Reproducible VM crash when loading FFI

 
Eliot, list, I've uploaded an image created with the code I posted and uploaded to http://www.filebox.com/d4t6rnpa0fqv

Hope it helps,

Guille

On Sun, Mar 24, 2013 at 10:10 PM, Nicolas Cellier <[hidden email]> wrote:

2013/3/23 Guillermo Polito <[hidden email]>:

>
> Hi!
>
> In my quest to crash the vm, i've found an ugly common case :(.
> I am trying to port the opendbx driver to 2.0, but I'm getting vm crashes when my configuration loads FFI :(. I updated my configuration to load version 1.7 of FFI (which I assume is the latest).
>
> I tried to do it in Pharo 2.0 with latest pharovm, and with eliot's Cog 2701 from his website, failing in both cases.
> The snippet of code that gets a sistematic crash is:
>
> Gofer it
> smalltalkhubUser: 'DBXTalk' project: 'DBXTalkDriver';
> package: 'ConfigurationOfOpenDBXDriver';
> load.
> ((Smalltalk at: #ConfigurationOfOpenDBXDriver) project version:#stable) load
>

Same thing with ConfigurationOfSmallapack.
You should then execute

Smalltalk saveAs: 'crash.image' thenQuit: true.

Gofer it
   smalltalkhubUser: 'DBXTalk' project: 'DBXTalkDriver';
    package: 'ConfigurationOfOpenDBXDriver';
    load.

and send the image to Eliot.

>
> This snippet crashes always with a segmentation fault (at least the fifteen times i tried :), with several different output in the console...
>
> Surprisingly, if I load FFI alone and not from the OpenDBXDriver configuration, it loads well... :/
>
> So I'm deferring the OpenDBX port a bit longer :(
>
> Thanks!
> Guille
>

Thanks for the findings!

I opened an issue http://bugs.pharo.org/issues/id/10134 with the contents
of your solution.

I think for completeness we should implement both solution, and of course
use the one that does not #becomeForward: in the standard case.

On 2013-03-25, at 19:52, Eliot Miranda <[hidden email]> wrote:
> Hi Guille,
>
>    thanks for this.  The problem is that Pharo's
> recreateSpecialObjectsArray uses newCompactClassesArray to recreate the
> compact classes array and on Cog that is not supported.  Cog caches the
> compactClassesArray (specialObjectsArray at: 29) in every jitted method
> prolog to reduce the time taken to derive the receiver's class in message
> lookup.  The Pharo 2.0 code for recreating the specialObjectsArray can
> therefore create a dangling pointer where the only reference to the old
> compactClassesArray is in machine code, and the machine code GC doesn't
> cope with there being roots in machine code.  Note that Cog also caches the
> characterTable and class SmallInteger in machine code.
>
> There are two solutions to this.
> One would be to reuse the compactClassesArray (my recommendation).  So that
> recreateSpecialObjectsArray looks like
>
> ...
> "An array of the 255 Characters in ascii order.
> Cog inlines table into machine code at: prim so do not regenerate it."
> newArray at: 25 put: Character characterTable.
> ...
> "A 32-element array with up to 32 classes that have compact instances.
> Cog inlines table into machine code class lookup so do not regenerate it."
> newArray at: 29 put: self compactClassesArray.
> ...
>
> where compactClassesArray, like Character characterTable, answers the
> existing object.
>
> The second solution (rather hackish) is to void all machine code on
> installing the new specialObjectsArray.  e.g.
> recreateSpecialObjectsArray
> "Smalltalk recreateSpecialObjectsArray"
> "To external package developers:
> **** DO NOT OVERRIDE THIS METHOD.  *****
> If you are writing a plugin and need additional special object(s) for your
> own use,
> use addGCRoot() function and use own, separate special objects registry "
> "The Special Objects Array is an array of objects used by the Squeak
> virtual machine.
> Its contents are critical and accesses to it by the VM are unchecked, so
> don't even
> think of playing here unless you know what you are doing."
> "Replace the interpreter's reference in one atomic operation.
> Void machine code to avoid crashing Cog."
> | newSpecialObjectsArray |
> newSpecialObjectsArray := self newSpecialObjectsArray.
> self specialObjectsArray becomeForward: newSpecialObjectsArray.
> Smalltalk vm voidCogVMState
>
>
> this is my fault for not ensuring recreateSpecialObjectsArray was properly
> commented.  I had commented the inlining of Character table, but not the
> inlining of the CompactClasses array. Apologies.
>
> HTH,
> Eliot
>
> On Sat, Mar 23, 2013 at 10:19 AM, Guillermo Polito <
> [hidden email]> wrote:
>
>> Hi!
>>
>> In my quest to crash the vm, i've found an ugly common case :(.
>> I am trying to port the opendbx driver to 2.0, but I'm getting vm crashes
>> when my configuration loads FFI :(. I updated my configuration to load
>> version 1.7 of FFI (which I assume is the latest).
>>
>> I tried to do it in Pharo 2.0 with latest pharovm, and with eliot's Cog
>> 2701 from his website, failing in both cases.
>> The snippet of code that gets a sistematic crash is:
>>
>> Gofer it
>> smalltalkhubUser: 'DBXTalk' project: 'DBXTalkDriver';
>> package: 'ConfigurationOfOpenDBXDriver';
>> load.
>> ((Smalltalk at: #ConfigurationOfOpenDBXDriver) project version:#stable)
>> load
>>
>>
>> This snippet crashes always with a segmentation fault (at least the
>> fifteen times i tried :), with several different output in the console...
>>
>> Surprisingly, if I load FFI alone and not from the OpenDBXDriver
>> configuration, it loads well... :/
>>
>> So I'm deferring the OpenDBX port a bit longer :(
>>
>> Thanks!
>> Guille
>>
>
>
>
> --
> best,
> Eliot

