*It appears that this accidentally broken in PR 8775.*
Note that `PDFHistory.forward` is only used with certain named actions, and these aren't that commonly used, which ought to explain why this error managed to sneak in.
Steps to reproduce the issue (and verify the fix):
1. Navigate to e.g. http://mirrors.ctan.org/info/lshort/english/lshort.pdf
2. Click on a couple of links, or outline items, such that the history is populated with a few entries.
3. In the console, execute `PDFViewerApplication.pdfHistory.back()` one or more times, thus navigating back to a previous viewer position.
4. In the console, execute `PDFViewerApplication.pdfHistory.forward() one or more times.
At the last step above, no (forward) navigation happens with the current `master`; now compare with this patch.
When testing the new `PDFHistory` implementation in practice, I felt that the current value of `UPDATE_VIEWAREA_TIMEOUT` is too large to be truly useful.
The purpose of the timeout is to attempt to address (the PDF.js part of) https://bugzilla.mozilla.org/show_bug.cgi?id=1153393, and it's currently fairly easy for the user e.g. close the browser before the timeout had a change to finish.
Obviously, the timeout is a best-effort solution, but with the current value of `UPDATE_VIEWAREA_TIMEOUT` it's not as useful as one would want.
Please note that lowering it shouldn't be a problem, since it still prevents the browser history from updating at *every* 'updateviewarea' event or during (quick) scrolling, which is all that's really needed to not impact the UX negatively.
---
Furthermore, with this lower timeout, we can also simplify the part of the 'popstate' event handler that attempted to update the browser history with the current position before moving back. In most cases, the current position will now already exist in the history, and this *greatly* decreases the complexity of this code path.
The main impetus for this change though, is that I unfortunately found that given the asynchronous nature of updating the browser history, there is some *edge* cases where that code could cause history corruption.
In practice, the user could thus get "stuck" at a particular history entry and not be able to move back. I haven't got any reliable STR for this, since it's so difficult to trigger, but it involved navigating around in a document such that a number of destinations are added to the browser history and then changing the rotation before going back/forward in the history.
Rather that attempting to patch this code, and making it even more difficult to understand than it already is or adding more asynchronous behaviour, by far the easiest solution is to remove it and simply rely on the (lowered) `UPDATE_VIEWAREA_TIMEOUT` instead.
Since e.g. zooming can occur when navigating to a new destionation, ensure that a resulting 'updateviewarea' event doesn't trigger adding of a *temporary* position to the browser history at a bad time.
By using the (heuristic) `POSITION_UPDATED_THRESHOLD` constant, we can ensure that the current document position will be added to the browser history when a sufficiently "large" number of `updateviewarea` events have been dispatched.
This patch attempts to address an issue in the old `PDFHistory` implementation, where the current position wouldn't be correctly saved when the browser was closed.
In theory this *should* already be working, however as the discussion in https://bugzilla.mozilla.org/show_bug.cgi?id=1153393 showed, it seems that both `pagehide` and `beforeunload` arrive to late to successfully update the history during closing.
Hence a timeout is used to *temporarily* add the current position to the browser history when the viewer is idle.
Note that we need to take care not to update the browser history too often, since that would render the viewer more or unusable. Furthermore, if the timeout is *too* long it may end up effectively disable this whole functionality.
The `UPDATE_VIEWAREA_TIMEOUT` constant is thus a heuristic value, which we may need to tweak taking the above into account.
This patch completely re-implements `PDFHistory` to get rid of various bugs currently present, and to hopefully make maintenance slightly easier. Most of the interface is similar to the existing one, but it should be somewhat simplified.
The new implementation should be more robust against failure, compared to the old one. Previously, it was too easy to end up in a state which basically caused the browser history to lock-up, preventing the user from navigating back/forward. (In the new implementation, the browser history should not be updated rather than breaking if things go wrong.)
Given that the code has to deal with various edge-cases, it's still not as simple as I would have liked, but it should now be somewhat easier to deal with.
The main source of complication in the code is actually that we allow the user to change the hash of a already loaded document (we'll no longer try to navigate back-and-forth in this case, since the next commit contains a workaround).
In the new code, there's also *a lot* more comments (perhaps too many?) to attempt to explain the logic. This is something that the old implementation was serverly lacking, which is a one of the reasons why it was so difficult to maintain.
