On 06/27/2013 10:04 AM, Kumar Srinivasan wrote:
Hi Sherman,
I started looking at this, my initial comment, the Unpacker.unpack
does not close its output and we allow multiple pack files to be concatenated,
I am assuming out.finish() will allow further jar files to be appended ?
or would this cause a problem ?
No, out.finish() will not allow further entry appending. Then, it appears
we need to have a different approach to "finish" the Jar/ZipOutputStream.
What need to be done here is that either out.close/finish() need to be
invoked under the UTC locale as well (to output the time stamps in cen
table correctly). This is another "incompatible" change of the previous
change, in which the msdosTime<->javaTime conversion no longer
occurs during the ZipEntry.set/getTime(), but during the read and write
the ZipEntry from/to the zip file.
-Sherman
Kumar
Hi,
The zip time related changes[1] I pushed back last month appears
to have the compatibility risk of breaking existing code. The main
idea in that changeset is to use the more accurate and timezone
insensitive utc time stored in the extra field for the ZipEntry.set/getTime()
if possible. However it turns out the reality is that the code out there
might have already had some interesting workaround/hack solution
to workaround the problem that the time stamp stored in the "standard'
zip entry header is a MS-DOS standard date/time, which is a "local
date/time" and sensitive to timezone, in which, if the zip is archived
in time zone A (our implementation converts the "java" time to dos
time by using the default tz A) and then transferred/un-archived in
a different zone B (use default tz B to convert back to java time), we
have a time stamp mess up. The "workaround" from pack200 for this
issue when pack/unpacking a jar file is to "specify/recommend/suggest"
in its spec that the "time zone" in a jar file entry is assumed to be "UTC",
so the pack/unpack200 implementation set the "default time" to utc
before the pack/unpack and set it back to the original after that. It worked
"perfectly" for a roundtrip pack/unpacking, until the changeset [2], in
which ZipEntry.getTime() (packing) returns a real utc time and the following
ZipEntry.setTime() (unpacking), then mess up the MS-DOS date/time
entry, this is the root cause of this regression.
Given the facts that
(1) there are actually two real physical time stamps in a zip file header,
one is in the date/time fields, which is MS-DOS-local-date/time-with-2-
seconds-granularity , one is in the extra data field, which is UTC-1-second
-granularity
(2) and there are applications over there that have dependency on the
MS-DOS date/time stamp.
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.
-Sherman
[1] http://hg.openjdk.java.net/jdk8/tl/jdk/rev/90df6756406f