 
Ok, first I'm sorry for the evil upload. That was the first at hand :$. I'll get a more trustful way to do it the next one...

Second, thanks for taking the time to look at it :).

Now, I've some questions:

- this "bug" is particular to Cog, isn't it? I mean, only related to inlining with JIT, so a Stack vm does not play with these rules, right?

- Why is voiding the cache hackish? I mean, replacing the special objects array is a very low level operation, and a very special one which is not commonly performed. Voiding the cache looks like normal to me in such a case...

Tx!

Guille

On Mon, Mar 25, 2013 at 11:56 PM, Eliot Miranda <[hidden email]> wrote:

 

On Mon, Mar 25, 2013 at 3:22 PM, Camillo Bruni <[hidden email]> wrote:

Thanks for the findings!

you're welcome :)

 

I opened an issue http://bugs.pharo.org/issues/id/10134 with the contents
of your solution.

I think for completeness we should implement both solution, and of course
use the one that does not #becomeForward: in the standard case.

To be clear, becomeForward: is the correct way to install the specialObjectsArray.  The issue is what entries in the specialObjectsArray need to remain constant.  The Character table, the compactClassesArray, all the classes and a few other entries (semaphores, etc) must remain the same.  This could do with better documenting of the method :-/

On 2013-03-25, at 19:52, Eliot Miranda <[hidden email]> wrote:
> Hi Guille,
>
>    thanks for this.  The problem is that Pharo's
> recreateSpecialObjectsArray uses newCompactClassesArray to recreate the
> compact classes array and on Cog that is not supported.  Cog caches the
> compactClassesArray (specialObjectsArray at: 29) in every jitted method
> prolog to reduce the time taken to derive the receiver's class in message
> lookup.  The Pharo 2.0 code for recreating the specialObjectsArray can
> therefore create a dangling pointer where the only reference to the old
> compactClassesArray is in machine code, and the machine code GC doesn't
> cope with there being roots in machine code.  Note that Cog also caches the
> characterTable and class SmallInteger in machine code.
>
> There are two solutions to this.
> One would be to reuse the compactClassesArray (my recommendation).  So that
> recreateSpecialObjectsArray looks like
>
> ...
> "An array of the 255 Characters in ascii order.
> Cog inlines table into machine code at: prim so do not regenerate it."
> newArray at: 25 put: Character characterTable.
> ...
> "A 32-element array with up to 32 classes that have compact instances.
> Cog inlines table into machine code class lookup so do not regenerate it."
> newArray at: 29 put: self compactClassesArray.
> ...
>
> where compactClassesArray, like Character characterTable, answers the
> existing object.
>
> The second solution (rather hackish) is to void all machine code on
> installing the new specialObjectsArray.  e.g.
> recreateSpecialObjectsArray
> "Smalltalk recreateSpecialObjectsArray"
> "To external package developers:
> **** DO NOT OVERRIDE THIS METHOD.  *****
> If you are writing a plugin and need additional special object(s) for your
> own use,
> use addGCRoot() function and use own, separate special objects registry "
> "The Special Objects Array is an array of objects used by the Squeak
