By default, i.e. with workers enabled, it's *purposely* not possible to send `Dict`s and `Stream`s from the worker-thread. This is achieved by defining a `function` on every `Dict` instance, since that ensures that [the structured clone algoritm](https://developer.mozilla.org/en-US/docs/Web/API/Web_Workers_API/Structured_clone_algorithm) will throw an Error on `postMessage`.
However, with workers *disabled* we fall-back to the `LoopbackPort` implementation which just ignores any `function`s, thus incorrectly allowing sending of data which *should* be unclonable.
With this patch we're finally able to abort worker-thread parsing of the `OperatorList`, rather than *only* aborting the main-thread rendering itself, when the `RenderTask.cancel` method is being called.
This will help improve perceived performance in the default viewer, especially when reading longer and more complex documents, since pages that've been scrolled out-of-view (and thus evicted from the cache) will no longer compete for parsing resources on the worker-thread.
*Please note:* With the implementation in this patch we're *not* aborting worker-thread parsing immediately on `RenderTask.cancel`, since that would lead to *worse* performance in many cases. For example: When zoom/rotation occurs in the viewer, while parsing/rendering is still ongoing, a `cancel` call will usually be (almost) immediately folled by a new `PDFPageProxy.render` call. In that case you obviously don't want to abort parsing on the worker-thread, since that would risk throwing away a partially parsed `OperatorList` and thus force unnecessary re-parsing which will regress perceived performance (especially for more complex documents).
When choosing a reasonable delay, before cancelling `getOperatorList` on the worker-thread when `RenderTask.cancel` is called, two different positions need to be considered:
1. The delay needs to be short enough, since a timeout in the multiple seconds range would essentially make this entire functionality meaningless (by always allowing most/all pages enough time to finish parsing).
2. The delay cannot be *too* short, since that would actually *reduce* performance in the zoom/rotation case outlined above. Furthermore, the time between `RenderTask.cancel` and `PDFPageProxy.render` calls will obviously be affected by both general computer performance and current CPU load.
It's certainly possible that the timeout may require some further tweaks, however the value settled on in this patch was easily *one order* of magnitude larger than the delta between cancel/render in my tests.
With the changes made in PR 11069, it's no longer necessary to include the `pageIndex`/`intent` parameters when sending 'GetOperatorList' data. In the previous implementation these properties were used to associate the `OperatorList` with the correct `RenderTask`, however now that `ReadableStream`s are used that's handled automatically and it's thus dead code at this point.
By transfering, rather than copying, `ArrayBuffer`s between the main- and worker-threads, you can avoid unnecessary allocations by only having *one* copy of the same data.
Hence manually setting `postMessageTransfers: false`, when calling `getDocument`, is a performance footgun[1] which will do nothing but waste memory.
Given that every reasonably modern browser supports `postMessage` transfers[2], I really don't see why it should be possible to force-disable this functionality.
Looking at the browser support, for `postMessage` transfers[2], it's highly unlikely that PDF.js is even usable in browsers without it. However, the feature testing of `postMessage` transfers is kept for the time being just to err on the safe side.
---
[1] This is somewhat similar to the, now removed, `disableWorker` parameter which also provided API users a much too simple way of reducing performance.
[2] See e.g. https://developer.mozilla.org/en-US/docs/Web/API/MessagePort/postMessage#Browser_compatibility and https://developer.mozilla.org/en-US/docs/Web/API/Transferable#Browser_compatibility
Note how the sent values have inconsistent types, with a boolean in one case and an object in the other (normal) case.
Furthermore, explicitly sending a `supportTypedArray: true` property seems superfluous at least to me.
It recently occurred to me that the CMap data should be an excellent candidate for transfering.
This will help reduce peak memory usage for PDF documents using CMaps, since transfering of data avoids duplicating it on both the main- and worker-threads.
Unfortunately it's not possible to actually transfer data when *returning* data through `sendWithPromise`, and another solution had to be used.
Initially I looked at using one message for requesting the data, and another message for returning the actual CMap data. While that should have worked, it would have meant adding a lot more complexity particularly on the worker-thread.
Hence the simplest solution, at least in my opinion, is to utilize `sendWithStream` since that makes it *really* easy to transfer the CMap data. (This required PR 11115 to land first, since otherwise CMap fetch errors won't propagate correctly to the worker-thread.)
Please note that the patch *purposely* only changes the API to Worker communication, and not the API *itself* since changing the interface of `CMapReaderFactory` would be a breaking change.
Furthermore, given the relatively small size of the `.bcmap` files (the largest one is smaller than the default range-request size) streaming doesn't really seem necessary either.
At this point in time it's easy to convert the `MessageHandler.on` call-sites to use arrow functions, and thus let the JavaScript engine handle scopes for us, rather than having to manually keep references to the relevant scopes in `MessageHandler`.[1]
An additional benefit of this is that a couple of `Function.prototype.call()` instances can now be converted into "normal" function calls, which should be a tiny bit more efficient.
All in all, I don't see any compelling reason why it'd be necessary to keep supporting custom `scope`s in the `MessageHandler` implementation.
---
[1] In the event that a custom scope is ever needed, simply using `bind` on the handler function when calling `MessageHandler.on` ought to work as well.
*Please note:* The majority of this patch was written by Yury, and it's simply been rebased and slightly extended to prevent issues when dealing with `RenderingCancelledException`.
