*This is an existing issue that I noticed while testing PR 8552.*
When zooming or rotation occurs, we'll try to use the current canvas as a (CSS transformed) preview until the page has been completely re-drawn.
If you manage to change the scale (or rotation) *very* quickly, it's possible that `PDFPageView.update` can be called *before* a previous `render` operation has progressed far enough to remove the `hidden` property from the canvas.
The result is thus that a page may be *entirely* black during zooming or rotation, which doesn't look very good. This effect can be a bit difficult to spot, but it does manifest even in the default viewer.
In the PDF from issue 8527, the clip operator (W) shows up before a path
is defined. The current SVG backend however expects a path to exist
before generating a `<svg:clipPath>` element.
In the example, the path was defined after the clip, followed by a
endPath operator (n).
So this commit fixes the bug by moving the path generation logic from
clip to endPath.
Our canvas backend appears to use similar logic:
`CanvasGraphics_endPath` calls `consumePath`, which in turn draws the
clip and resets the `pendingClip` state. The canvas backend calls
`consumePath` from multiple other places, so we probably need to check
whether doing so is also necessary for the SVG backend.
I scanned our corpus of PDF files in test/pdfs, and found that in every
instance (except for one), the "W" PDF operator (clip) is immediately
followed by "n" (endPath). The new test from this commit (clippath.pdf)
starts with "W", followed by a path definition and then "n".
# Commands used to find some of the clipping commands:
grep -ra '^W$' -C7 | less -S
grep -ra '^W ' -C7 | less -S
grep -ra ' W$' -C7 | less -S
test/pdfs/issue6413.pdf is the only file where "W" (a tline 55) is not
followed by "n". In fact, the "W" is the last operation of a series of
XObject painting operations, and removing it does not have any effect
on the rendered PDF (confirmed by looking at the output of PDF.js's
canvas backend, and ImageMagick's convert command).
Part of the rotation handling code, in what's now `web/app.js`, hasn't really changed since before the viewer was split into multiple files/components.
Similar to other properties, such as current page/scale, we should probably avoid tracking state in multiple places. Hence I'm suggesting that we don't store the rotation in `PDFViewerApplication`, and access the value in `PDFViewer` instead.
Since `PDFViewerApplication.pageRotation` has existed for a very long time, a getter was added to avoid outright breaking third-party code that may depend on it.
Currently a number of these properties do not work correctly if set *before* calling `setDocument`; please refer to the discussion starting in https://github.com/mozilla/pdf.js/pull/8539#issuecomment-309706629.
Rather than trying to have *some* of these methods working, but not others, it seems much more consistent to simply always require that `setDocument` has been called.
Wait for the completion of writing the generated SVG file before
processing the next page. This is to enable the garbage collector to
garbage-collect the (potentially large) SVG string before trying to
allocate memory again for the next page.
Note that since the PDF-to-SVG conversion is now sequential instead of
parallel, the time to generate all pages increases.
Test case:
node --max_old_space_size=200 examples/node/pdf2svg.js /tmp/FatalProcessOutOfMemory.pdf
Before this patch:
- Node.js crashes due to OOM after processing 20 pages.
After this patch:
- Node.js is able to convert all 203 PDFs to SVG without crashing.
Test case:
Using the PDF file from https://github.com/mozilla/pdf.js/issues/8534
node --max_old_space_size=200 examples/node/pdf2svg.js /tmp/FatalProcessOutOfMemory.pdf
Before this patch:
Node.js crashes due to OOM after processing 10 pages.
After this patch:
Node.js crashes due to OOM after processing 19 pages.
Since this call occurs *before* the `PDFViewer.setDocument` call, it won't actually cause any scale change.
Furthermore, moving it should not be necessary, since the `scale` is already used as the fallback case in `PDFViewerApplication.setInitialView` (provided it's non-zero, which isn't even the case in the default viewer).
Hence this patch should cause no functional changes at all, since it simply removes a piece of unnecessary code.
This patch adds Streams API support in getTextContent
so that we can stream data in chunks instead of fetching
whole data from worker thread to main thread. This patch
supports Streams API without changing the core functionality
of getTextContent.
Enqueue textContent directly at getTextContent in partialEvaluator.
Adds desiredSize and ready property in streamSink.
Currently, these properties are reset in what appears to be somewhat arbitrary locations (within the `load` and `open` methods respectively). The explanation is probably that both of these properties predates the existence of any centralized clean-up code in the viewer.
Hence I think that it makes sense to move the resetting of these properties to the `close` method, since that improves the overview of what's actually cleaned-up/reset when changing documents in the viewer.
The `ObjectLoader` currently takes an Object as input, despite actually working with `Dict`s internally. This means that at the (two) existing call-sites, we're passing in the "private" `Dict.map` property directly.
Doing this seems like an anti-pattern, and we could (and even should) simply provide the actual `Dict` when creating an `ObjectLoader` instance.