> virtual machine.
> Its contents are critical and accesses to it by the VM are unchecked, so
> don't even
> think of playing here unless you know what you are doing."
> "Replace the interpreter's reference in one atomic operation.
> Void machine code to avoid crashing Cog."
> | newSpecialObjectsArray |
> newSpecialObjectsArray := self newSpecialObjectsArray.
> self specialObjectsArray becomeForward: newSpecialObjectsArray.
> Smalltalk vm voidCogVMState
>
>
> this is my fault for not ensuring recreateSpecialObjectsArray was properly
> commented.  I had commented the inlining of Character table, but not the
> inlining of the CompactClasses array. Apologies.
>
> HTH,
> Eliot
>
> On Sat, Mar 23, 2013 at 10:19 AM, Guillermo Polito <
> [hidden email]> wrote:
>
>> Hi!
>>
>> In my quest to crash the vm, i've found an ugly common case :(.
>> I am trying to port the opendbx driver to 2.0, but I'm getting vm crashes
>> when my configuration loads FFI :(. I updated my configuration to load
>> version 1.7 of FFI (which I assume is the latest).
>>
>> I tried to do it in Pharo 2.0 with latest pharovm, and with eliot's Cog
>> 2701 from his website, failing in both cases.
>> The snippet of code that gets a sistematic crash is:
>>
>> Gofer it
>> smalltalkhubUser: 'DBXTalk' project: 'DBXTalkDriver';
>> package: 'ConfigurationOfOpenDBXDriver';
>> load.
>> ((Smalltalk at: #ConfigurationOfOpenDBXDriver) project version:#stable)
>> load
>>
>>
>> This snippet crashes always with a segmentation fault (at least the
>> fifteen times i tried :), with several different output in the console...
>>
>> Surprisingly, if I load FFI alone and not from the OpenDBXDriver
>> configuration, it loads well... :/
>>
>> So I'm deferring the OpenDBX port a bit longer :(
>>
>> Thanks!
>> Guille
>>
>
>
>
> --
> best,
> Eliot

--
best,

Eliot

 
Ok, as a summary. I've tried

1) not recreating the compact classes array without voiding vm caches: works ok.

2) making another compact classes array, voiding the vm caches (primitive 214): random crashes :/. Sometimes it works, some times it does not...

I've tried, just for the record, to void the cache before, after and (before and after) the become forward of the special objects array:

a)

| newSpecialObjectsArray |

self vm voidCogVMState.

newSpecialObjectsArray := self newSpecialObjectsArray.

self specialObjectsArray becomeForward: newSpecialObjectsArray.

b)

| newSpecialObjectsArray |

newSpecialObjectsArray := self newSpecialObjectsArray.

self specialObjectsArray becomeForward: newSpecialObjectsArray.

self vm voidCogVMState.

c)

| newSpecialObjectsArray |

self vm voidCogVMState.

newSpecialObjectsArray := self newSpecialObjectsArray.

self specialObjectsArray becomeForward: newSpecialObjectsArray.

self vm voidCogVMState.

All with different results when executing several times... :/

voidCogVMState is the following in Pharo:

VirtualMachine class>>voidCogVMState

"Void any internal caches the VM maintains other than the method lookup caches.

These comprise

- the stack zone, where method activations are stored, and

- the machine code zone, where the machine code form of CompiledMethods is held."

<primitive: 214>

^self primitiveFailed

The comment says it does not clean method lookup caches...

any ideas?

