New subject: Squeak 5 Image Segment Work [was [squeak-dev] [ANN] Squeak 5]

19 Aug 2015

      Hi All, but especially Tobias,
as we know, ImageSegments are broken in Spur. They work occasionally,
but when stressed the VM ends up crashing.  The fundamental issue is that
the ImageSegment code makes assumptions about object ordering that Spur
violates.  For example, here's ImageSegment's install:
install
"This operation retrieves the segment if necessary from file storage,
installs it in memory, and replaces (using become:) all the root stubs with
the reconstructed roots of the segment."
| newRoots |
state = #onFile ifTrue: [self readFromFile].
state = #onFileWithSymbols ifTrue: [self readFromFileWithSymbols.
endMarker := segment nextObject. "for enumeration of objects"
endMarker == 0 ifTrue: [endMarker := 'End' clone]].
(state = #active) | (state = #imported) ifFalse: [self errorWrongState].
newRoots := self loadSegmentFrom: segment outPointers: outPointers.
state = #imported
ifTrue: ["just came in from exported file"
arrayOfRoots := newRoots]
ifFalse: [
arrayOfRoots elementsForwardIdentityTo: newRoots].
state := #inactive.
Beeper beepPrimitive
So before the image segment bytes (the segment inst var) is loaded, the
object after it is assigned to endMarker, and if there isn't an object
after segment, a new object ('End' clone) is assigned to endMarker.
This makes the assumption that objects are allocated in a strict order, and
therefore endMarker will always be the object after segment.
Loading the segment via "newRoots := self loadSegmentFrom: segment
outPointers: outPointers" then turns segment into a zero-length WordArray
and its contents into the objects loaded by the segment.  Therefore, in the
V3 system, the objects loaded from segment can be enumerated starting at
segment nextObject and repeating until endMarker is found:
allObjectsDo: aBlock
"Enumerate all objects that came from this segment.  NOTE this assumes that
the segment was created (and extracted).  After the segment has been
installed (install), this method allows you to enumerate its objects."
| obj |
endMarker == nil ifTrue: [
^ self error: 'Just extract and install, don''t writeToFile:'].
segment size ~= 1 ifTrue: [
^ self error: 'Vestigial segment size must be 1 (version word)'].
obj := segment nextObject.  "Start with the next object after the vestigial
header"
[obj == endMarker] whileFalse:  "Stop at the next object after the full
segment"
[aBlock value: obj.
obj := obj nextObject].  "Step through the objects installed from the
segment."
Now, as written, this just won't work in Spur.
a) the only place where there is any kind of stable order to objects is in
oldSpace, so segment /has/ to be forced to old space to have any chance of
its objects being in order when it gets converted from bytes to objects.
b) 'End' clone will be in newSpace and so endMarker isn't reliable unless
it was obtained by segment nextObject when segment was already in oldSpace.
So it is perhaps possible to fix ImageSegments in Spur by forcing segment
to oldSpace and being more careful with endMarker.  But I think there is a
better way.
If the set of objects the segment contains can be obtained some how then
this set can be simply enumerated, not depending on nextObject.  The
primitive has to answer the array of roots, so its result can't be changed
to be the entire array.  But segment could be becomed into an Array of all
the objects in segment prior to it being loaded, in which case the above
would become
install
"This operation retrieves the segment if necessary from file storage,
installs it in memory, and replaces (using become:) all the root stubs with
the reconstructed roots of the segment."
| newRoots |
state = #onFile ifTrue: [self readFromFile].
state = #onFileWithSymbols ifTrue:
[self readFromFileWithSymbols].
(state = #active) | (state = #imported) ifFalse: [self errorWrongState].
newRoots := self loadSegmentFrom: segment outPointers: outPointers.
state = #imported
ifTrue: "just came in from exported file"
[arrayOfRoots := newRoots]
ifFalse:
[arrayOfRoots elementsForwardIdentityTo: newRoots].
state := #inactive.
Beeper beepPrimitive
allObjectsDo: aBlock
"Enumerate all objects that came from this segment.  NOTE this assumes that
the segment was created (and extracted).  After the segment has been
installed (install), this method allows you to enumerate its objects."
| obj |
segment isArray ifFalse:
[^ self error: 'Segment hasn''t been loaded?'].
segment do: aBlock
and the endMarker instVar would be deleted.
I am willing and ready to modify the primitive to convert the segment
correctly.  Who will volunteer to rewrite the image-level ImageSegment code
to use the new primitive?
On Wed, Aug 12, 2015 at 10:40 PM, Eliot Miranda eliot.miranda@gmail.com
wrote:
...
Hi Tobias,
On Wed, Aug 12, 2015 at 10:18 PM, Tobias Pape Das.Linux@gmx.de wrote:
...
Hi all
On 13.08.2015, at 02:15, Eliot Miranda eliot.miranda@gmail.com wrote:
...
Hi Tobias,
forget that. I found it.  THis right?
Trunk test suite for Spur
Using existing cogspur r.3410
cp -r /var/lib/jenkins/workspace/Trunk/default/target/cogspur.r3410
/tmp/d20150812-28620-etbikj
...
image test suite
VM: /tmp/d20150812-28620-etbikj/cogspur.r3410/cogspurlinuxht/bin/squeak
/tmp/d20150812-28620-etbikj/cogspur.r3410/cogspurlinuxht/bin/squeak
-version
...
/var/lib/jenkins/workspace/Trunk/default/tests.st
spawning command 0 with timeout 1800 seconds:
"/tmp/d20150812-28620-etbikj/cogspur.r3410/cogspurlinuxht/bin/squeak"
"-vm-sound-null" "-vm-display-null"
"/var/lib/jenkins/workspace/Trunk/default/target/SpurPostTestTrunkImage.image"
"../tests.st"
...
(Command started with PID 28643)
2015-08-12T23:31:48.17+01:00: Loading Hudson build tools... from
/var/lib/jenkins/workspace/Trunk/default/target/HudsonBuildTools.st
...
2015-08-12T23:31:48.388+01:00: Running tests...
setsockopt: Protocol not available
setsockopt: Protocol not available
28646:error:140770FC:SSL routines:SSL23_GET_SERVER_HELLO:unknown
protocol:s23_clnt.c:612:
...
Recursive not understood error encountered
I bet this is ImageSegment related.
It sure is.
    BitmapStreamTests>testMatrixTransform2x3WithImageSegment

