Re: Does anyone recognize this glitch in our trunk update processing? (was: [Vm-dev] buildspurtrunkvmmakerimage.sh fails on Mac)
I've definitely experienced the exact same problem, and got around it the same way, restarting the method and proceeding. Very strange.
On Sun, Aug 21, 2016 at 3:13 PM, David T. Lewis lewis@mail.msen.com wrote:
Moving from vm-dev to squeak-dev because it is a trunk update stream question. Follow ups to squeak-dev please.
Does anyone recognize this problem?
We have an issue in trunk update processing that prevents a full update without manual intervention. This does not affect the 5.1 release, but it would be good to fix it so that the update from 5.0 to 5.1 can be reproduced in a continuous integration test.
To reproduce: Start with Squeak5.0-15113.image, set the update URL preference to http://source.squeak.org/trunk, and do an "update from server".
The error first appears when beginning to process the update-mt.378.mcm update map. The problem is related to updating the progress bar with ProgressInitiationException.
When the failure occurs, the debugger (see attached image) seems to show an array access being interpreted as if the array was an instance of StrikeFont. Per Nicolas' note below, it may be related to showing the progress bar while updating it.
I can restart the update process from the debugger from SystemProgressMorph>>position:label:min:max: in the stack frame, and the update proceeds. The error happens several more times, then at some point it seems to go away, so apparently the problem was fixed at some later point in the update stream.
Does anyone recognize this problem? Can we work around it by changing one or more of the update maps?
Thanks, Dave
On Fri, Aug 19, 2016 at 10:30:30PM +0200, Nicolas Cellier wrote:
yes, this is a problem I encountered both in 32 and 64 bits squeak while updating.
I think it's related to showing the progress bar while updating the progress bar. It's a Squeak trunk update problem, not a VM problem.
If some class layout change, but some block is still active (yeah, did you see how many blocks we traverse just to change display of a bar...), then the old block has invalid inst var offsets (offsets to the old object that now has become a new object with new layout).
In an interactive image, just close the debugger and relaunch the update. In an headless automated process, well... We should have fixed the update or start from a newer artifact...
2016-08-19 21:32 GMT+02:00 tim Rowledge tim@rowledge.org:
I just tried running the vmmaker-build script on my iMac with weirdly bad results.
The vm it downloads is the CogSpur.app-16.18.3692 version. Whilst trying to do the update part of the process I had several *very* odd failures where an object in a block was mistaken for nil during a message send. For example in a progress bar update (which I can???t provide a log for since the log got overwritten later) a line (bars at: index) lead to a nil is not indexable halt. Yet in inspecting the prior stackframe the bars object was a perfectly normal array of progress bars. A similar but not identical problem happened during another attempt.
Attempting to filein the BuildSqueakSpurTrunkVMMaker.st file failed when the vm exploded with no visible log.
Now I???m attempting to repeat that with a new 5.1rc1 vm & already up to date 5.1rc1 image. Which has sorta-worked, in a rather odd manner. It got to the final save and quit but when I start the supposedly prepared image there is a problem with it trying to filein that file. I can???t tell right now if it relates to the manual filein I started or if something got added as a deferred action? Or maybe it???s a side-effect of the filein having a save & quit? Not come across this before.
So far as I can tell it did complete the filein, so maybe all is well.
Same here. I think the code changes during the update and the old code on the stack causes the problem.
Levente
On Sun, 21 Aug 2016, Chris Muller wrote:
I've definitely experienced the exact same problem, and got around it the same way, restarting the method and proceeding. Very strange.
On Sun, Aug 21, 2016 at 3:13 PM, David T. Lewis lewis@mail.msen.com wrote:
Moving from vm-dev to squeak-dev because it is a trunk update stream question. Follow ups to squeak-dev please.
Does anyone recognize this problem?
We have an issue in trunk update processing that prevents a full update without manual intervention. This does not affect the 5.1 release, but it would be good to fix it so that the update from 5.0 to 5.1 can be reproduced in a continuous integration test.
To reproduce: Start with Squeak5.0-15113.image, set the update URL preference to http://source.squeak.org/trunk, and do an "update from server".
The error first appears when beginning to process the update-mt.378.mcm update map. The problem is related to updating the progress bar with ProgressInitiationException.
When the failure occurs, the debugger (see attached image) seems to show an array access being interpreted as if the array was an instance of StrikeFont. Per Nicolas' note below, it may be related to showing the progress bar while updating it.
I can restart the update process from the debugger from SystemProgressMorph>>position:label:min:max: in the stack frame, and the update proceeds. The error happens several more times, then at some point it seems to go away, so apparently the problem was fixed at some later point in the update stream.
Does anyone recognize this problem? Can we work around it by changing one or more of the update maps?
Thanks, Dave
On Fri, Aug 19, 2016 at 10:30:30PM +0200, Nicolas Cellier wrote:
yes, this is a problem I encountered both in 32 and 64 bits squeak while updating.
I think it's related to showing the progress bar while updating the progress bar. It's a Squeak trunk update problem, not a VM problem.
If some class layout change, but some block is still active (yeah, did you see how many blocks we traverse just to change display of a bar...), then the old block has invalid inst var offsets (offsets to the old object that now has become a new object with new layout).
In an interactive image, just close the debugger and relaunch the update. In an headless automated process, well... We should have fixed the update or start from a newer artifact...
2016-08-19 21:32 GMT+02:00 tim Rowledge tim@rowledge.org:
I just tried running the vmmaker-build script on my iMac with weirdly bad results.
The vm it downloads is the CogSpur.app-16.18.3692 version. Whilst trying to do the update part of the process I had several *very* odd failures where an object in a block was mistaken for nil during a message send. For example in a progress bar update (which I can???t provide a log for since the log got overwritten later) a line (bars at: index) lead to a nil is not indexable halt. Yet in inspecting the prior stackframe the bars object was a perfectly normal array of progress bars. A similar but not identical problem happened during another attempt.
Attempting to filein the BuildSqueakSpurTrunkVMMaker.st file failed when the vm exploded with no visible log.
Now I???m attempting to repeat that with a new 5.1rc1 vm & already up to date 5.1rc1 image. Which has sorta-worked, in a rather odd manner. It got to the final save and quit but when I start the supposedly prepared image there is a problem with it trying to filein that file. I can???t tell right now if it relates to the manual filein I started or if something got added as a deferred action? Or maybe it???s a side-effect of the filein having a save & quit? Not come across this before.
So far as I can tell it did complete the filein, so maybe all is well.
Seen it, discussed with Marcel but we don't have a good idea yet how to work around it.
The only thing I can think of is to walk up the sender chain and restart some context above the one referencing the old instance layout. Maybe an exception handler in the updater that restarts updating in this specific case?
- Bert -
On Sun, Aug 21, 2016 at 10:55 PM, Levente Uzonyi leves@caesar.elte.hu wrote:
Same here. I think the code changes during the update and the old code on the stack causes the problem.
Levente
On Sun, 21 Aug 2016, Chris Muller wrote:
I've definitely experienced the exact same problem, and got around it the same way, restarting the method and proceeding. Very strange.
On Sun, Aug 21, 2016 at 3:13 PM, David T. Lewis lewis@mail.msen.com wrote:
Moving from vm-dev to squeak-dev because it is a trunk update stream question. Follow ups to squeak-dev please.