One particular thing to note is that the new code uses the `pagehide` event rather than `beforeunload`, since the latter seems to be a bad idea based on https://bugzilla.mozilla.org/show_bug.cgi?id=1336763.
The current implementation of `PDFHistory` contains a number of smaller bugs, which are *very* difficult to address without breaking other parts of its code.
Possibly the main issue with the current implementation, is that I wrote it quite some time ago, and at the time my understanding of the various edge-cases the code has to deal with was quite limited.
Currently `PDFHistory` may, despite most of those cases being fixed, in certain edge-cases lock-up the browser history, essentially preventing the user from navigating back/forward.
Hence rather than trying to iterate on `PDFHistory` to make it better, the only viable approach is unfortunately rip it out in its entirety and re-write it from scratch.
Note that as discussed on IRC, this makes the viewer slightly slower to load *only* in `gulp server` mode, however the difference seem slight enough that I think it will be fine.
Using `else` after `return` is not necessary, and can often lead to unnecessarily cluttered code. By using the `no-else-return` rule in ESLint we can avoid this pattern, see http://eslint.org/docs/rules/no-else-return.
When the user edits the URL and changes the reference fragment (hash),
PDF.js intercepts this action, and saves the then-current history state
in the previous history entry. This is implemented by navigating back,
editing the history and navigating forward again.
The current logic has a flaw: It assumes that calling history.back() and
history.forward() immediately updates the history state. This is however
not guaranteed by the web standards, which states that calling e.g.
history.back "must traverse the history by a delta -1", which means that
the browser must QUEUE a task to traverse the session history, per spec:
http://w3.org/TR/2011/WD-html5-20110113/history.html#dom-history-backhttps://html.spec.whatwg.org/multipage/browsers.html#dom-history-back
Firefox and Internet Explorer deviate from the standards by immediately
changing the history state instead of queuing the navigation.
WebKit derived browsers (Chrome, Opera, Safari) and Opera presto do not.
The user-visible consequence of strictly adhering to the standards in
PDF.js can be shown as follows:
1. Edit the URL.
2. Append #page=2 for example.
3. Press Enter.
-> Presto and WebKit: PDF.js reverts to the previous URL.
-> Gecko and Trident: PDF.js keeps the new URL, as expected.
To fix the issue, modification of the previous history item happens in
a few asynchronous steps, guided by the popstate event to detect when
the history navigation request has been committed.
--
Some more implementation notes:
I have removed the preventDefault and stopPropagation calls, because
popstate is not cancelable, and window is already the last target of the
event propagation.
The previous allowHashChange logic was hard to follow, because it did
not explain that hashchange will be called twice; once during the
popstate handler for history.back() (which will reset allowHashChange),
and again for history.forward() (where allowHashChange will be false).
The purpose of allowHashChange is now more explicit, by incorporating
the logic in the replacePreviousHistoryState helper function.
This patch:
- Adds a couple of utility methods to `PDFViewer` to enable checking `presentationModeState` without cumbersome comparisons.
- Disentangles `PresentationMode` from `PDFHistory` and `HandTool`, by adding event listeners for `presentationmodechanged` to both of them.
- Adds a utility function to `PDFViewerApplication` for requesting PresentationMode.
- Prevents initialization of `PresentationMode` if the browser does not support the fullscreen API.
In Chromium extensions, the viewer's URL looks like this:
chrome-extension://oemmndcbldboiebfnladdacbdfmadadm/http://example.com/file.pdf
Furthermore, the PDF Viewer itself can also add something to the reference fragment:
chrome-extension://oemmndcbldboiebfnladdacbdfmadadm/http://example.com/file.pdf#page=2
Consequently, it is difficult to copy a clean URL (e.g. for sharing over mail)
without having to tidy-up the URL manually.
This commit solves this issue by adding a button to the omnibox,
which shows the clean PDF URL on click.
When <base href> is present, history.replaceState and
history.pushState behave inconsistent with relative URLs.
http://code.google.com/p/chromium/issues/detail?id=274024
Contrary to what one expect, passing '' as the URL parameter to
replaceState/pushState does not associate the currently active
URL with the history entry, but a path relative to <base href>.
To fix the issue, explicitly associate the current active URL
with the history's state.