sends
        BitmapStreamTests>validateImageSegment
Basically the ImageSegment code has been working in Spur on a hope and a
prayer.  The code assumes objects are allocated in chronological order and
that this order is preserved, not so with Spur.  So post image segment
loading someObject/nextObject is used to enumerate the objects loaded.
This can't work reliably in Spur.  I *think* (ok I hope) that I've
implemented the segment load primitive in Spur to answer an Array of the
objects loaded, so that these can be explicitly enumerated.
So the job is a) to check that I have indeed implemented the primitive to
do this and b) to rewrite the image segment loading code in the light of
this.
David, this is an example of something that isn't back portable and should
not be back-ported.
The most strange thing is that I cannot reproduce this on my Mac…
...
There the test does not crash the image…
Well then it may be a signed/unsigned bug in image loading instead.  On
linux the image typically gets loaded quite high in the address space and
so a good portion of the heap ends up above 0x7FFFFFFF, or negative
territory if misinterpreted as signed integers.  On Mac and Windows the
image tends to get loaded quite low and so has to be pretty big to fall
foul of signedness issues.
Busy right now but will check tomorrow.
...
Best regards
        -Tobias
...
On Wed, Aug 12, 2015 at 5:11 PM, Eliot Miranda eliot.miranda@gmail.com
wrote:
...
Hi Tobias,
On Wed, Aug 12, 2015 at 12:49 PM, Tobias Pape Das.Linux@gmx.de wrote:
On 12.08.2015, at 20:55, Eliot Miranda eliot.miranda@gmail.com wrote:
...
Hi All,
Fabio's kindly done most of the changes.  But some questions

