Hi all,

Now I had some minutes to take a further look at this.

** Mariano, yes, I remember that conversation. What we propose is aligned to that:

FileSystem workingDirectory shouldBe == getcwd

And **almost always**

FileSystem workingDirectory shouldBe == ��$PWD

This is because the real working directory is the one held by the process. $PWD is managed by bash and it can get out of sync with getcwd. For more info take a look at rajula's blogpost in the first email.

** About security, I don't really buy the argument.

��- There are many aspects of security: internal in the image, external. FFI is not the only thing to *care about*

��- right now we can do stuff like

�� - FileSystem root ensureDeleteAll (and given pharo started with sudo...)

�� - nil become: true

�� - OpalCompiler compile: 'someMethodThatOverriddenBreaksEverything ...'

��- FFI plugin is shipped with the VM

�� - so removing FFI support is as easy as removing the plugin

�� - but this means that even if that FFI support is not in the image, any piece of code can do:

�� - Myclass compile: 'test Metacello install: ''FFI''; please'; test.

�� - or create a method calling the FFI primitives

��

��This does not mean I don't think about security, but I believe it should be managed in some other way. A security solution for real needs to manage reflection as well as external resources, and control how dependencies and packages are loaded. Because, moreover, I believe that while you may want to forbid FFI calls in some part of your app, you may want that some other parts of the system can use FFI.

** About modularity. But I'm sure I'm one of the guys that cares the most about modularity and cries aloud every time we introduce bad dependencies. Now, I'm proposing the introduction of FFI because I think its manageable and it's being controlled. It does not affect the bootstrap, and I don't think it will forbid us to get even smaller kernels.

** The most important argument, that at least I did not see anybody complain about, is backwards compatibility. This implementation may break backwards compatibility, because this does not hold anymore:

FileSystem imageDirectory == FileSystem workingDirectory

Well, yes, it holds if you start your vm/image from the same directory where the image is located. Which is what most users do.

** I also talked with Esteban about his reasons. His concern was about consistency. If you take a look at how getEnv is resolved:

getEnv: varName

"This method calls the Standard C Library getenv() function."

"OSEnvironment current getEnv: 'HOME' "

<primitive: 'primitiveGetenv' module: '' error: ec>

ec ifNil: [ ^self getEnvViaFFI: varName ].

ec == #'bad argument' ifTrue: [

varName isString��

ifFalse: [ ^self getEnv: varName asString ] ].

self primitiveFail

If first tries to call a primitive and then, as fallback, it calls FFI by default.

I changed the pull request to do the same thing about the working directory. That way, as soon as we have implemented the primitive, that will be in usage instead of the FFI call.

Guille

On Wed, Oct 11, 2017 at 3:07 PM, Denis Kudriashov <dionisiydk@gmail.com> wrote:

2017-10-11 14:14 GMT+02:00 Alistair Grant <akgrant0710@gmail.com>:
On Wed, Oct 11, 2017 at 01:32:52PM +0200, Denis Kudriashov wrote:
> 2017-10-11 10:54 GMT+02:00 Esteban Lorenzano <estebanlm@gmail.com>:
>
>�� in general, this core-core functions is better to implement them as
>�� primitives.
>�� and you put a fallback using UFFI (in case the primitive is not there).
>�� of course, your solution works? but it adds dependency with FFI-Kernel,
>�� which is also not good to have it in the bootstrap IMO.
>
>�� take a look at OSEnvironment>>getEnv implementation.
>
>�� we can go that direction and I think it will be better.
>
>
> +1 for new primitive. Because instead any new platform will require implemented
> FFI which is not easy.

Would it make sense to add this to FilePlugin?�� The plugin has to be
working for general file system access anyway.

I think it is good idea
��

Cheers,
Alistair

>�� Esteban
>
>
>�� On 11 Oct 2017, at 10:27, Guillermo Polito <guillermopolito@gmail.com>
>�� wrote:
>
>�� Hi all,
>
>�� I'd like to push a really core change in file management: the working
>�� directory. This is really needed for command line apps, when you have
>�� your app deployed in some directory and you're launching it from
>�� another one. The current implementation, where workingDirectory =
>�� imageDirectory, forces to have absolute paths or extra handling all
>�� over the place to manage this complexity.
>
>�� Rajula proposed a couple of months ago a solution for this based on the
>�� getcwd functions. You can read in his blogpost why using getcwd is
>�� better than $PWD in general here:
>
>�� https://vineetreddy.wordpress.com/2017/05/17/pwd-vs-getcwd/
>
>�� Now, since accessing the working directory is a core part of Pharo but
>�� based on UFFI, his implementation was breaking the build process. We
>�� cannot and we will not integrate UFFI in the bootstrap because it
>�� depends mainly on the compiler which is a big beast. Instead, I propose
>�� that only for this core-core-core feature, we use directly FFI.
>
>�� In other words, the bootstrap will include just a couple of classes to
>�� manage the basics of FFI. And the working directory will be fetched by
>�� using this low level API. Such a call looks like this:
>
>�� (ExternalLibraryFunction
>�� name: 'getcwd'
>�� module: 'libc.dylib'
>�� callType: 1
>�� returnType: ExternalType char asPointerType
>�� argumentTypes: {
>�� ExternalType char asPointerType.
>�� ExternalType long })
>�� invokeWith: buffer with: bufferSize.
>
>�� We reviewed it with Pablo two days ago. The build process works with
>�� this implementation and tests are still running. The pull request is in
>�� here:
>
>�� https://github.com/pharo-project/pharo/pull/92
>
>
>�� Thanks,
>�� Guille, Pablo and Rajula
>
>�� --

>�� [presentation]�� Guille Polito
>�� [CNRS-filaire]�� Research Engineer

>
>
>�� Centre de Recherche en Informatique, Signal et
>�� Automatique de Lille
>�� CRIStAL - UMR 9189
>�� French National Center for Scientific Research -
>�� http://www.cnrs.fr
>
>�� Web: http://guillep.github.io
>�� Phone: +33 06 52 70 66 13
>
>
>
>

��	Guille Polito Research Engineer Centre de Recherche en Informatique, Signal et Automatique de Lille CRIStAL - UMR 9189 French National Center for Scientific Research - http://www.cnrs.fr Web: http://guillep.github.io Phone: +33 06 52 70 66 13