[gentoo-dev] default php-ext status

[gentoo-dev] default php-ext status

[Jaco Kroon]

Hi All,

For PHP modules that's getting installed, by default they all link 
ext/${modulename}.ini into ext-active/ such that all modules by default 
is active for all SAPIs.

Short of INSTALL_MASK, is there way to better control for sysadmins?

If I rm the symlinks, on next remerge they restore themselves:

plastiekpoot [16:43:09] /etc/php (master) # ls -lah */ext-active/xdebug.ini
lrwxrwxrwx 1 root root 17 Apr 22 15:36 
apache2-php8.1/ext-active/xdebug.ini -> ../ext/xdebug.ini
lrwxrwxrwx 1 root root 17 Apr 22 15:36 cli-php8.1/ext-active/xdebug.ini 
-> ../ext/xdebug.ini
lrwxrwxrwx 1 root root 17 Apr 22 15:36 fpm-php8.1/ext-active/xdebug.ini 
-> ../ext/xdebug.ini
plastiekpoot [16:43:15] /etc/php (master) # rm 
apache2-php8.1/ext-active/xdebug.ini
plastiekpoot [16:43:21] /etc/php (master) # emerge -av xdebug
...
plastiekpoot [16:43:52] /etc/php (master) # ls -lah */ext-active/xdebug.ini
lrwxrwxrwx 1 root root 17 Apr 22 16:43 
apache2-php8.1/ext-active/xdebug.ini -> ../ext/xdebug.ini
lrwxrwxrwx 1 root root 17 Apr 22 16:43 cli-php8.1/ext-active/xdebug.ini 
-> ../ext/xdebug.ini
lrwxrwxrwx 1 root root 17 Apr 22 16:43 fpm-php8.1/ext-active/xdebug.ini 
-> ../ext/xdebug.ini

Which in my *opinion* is not desirable behaviour, but I'm open for 
discussion.

xdebug.mode = off by default, but the extension still gets loaded.

Not sure if there is a sensible "solution", further expanding USE flags 
certainly is not a desirable option, so perhaps INSTALL_MASK 
(per-package env) is the best solution ... ?

Kind regards,
Jaco

Re: [gentoo-dev] default php-ext status

[Michael Orlitzky]

On Mon, 2024-04-22 at 16:46 +0200, Jaco Kroon wrote:
> 
> Which in my *opinion* is not desirable behaviour, but I'm open for 
> discussion.
> 
> xdebug.mode = off by default, but the extension still gets loaded.
> 
> Not sure if there is a sensible "solution", further expanding USE flags 
> certainly is not a desirable option, so perhaps INSTALL_MASK 
> (per-package env) is the best solution ... ?

You can replace the symlinks with empty files. The package manager
should then refuse to overwrite them. It _looks_ wrong (why are these
non-active extensions in ext-active?), but it's the easiest solution.

I think the default should be for the extensions to be per-SAPI
disabled and for users to enable them by creating the symlinks. (The
"ext" and "ext-active" design screams this.) That would be an easy
change to php-ext-source-r3.eclass but risks breaking everyone's web
servers if we do it without news items and/or ewarnings.

There's nobody in the PHP project at the moment. If you want to make an
-r4 of the eclass and handle the paperwork, by all means...

Re: [gentoo-dev] default php-ext status

[Jaco Kroon]

[-- Attachment #1: Type: text/plain, Size: 3663 bytes --]

Hi,

On 2024/04/22 18:49, Michael Orlitzky wrote:
> On Mon, 2024-04-22 at 16:46 +0200, Jaco Kroon wrote:
>> Which in my *opinion* is not desirable behaviour, but I'm open for
>> discussion.
>>
>> xdebug.mode = off by default, but the extension still gets loaded.
>>
>> Not sure if there is a sensible "solution", further expanding USE flags
>> certainly is not a desirable option, so perhaps INSTALL_MASK
>> (per-package env) is the best solution ... ?
> You can replace the symlinks with empty files. The package manager
> should then refuse to overwrite them. It _looks_ wrong (why are these
> non-active extensions in ext-active?), but it's the easiest solution.
>
> I think the default should be for the extensions to be per-SAPI
> disabled and for users to enable them by creating the symlinks. (The
> "ext" and "ext-active" design screams this.) That would be an easy
> change to php-ext-source-r3.eclass but risks breaking everyone's web
> servers if we do it without news items and/or ewarnings.

eselect php-ext perhaps?Both to configure the defaults for new 
modules/sapi/version kind of thing (perhaps layered), and then hook into 
that from the eclass?

