start: Make start-browser synchronous.

[Ambrevar]

Description

There is no point in having it asynchronous, since anything run in parallel before browser is ready is likely to trigger a race conditions or undefined behaviour.

It simplifies testing a lot.

Fixes #2718 (comment).

Discussion

• Edit tests to not use nyxt:after-startup-hook since it's not needed anymore.
• Better name for the promise? @aartaka ?
• I tried to see if we could generalize the concept, for instance come up with a macro that would automatically synchronize asynchronous functions. It's difficult, because we don't really have a convention for these async functions and it heavily depends on WebKitGTK. For instance, buffer-load is "finished" when the on-signal-load-finished signal is triggered, in which we call the buffer-loaded-hook. This is completely different from start-browser. Can we devise a convention somehow?

Checklist:

Everything in this checklist is required for each PR. Please do not approve a PR that does not have all of these items.

• I have pulled from master before submitting this PR
• There are no merge conflicts.
• I've added the new dependencies as:
  • ASDF dependencies,
  • Git submodules,

    cd /path/to/nyxt/checkout
    git submodule add https://gitlab.common-lisp.net/nyxt/py-configparser _build/py-configparser

  • and Guix dependencies.
• My code follows the style guidelines for Common Lisp code. See:
  • Norvig & Pitman's Tutorial on Good Lisp Programming Style (PDF)
  • Google Common Lisp Style Guide
• I have performed a self-review of my own code.
• My code has been reviewed by at least one peer. (The peer review to approve a PR counts. The reviewer must download and test the code.)
• Documentation:
  • All my code has docstrings and :documentations written in the aforementioned style. (It's OK to skip the docstring for really trivial parts.)
  • I have updated the existing documentation to match my changes.
  • I have commented my code in hard-to-understand areas.
  • I have updated the changelog.lisp with my changes if it's anything user-facing (new features, important bug fix, compatibility breakage).
  • I have added a migration.lisp entry for all compatibility-breaking changes.
  • (If this changes something about the features showcased on Nyxt website) I have these changes described in the new/existing article at Nyxt website or will notify one of maintainters to do so.
• Compilation and tests:
  • My changes generate no new warnings.
  • I have added tests that prove my fix is effective or that my feature works. (If possible.)
  • New and existing unit tests pass locally with my changes.

[aartaka]

I like how it looks! But I want to see the changes to tests too, so that we see how it works in headless scenarios.

• Better name for the promise? @aartaka ?

No, seems fine. Along the lines of after-startup-hook etc.

• I tried to see if we could generalize the concept, for instance come up with a macro that would automatically synchronize asynchronous functions. It's difficult, because we don't really have a convention for these async functions and it heavily depends on WebKitGTK. For instance, buffer-load is "finished" when the on-signal-load-finished signal is triggered, in which we call the buffer-loaded-hook. This is completely different from start-browser. Can we devise a convention somehow?

Hmmmmm. The case of buffer-load is even more complex, though: is it finished when on-signal-load-finished fires, or when all the resources are loaded? The latter is not hookable, although it's only then that the page is fully interactive and the loading is done, as far as user point of view is concerned...

So there seems to be no solid convention due to lack of clear meaning for things. That's the same problem I've encountered when trying to abstract our concurrency patterns: we don't have many patterns at all. There are lots of unique cases which are not abstractable because of their.

The only reliable option I can think of for this (a)synchronization is to store the promise/completion predicate inside the function itself. Completion predicate would mean less lpara:fulfill across the code and more autonomy, but more evaluation issues. In any case, storing things inside a function is complicated: MOP, our own function class, define-function macro as a sibling to define-class smartness. That's kinda scary, but curious in the possibilities it opens :D

[aadcg]

If nyxt:start is synchronous shouldn't the following work?

CL-USER> (progn (nyxt:start :failsafe t) (nyxt:quit))
Nyxt version 3-pre-release-2-812-g007b59d1b
<INFO> [11:10:28] Source location: #P"/home/aadcg/common-lisp/nyxt/"
<INFO> [11:10:28] Profile: "nofile"
<INFO> [11:10:28] GTK extensions directory: #P"/home/aadcg/common-lisp/nyxt/libraries/web-extensions/"
<INFO> [11:10:29] GTK extensions directory: #P"/home/aadcg/common-lisp/nyxt/libraries/web-extensions/"
<INFO> [11:10:29] Loading #P"/home/aadcg/.local/share/nyxt/history/default.lisp".
NIL
<WARN> [11:10:30] Warning: decide policy error: There is no applicable method for the generic function
                       #<STANDARD-GENERIC-FUNCTION NYXT/RENDERER/GTK::ON-SIGNAL-DECIDE-POLICY (1)>
                     when called with arguments
                       (#<WEB-BUFFER 1269 {100AF24363}>
                        #<CL-WEBKIT2:WEBKIT-NAVIGATION-POLICY-DECISION {100CB77AD3}>
                        :WEBKIT-POLICY-DECISION-TYPE-NAVIGATION-ACTION).
See also:
  The ANSI Standard, Section 7.6.6
<WARN> [11:10:30] Warning: decide policy error: There is no primary method for the generic function
                       #<STANDARD-GENERIC-FUNCTION NYXT:BUFFERS (4)>
                     when called with arguments
                       (NIL).
See also:
  The ANSI Standard, Section 7.6.6.2

[aadcg]

Also, I see little sense in having a separate PR for this. Can I move this commit and the thread to #2718? @aartaka

[aartaka]

Hmmmm. I'd rather separate those two, because synchronous start is quite a set of concepts and problems. Would be beneficial to resolve this PR on its own to not bring more issues into the same context.

[aadcg]

Ok, then I'll wait for developments on this PR.

[Ambrevar]

If nyxt:start is synchronous shouldn't the following work?

Yes it should. Will investigate.

[Ambrevar]

There are lots of unique cases which are not abstractable because of their.

@aartaka Looks like you forgot a word :)

[Ambrevar]

(progn (nyxt:start :failsafe t) (nyxt:quit)) fails because threads are terminated out of order, and *browser* is set to NIL.

The fundamental reason behind this is given by SB-THREAD:TERMINATE-THREAD:

The unwind caused by TERMINATE-THREAD is asynchronous [...]

We could use the unportable sb-sys:without-interrupts. Not so great...

Turns out that quiting is really hard... Any better idea?

In any case, this is a general issue and goes beyond this pull request, so I don't think it's a blocker to merge this.

[Ambrevar]

@aartaka Tests don't use after-startup-hook anymore.

[aadcg]

Thanks @Ambrevar. Ok to merge?

[aartaka]

Neat! This PR looks just perfect to me now 🤩

[Ambrevar]

Done!

[aartaka]

Thanks, Pierre (´∀｀)♡