remain for more general discussion.  See below.
...
...
Fabio,  thanks so much for doing this, and so quickly!
On Tue, Aug 11, 2015 at 8:51 PM, Eliot Miranda <
eliot.miranda@gmail.com> wrote:
...
...
Hi All,
who will update http://squeak.org/downloads/ to include the 5.0

release?
...
...
a) I suggest that the 5.0 all-in-one have a line in the left-hand
table that includes the 4.6 release and that it precede the 4.6 release in
the list.
...
...
Done.
b) the Trunk link points to TrunkImage.zip which contains a non-Spur
image.
...
...
So this is the biggie.  What should we do?

add a Trunk 5.0 build with a different link?  (Noooooo)

change  http://build.squeak.org/job/SqueakTrunk to build what it

says (Yessss please, who can do this?)
...
Done: build.squeak.org/job/Trunk/
But as I said several times, Spur/trunk test just crash for month…
can you point me to the crash?  I'm looking at
http://build.squeak.org/job/Trunk/default/lastBuild/#showFailuresLink
and see 8 test failures but no crash.
...
...

add a Squeak 4.6 build job?  Is that even worth it any more

considering 4.6 is released and stable?  If it is then what should it
build?  David?
...
...
c) Spur VMs need to be linked to by an added line in the Virtual
Machines list.
...
...
Not applicable.  the VM links point to the root of my site so they
effectively point to both old and Spur VMs.
...
...
d) Spur imagers need to be included in the Image and Changes list
under Custom Installation
...
...
Done.
e) The SqueakV50.sources file also needs to be included in the
Current Sources list.
...
...
Done.
f) we could start a History list for the 5.0 line
This is probably quite a big reorg on the page and not urgent.
So there are a few things to fix before 5.0 is freely downloadable.
On Tue, Aug 11, 2015 at 8:23 PM, Chris Muller ma.chris.m@gmail.com
wrote:
...
...
In the 17 months since Squeak 4.5 was released, a huge development
effort took place to create the next generation virtual-machine for
the Squeak / Pharo / Newspeak family of programming systems.  Squeak
is the modern incarnation of the Smalltalk-80 programming environment
originally developed at the Xerox PARC.
"Squeak 5" introduces this new VM and associated new memory model,
collectively referred to as "Spur".  Presented [1] by Eliot Miranda
and Clément Béra at the 2015 International Symposium on Memory
Management, this new VM affords Squeak applications a significant
boost in performance and memory management.  Among other
optimizations, the #become operation no longer requires a memory scan.
Object pinning and ephemerons are also now supported.  The release
notes [2] provide more details.
The new memory model requires a new image file format.  Although this
new format results in about a 15% increased memory requirement for the
same number of 4.x objects, a new segmented heap allows memory to be
given back to the OS when its no longer needed, a great benefit for
application servers.
As forward compatibility is as important to the Squeak community as
backward compatibility, Squeak 5 is delivers an image with identical
content as the recent 4.6 release.  Although this new Squeak 5 VM
cannot open images saved under the prior 4.x Cog format, objects and
code can be easily exported from the 4.x image and then imported into
Squeak 5.  Applications whose code runs strictly above the Smalltalk
meta layer will prove remarkably compatible with the new format, most
applications will require no changes whatsotever.
Squeak 5 is the result of monumental effort by a tiny group of very
talented people, but its also just the beginning of yet a new effort;
Spur is just a stepping stone to a more ambitious goals planned over
the next five years.
[1] -- A Partial Read Barrier for Efficient Support of Live
Object-oriented Programming
http://conf.researchr.org/event/ismm-2015/ismm-2015-papers-a-partial-read-ba...
...
...
[2] -- Squeak 5 Release Notes
http://wiki.squeak.org/squeak/6207
--
_,,,^..^,,,_
best, Eliot
-- 
_,,,^..^,,,_
best, Eliot