Does anyone recognize this problem?
We have an issue in trunk update processing that prevents a full update without manual intervention. This does not affect the 5.1 release, but it would be good to fix it so that the update from 5.0 to 5.1 can be reproduced in a continuous integration test.
To reproduce: Start with Squeak5.0-15113.image, set the update URL preference to http://source.squeak.org/trunk, and do an "update from server".
The error first appears when beginning to process the update-mt.378.mcm update map. The problem is related to updating the progress bar with ProgressInitiationException.
When the failure occurs, the debugger (see attached image) seems to show an array access being interpreted as if the array was an instance of StrikeFont. Per Nicolas' note below, it may be related to showing the progress bar while updating it.
I can restart the update process from the debugger from SystemProgressMorph>>position:label:min:max: in the stack frame, and the update proceeds. The error happens several more times, then at some point it seems to go away, so apparently the problem was fixed at some later point in the update stream.
Does anyone recognize this problem? Can we work around it by changing one or more of the update maps?
Thanks, Dave
On Fri, Aug 19, 2016 at 10:30:30PM +0200, Nicolas Cellier wrote:
yes, this is a problem I encountered both in 32 and 64 bits squeak while updating.
I think it's related to showing the progress bar while updating the progress bar. It's a Squeak trunk update problem, not a VM problem.
If some class layout change, but some block is still active (yeah, did you see how many blocks we traverse just to change display of a bar...), then the old block has invalid inst var offsets (offsets to the old object that now has become a new object with new layout).
In an interactive image, just close the debugger and relaunch the update. In an headless automated process, well... We should have fixed the update or start from a newer artifact...
2016-08-19 21:32 GMT+02:00 tim Rowledge tim@rowledge.org:
I just tried running the vmmaker-build script on my iMac with weirdly bad results.
The vm it downloads is the CogSpur.app-16.18.3692 version. Whilst trying to do the update part of the process I had several *very* odd failures where an object in a block was mistaken for nil during a message send. For example in a progress bar update (which I can???t provide a log for since the log got overwritten later) a line (bars at: index) lead to a nil is not indexable halt. Yet in inspecting the prior stackframe the bars object was a perfectly normal array of progress bars. A similar but not identical problem happened during another attempt.
Attempting to filein the BuildSqueakSpurTrunkVMMaker.st file failed when the vm exploded with no visible log.
Now I???m attempting to repeat that with a new 5.1rc1 vm & already up to date 5.1rc1 image. Which has sorta-worked, in a rather odd manner. It got to the final save and quit but when I start the supposedly prepared image there is a problem with it trying to filein that file. I can???t tell right now if it relates to the manual filein I started or if something got added as a deferred action? Or maybe it???s a side-effect of the filein having a save & quit? Not come across this before.
So far as I can tell it did complete the filein, so maybe all is well.
This isn't the first time our changes have caused update hiccups. I think it just comes with the territory. Easiest to just save off another image that is beyond the problem and move on from there.
On Mon, Aug 22, 2016 at 7:36 AM, Bert Freudenberg bert@freudenbergs.de wrote:
Seen it, discussed with Marcel but we don't have a good idea yet how to work around it.
The only thing I can think of is to walk up the sender chain and restart some context above the one referencing the old instance layout. Maybe an exception handler in the updater that restarts updating in this specific case?
- Bert -
On Sun, Aug 21, 2016 at 10:55 PM, Levente Uzonyi leves@caesar.elte.hu wrote:
Same here. I think the code changes during the update and the old code on the stack causes the problem.
Levente
On Sun, 21 Aug 2016, Chris Muller wrote:
I've definitely experienced the exact same problem, and got around it the same way, restarting the method and proceeding. Very strange.
On Sun, Aug 21, 2016 at 3:13 PM, David T. Lewis lewis@mail.msen.com wrote:
Moving from vm-dev to squeak-dev because it is a trunk update stream question. Follow ups to squeak-dev please.
Does anyone recognize this problem?
We have an issue in trunk update processing that prevents a full update without manual intervention. This does not affect the 5.1 release, but it would be good to fix it so that the update from 5.0 to 5.1 can be reproduced in a continuous integration test.
To reproduce: Start with Squeak5.0-15113.image, set the update URL preference to http://source.squeak.org/trunk, and do an "update from server".
The error first appears when beginning to process the update-mt.378.mcm update map. The problem is related to updating the progress bar with ProgressInitiationException.
When the failure occurs, the debugger (see attached image) seems to show an array access being interpreted as if the array was an instance of StrikeFont. Per Nicolas' note below, it may be related to showing the progress bar while updating it.
I can restart the update process from the debugger from SystemProgressMorph>>position:label:min:max: in the stack frame, and the update proceeds. The error happens several more times, then at some point it seems to go away, so apparently the problem was fixed at some later point in the update stream.
Does anyone recognize this problem? Can we work around it by changing one or more of the update maps?
Thanks, Dave
On Fri, Aug 19, 2016 at 10:30:30PM +0200, Nicolas Cellier wrote:
yes, this is a problem I encountered both in 32 and 64 bits squeak while updating.
I think it's related to showing the progress bar while updating the progress bar. It's a Squeak trunk update problem, not a VM problem.
If some class layout change, but some block is still active (yeah, did you see how many blocks we traverse just to change display of a bar...), then the old block has invalid inst var offsets (offsets to the old object that now has become a new object with new layout).
In an interactive image, just close the debugger and relaunch the update. In an headless automated process, well... We should have fixed the update or start from a newer artifact...
2016-08-19 21:32 GMT+02:00 tim Rowledge tim@rowledge.org:
I just tried running the vmmaker-build script on my iMac with weirdly bad results.
The vm it downloads is the CogSpur.app-16.18.3692 version. Whilst trying to do the update part of the process I had several *very* odd failures where an object in a block was mistaken for nil during a message send. For example in a progress bar update (which I can???t provide a log for since the log got overwritten later) a line (bars at: index) lead to a nil is not indexable halt. Yet in inspecting the prior stackframe the bars object was a perfectly normal array of progress bars. A similar but not identical problem happened during another attempt.
Attempting to filein the BuildSqueakSpurTrunkVMMaker.st file failed when the vm exploded with no visible log.
Now I???m attempting to repeat that with a new 5.1rc1 vm & already up to date 5.1rc1 image. Which has sorta-worked, in a rather odd manner. It got to the final save and quit but when I start the supposedly prepared image there is a problem with it trying to filein that file. I can???t tell right now if it relates to the manual filein I started or if something got added as a deferred action? Or maybe it???s a side-effect of the filein having a save & quit? Not come across this before.
So far as I can tell it did complete the filein, so maybe all is well.
There might be another possibility: destroy the progress bar in the preamble. How to detect the progress bar is left as an exercize (translate: I just don't know, maybe also by walking the stack). Reopen the progress bar in postscript.
2016-08-22 17:57 GMT+02:00 Chris Muller asqueaker@gmail.com:
This isn't the first time our changes have caused update hiccups. I think it just comes with the territory. Easiest to just save off another image that is beyond the problem and move on from there.
On Mon, Aug 22, 2016 at 7:36 AM, Bert Freudenberg bert@freudenbergs.de wrote:
Seen it, discussed with Marcel but we don't have a good idea yet how to
work
around it.
