When I first announced that I was going to be looking at the roots of DocShell, and how it has changed over the years, I thought I was going to be leafing through a lot of old CVS commits in order to see what went on before the switch to Mercurial.
I thought it’d be so, and indeed it was so. And it was painful. Having worked with DVCS’s like Mercurial and Git so much over the past couple of years, my brain was just not prepared to deal with CVS.
My solution? Take the old CVS tree, and attempt to make a Git repo out of it. Git I can handle.
“But why isn’t the CVS history in the Mercurial tree?” I hear you ask. And that’s a good question. It might have to do with the fact that converting the CVS history over is bloody hard – or at least that was my experience. cvs2git has the unfortunate habit of analyzing the entire repository / history and spitting out any errors or corruptions it found at the end.1 This is fine for small repositories, but the Mozilla CVS repo (even back in 1999) was quite substantial, and had quite a bit of history.
So my experience became: run cvs2git, wait 25 minutes, glare at an error message about corruption, scour the web for solutions to the issue, make random stabs at a solution, and repeat.
Not the greatest situation. I did what most people in my position would do, and cast my frustration into the cold, unfeeling void that is Twitter.
But, lo and behold, somebody on the other side of the planet was listening. Unfocused informed me that whoever created the gecko-dev Github mirror somehow managed to type in the black-magic incantations that would import all of the old CVS history into the Git mirror. I simply had to clone gecko-dev, and I was golden.
Thanks Unfocused. 🙂
So I had my tree. I cracked open Gitx, put some tea on, and started pouring through the commits from the initial creation of the docshell folder (October 15, 1999) to the last change in that folder just before the switch over to 2005 (December 15, 2004)2.
The following are my notes as I peered through those commits.
That’s the message for the first commit when the docshell/ folder was first created by Travis Bogard. Correction (Jan 3rd, 2015) – this is actually not true at all. The commit I’ve linked to here is the first commit for nsDocShell.cpp, but the docshell/ folder was actually created over a year earlier. It looks like the “docshell” monicker was selected before any of Travis’ refactoring had even begun3.
Without even looking at the code, that’s a pretty strange commit just by the message alone. No bug number, no reviewer, no approval, nothing even approximating a vague description of what was going on.
Leafing through these early commits, I was surprised to find that quite common. In fact, I learned that it was about a year after this work started that code review suddenly became mandatory for commits.
So, for these first bits of code, nobody seems to be reviewing it – or at least, nobody is signing off on it in commit messages.
Like I mentioned, the date for this commit is October 15, 1999. If the timeline in this Wikipedia article about the history of the Mozilla Suite is reliable, that puts us somewhere between Milestone 10 and Milestone 11 of the first 1.0 Mozilla Suite release.4
Before we continue, who is this intrepid Travis Bogard who is spearheading this embedding initiative and the DocShell / WebShell rewrite?
At the time, according to his LinkedIn page, he worked for America Online (which at this point in time owned Netscape.5) He’d been working for AOL since 1996, working his way up the ranks from lowly intern all the way to Principal Software Engineer.
Travis was the one who originally wrote the wiki page about how painful it was embedding the web engine, and how unwieldy nsWebShell was.6 It was Travis who led the charge to strip away all of the complexity and mess inside of WebShell, and create smaller, more specialized interfaces for the behind the scenes DocShell class, which would carry out most of the work that WebShell had been doing up until that point.7
So, for these first few months, it was Travis who would be doing most of the work on DocShell.
These first few months, Travis puts the pedal to the metal moving things out of WebShell and into DocShell. Remember – the idea was to have a thin, simple nsWebBrowser that embedders could touch, and a fat, complex DocShell that was privately used within Gecko that was accessible via many small, specialized interfaces.
Wholesale replacing or refactoring a major part of the engine is no easy task, however – and since WebShell was core to the very function of the browser (and the mail/news client, and a bunch of other things), there were two copies of WebShell made.
The original WebShell existed in webshell/ under the root directory. The second WebShell, the one that would eventually replace it, existed under docshell/base. The one under docshell/base is the one that Travis was stripping down, but nobody was using it until it was stable. They’d continue using the one under webshell/, until they were satisfied with their implementation by both manual and automated testing.
