I don't see any reasonable way to avoid this "problem".
Basically, the problem boils down to this:
when an HS version is forked off for stabilization into a Update
release, there exists a set of bugs (to be found) that exist in that
forked HS codebase (let's call that HSa). Naturally, those bugs exist in
the ongoing 7u-dev HS codebase as well (let's call that HSb).
Now, when one of those bugs is found and fixed in HSb as important for
the stabilization release, there's a bunch of possibilities:
(1) immediately sync the HSa repository into HSb
(2) delay the HSa -> HSb sync until HSa is completed for release
(3) transplant the fix from HSa to HSb (resulting in the same
BugID, but a different ChangesetID)
(4) create a whole new BugID for the fix, and push that changeset
into HSb
(5) recognize that current (mainline) work is making the fix
irrelevant, and thus, we don't want it at all in HSb
Option 1 and 2 produce problems with the situation put forth in Option
5, as a full repo sync from HSa -> HSb may introduce fixes that are no
longer desirable in the mainline development repo, so, frankly, they're
not something that we should do.
Option 3 combined with a general HSa -> HSb sync at the end of the HSa
stabilization release breaks a rule we've set up for ourselves, in that
repositories should NOT multiple changesets with the same BugID. This of
course is merely a guideline, and we can break it at the community's desire.
Option 4 is a bit messy, since it requires that people recognize that
two different BugIDs actually address the exact same code flaw. It gets
messier, since when doing a HSa -> HSb sync at the end of a HSa
stabilization period, whomever is doing the sync has to actually look at
the code for each potential BugID fix being sync'd back into HSb, to see
if that BugID is already fixed in the HSa source base. Normally, if
this isn't done, it's no big deal, since it merely means that there are
two changesets that fix the exact same code portion. However, in the
case where we recognized that we actually DON'T want to fix the Bug in
the HSb mainline code, it requires a bit of extra vigilance to make sure
it doesn't slip back in from the HSa release.
There were very significant issues around merging back changes from 7u2
into 7u-dev Hostpot, mostly due to the unusual way 7u1 Hotspot was
managed. In the case of 7u4, the decision to push in a hs23 version
from mailine, rather than use the hs22 version from 7u3, caused
additional grief.
Honestly, the only way to avoid these issue is to do the following:
(a) after a fork of Hotspot for a stabilization release, declare that
there will NEVER be an actual repository sync from that fork back into
the mainline development tree, OR vice versa. This is hard to do, as
circumstances always interfere with the best intentions.
(b) require all developers who fix a bug in the stabilization release to
separately push into the mainline repository; however the developer much
check that the bugfix is still applicable to the mainline repo (i.e.
check that other development hasn't obsoleted the fix, or that the fix
doesn't break something else).
In theory the above should be simple, but like everything else in life,
the devil is in the details. It creates a whole bunch of extra work (vs
bulk, end-of-stabilization re-integrations) for both developers and QA
people, and also requires a bit of crystal-ball-gazing around the
applicability of a fix, whereas bulk re-integrations are a bit more
brutal in application (i.e. in a bulk integration, the general rule is:
if its still broke in the mainline, re-integrate the fix from the
stabilization release even if there might be a future redesign
obsoleting the fix).
There is no silver bullet.
:-(
-Erik
On 4/27/2012 3:23 PM, David Holmes wrote:
Hi Volker,
On 27/04/2012 11:15 PM, Volker Simonis wrote:
Hi David,
thank you very much for the detailed explanation and the links.
I missed that discussion (there are just too many mailing lists to
monitor) and now it's obviously too late to complain:)
But as far as I understood, Eriks mail "Community Input on Hotspot 22
integration into 7u"
(http://mail.openjdk.java.net/pipermail/jdk7u-dev/2011-August/000161.html)
was about problems integrating hs22 into 7u.
Did these same problems also occur for hs23 as well? I assume yes and
if so your explanation would sound perfectly clear to me.
I assume that the same "merge" based model has been followed since
that point.
On the other hand I saw no definitive answer to Edvard's mail "
Proposal on how we should handle syncing between stabilisation forests
and always open 7u"
(http://mail.openjdk.java.net/pipermail/jdk7u-dev/2011-August/000161.html).
Does this mean that we will constantly see these kind of problems
(i.e. duplicate changesets with the same bug ID or even worse
duplicate changesets with different bug IDs) in the future?
It will occur as long as we are dealing with hsN-1 and hsN and we
merge the way we are.
Is there any "official" document available which explains the syncing
between stabilisation forests and 7u?
I'm not aware of anything. I'm also not sure this is specifically an
issue with stabilization forests - the 7u-dev forest will have the
same issue.
David
-----
Thank you and best regards,
Volker
On Fri, Apr 27, 2012 at 12:45 PM, David
Holmes<david.hol...@oracle.com> wrote:
Hi Volker,
On 27/04/2012 6:59 PM, Volker Simonis wrote:
could anybody please explain why it was necessary to reintegrate a
whole lot of changes into hs23 of jdk7u/hotspot which have been in the
hotspot repository since hs22 anyway?
This was discussed last year (with thanks to John Rose for these
references):
http://mail.openjdk.java.net/pipermail/jdk7u-dev/2011-August/000161.html
http://mail.openjdk.java.net/pipermail/jdk7u-dev/2011-October/000582.html
http://mail.openjdk.java.net/pipermail/jdk7u-dev/2011-November/000748.html
Basically what you are seeing are fixes that went into hs22 and hs23
after
hs23 forked. hs22 went into 7u2 which became the basis for 7u3 then
7u4.
Then hs23 had to be integrated into 7u4 and so you end up with
different
changesets that duplicate the same actual changes.
David
-----
This all happened with the following bulk integrations:
http://mail.openjdk.java.net/pipermail/jdk7u-dev/2012-April/002588.html
http://mail.openjdk.java.net/pipermail/jdk7u-dev/2012-April/002651.html
and it affects (among others):
7059019: G1: add G1 support to the SA
7092412: G1: Some roots not marked during an initial mark that gets an
evacuation failure
7045232: G1: pool names are inconsistent with other collectors (don't
have 'Space')
7068215: G1: Print reference processing time during remark
7091032: G1: assert failure when NewRatio is used
7092245: G1: Wrong format specifier in G1PrintRegionLivenessInfo
header
output
7092238: G1: Uninitialized field gc_efficiency in
G1PrintRegionLivenessInfo output
6484982: G1: process references during evacuation pauses
7075646: G1: fix inconsistencies in the monitoring data
7086533: G1: assert(!_g1->is_obj_dead(obj)): We should not be
preserving dead objs: g1CollectedHeap.cpp:3835
7097053: G1: assert(da ? referent->is_oop() :
referent->is_oop_or_null()) failed: referenceProcessor.cpp:1054
7097048: G1: extend the G1 SA changes to print per-heap space
information
7092236: java/util/EnumSet/EnumSetBash.java fails
7092278: "jmap -finalizerinfo" throws
"sun.jvm.hotspot.utilities.AssertionFailure: invalid cp index 0 137"
7096366: PPC: corruption of floating-point values with DeoptimizeALot
7100165: JSR 292: leftover printing code in methodHandleWalk.cpp
Looking for each of this changes will show you that they are present
two times in the repository with exactly the same changesets. E.g.:
http://hg.openjdk.java.net/jdk7u/jdk7u4/hotspot/log?rev=7059019
Regards,
Volker