The only thing I can think of is to walk up the sender chain and restart some context above the one referencing the old instance layout. Maybe an exception handler in the updater that restarts updating in this specific case?
- Bert -
On Sun, Aug 21, 2016 at 10:55 PM, Levente Uzonyi leves@caesar.elte.hu wrote:
Same here. I think the code changes during the update and the old code
on
the stack causes the problem.
Levente
On Sun, 21 Aug 2016, Chris Muller wrote:
I've definitely experienced the exact same problem, and got around it the same way, restarting the method and proceeding. Very strange.
On Sun, Aug 21, 2016 at 3:13 PM, David T. Lewis lewis@mail.msen.com wrote:
Moving from vm-dev to squeak-dev because it is a trunk update stream question. Follow ups to squeak-dev please.
Does anyone recognize this problem?
We have an issue in trunk update processing that prevents a full
update
without manual intervention. This does not affect the 5.1 release, but it would be good to fix it so that the update from 5.0 to 5.1 can be reproduced in a continuous integration test.
To reproduce: Start with Squeak5.0-15113.image, set the update URL preference to http://source.squeak.org/trunk, and do an "update from server".
The error first appears when beginning to process the
update-mt.378.mcm
update map. The problem is related to updating the progress bar with ProgressInitiationException.
When the failure occurs, the debugger (see attached image) seems to show an array access being interpreted as if the array was an instance of StrikeFont. Per Nicolas' note below, it may be related to showing the progress bar while updating it.
I can restart the update process from the debugger from SystemProgressMorph>>position:label:min:max: in the stack frame, and
the
update proceeds. The error happens several more times, then at some point it seems to go away, so apparently the problem was fixed at some later point in the update stream.
Does anyone recognize this problem? Can we work around it by changing one or more of the update maps?
Thanks, Dave
On Fri, Aug 19, 2016 at 10:30:30PM +0200, Nicolas Cellier wrote:
yes, this is a problem I encountered both in 32 and 64 bits squeak while updating.
I think it's related to showing the progress bar while updating the progress bar. It's a Squeak trunk update problem, not a VM problem.
If some class layout change, but some block is still active (yeah,
did
you see how many blocks we traverse just to change display of a bar...), then the old block has invalid inst var offsets (offsets to the old object that now has become a new object with new layout).
In an interactive image, just close the debugger and relaunch the update. In an headless automated process, well... We should have fixed the update or start from a newer artifact...
2016-08-19 21:32 GMT+02:00 tim Rowledge tim@rowledge.org: > > > I just tried running the vmmaker-build script on my iMac with
weirdly
> bad > results. > > The vm it downloads is the CogSpur.app-16.18.3692 version. Whilst > trying > to do the update part of the process I had several *very* odd
failures
> where an object in a block was mistaken for nil during a message
send.
> For > example in a progress bar update (which I can???t provide a log for > since the > log got overwritten later) a line > (bars at: index) > lead to a nil is not indexable halt. Yet in inspecting the prior > stackframe the bars object was a perfectly normal array of progress > bars. A > similar but not identical problem happened during another attempt. > > Attempting to filein the BuildSqueakSpurTrunkVMMaker.st file failed > when > the vm exploded with no visible log. > > Now I???m attempting to repeat that with a new 5.1rc1 vm & already
up
> to > date 5.1rc1 image. Which has sorta-worked, in a rather odd manner.
It
> got > to the final save and quit but when I start the supposedly prepared > image > there is a problem with it trying to filein that file. I can???t
tell
> right > now if it relates to the manual filein I started or if something got > added > as a deferred action? Or maybe it???s a side-effect of the filein > having a > save & quit? Not come across this before. > > > > So far as I can tell it did complete the filein, so maybe all is
well.
>
On Mon, Aug 22, 2016 at 02:36:09PM +0200, Bert Freudenberg wrote:
Seen it, discussed with Marcel but we don't have a good idea yet how to work around it.
The only thing I can think of is to walk up the sender chain and restart some context above the one referencing the old instance layout. Maybe an exception handler in the updater that restarts updating in this specific case?
- Bert -
I don't see a good place to put an exception handler, at least not without making things even more complicated than they already are.
It looks like the the direct reference to instance var 'bars' is where we encounter the problem. Presumably the old code in a block is refering to an ivar slot in SystemProgressMorph that used to point to the 'vars' array, but that is now pointing to the 'font' ivar.
And indeed I see now that Marcel found and fixed this issue in Morphic-mt.1198, so that accessor methods will now be used instead of direct references to the variables.
But we still have the problem of a SystemProgressMorph evaluating an older block that was created with an earlier version of the SystemProgressMorph>>position:label:min:max: method, so the update stream processing still fails if the update processing was started before Morphic-mt.1198 was loaded.
A partial workaround is to modify #position:label:min:max: to use an accessor for 'bars' prior to starting the update processing. This prevents the error condition that we are discussing here, but unfortunately it also results in a couple of merge dialogs for the modified methods, so the update still requires manual intervention.
Dave
On Sun, Aug 21, 2016 at 10:55 PM, Levente Uzonyi leves@caesar.elte.hu wrote:
Same here. I think the code changes during the update and the old code on the stack causes the problem.
Levente
On Sun, 21 Aug 2016, Chris Muller wrote:
I've definitely experienced the exact same problem, and got around it the same way, restarting the method and proceeding. Very strange.
On Sun, Aug 21, 2016 at 3:13 PM, David T. Lewis lewis@mail.msen.com wrote:
Moving from vm-dev to squeak-dev because it is a trunk update stream question. Follow ups to squeak-dev please.
Does anyone recognize this problem?
We have an issue in trunk update processing that prevents a full update without manual intervention. This does not affect the 5.1 release, but it would be good to fix it so that the update from 5.0 to 5.1 can be reproduced in a continuous integration test.
To reproduce: Start with Squeak5.0-15113.image, set the update URL preference to http://source.squeak.org/trunk, and do an "update from server".
The error first appears when beginning to process the update-mt.378.mcm update map. The problem is related to updating the progress bar with ProgressInitiationException.
When the failure occurs, the debugger (see attached image) seems to show an array access being interpreted as if the array was an instance of StrikeFont. Per Nicolas' note below, it may be related to showing the progress bar while updating it.
I can restart the update process from the debugger from SystemProgressMorph>>position:label:min:max: in the stack frame, and the update proceeds. The error happens several more times, then at some point it seems to go away, so apparently the problem was fixed at some later point in the update stream.
Does anyone recognize this problem? Can we work around it by changing one or more of the update maps?
Thanks, Dave
On Fri, Aug 19, 2016 at 10:30:30PM +0200, Nicolas Cellier wrote:
yes, this is a problem I encountered both in 32 and 64 bits squeak while updating.
I think it's related to showing the progress bar while updating the progress bar. It's a Squeak trunk update problem, not a VM problem.
If some class layout change, but some block is still active (yeah, did you see how many blocks we traverse just to change display of a bar...), then the old block has invalid inst var offsets (offsets to the old object that now has become a new object with new layout).
In an interactive image, just close the debugger and relaunch the update. In an headless automated process, well... We should have fixed the update or start from a newer artifact...