Guille

On Tue, Mar 26, 2013 at 11:48 AM, Guillermo Polito <[hidden email]> wrote:

Ok, first I'm sorry for the evil upload. That was the first at hand :$. I'll get a more trustful way to do it the next one...

Second, thanks for taking the time to look at it :).

Now, I've some questions:

- this "bug" is particular to Cog, isn't it? I mean, only related to inlining with JIT, so a Stack vm does not play with these rules, right?

- Why is voiding the cache hackish? I mean, replacing the special objects array is a very low level operation, and a very special one which is not commonly performed. Voiding the cache looks like normal to me in such a case...

Tx!

Guille

On Mon, Mar 25, 2013 at 11:56 PM, Eliot Miranda <[hidden email]> wrote:

 

On Mon, Mar 25, 2013 at 3:22 PM, Camillo Bruni <[hidden email]> wrote:

Thanks for the findings!

you're welcome :)

 

I opened an issue http://bugs.pharo.org/issues/id/10134 with the contents
of your solution.

I think for completeness we should implement both solution, and of course
use the one that does not #becomeForward: in the standard case.

To be clear, becomeForward: is the correct way to install the specialObjectsArray.  The issue is what entries in the specialObjectsArray need to remain constant.  The Character table, the compactClassesArray, all the classes and a few other entries (semaphores, etc) must remain the same.  This could do with better documenting of the method :-/

On 2013-03-25, at 19:52, Eliot Miranda <[hidden email]> wrote:
> Hi Guille,
>
>    thanks for this.  The problem is that Pharo's
> recreateSpecialObjectsArray uses newCompactClassesArray to recreate the
> compact classes array and on Cog that is not supported.  Cog caches the
> compactClassesArray (specialObjectsArray at: 29) in every jitted method
> prolog to reduce the time taken to derive the receiver's class in message
> lookup.  The Pharo 2.0 code for recreating the specialObjectsArray can
> therefore create a dangling pointer where the only reference to the old
> compactClassesArray is in machine code, and the machine code GC doesn't
> cope with there being roots in machine code.  Note that Cog also caches the
> characterTable and class SmallInteger in machine code.
>
> There are two solutions to this.
> One would be to reuse the compactClassesArray (my recommendation).  So that
> recreateSpecialObjectsArray looks like
>
> ...
> "An array of the 255 Characters in ascii order.
> Cog inlines table into machine code at: prim so do not regenerate it."
> newArray at: 25 put: Character characterTable.
> ...
> "A 32-element array with up to 32 classes that have compact instances.
> Cog inlines table into machine code class lookup so do not regenerate it."
> newArray at: 29 put: self compactClassesArray.
> ...
>
> where compactClassesArray, like Character characterTable, answers the
> existing object.
>
> The second solution (rather hackish) is to void all machine code on
> installing the new specialObjectsArray.  e.g.
> recreateSpecialObjectsArray
> "Smalltalk recreateSpecialObjectsArray"
> "To external package developers:
> **** DO NOT OVERRIDE THIS METHOD.  *****
> If you are writing a plugin and need additional special object(s) for your
> own use,
> use addGCRoot() function and use own, separate special objects registry "
> "The Special Objects Array is an array of objects used by the Squeak
> virtual machine.
> Its contents are critical and accesses to it by the VM are unchecked, so
> don't even
> think of playing here unless you know what you are doing."
> "Replace the interpreter's reference in one atomic operation.
> Void machine code to avoid crashing Cog."
> | newSpecialObjectsArray |
> newSpecialObjectsArray := self newSpecialObjectsArray.
> self specialObjectsArray becomeForward: newSpecialObjectsArray.
> Smalltalk vm voidCogVMState
>
>
> this is my fault for not ensuring recreateSpecialObjectsArray was properly
> commented.  I had commented the inlining of Character table, but not the
> inlining of the CompactClasses array. Apologies.
>
> HTH,
> Eliot
>
> On Sat, Mar 23, 2013 at 10:19 AM, Guillermo Polito <
> [hidden email]> wrote:
>
>> Hi!
>>
>> In my quest to crash the vm, i've found an ugly common case :(.
>> I am trying to port the opendbx driver to 2.0, but I'm getting vm crashes
>> when my configuration loads FFI :(. I updated my configuration to load
>> version 1.7 of FFI (which I assume is the latest).
>>
>> I tried to do it in Pharo 2.0 with latest pharovm, and with eliot's Cog
>> 2701 from his website, failing in both cases.
>> The snippet of code that gets a sistematic crash is:
>>
>> Gofer it
>> smalltalkhubUser: 'DBXTalk' project: 'DBXTalkDriver';
>> package: 'ConfigurationOfOpenDBXDriver';
>> load.
>> ((Smalltalk at: #ConfigurationOfOpenDBXDriver) project version:#stable)
>> load
>>
>>
>> This snippet crashes always with a segmentation fault (at least the
>> fifteen times i tried :), with several different output in the console...
>>
>> Surprisingly, if I load FFI alone and not from the OpenDBXDriver
>> configuration, it loads well... :/
>>
>> So I'm deferring the OpenDBX port a bit longer :(
>>
>> Thanks!
>> Guille
>>
>
>
>
> --
> best,
> Eliot