When they were satisfied, they’d branch off of the main development line, and start removing occurances of WebShell where they didn’t need to be, and replace them with nsWebBrowser or DocShell where appropriate. When they were done that, they’d merge into main line, and celebrate!
At least, that was the plan.
That plan is spelled out here in the Plan of Attack for the redesign. That plan sketches out a rough schedule as well, and targets November 30th, 1999 as the completion point of the project.
This parallel development means that any bugs that get discovered in WebShell during the redesign needs to get fixed in two places – both under webshell/ and docshell/base.
Breaking up is so hard to do
So what was actually changing in the code? In Travis’ first commit, he adds the following interfaces:
along with some build files. Something interesting here is this nsIHTMLDocShell – where it looked like at this point, the plan was to have different DocShell interfaces depending on the type of document being displayed. Later on, we see this idea dissipate.
If DocShell was a person, these are its baby pictures. At this point, nsIDocShell has just two methods: LoadDocument, LoadDocumentVia, and a single nsIDOMDocument attribute for the document.
And here’s the interface for WebShell, which Travis was basing these new interfaces off of. Note that there’s no LoadDocument, or LoadDocumentVia, or an attribute for an nsIDOMDocument. So it seems this wasn’t just a straight-forward breakup into smaller interfaces – this was a rewrite, with new interfaces to replace the functionality of the old one.8
This is consistent with the remarks in this wikipage where it was mentioned that the new DocShell interfaces should have an API for the caller to supply a document, instead of a URI – thus taking the responsibility of calling into the networking library away from DocShell and putting it on the caller.
nsIDocShellEdit seems to be a replacement for some of the functions of the old nsIClipboardCommands methods that WebShell relied upon. Specifically, this interface was concerned with cutting, copying and pasting selections within the document. There is also a method for searching. These methods are all just stubbed out, and don’t do much at this point.
nsIDocShellFile seems to be the interface used for printing and saving documents.
nsIGenericWindow (which I believe is the ancestor of nsIBaseWindow), seems to be an interface that some embedding window must implement in order for the new nsWebBrowser / DocShell to be embedded in it. I think. I’m not too clear on this. At the very least, I think it’s supposed to be a generic interface for windows supplied by the underlying operating system.
nsIGlobalHistory is an interface for, well, browsing history.
This was before tabs, so we had just a single, linear global history to maintain, and I guess that’s what this interface was for. Correction (Jan 3rd, 2015) – global history is really just a recording of everywhere you’ve ever been, ever. It looks like I was getting confused between global history (all the places I’ve been) and session history (all the places I’ve been this session).
nsIScrollable is an interface for manipulating the scroll position of a document.
So these magnificent seven new interfaces were the first steps in breaking up WebShell… what was next?
Enter the Container
nsIDocShellContainer was created so that the DocShells could be formed into a tree and enumerated, and so that child DocShells could be named. It was introduced in this commit.
In this commit, only five days after the first landing, Travis appears to reverse the decision to pass the responsibility of loading the document onto the caller of DocShell. LoadDocument and LoadDocumentVia are replaced by LoadURI and LoadURIVia. Steve Clark (aka “buster”) is also added to the authors list of the nsIDocShell interface. It’s not clear to me why this decision was reversed, but if I had to guess, I’d say it proved to be too much of a burden on the callers to load all of the documents themselves. Perhaps they punted on that goal, and decided to tackle it again later (though I will point out that today’s nsIDocShell still has LoadURI defined in it).
The first implementation of nsIDocShell showed up on October 25, 1999. It was nsHTMLDocShell, and with the exception of nsIGlobalHistory, it implemented all of the other interfaces that I listed in Travis’ first landing.
The base implementation
On October 25th, the stubs of a DocShell base implementation showed up in the repository. The idea, I would imagine, is that for each of the document types that Gecko can display, we’d have a DocShell implementation, and each of these DocShell implementations would inherit from this DocShell base class, and only override the things that they need specialized for their particular document type.
Later on, when the idea of having specialized DocShell implementations evaporates, this base class will end up being nsDocShell.cpp.