2016-08-19 21:32 GMT+02:00 tim Rowledge tim@rowledge.org:
I just tried running the vmmaker-build script on my iMac with weirdly bad results.
The vm it downloads is the CogSpur.app-16.18.3692 version. Whilst trying to do the update part of the process I had several *very* odd failures where an object in a block was mistaken for nil during a message send. For example in a progress bar update (which I can???t provide a log for since the log got overwritten later) a line (bars at: index) lead to a nil is not indexable halt. Yet in inspecting the prior stackframe the bars object was a perfectly normal array of progress bars. A similar but not identical problem happened during another attempt.
Attempting to filein the BuildSqueakSpurTrunkVMMaker.st file failed when the vm exploded with no visible log.
Now I???m attempting to repeat that with a new 5.1rc1 vm & already up to date 5.1rc1 image. Which has sorta-worked, in a rather odd manner. It got to the final save and quit but when I start the supposedly prepared image there is a problem with it trying to filein that file. I can???t tell right now if it relates to the manual filein I started or if something got added as a deferred action? Or maybe it???s a side-effect of the filein having a save & quit? Not come across this before.
So far as I can tell it did complete the filein, so maybe all is well.
There's yet another possibility, in a preamble: - clone the old SystemProgressMorph class, - mutate all instances to the clone class
Don't use class rename and class removal, because the system will do too much clean up. instead hack at a lower level
2016-08-23 4:03 GMT+02:00 David T. Lewis lewis@mail.msen.com:
On Mon, Aug 22, 2016 at 02:36:09PM +0200, Bert Freudenberg wrote:
Seen it, discussed with Marcel but we don't have a good idea yet how to work around it.
The only thing I can think of is to walk up the sender chain and restart some context above the one referencing the old instance layout. Maybe an exception handler in the updater that restarts updating in this specific case?
- Bert -
I don't see a good place to put an exception handler, at least not without making things even more complicated than they already are.
It looks like the the direct reference to instance var 'bars' is where we encounter the problem. Presumably the old code in a block is refering to an ivar slot in SystemProgressMorph that used to point to the 'vars' array, but that is now pointing to the 'font' ivar.
And indeed I see now that Marcel found and fixed this issue in Morphic-mt.1198, so that accessor methods will now be used instead of direct references to the variables.
But we still have the problem of a SystemProgressMorph evaluating an older block that was created with an earlier version of the SystemProgressMorph>>position:label:min:max: method, so the update stream processing still fails if the update processing was started before Morphic-mt.1198 was loaded.
A partial workaround is to modify #position:label:min:max: to use an accessor for 'bars' prior to starting the update processing. This prevents the error condition that we are discussing here, but unfortunately it also results in a couple of merge dialogs for the modified methods, so the update still requires manual intervention.
Dave
On Sun, Aug 21, 2016 at 10:55 PM, Levente Uzonyi leves@caesar.elte.hu wrote:
Same here. I think the code changes during the update and the old code
on
the stack causes the problem.
Levente
On Sun, 21 Aug 2016, Chris Muller wrote:
I've definitely experienced the exact same problem, and got around it the same way, restarting the method and proceeding. Very strange.
On Sun, Aug 21, 2016 at 3:13 PM, David T. Lewis lewis@mail.msen.com wrote:
Moving from vm-dev to squeak-dev because it is a trunk update stream question. Follow ups to squeak-dev please.
Does anyone recognize this problem?
We have an issue in trunk update processing that prevents a full
update
without manual intervention. This does not affect the 5.1 release,
but
it would be good to fix it so that the update from 5.0 to 5.1 can be reproduced in a continuous integration test.
To reproduce: Start with Squeak5.0-15113.image, set the update URL preference to http://source.squeak.org/trunk, and do an "update from server".
The error first appears when beginning to process the
update-mt.378.mcm
update map. The problem is related to updating the progress bar with ProgressInitiationException.
When the failure occurs, the debugger (see attached image) seems to show an array access being interpreted as if the array was an
instance
of StrikeFont. Per Nicolas' note below, it may be related to showing the progress bar while updating it.
I can restart the update process from the debugger from SystemProgressMorph>>position:label:min:max: in the stack frame,
and the
update proceeds. The error happens several more times, then at some
point
it seems to go away, so apparently the problem was fixed at some
later
point in the update stream.
Does anyone recognize this problem? Can we work around it by
changing one
or more of the update maps?
Thanks, Dave
On Fri, Aug 19, 2016 at 10:30:30PM +0200, Nicolas Cellier wrote:
yes, this is a problem I encountered both in 32 and 64 bits squeak
while
updating.
I think it's related to showing the progress bar while updating the progress bar. It's a Squeak trunk update problem, not a VM problem.
If some class layout change, but some block is still active (yeah,
did
you see how many blocks we traverse just to change display of a bar...), then the old block has invalid inst var offsets (offsets to the old
object
that now has become a new object with new layout).
In an interactive image, just close the debugger and relaunch the update. In an headless automated process, well... We should have fixed the update or start from a newer artifact...
2016-08-19 21:32 GMT+02:00 tim Rowledge tim@rowledge.org:
> > I just tried running the vmmaker-build script on my iMac with
weirdly
> bad > results. > > The vm it downloads is the CogSpur.app-16.18.3692 version. Whilst > trying > to do the update part of the process I had several *very* odd
failures
> where an object in a block was mistaken for nil during a message
send.
> For > example in a progress bar update (which I can???t provide a log for > since the > log got overwritten later) a line > (bars at: index) > lead to a nil is not indexable halt. Yet in inspecting the prior > stackframe the bars object was a perfectly normal array of progress > bars. A > similar but not identical problem happened during another attempt. > > Attempting to filein the BuildSqueakSpurTrunkVMMaker.st file failed > when > the vm exploded with no visible log. > > Now I???m attempting to repeat that with a new 5.1rc1 vm & already
up
> to > date 5.1rc1 image. Which has sorta-worked, in a rather odd manner.
It
> got > to the final save and quit but when I start the supposedly prepared > image > there is a problem with it trying to filein that file. I can???t
tell
> right > now if it relates to the manual filein I started or if something
got
> added > as a deferred action? Or maybe it???s a side-effect of the filein > having a > save & quit? Not come across this before. > > > > So far as I can tell it did complete the filein, so maybe all is
well.
> >
Nicolas Cellier wrote
There's yet another possibility, in a preamble:
- clone the old SystemProgressMorph class,
- mutate all instances to the clone class
Don't use class rename and class removal, because the system will do too much clean up. instead hack at a lower level
2016-08-23 4:03 GMT+02:00 David T. Lewis <
lewis@.msen
>:
On Mon, Aug 22, 2016 at 02:36:09PM +0200, Bert Freudenberg wrote:
Seen it, discussed with Marcel but we don't have a good idea yet how to work around it.
The only thing I can think of is to walk up the sender chain and
restart
some context above the one referencing the old instance layout. Maybe
an
exception handler in the updater that restarts updating in this
specific
case?
- Bert -
I don't see a good place to put an exception handler, at least not without making things even more complicated than they already are.
It looks like the the direct reference to instance var 'bars' is where we encounter the problem. Presumably the old code in a block is refering to an ivar slot in SystemProgressMorph that used to point to the 'vars' array, but that is now pointing to the 'font' ivar.