Accessing properties stored in the `Dict` is now done using the intended methods instead, in particular `getRaw` which (as the name suggests) doesn't do any de-referencing, thus maintaining the current functionality of the code.
The only functional change in this patch is that `ObjectLoader.load` will now ignore empty nodes, such that `ObjectLoader._walk` only needs to deal with nodes that are known to contain data. (This lets us skip, among other checks, meaningless `addChildren` function calls.)
This method is currently called from `PDFViewer._scrollUpdate` on *every* scroll event in the viewer.
However, I cannot see why this code is now necessary (assuming that it once was), since text-selection and searching still works *exactly* the same way with this patch as with the current `master`.
When `PDFPageView.updatePosition` is called, the page can be in either of these states:
1. The page hasn't been rendered, in which case the `textLayer` property doesn't exist yet.
2. The page is currently rendering, meaning that the `textLayer` property exists. Given that the `textContent` won't be fetched until the page has been successfully rendered, `TextLayerBuilder.render` will return immediately and effectively be a no-op (since there's nothing to render yet).
3. The has been been rendered, and the `textLayer` is currently rendering.
4. The page, and its `textLayer`, has been completely rendered. In this case, `TextLayerBuilder.render` will return immediately and effectively be a no-op.
Here, only the *third* case seem to require any further analysis:
When scrolling occurs while the `textLayer` is rendering, `TextLayerBuilder.render` will via a helper method call `TextLayerRenderTask.cancel` (in src/display/text_layer.js) to stop processing.
However, due to the run-to-completion nature of JavaScript, once `TextLayerRenderTask._render` has been invoked `appendText` will always run.[1]
So even though we cancel rendering of pending `textLayer`s during scrolling, via the repeated `TextLayerBuilder.render` calls from within the `PDFPageView.updatePosition` method, that does *not* prevent us from running the code inside of `TextLayerRenderTask._render` over and over for the *same* page; which all seems *very* inefficient to me.[2]
All this will thus have the effect of delaying the *actual* rendering of a `textLayer` ever so slightly while scrolling in the viewer. However, it does so at the expense of potentially hundreds of unnecessary `appendText` calls.[3]
Hence it seems to me that it's less resource intensive overall to simply let rendering of the `textLayer` complete, once it has started. Obviously, we still abort all rendering of a page, and its `textLayer`, when it's being destroyed (e.g. by being evicted from the page cache).
In case that there's any worry that the patch could affect e.g. highlighting of search results, please note that the existing code in `TextLayerBuilder.render` already calls `updateMatches` when the `TextLayerTask` resolves successfully.
*I'm sorry that this became quite long, but to try and summarize:*
`PDFPageView.updatePosition` doesn't actually do anything in *most* cases. In the one case where it matters, it seems that it's actually doing more harm than good; which is why I'm proposing that we just remove it.
---
[1] Although we may be able to skip the `render` call, provided that it happens *after* a `timeout` (as is the case in the default viewer).
[2] With current work being done to support streaming of `TextContent`, we'd also need to add just as many duplicate API calls to `PDFPageView.updatePosition`.
[3] The number of duplicate `appendText` calls is directly proportional not only to the scroll speed, but also to the number of pages in the document.
Since the localization service is now asynchronous, depending on the load the browser is under, there's a small risk that the lookup of the 'page_scale_percent' string could be delayed slightly.
If the scale would change a couple of times in *very* quick succession, there's perhaps a *theoretical* possibility that the Zoom dropdown would display an incorrect value.
Consider the following, somewhat contrived, theoretical example of two zoom commands being executed *right* after one another:
```javascript
PDFViewerApplication.pdfViewer.currentScale = 1.23;
PDFViewerApplication.pdfViewer.currentScaleValue = 'page-width';
```
Only the `currentScale` call will currently trigger a l10n lookup in `selectScaleOption`. However, as far as I understand, there's no *guarantee* that the l10n string is resolved *before* `selectScaleOption` is called again as a result of the `currentScaleValue` call.
This thus has the possibility of putting the Zoom dropdown into an inconsistent state, since it's currently updated synchronously for one code-path and asynchronously for another.
To avoid these issues, I'm proposing that we *always* update the Zoom dropdown asynchronously, such that we can guarantee that the ordering is correct.
After PR 8394, at least in Firefox, the Zoom dropdown now frequently displays an old custom scale instead of the correct one. To see this behaviour, the following STR works for me:
1. Open https://mozilla.github.io/pdf.js/web/viewer.html.
2. Zoom in, by clicking on the corresponding button in the toolbar.
3. Run `PDFViewerApplication.pdfViewer.currentScaleValue` in the console.
4. Compare what's displayed in the Zoom dropdown with what's printed in the console. (If no difference can be observed, try repeating steps 2 through 4 a couple of times.)
I really don't understand why this happens, but it seems that waiting until a custom scale has been set *before* selecting it fixes things in Firefox (and works fine in e.g. Chrome as well).
Note that this patch thus makes this particular piece of the code consistent with the state prior to PR 8394.