--
best,

Eliot

 
Hi Guille,

On Tue, Mar 26, 2013 at 9:33 AM, Guillermo Polito <[hidden email]> wrote:

 
Ok, as a summary. I've tried

1) not recreating the compact classes array without voiding vm caches: works ok.

2) making another compact classes array, voiding the vm caches (primitive 214): random crashes :/. Sometimes it works, some times it does not...

I've tried, just for the record, to void the cache before, after and (before and after) the become forward of the special objects array:

You need to void the cache *after* creating the new array and *immediately before* installing it.  Voiding the cache before creating it gives the VM too much time to run and recache.  So try

d)

| newSpecialObjectsArray |

newSpecialObjectsArray := self newSpecialObjectsArray.

self vm voidCogVMState.

self specialObjectsArray becomeForward: newSpecialObjectsArray

a)

| newSpecialObjectsArray |

self vm voidCogVMState.

newSpecialObjectsArray := self newSpecialObjectsArray.

self specialObjectsArray becomeForward: newSpecialObjectsArray.

b)

| newSpecialObjectsArray |

newSpecialObjectsArray := self newSpecialObjectsArray.

self specialObjectsArray becomeForward: newSpecialObjectsArray.

self vm voidCogVMState.

c)

| newSpecialObjectsArray |

self vm voidCogVMState.

newSpecialObjectsArray := self newSpecialObjectsArray.

self specialObjectsArray becomeForward: newSpecialObjectsArray.

self vm voidCogVMState.

All with different results when executing several times... :/

voidCogVMState is the following in Pharo:

VirtualMachine class>>voidCogVMState

"Void any internal caches the VM maintains other than the method lookup caches.

These comprise

- the stack zone, where method activations are stored, and

- the machine code zone, where the machine code form of CompiledMethods is held."

<primitive: 214>

^self primitiveFailed

The comment says it does not clean method lookup caches...

Right, and that's fine.  We're only wanting to discard cached machine code etc. 

any ideas?

As explained, try d).

 

Guille

On Tue, Mar 26, 2013 at 11:48 AM, Guillermo Polito <[hidden email]> wrote:

Ok, first I'm sorry for the evil upload. That was the first at hand :$. I'll get a more trustful way to do it the next one...

Second, thanks for taking the time to look at it :).

Now, I've some questions:

- this "bug" is particular to Cog, isn't it? I mean, only related to inlining with JIT, so a Stack vm does not play with these rules, right?

- Why is voiding the cache hackish? I mean, replacing the special objects array is a very low level operation, and a very special one which is not commonly performed. Voiding the cache looks like normal to me in such a case...

Tx!

Guille

On Mon, Mar 25, 2013 at 11:56 PM, Eliot Miranda <[hidden email]> wrote:

 

On Mon, Mar 25, 2013 at 3:22 PM, Camillo Bruni <[hidden email]> wrote:

Thanks for the findings!

you're welcome :)

 

I opened an issue http://bugs.pharo.org/issues/id/10134 with the contents
of your solution.

I think for completeness we should implement both solution, and of course
use the one that does not #becomeForward: in the standard case.