And indeed I see now that Marcel found and fixed this issue in Morphic-mt.1198, so that accessor methods will now be used instead of direct references to the variables.
But we still have the problem of a SystemProgressMorph evaluating an older block that was created with an earlier version of the SystemProgressMorph>>position:label:min:max: method, so the update stream processing still fails if the update processing was started before Morphic-mt.1198 was loaded.
A partial workaround is to modify #position:label:min:max: to use an accessor for 'bars' prior to starting the update processing. This prevents the error condition that we are discussing here, but unfortunately it also results in a couple of merge dialogs for the modified methods, so the update still requires manual intervention.
Dave
On Sun, Aug 21, 2016 at 10:55 PM, Levente Uzonyi <
leves@.elte
>
wrote:
Same here. I think the code changes during the update and the old
code on
the stack causes the problem.
Levente
On Sun, 21 Aug 2016, Chris Muller wrote:
I've definitely experienced the exact same problem, and got around
it
the same way, restarting the method and proceeding. Very strange.
On Sun, Aug 21, 2016 at 3:13 PM, David T. Lewis <
lewis@.msen
>
wrote:
Moving from vm-dev to squeak-dev because it is a trunk update
stream
question. Follow ups to squeak-dev please.
Does anyone recognize this problem?
We have an issue in trunk update processing that prevents a full
update
without manual intervention. This does not affect the 5.1 release,
but
it would be good to fix it so that the update from 5.0 to 5.1 can
be
reproduced in a continuous integration test.
To reproduce: Start with Squeak5.0-15113.image, set the update URL preference to http://source.squeak.org/trunk, and do an "update from server".
The error first appears when beginning to process the
update-mt.378.mcm
update map. The problem is related to updating the progress bar
with
ProgressInitiationException.
When the failure occurs, the debugger (see attached image) seems to show an array access being interpreted as if the array was an
instance
of StrikeFont. Per Nicolas' note below, it may be related to
showing
the progress bar while updating it.
I can restart the update process from the debugger from SystemProgressMorph>>position:label:min:max: in the stack frame,
and the
update proceeds. The error happens several more times, then at some
point
it seems to go away, so apparently the problem was fixed at some
later
point in the update stream.
Does anyone recognize this problem? Can we work around it by
changing one
or more of the update maps?
Thanks, Dave
On Fri, Aug 19, 2016 at 10:30:30PM +0200, Nicolas Cellier wrote:
> > yes, this is a problem I encountered both in 32 and 64 bits squeak
while
> updating. > > I think it's related to showing the progress bar while updating
the
> progress bar. It's a Squeak trunk update problem, not a VM
problem.
> > If some class layout change, but some block is still active (yeah,
did
> you > see how many blocks we traverse just to change display of a
bar...),
> then > the old block has invalid inst var offsets (offsets to the old
object
> that > now has become a new object with new layout). > > In an interactive image, just close the debugger and relaunch the > update. > In an headless automated process, well... We should have fixed the > update > or start from a newer artifact... > > 2016-08-19 21:32 GMT+02:00 tim Rowledge <
tim@
>:
> >> >> I just tried running the vmmaker-build script on my iMac with
weirdly
>> bad >> results. >> >> The vm it downloads is the CogSpur.app-16.18.3692 version. Whilst >> trying >> to do the update part of the process I had several *very* odd
failures
>> where an object in a block was mistaken for nil during a message
send.
>> For >> example in a progress bar update (which I can???t provide a log
for
>> since the >> log got overwritten later) a line >> (bars at: index) >> lead to a nil is not indexable halt. Yet in inspecting the prior >> stackframe the bars object was a perfectly normal array of
progress
>> bars. A >> similar but not identical problem happened during another
attempt.
>> >> Attempting to filein the BuildSqueakSpurTrunkVMMaker.st file
failed
>> when >> the vm exploded with no visible log. >> >> Now I???m attempting to repeat that with a new 5.1rc1 vm &
already up
>> to >> date 5.1rc1 image. Which has sorta-worked, in a rather odd
manner. It
>> got >> to the final save and quit but when I start the supposedly
prepared
>> image >> there is a problem with it trying to filein that file. I can???t
tell
>> right >> now if it relates to the manual filein I started or if something
got
>> added >> as a deferred action? Or maybe it???s a side-effect of the filein >> having a >> save & quit? Not come across this before. >> >> >> >> So far as I can tell it did complete the filein, so maybe all is
well.
>> >>
Hi, there.
It is, unfortunately, not possible to turn back time and rewrite history. ;-) There is several important code that makes excessive use of instVar accesses such as in SystemProgressMorph or HandMorph. Making any change in their class layout while using some of their methods will provoke a strange error. This could only be changed by a VM that keeps old class layouts/schemata around for some longer time until there is no old compiled method on any stack anymore.
Usually, restarting the update process helps here. There is no way to fix that in order to interactively update from 5.0 to 5.1 without any errors.
However, we have an update script that updates build 15113, which is effectively Squeak 5.0 to the latest trunk. Here is the script: https://github.com/squeak-smalltalk/squeak-app/blob/master/prepare_image.st And here is the image build 15113: http://files.squeak.org/base/Squeak-trunk/base.zip
We use that the create bundles with the latest alpha, beta, rc versions.
Best, Marcel
-- View this message in context: http://forum.world.st/Re-Does-anyone-recognize-this-glitch-in-our-trunk-upda... Sent from the Squeak - Dev mailing list archive at Nabble.com.
It would be possible to start the update with the required error handlers around it, but unless you really had to do it, you're better off using a fixed image.
Levente
On Tue, 23 Aug 2016, marcel.taeumel wrote:
Nicolas Cellier wrote
There's yet another possibility, in a preamble:
- clone the old SystemProgressMorph class,
- mutate all instances to the clone class
Don't use class rename and class removal, because the system will do too much clean up. instead hack at a lower level
2016-08-23 4:03 GMT+02:00 David T. Lewis <
lewis@.msen
>:
On Mon, Aug 22, 2016 at 02:36:09PM +0200, Bert Freudenberg wrote:
Seen it, discussed with Marcel but we don't have a good idea yet how to work around it.
The only thing I can think of is to walk up the sender chain and
restart
some context above the one referencing the old instance layout. Maybe
an
exception handler in the updater that restarts updating in this
specific
case?
- Bert -
I don't see a good place to put an exception handler, at least not without making things even more complicated than they already are.
It looks like the the direct reference to instance var 'bars' is where we encounter the problem. Presumably the old code in a block is refering to an ivar slot in SystemProgressMorph that used to point to the 'vars' array, but that is now pointing to the 'font' ivar.
And indeed I see now that Marcel found and fixed this issue in Morphic-mt.1198, so that accessor methods will now be used instead of direct references to the variables.
But we still have the problem of a SystemProgressMorph evaluating an older block that was created with an earlier version of the SystemProgressMorph>>position:label:min:max: method, so the update stream processing still fails if the update processing was started before Morphic-mt.1198 was loaded.
A partial workaround is to modify #position:label:min:max: to use an accessor for 'bars' prior to starting the update processing. This prevents the error condition that we are discussing here, but unfortunately it also results in a couple of merge dialogs for the modified methods, so the update still requires manual intervention.
Dave
On Sun, Aug 21, 2016 at 10:55 PM, Levente Uzonyi <
leves@.elte
>
wrote:
Same here. I think the code changes during the update and the old
code on
the stack causes the problem.
