a/ctime, which normally people don't care. Another 
footprint concern
is the minimum 18 bytes increase (for mtime alone, in 9 for loc and 9 
for cen) for each
file in jar file, in case of rt.jar, it is relatively a big increase. A 
possible solution is to
have explicit API for ZOS and explicit command option for jar command to 
"turn" this
one explicitly by developer/user, instead of the default. Or, as an 
alternative to have
some mechanism to turn them off, if the size increase is not desired.

>
> I spotted one or two "modificatin" in the javadoc, but I'll make time 
> to do a proof-read the javadoc once the approach is agreed.
>
> -Alan.
>
>
>
>