“Does not compile yet”
added a bunch of initial implementation. does not compile yet, but that’s ok because docshell isn’t part of the build yet.
So not only are none of these patches being reviewed (as far as I can tell), and are not mapped to any bugs in the bug tracker, but the patches themselves just straight-up do not build. They are not building on tinderbox.
This is in pretty stark contrast to today’s code conventions. While it’s true that we might land code that is not entered for most Nightly users, we usually hide such code behind an about:config pref so that developers can flip it on to test it. And I think it’s pretty rare (if it ever occurs) for us to land code in mozilla-central that’s not immediately put into the build system.
Perhaps the WebShell tests that were part of the Plan of Attack were being written in parallel and just haven’t landed, but I suspect that they haven’t been written at all at this point. I suspect that the team was trying to stand something up and make it partially work, and then write tests for WebShell and try to make them pass for both old WebShell and DocShell. Or maybe just the latter.
These days, I think that’s probably how we’d go about such a major re-architecture / redesign / refactor; we’d write tests for the old component, land code that builds but is only entered via an about:config pref, and then work on porting the tests over to the new component. Once the tests pass for both, flip the pref for our Nightly users and let people test the new stuff. Once it feels stable, take it up the trains. And once it ships and it doesn’t look like anything is critically wrong with the new component, begin the process of removing the old component / tests and getting rid of the about:config pref.
Note that I’m not at all bashing Travis or the other developers who were working on this stuff back then – I’m simply remarking on how far we’ve come in terms of development practices.
Remember AOL keywords?
Tangentially, I’m seeing some patches go by that have to do with hooking up some kind of “Keyword” support to WebShell.
Remember those keywords? This was the pre-Google era where there were only a few simplistic search engines around, and people were still trying to solve discoverability of things on the web. Keywords was, I believe, AOL’s attempt at a solution.
You can read up on AOL Keywords here. I just thought it was interesting to find some Keywords support being written in here.
One DocShell to rule them all
Now that we have decided that there is only one docshell for all content types, we needed to get rid of the base class/ content type implementation. This checkin takes and moves the nsDocShellBase to be nsDocShell. It now holds the nsIHTMLDocShell stuff. This will be going away. nsCDocShell was created to replace the previous nsCHTMLDocShell.
This commit lands on November 12th (almost a month from the first landing), and is the point where the DocShell-implementation-per-document-type plan breaks down. nsDocShellBase gets renamed to nsDocShell, and the nsIHTMLDocShell interface gets moved into nsIDocShell.idl, where a comment above it indicates that the interface will soon go away.
We have nsCDocShell.idl, but this interface will eventually disappear as well.
pork jockey paint fixes. bug=18140, r=kmcclusk,pavlov
What the hell is a “pork jockey”? A quick search around, and I see yet another reference to it in Bugzilla on bug 14928. It seems to be some kind of project… or code name…
I eventually found this ancient wiki page that documents some of the language used in bugs on Bugzilla, and it has an entry for “pork jockey”: a pork jockey bug is a “bug for work needed on infrastructure/architecture”.
I mentioned this in #developers, and dmose (who was hacking on Mozilla code at the time), explained:
16:52 (dmose) mconley: so, porkjockeys
16:52 (mconley) let’s hear it
16:52 (dmose) mconley: at some point long ago, there was some infrastrcture work that needed to happen
16:52 mconley flips on tape recorder
16:52 (dmose) and when people we’re talking about it, it seemed very hard to carry off
16:52 (dmose) somebody said that that would happen on the same day he saw pigs fly
16:53 (mconley) ah hah
16:53 (dmose) so ultimately the group of people in charge of trying to make that happen were…
16:53 (dmose) the porkjockeys
16:53 (dmose) which was the name of the mailing list too
On November 17th, the nsIGenericWindow interface was removed because it was being implemented in widget/base/nsIBaseWindow.idl.
On November 27th, nsWebShell started to implement nsIBaseWindow, which helped pull a bunch of methods out of the WebShell implementations.