Levente
On Sun, 21 Aug 2016, Chris Muller wrote:
I've definitely experienced the exact same problem, and got around
it
the same way, restarting the method and proceeding. Very strange.
On Sun, Aug 21, 2016 at 3:13 PM, David T. Lewis <
lewis@.msen
>
wrote:
> > Moving from vm-dev to squeak-dev because it is a trunk update
stream
> question. > Follow ups to squeak-dev please. > > Does anyone recognize this problem? > > We have an issue in trunk update processing that prevents a full
update
> without manual intervention. This does not affect the 5.1 release,
but
> it would be good to fix it so that the update from 5.0 to 5.1 can
be
> reproduced in a continuous integration test. > > To reproduce: Start with Squeak5.0-15113.image, set the update URL > preference > to http://source.squeak.org/trunk, and do an "update from server". > > The error first appears when beginning to process the
update-mt.378.mcm
> update map. The problem is related to updating the progress bar
with
> ProgressInitiationException. > > When the failure occurs, the debugger (see attached image) seems to > show an array access being interpreted as if the array was an
instance
> of StrikeFont. Per Nicolas' note below, it may be related to
showing
> the progress bar while updating it. > > I can restart the update process from the debugger from > SystemProgressMorph>>position:label:min:max: in the stack frame,
and the
> update proceeds. The error happens several more times, then at some
point
> it seems to go away, so apparently the problem was fixed at some
later
> point in the update stream. > > Does anyone recognize this problem? Can we work around it by
changing one
> or more of the update maps? > > Thanks, > Dave > > > On Fri, Aug 19, 2016 at 10:30:30PM +0200, Nicolas Cellier wrote: > >> >> yes, this is a problem I encountered both in 32 and 64 bits squeak
while
>> updating. >> >> I think it's related to showing the progress bar while updating
the
>> progress bar. It's a Squeak trunk update problem, not a VM
problem.
>> >> If some class layout change, but some block is still active (yeah,
did
>> you >> see how many blocks we traverse just to change display of a
bar...),
>> then >> the old block has invalid inst var offsets (offsets to the old
object
>> that >> now has become a new object with new layout). >> >> In an interactive image, just close the debugger and relaunch the >> update. >> In an headless automated process, well... We should have fixed the >> update >> or start from a newer artifact... >> >> 2016-08-19 21:32 GMT+02:00 tim Rowledge <
tim@
>:
>> >>> >>> I just tried running the vmmaker-build script on my iMac with
weirdly
>>> bad >>> results. >>> >>> The vm it downloads is the CogSpur.app-16.18.3692 version. Whilst >>> trying >>> to do the update part of the process I had several *very* odd
failures
>>> where an object in a block was mistaken for nil during a message
send.
>>> For >>> example in a progress bar update (which I can???t provide a log
for
>>> since the >>> log got overwritten later) a line >>> (bars at: index) >>> lead to a nil is not indexable halt. Yet in inspecting the prior >>> stackframe the bars object was a perfectly normal array of
progress
>>> bars. A >>> similar but not identical problem happened during another
attempt.
>>> >>> Attempting to filein the BuildSqueakSpurTrunkVMMaker.st file
failed
>>> when >>> the vm exploded with no visible log. >>> >>> Now I???m attempting to repeat that with a new 5.1rc1 vm &
already up
>>> to >>> date 5.1rc1 image. Which has sorta-worked, in a rather odd
manner. It
>>> got >>> to the final save and quit but when I start the supposedly
prepared
>>> image >>> there is a problem with it trying to filein that file. I can???t
tell
>>> right >>> now if it relates to the manual filein I started or if something
got
>>> added >>> as a deferred action? Or maybe it???s a side-effect of the filein >>> having a >>> save & quit? Not come across this before. >>> >>> >>> >>> So far as I can tell it did complete the filein, so maybe all is
well.
>>> >>> >
Hi, there.
It is, unfortunately, not possible to turn back time and rewrite history. ;-) There is several important code that makes excessive use of instVar accesses such as in SystemProgressMorph or HandMorph. Making any change in their class layout while using some of their methods will provoke a strange error. This could only be changed by a VM that keeps old class layouts/schemata around for some longer time until there is no old compiled method on any stack anymore.
Usually, restarting the update process helps here. There is no way to fix that in order to interactively update from 5.0 to 5.1 without any errors.
However, we have an update script that updates build 15113, which is effectively Squeak 5.0 to the latest trunk. Here is the script: https://github.com/squeak-smalltalk/squeak-app/blob/master/prepare_image.st And here is the image build 15113: http://files.squeak.org/base/Squeak-trunk/base.zip
We use that the create bundles with the latest alpha, beta, rc versions.
Best, Marcel
-- View this message in context: http://forum.world.st/Re-Does-anyone-recognize-this-glitch-in-our-trunk-upda... Sent from the Squeak - Dev mailing list archive at Nabble.com.
On Tue, Aug 23, 2016 at 11:21 AM, marcel.taeumel Marcel.Taeumel@hpi.de wrote:
Hi, there.
It is, unfortunately, not possible to turn back time and rewrite history. ;-)
... unless we're dealing with software. Oh wait, we are ;)
There is several important code that makes excessive use of instVar
accesses such as in SystemProgressMorph or HandMorph. Making any change in their class layout while using some of their methods will provoke a strange error. This could only be changed by a VM that keeps old class layouts/schemata around for some longer time until there is no old compiled method on any stack anymore.
Class layout changes and instance migration is handled fully in the image by class ClassBuilder. The VM is not at fault here.
Usually, restarting the update process helps here. There is no way to fix that in order to interactively update from 5.0 to 5.1 without any errors.
We can retroactively change the config map that introduced the problem. I'm certain it would be possible to fix, but I am not entirely sure it's worth the trouble.
- Bert -
Bert Freudenberg wrote
On Tue, Aug 23, 2016 at 11:21 AM, marcel.taeumel <
Marcel.Taeumel@
> wrote:
Hi, there.
It is, unfortunately, not possible to turn back time and rewrite history. ;-)
... unless we're dealing with software. Oh wait, we are ;)
There is several important code that makes excessive use of instVar
accesses such as in SystemProgressMorph or HandMorph. Making any change in their class layout while using some of their methods will provoke a strange error. This could only be changed by a VM that keeps old class layouts/schemata around for some longer time until there is no old compiled method on any stack anymore.
Class layout changes and instance migration is handled fully in the image by class ClassBuilder. The VM is not at fault here.
Usually, restarting the update process helps here. There is no way to fix that in order to interactively update from 5.0 to 5.1 without any errors.
We can retroactively change the config map that introduced the problem. I'm certain it would be possible to fix, but I am not entirely sure it's worth the trouble.
- Bert -
Hi Bert,
the config map will not help, we would have to change numerous MCZs since point X in time to rewrite history. :) "Oh, would we have always been using accessors here instead of instVar accessess..."
Best, Marcel
-- View this message in context: http://forum.world.st/Re-Does-anyone-recognize-this-glitch-in-our-trunk-upda... Sent from the Squeak - Dev mailing list archive at Nabble.com.
marcel.taeumel wrote
Bert Freudenberg wrote
On Tue, Aug 23, 2016 at 11:21 AM, marcel.taeumel <
Marcel.Taeumel@
> wrote:
Hi, there.