> There's nobody in the PHP project at the moment. If you want to make an
> -r4 of the eclass and handle the paperwork, by all means...

What?!?

Surprise!  Wish I had capacity for volunteering more time.  Happy to do 
drive-by's for PHP stuff we use, but definitely not in a position to "be 
the project" kind of thing.

Perhaps I can start off on an eselect-php-extensions package (or expand 
eselect-php?), and once at least one other party is happy with the *UI* 
component and how that works, we can update the eclass to use that 
rather than blindly installing the symlinks?

eselect php-ext list [version]
eselect php-ext enable {extension} [version] [sapi]
eselect php-ext disable {extension} [version] [sapi]

Where we have one "special extention" called _default_ and one special 
version called "all" (and sapi "all").  These shall be used for any 
tuple for which there is no explicit configuration, so for any specific 
(pkg, version, sapi) tuple we check in this order for available configs:

1. (pkg, version, sapi)
2. (pkg, all, sapi)
3. (pkg, version, all)
4. (pkg, all, all)
5-8. repeat of 1-4 but with _default_.

It's unclear what the preferred order for 2,3 should be, however, based 
on our requirements sapi should take priority over version as this will 
be the most common enable/disable in my experience.

Once that works, then eclass updates such that eclass doesn't install 
any symlinks for ext-active/ as part of pkg_install.

postinstall in eclass enumerates the relevant php versions for which the 
package was installed, along with sapi's and check which symlinks should 
be created.
prerm does opposite (assuming we're not replacing a version, since I 
believe postinstall would invoke prior to the old version's prerm?).

Possibly easier to just not remove symlinks in prerm but provide a 
eselect php-ext cleanup which will cleanup broken symlinks in /etc/php 
in postrm?

Also, a way to cleanup the stored configuration for enabled/disabled.

eclass maintains current behaviour if eselect-php-extensions is not 
installed (which may become a requirement later, but how do we handle 
users making manual additions/modifications to /etc/php/ or have some 
form of "fsck" for that?)

SLOT'ing here may require some careful handling since for example 
pecl-parallel had a version for php7* (pecl-parallel 1.1.*) and another 
fo php8* (1.2.*).

Don't think this is undo-able but it'll be a first for me so may be a 
few rounds of back&forth.

Kind regards,
Jaco

[-- Attachment #2: Type: text/html, Size: 4750 bytes --]

Re: [gentoo-dev] default php-ext status

[Michael Orlitzky]

On Tue, 2024-04-23 at 13:05 +0200, Jaco Kroon wrote:
> but how do we handle 
> users making manual additions/modifications to /etc/php/ or have some 
> form of "fsck" for that?)

We'd have to use a new location that doesn't collide with anything that
the user might edit, like /etc/php/*/ext-eselect, the contents of which
would then be included into the active config.

The rest of the design sounds reasonable, but I would caution you to
take it one step at a time. For now, managing extensions is a matter of
editing one-line text files under /etc. They can be stored in git and
don't require any special tooling. Updates are protected by
CONFIG_PROTECT. Aside from the issue that started this thread, it works
pretty well.

Having an eselect-foo that is tied to certain versions of packages is
no fun. To some extent, eselect-php is like this. If we want to add a
new major version of php, eselect-php needs to know about it.
Conversely, if we want to change some aspect of eselect-php, then all
versions of php need to tolerate it. You can put blockers in the
ebuilds, or maintain compatibility until it's no longer needed (years,
for PHP), but no matter what you do it's more work than if you didn't
have to do it.

In any case, the default for the eclass will have to switch to
extensions being off-by-default. If you want to dive in to this, I
would suggest doing that first, and then giving it a month or so to
decide whether or not you still care. IMO writing a new eselect script
will be a lot of work for a little gain at that point.