On November 29th, nsWebShell now implements nsIDocShell – so this seems to be the first point where the DocShell work gets brought into code paths that might be hit. This is particularly interesting, because he makes this change to both the WebShell implementation that he has under the docshell/ folder, as well as the WebShell implementation under the webshell/ folder. This means that some of the DocShell code is now actually being used.
On November 30th, Travis lands a patch to remove some old commented out code. The commit message mentions that the nsIDocShellEdit and nsIDocShellFile interfaces introduced in the first landing are now defunct. It doesn’t look like anything is diving in to replace these interfaces straight away, so it looks like he’s just not worrying about it just yet. The defunct interfaces are removed in this commit one day later.
One day later, nsIDocShellTreeNode interface is added to replace nsIDocShellContainer. The interface is almost identical to nsIDocShellContainer, except that it allows the caller to access child DocShells at particular indices as opposed to just returning an enumerator.
December 3rd was a very big day! Highlights include:
- WebShell stops implementing nsIScriptContextOwner in favour of the shiny new nsIScriptGlobalObjectOwner.
- The (somewhat nebulous) distinction of DocShell “treeItems” and “treeNodes” is made. At this point, the difference between the two is that nsIDocShellTreeItem must be implemented by anything that wishes to be a leaf or middle node of the DocShell tree. The interface itself provides accessors to various attributes on the tree item. nsIDocShellTreeNode, on the other hand, is for manipulating one of these items in the tree – for example, finding, adding or removing children. I’m not entirely sure this distinction is useful, but there you have it. WebShell (both the fork and “live” version) are set to implement these two new interfaces.
Noticing something? A lot of these changes are getting dumped straight into the live version of WebShell (under the webshell/ directory). That’s not really what the Plan of Attack had spelled out, but that’s what appears to be happening. Perhaps all of this stuff was trivial enough that it didn’t warrant waiting for the WebShell fork to switch over.
On December 15th, buster re-lands the nsIDocShellEdit and nsIDocShellFile interfaces that were removed on November 30th, but they’re called nsIContentViewerEdit and nsIContentViewerFile, respectively. Otherwise, they’re identical.
On December 17th, WebShell becomes a subclass of DocShell. This means that a bunch of things can get removed from WebShell, since they’re being taken care of by the parent DocShell class. This is a pretty significant move in the whole “replacing WebShell” strategy.
Similar work occurs on December 20th, where even more methods inside WebShell start to forward to the base DocShell class.
That’s the last bit of notable work during 1999. These next bits show up in the new year, and provides further proof that we didn’t all blow up during Y2K.
On Feburary 2nd, 2000, a new interface called nsIWebNavigation shows up. This interface is still used to this day to navigate a browser, and to get information about whether it can go “forwards” or “backwards”.
On February 8th, a patch lands that makes nsGlobalWindow deal entirely in DocShells instead of WebShells. nsIScriptGlobalObject also now deals entirely with DocShells. This is a pretty big move, and the patch is sizeable.
On February 11th, more methods are removed from WebShell, since the refactorings and rearchitecture have made them obsolete.
On February 14th, for Valentine’s day, Travis lands a patch to have DocShell implement the nsIWebNavigation interface. Later on, he lands a patch that relinquishes further control from WebShell, and puts the DocShell in control of providing the script environment and providing the nsIScriptGlobalObjectOwner interface. Not much later, he lands a patch that implements the Stop method from the nsIWebNavigation interface for DocShell. It’s not being used yet, but it won’t be long now. Valentine’s day was busy!
On February 24th, more stuff (like the old Stop implementation) gets gutted from WebShell. Some, if not all, of those methods get forwarded to the underlying DocShell, unsurprisingly.
Similar story on February 29th, where a bunch of the scroll methods are gutted from WebShell, and redirected to the underlying DocShell. This one actually has a bug and some reviewers!9 Travis also landed a patch that gets DocShell set up to be able to create its own content viewers for various documents.