It is, unfortunately, not possible to turn back time and rewrite history. ;-)
... unless we're dealing with software. Oh wait, we are ;)
There is several important code that makes excessive use of instVar
accesses such as in SystemProgressMorph or HandMorph. Making any change in their class layout while using some of their methods will provoke a strange error. This could only be changed by a VM that keeps old class layouts/schemata around for some longer time until there is no old compiled method on any stack anymore.
Class layout changes and instance migration is handled fully in the image by class ClassBuilder. The VM is not at fault here.
Usually, restarting the update process helps here. There is no way to fix that in order to interactively update from 5.0 to 5.1 without any errors.
We can retroactively change the config map that introduced the problem. I'm certain it would be possible to fix, but I am not entirely sure it's worth the trouble.
- Bert -
Hi Bert,
the config map will not help, we would have to change numerous MCZs since point X in time to rewrite history. :) "Oh, would we have always been using accessors here instead of instVar accessess..."
Best, Marcel
...even then, how do you update the code in image Y, which is going to be updated before it is updated? This is not even possible for every user's images out there... :-/
The Java Hotspot VM has a mapping table for changed instVar accesses. Maybe we could use this trick, too? Still, would only applies to future versions. We cannot rewrite history because it is not only one piece of software in a single version control system with a single version checked out. :)
Best, Marcel
-- View this message in context: http://forum.world.st/Re-Does-anyone-recognize-this-glitch-in-our-trunk-upda... Sent from the Squeak - Dev mailing list archive at Nabble.com.
Or would this even more simple hack work? Put in the preamble something like:
Smalltalk at: SystemProgressMorph put: SystemProgressMorph copy.
This way, the old instances should not be mutated...
2016-08-24 7:06 GMT+02:00 marcel.taeumel Marcel.Taeumel@hpi.de:
marcel.taeumel wrote
Bert Freudenberg wrote
On Tue, Aug 23, 2016 at 11:21 AM, marcel.taeumel <
Marcel.Taeumel@
> wrote:
Hi, there.
It is, unfortunately, not possible to turn back time and rewrite history. ;-)
... unless we're dealing with software. Oh wait, we are ;)
There is several important code that makes excessive use of instVar
accesses such as in SystemProgressMorph or HandMorph. Making any change in their class layout while using some of their methods will provoke a strange error. This could only be changed by a VM that keeps old class layouts/schemata around for some longer time until there is no old compiled method on any stack anymore.
Class layout changes and instance migration is handled fully in the
image
by class ClassBuilder. The VM is not at fault here.
Usually, restarting the update process helps here. There is no way to fix that in order to interactively update from 5.0 to 5.1 without any errors.
We can retroactively change the config map that introduced the problem. I'm certain it would be possible to fix, but I am not entirely sure it's worth the trouble.
- Bert -
Hi Bert,
the config map will not help, we would have to change numerous MCZs since point X in time to rewrite history. :) "Oh, would we have always been using accessors here instead of instVar accessess..."
Best, Marcel
...even then, how do you update the code in image Y, which is going to be updated before it is updated? This is not even possible for every user's images out there... :-/
The Java Hotspot VM has a mapping table for changed instVar accesses. Maybe we could use this trick, too? Still, would only applies to future versions. We cannot rewrite history because it is not only one piece of software in a single version control system with a single version checked out. :)
Best, Marcel
-- View this message in context: http://forum.world.st/Re-Does- anyone-recognize-this-glitch-in-our-trunk-update-processing-was-Vm-dev- buildspurtrunkvmmaker-tp4912130p4912403.html Sent from the Squeak - Dev mailing list archive at Nabble.com.
Due to read only ClassBinding it would be just a little bit more hackish:
(Smalltalk globals associationAt: #SystemProgressMorph) instVarAt: 2 put: SystemProgressMorph copy.
2016-09-01 1:12 GMT+02:00 Nicolas Cellier < nicolas.cellier.aka.nice@gmail.com>:
Or would this even more simple hack work? Put in the preamble something like:
Smalltalk at: SystemProgressMorph put: SystemProgressMorph copy.
This way, the old instances should not be mutated...
2016-08-24 7:06 GMT+02:00 marcel.taeumel Marcel.Taeumel@hpi.de:
marcel.taeumel wrote
Bert Freudenberg wrote
On Tue, Aug 23, 2016 at 11:21 AM, marcel.taeumel <
Marcel.Taeumel@
> wrote:
Hi, there.
It is, unfortunately, not possible to turn back time and rewrite history. ;-)
... unless we're dealing with software. Oh wait, we are ;)
There is several important code that makes excessive use of instVar
accesses such as in SystemProgressMorph or HandMorph. Making any
change
in their class layout while using some of their methods will provoke a strange error. This could only be changed by a VM that keeps old class layouts/schemata around for some longer time until there is no old compiled method on any stack anymore.
Class layout changes and instance migration is handled fully in the
image
by class ClassBuilder. The VM is not at fault here.
Usually, restarting the update process helps here. There is no way to fix that in order to interactively update from 5.0 to 5.1 without any errors.
We can retroactively change the config map that introduced the problem. I'm certain it would be possible to fix, but I am not entirely sure it's worth the trouble.
- Bert -
Hi Bert,
the config map will not help, we would have to change numerous MCZs
since
point X in time to rewrite history. :) "Oh, would we have always been using accessors here instead of instVar accessess..."
Best, Marcel
...even then, how do you update the code in image Y, which is going to be updated before it is updated? This is not even possible for every user's images out there... :-/
The Java Hotspot VM has a mapping table for changed instVar accesses. Maybe we could use this trick, too? Still, would only applies to future versions. We cannot rewrite history because it is not only one piece of software in a single version control system with a single version checked out. :)
Best, Marcel
-- View this message in context: http://forum.world.st/Re-Does- anyone-recognize-this-glitch-in-our-trunk-update-processing- was-Vm-dev-buildspurtrunkvmmaker-tp4912130p4912403.html Sent from the Squeak - Dev mailing list archive at Nabble.com.
The copy is not completely clean because this assertion then fails:
(SystemProgressMorph superclass subclasses detect: [:e | e name = #SystemProgressMorph]) == SystemProgressMorph
IOW, the superclass is still pointing to the old class, so any change to superclass layout won't propagate to the new subclass... Bad.
But there's also this possibility:
| oldClass | oldClass := SystemProgressMorph copy become: SystemProgressMorph. SystemProgressMorph allInstancesDo: [:e | oldClass adoptInstance: e].
We mutate all the instances to a class detached from Smalltalk and immune to some changes... Beware: don't you modify the layout of one superclass...
2016-09-01 1:17 GMT+02:00 Nicolas Cellier < nicolas.cellier.aka.nice@gmail.com>:
Due to read only ClassBinding it would be just a little bit more hackish:
(Smalltalk globals associationAt: #SystemProgressMorph) instVarAt: 2 put: SystemProgressMorph copy.
2016-09-01 1:12 GMT+02:00 Nicolas Cellier <nicolas.cellier.aka.nice@ gmail.com>:
Or would this even more simple hack work? Put in the preamble something like:
Smalltalk at: SystemProgressMorph put: SystemProgressMorph copy.