To be clear, becomeForward: is the correct way to install the specialObjectsArray.  The issue is what entries in the specialObjectsArray need to remain constant.  The Character table, the compactClassesArray, all the classes and a few other entries (semaphores, etc) must remain the same.  This could do with better documenting of the method :-/

On 2013-03-25, at 19:52, Eliot Miranda <[hidden email]> wrote:
> Hi Guille,
>
>    thanks for this.  The problem is that Pharo's
> recreateSpecialObjectsArray uses newCompactClassesArray to recreate the
> compact classes array and on Cog that is not supported.  Cog caches the
> compactClassesArray (specialObjectsArray at: 29) in every jitted method
> prolog to reduce the time taken to derive the receiver's class in message
> lookup.  The Pharo 2.0 code for recreating the specialObjectsArray can
> therefore create a dangling pointer where the only reference to the old
> compactClassesArray is in machine code, and the machine code GC doesn't
> cope with there being roots in machine code.  Note that Cog also caches the
> characterTable and class SmallInteger in machine code.
>
> There are two solutions to this.
> One would be to reuse the compactClassesArray (my recommendation).  So that
> recreateSpecialObjectsArray looks like
>
> ...
> "An array of the 255 Characters in ascii order.
> Cog inlines table into machine code at: prim so do not regenerate it."
> newArray at: 25 put: Character characterTable.
> ...
> "A 32-element array with up to 32 classes that have compact instances.
> Cog inlines table into machine code class lookup so do not regenerate it."
> newArray at: 29 put: self compactClassesArray.
> ...
>
> where compactClassesArray, like Character characterTable, answers the
> existing object.
>
> The second solution (rather hackish) is to void all machine code on
> installing the new specialObjectsArray.  e.g.
> recreateSpecialObjectsArray
> "Smalltalk recreateSpecialObjectsArray"
> "To external package developers:
> **** DO NOT OVERRIDE THIS METHOD.  *****
> If you are writing a plugin and need additional special object(s) for your
> own use,
> use addGCRoot() function and use own, separate special objects registry "
> "The Special Objects Array is an array of objects used by the Squeak
> virtual machine.
> Its contents are critical and accesses to it by the VM are unchecked, so
> don't even
> think of playing here unless you know what you are doing."
> "Replace the interpreter's reference in one atomic operation.
> Void machine code to avoid crashing Cog."
> | newSpecialObjectsArray |
> newSpecialObjectsArray := self newSpecialObjectsArray.
> self specialObjectsArray becomeForward: newSpecialObjectsArray.
> Smalltalk vm voidCogVMState
>
>
> this is my fault for not ensuring recreateSpecialObjectsArray was properly
> commented.  I had commented the inlining of Character table, but not the
> inlining of the CompactClasses array. Apologies.
>
> HTH,
> Eliot
>
> On Sat, Mar 23, 2013 at 10:19 AM, Guillermo Polito <
> [hidden email]> wrote:
>
>> Hi!
>>
>> In my quest to crash the vm, i've found an ugly common case :(.
>> I am trying to port the opendbx driver to 2.0, but I'm getting vm crashes
>> when my configuration loads FFI :(. I updated my configuration to load
>> version 1.7 of FFI (which I assume is the latest).
>>
>> I tried to do it in Pharo 2.0 with latest pharovm, and with eliot's Cog
>> 2701 from his website, failing in both cases.
>> The snippet of code that gets a sistematic crash is:
>>
>> Gofer it
>> smalltalkhubUser: 'DBXTalk' project: 'DBXTalkDriver';
>> package: 'ConfigurationOfOpenDBXDriver';
>> load.
>> ((Smalltalk at: #ConfigurationOfOpenDBXDriver) project version:#stable)
>> load
>>
>>
>> This snippet crashes always with a segmentation fault (at least the
>> fifteen times i tried :), with several different output in the console...
>>
>> Surprisingly, if I load FFI alone and not from the OpenDBXDriver
>> configuration, it loads well... :/
>>
>> So I'm deferring the OpenDBX port a bit longer :(
>>
>> Thanks!
>> Guille
>>
>
>
>
> --
> best,
> Eliot

--
best,

Eliot

--
best,

Eliot