March 10th saw Travis gut plenty of methods from WebShell and redirect to DocShell instead. These include Init, SetDocLoaderObserver, GetDocLoaderObserver, SetParent, GetParent, GetChildCount, AddChild, RemoveChild, ChildAt, GetName, SetName, FindChildWithName, SetChromeEventHandler, GetContentViewer, IsBusy, SetDocument, StopBeforeRequestingURL, StopAfterURLAvailable, GetMarginWidth, SetMarginWidth, GetMarginHeight, SetMarginHeight, SetZoom, GetZoom. A few follow-up patches did something similar. That must have been super satisfying.
March 11th, Travis removes the Back, Forward, CanBack and CanForward methods from WebShell. Consumers of those can use the nsIWebNavigation interface on the DocShell instead.
March 30th sees the nsIDocShellLoadInfo interface show up. This interface is for “specifying information used in a nsIDocShell::loadURI call”. I guess this is certainly better than adding a huge amount of arguments to ::loadURI.
During all of this, I’m seeing references to a “new session history” being worked on. I’m not really exploring session history (at this point), so I’m not highlighting those commits, but I do want to point out that a replacement for the old Netscape session history stuff was underway during all of this DocShell business, and the work intersected quite a bit.
On April 16th, Travis lands a commit that takes yet another big chunk out of WebShell in terms of loading documents and navigation. The new session history is now being used instead of the old.
The last 10% is the hardest part
We’re approaching what appears to be the end of the DocShell work. According to his LinkedIn profile, Travis left AOL in May 2000. His last commit to the repository before he left was on April 24th. Big props to Travis for all of the work he put in on this project – by the time he left, WebShell was quite a bit simpler than when he started. I somehow don’t think he reached the end state that he had envisioned when he’d written the original redesign document – the work doesn’t appear to be done. WebShell is still around (in fact, parts of it around were around until only recently!10 ). Still, it was a hell of chunk of work he put in.
And if I’m interpreting the rest of the commits after this correctly, there is a slow but steady drop off in large architectural changes, and a more concerted effort to stabilize DocShell, nsWebBrowser and nsWebShell. I suspect this is because everybody was buckling down trying to ship the first version of the Mozilla Suite (which finally occurred June 5th, 2002 – still more than 2 years down the road).
There are still some notable commits though. I’ll keep listing them off.
On June 22nd, a developer called “rpotts” lands a patch to remove the SetDocument method from DocShell, and to give the implementation / responsibility of setting the document on implementations of nsIContentViewer.
July 5th sees rpotts move the new session history methods from nsIWebNavigation to a new interface called nsIDocShellHistory. It’s starting to feel like the new session history is really heating up.
On July 18th, a developer named Judson Valeski lands a large patch with the commit message “webshell-docshell consolodation changes”. Paraphrasing from the bug, the point of this patch is to move WebShell into the DocShell lib to reduce the memory footprint. This also appears to be a lot of cleanup of the DocShell code. Declarations are moved into header files. The nsDocShellModule is greatly simplified with some macros. It looks like some dead code is removed as well.
On November 9th, a developer named “disttsc” moves the nsIContentViewer interface from the webshell folder to the docshell folder, and converts it from a manually created .h to an .idl. The commit message states that this work is necessary to fix bug 46200, which was filed to remove nsIBrowserInstance (according to that bug, nsIBrowserInstance must die).
That’s probably the last big, notable change to DocShell during the 2000’s.
2001: A DocShell Odyssey
On March 8th, a developer named “Dan M” moves the GetPersistence and SetPersistence methods from nsIWebBrowserChrome to nsIDocShellTreeOwner. He sounds like he didn’t really want to do it, or didn’t want to be pegged with the responsibility of the decision – the commit message states “embedding API review meeting made me do it.” This work was tracked in bug 69918.
On April 16th, Judson Valeski makes it so that the mimetypes that a DocShell can handle are not hardcoded into the implementation. Instead, handlers can be registered via the CategoryManager. This work was tracked in bug 40772.
On April 26th, a developer named Simon Fraser adds an implementation of nsISimpleEnumerator for DocShells. This implementation is called, unsurprisingly, nsDocShellEnumerator. This was for bug 76758. A method for retrieving an enumerator is added one day later in a patch that fixes a number of bugs, all related to the page find feature.