This way, the old instances should not be mutated...
2016-08-24 7:06 GMT+02:00 marcel.taeumel Marcel.Taeumel@hpi.de:
marcel.taeumel wrote
Bert Freudenberg wrote
On Tue, Aug 23, 2016 at 11:21 AM, marcel.taeumel <
Marcel.Taeumel@
> wrote:
Hi, there.
It is, unfortunately, not possible to turn back time and rewrite history. ;-)
... unless we're dealing with software. Oh wait, we are ;)
There is several important code that makes excessive use of instVar
accesses such as in SystemProgressMorph or HandMorph. Making any
change
in their class layout while using some of their methods will provoke a strange error. This could only be changed by a VM that keeps old class layouts/schemata around for some longer time until there is no old compiled method on any stack anymore.
Class layout changes and instance migration is handled fully in the
image
by class ClassBuilder. The VM is not at fault here.
Usually, restarting the update process helps here. There is no way to fix that in order to interactively update from 5.0 to 5.1 without any errors.
We can retroactively change the config map that introduced the
problem.
I'm certain it would be possible to fix, but I am not entirely sure it's worth the trouble.
- Bert -
Hi Bert,
the config map will not help, we would have to change numerous MCZs
since
point X in time to rewrite history. :) "Oh, would we have always been using accessors here instead of instVar accessess..."
Best, Marcel
...even then, how do you update the code in image Y, which is going to be updated before it is updated? This is not even possible for every user's images out there... :-/
The Java Hotspot VM has a mapping table for changed instVar accesses. Maybe we could use this trick, too? Still, would only applies to future versions. We cannot rewrite history because it is not only one piece of software in a single version control system with a single version checked out. :)
Best, Marcel
-- View this message in context: http://forum.world.st/Re-Does- anyone-recognize-this-glitch-in-our-trunk-update-processing- was-Vm-dev-buildspurtrunkvmmaker-tp4912130p4912403.html Sent from the Squeak - Dev mailing list archive at Nabble.com.
On Thu, Sep 1, 2016 at 12:33 AM, Nicolas Cellier < nicolas.cellier.aka.nice@gmail.com> wrote:
The copy is not completely clean because this assertion then fails:
(SystemProgressMorph superclass subclasses detect: [:e | e name = #SystemProgressMorph]) == SystemProgressMorph
IOW, the superclass is still pointing to the old class, so any change to superclass layout won't propagate to the new subclass... Bad.
But there's also this possibility:
| oldClass | oldClass := SystemProgressMorph copy become: SystemProgressMorph. SystemProgressMorph allInstancesDo: [:e | oldClass adoptInstance: e].
We mutate all the instances to a class detached from Smalltalk and immune to some changes... Beware: don't you modify the layout of one superclass...
+1
2016-09-01 1:17 GMT+02:00 Nicolas Cellier <nicolas.cellier.aka.nice@ gmail.com>:
Due to read only ClassBinding it would be just a little bit more hackish:
(Smalltalk globals associationAt: #SystemProgressMorph) instVarAt: 2 put: SystemProgressMorph copy.
2016-09-01 1:12 GMT+02:00 Nicolas Cellier <nicolas.cellier.aka.nice@gmai l.com>:
Or would this even more simple hack work? Put in the preamble something like:
Smalltalk at: SystemProgressMorph put: SystemProgressMorph copy.
This way, the old instances should not be mutated...
2016-08-24 7:06 GMT+02:00 marcel.taeumel Marcel.Taeumel@hpi.de:
marcel.taeumel wrote
Bert Freudenberg wrote
On Tue, Aug 23, 2016 at 11:21 AM, marcel.taeumel <
Marcel.Taeumel@
> wrote:
> > Hi, there. > > It is, unfortunately, not possible to turn back time and rewrite > history. > ;-) >
... unless we're dealing with software. Oh wait, we are ;)
There is several important code that makes excessive use of instVar > accesses such as in SystemProgressMorph or HandMorph. Making any
change
> in > their class layout while using some of their methods will provoke a > strange > error. This could only be changed by a VM that keeps old class > layouts/schemata around for some longer time until there is no old > compiled > method on any stack anymore.
Class layout changes and instance migration is handled fully in the
image
by class ClassBuilder. The VM is not at fault here.
> Usually, restarting the update process helps here. There is no way
to
> fix > that in order to interactively update from 5.0 to 5.1 without any > errors. >
We can retroactively change the config map that introduced the
problem.
I'm certain it would be possible to fix, but I am not entirely sure it's worth the trouble.
- Bert -
Hi Bert,
the config map will not help, we would have to change numerous MCZs
since
point X in time to rewrite history. :) "Oh, would we have always been using accessors here instead of instVar accessess..."
Best, Marcel
...even then, how do you update the code in image Y, which is going to be updated before it is updated? This is not even possible for every user's images out there... :-/
The Java Hotspot VM has a mapping table for changed instVar accesses. Maybe we could use this trick, too? Still, would only applies to future versions. We cannot rewrite history because it is not only one piece of software in a single version control system with a single version checked out. :)
Putting more brittle stuff into he VM is not the solution. We have all that we need above the VM that can be dynamically tailored to suit our needs.
Can we not put the progress morph in its own process, making it easier to tear down and restart?
Best, Marcel
-- View this message in context: http://forum.world.st/Re-Does- anyone-recognize-this-glitch-in-our-trunk-update-processing- was-Vm-dev-buildspurtrunkvmmaker-tp4912130p4912403.html Sent from the Squeak - Dev mailing list archive at Nabble.com.
_,,,^..^,,,_ best, Eliot
On Thursday, 1 September 2016, Nicolas Cellier < nicolas.cellier.aka.nice@gmail.com> wrote:
The copy is not completely clean because this assertion then fails:
(SystemProgressMorph superclass subclasses detect: [:e | e name = #SystemProgressMorph]) == SystemProgressMorph
IOW, the superclass is still pointing to the old class, so any change to superclass layout won't propagate to the new subclass... Bad.
But there's also this possibility:
| oldClass | oldClass := SystemProgressMorph copy become: SystemProgressMorph. SystemProgressMorph allInstancesDo: [:e | oldClass adoptInstance: e].
We mutate all the instances to a class detached from Smalltalk and immune to some changes...
This is an interesting idea.
Basically we want to prevent the existing instances to be migrated to the new class layout when we change the instance variables. One way is what Nicolas suggests, but another would be to just keep the migration from happening in the first place, in ClassBuilder.
Could this be a general approach? To never migrate instances that are receivers of a context? They would still be instances of the old (now obsolete) class version. They could even be kept in a weak array to be migrated later, once the contexts are no longer active. The processes could even be restarted to get rid of these contexts.
There are other Smalltalks (e.g. GemStone) that even allow different class versions to exist side-by-side. We do have class versions too, but they have only been used for migrating serialized instances when the instance variables did *not* change (which is how the need for migration is normally detected). Hmm, maybe that is actually irrelevant, just ran through my mind ;)
So what do you think about delayed migration of instances on the stack?
- Bert -
squeak-dev@lists.squeakfoundation.org
-
Bert Freudenberg -
Chris Muller -
David T. Lewis -
Eliot Miranda -
Levente Uzonyi -
marcel.taeumel -
Nicolas Cellier