April 27th saw the first of the NSPR logging for DocShell get added to the code by a developer named Chris Waterson. Work for that was tracked in bug 76898.
There’s a big gap here, where the majority of the landings are relatively minor bug fixes, cleanup, or only slightly related to DocShell’s purpose, and not worth mentioning.11
And beyond the infinite…
On January 8th, 2002, for bug 113970, Stephen Walker lands a patch that takes yet another big chunk out of WebShell, and added this to the nsIWebShell.h header:
/** * WARNING WARNING WARNING WARNING WARNING WARNING WARNING WARNING !!!! * * THIS INTERFACE IS DEPRECATED. DO NOT ADD STUFF OR CODE TO IT!!!! */
I’m actually surprised it too so long for something like this to get added to the nsIWebShell interface – though perhaps there was a shared understanding that nsIWebShell was shrinking, and such a notice wasn’t really needed.
On January 9th, 2003 (yes, a whole year later – I didn’t find much worth mentioning in the intervening time), I see the first reference to “deCOMtamination”, which is an effort to reduce the amount of XPCOM-style code being used. You can read up more on deCOMtamination here.
On January 13th, 2003, Nisheeth Ranjan lands a patch to use “machine learning” in order to order results in the urlbar autocomplete list. I guess this was the precursor to the frencency algorithm that the AwesomeBar uses today? Interestingly, this code was backed out again on February 21st, 2003 for reasons that aren’t immediately clear – but it looks like, according to this comment, the code was meant to be temporary in order to gather “weights” from participants in the Mozilla 1.3 beta, which could then be hard-coded into the product. The machine-learning bug got closed on June 25th, 2009 due to AwesomeBar and frecency.
On November 23rd, 2004, the parent and treeOwner attributes for nsIDocShellTreeItem are made scriptable. They are read-only for script.
And that’s probably the last notable patch related to DocShell in 2004.
Lord of the rings
Reading through all of these commits in the docshell/ and webshell/ folders is a bit like taking a core sample of a very mature tree, and reading its rings. I can see some really important events occurring as I flip through these commits – from the very birth of Mozilla, to the birth of XPCOM and XUL, to porkjockeys, and the start of the embedding efforts… all the way to the splitting out of Thunderbird, deCOMtamination and introduction of the MPL. I really got a sense of the history of the Mozilla project reading through these commits.
I feel like I’m getting a better and better sense of where DocShell came from, and why it does what it does. I hope if you’re reading this, you’re getting that same sense.
Stay tuned for the next bit, where I look at years 2005 to 2010.
Messages like: ERROR: A CVS repository cannot contain both cvsroot/mozilla/browser/base/content/metadata.js,v and cvsroot/mozilla/browser/base/content/Attic/metadata.js,v, for example ↩
Those commits have hashes 575e116287f99fbe26f54ca3f3dbda377002c5e7 and 60567bb185eb8eea80b9ab08d8695bb27ba74e15 if you want to follow along at home. ↩
Strangely enough, that docshell folder wasn’t even being built. It was just sitting there, containing a copy of nsWebShell.cpp. That copy mirrored the one under webshell/, and the two copies were kept in sync. Maybe DocShell was always a thing that somebody always wanted to do but never got around to until Travis came along. ↩
Mozilla Suite 1.0 eventually ships on June 5th, 2002 ↩
That wiki page was first posted on October 8th, 1999 – exactly a week before the first docshell work had started. ↩
I should point out that there is no mention of this embedding work in the first roadmap that was published for the Mozilla project. It was only in the second roadmap, published almost a year after the DocShell work began, that embedding was mentioned, albeit briefly. ↩
More clues as to what WebShell was originally intended for are also in that interface file:
/** * The web shell is a container for implementations of nsIContentViewer. * It is a content-viewer-container and also knows how to delegate certain * behavior to an nsIWebShellContainer. * * Web shells can be arranged in a tree. * * Web shells are also nsIWebShellContainer's because they can contain * other web shells. */
So that helps clear things up a bit. I think. ↩
travis reviewed it, and the patch was approved by rickg. Not sure what approval meant back then… was that like a super review? ↩
Hopefully I didn’t miss anything interesting in there! ↩