By leveraging streams this (finally) provides a simple way in which parsing can be aborted on the worker-thread, which will ultimately help save resources.
With this patch worker-thread parsing will *only* be aborted when the document is destroyed, and not when rendering is cancelled. There's a couple of reasons for this:
- The API currently expects the *entire* OperatorList to be extracted, or an Error to occur, once it's been started. Hence additional re-factoring/re-writing of the API code will be necessary to properly support cancelling and re-starting of OperatorList parsing in cases where the `lastChunk` hasn't yet been seen.
- Even with the above addressed, immediately cancelling when encountering a `RenderingCancelledException` will lead to worse performance in e.g. the default viewer. When zooming and/or rotation of the document occurs it's very likely that `cancel` will be (almost) immediately followed by a new `render` call. In that case you'd obviously *not* want to abort parsing on the worker-thread, since then you'd risk throwing away a partially parsed Page and thus be forced to re-parse it again which will regress perceived performance.
- This patch is already *somewhat* risky, given that it touches fundamentally important/critical code, and trying to keep it somewhat small should hopefully reduce the risk of regressions (and simplify reviewing as well).
Time permitting, once this has landed and been in Nightly for awhile, I'll try to work on the remaining points outlined above.
Co-Authored-By: Yury Delendik <ydelendik@mozilla.com>
Co-Authored-By: Jonas Jenwald <jonas.jenwald@gmail.com>
With the changes to the `StreamType`/`FontType` "enums" in PR 11029, one unfortunate result is that `getStats` now *always* returns empty Arrays. Something that everyone, myself included, apparently missed is that you obviously cannot index an Array with Strings :-)
I wrongly assumed that the unit-tests would catch any bugs, but they apparently suffered from the same issue as the code in `src/core/`.
Another possible option could perhaps be to use `Set`s, rather than objects, but that will require larger changes since `LoopbackPort` (in `src/display/api.js`) doesn't support them.
There's a number of spots in the current code, and tests, where `cancel` methods are not called with appropriate arguments (leading to Promises not being rejected with Errors as intended).
In some cases the cancel `reason` is implicitly set to `undefined`, and in others the cancel `reason` is just a plain String. To address this inconsistency, the patch changes things such that cancelling is done with `AbortException`s everywhere instead.
Note that, in the old code, there was a code-path which could prevent this from happening thus affecting future cleanup.
Furthermore, ensure that we'll always attempt to cleanup when handling the 'PageError' message, similar to the code in e.g. the `PDFPageProxy._renderPageChunk` method.
The `receivingOperatorList` property is currently tracked *twice* in the rendering code, both directly and inversely through the `intentState.operatorList.lastChunk` boolean. This type of double bookkeeping is never a good idea, since it's just too easy for the properties to accidentally fall out of sync.
In this case there's even a `cleanup`-related bug caused by this, which means that `PDFPageProxy._tryCleanup` will never be able to discard any data if there's an error on the worker-thread (as handled through the 'PageError' message).
Hence the simplest solution seems, at least to me, to update `PDFPageProxy._tryCleanup` to replace the `intentState.receivingOperatorList` check with a `!intentState.operatorList.lastChunk` check and completely remove the former property.
Given that `cleanupAfterRender` is already set for large images, when handling 'obj' messages, this patch *should* thus be safe in general (since otherwise there ought be existing bugs related to cleanup and printing).
Calling `someArray = []` will create a new Array, which seems completely unnecessary when it's sufficient to just call `someArray.length = 0` to achieve the same effect.
Even though I cannot imagine these particular cases having any noticeable performance impact, similar changes were made in `core/` code years ago since it's apparently more efficient memory wise.
Please see the specification, https://www.adobe.com/content/dam/acom/en/devnet/acrobat/pdfs/PDF32000_2008.pdf#M11.9.12864.1Heading.71.Viewer.Preferences
Furthermore, note that this patch *only* adds API support and unit-tests but does not attempt to integrate e.g. the `ViewerPreferences -> Direction` property into the viewer (which would be necessary to address issue 10736).
The reason for this is that it's not entirely clear to me exactly if/how that could be implemented; e.g. would it be as simple as setting the `dir` attribute on the `viewerContainer` DOM element, or will it be more complicated?
There's also the question of how the `ViewerPreferences -> Direction` value interacts with the `PageMode`, and this will generally require a fair bit of manual testing. Since the direction of the *entire* viewer depends on the browser locale, there's also a somewhat open question regarding what default value to use for different locales.
Finally, if the viewer supports `ViewerPreferences -> Direction` then I'm assuming that it will be necessary to allow users to override the default value, which will require (most likely) new `SecondaryToolbar` buttons and icons for those etc.
Hence this patch only lays the necessary foundation for eventually addressing issue 10736, but defers the actual implementation until later. (Time permitting, I'll try to look into the viewer part later.)
*Please note:* This patch purposely ignores `src/display/network.js`, since its support for progressive reading depends on the non-standard `moz-chunked-arraybuffer` responseType which is currently in the process of being removed.
With PR 10675 having fixed the completely broken `disableRange=true` setting in the Firefox version of PDF.js, I couldn't help but noticing that loading progress is never reported properly in that case.
Currently loading progress is only reported for the `rangeProgress` chrome-event, which obviously isn't dispatched with `disableRange=true` set. However, the `progressiveRead` chrome-event includes loading progress as well, but this information isn't being used in any way.
Furthermore, the `PDFDataRangeTransport.onDataProgress` method wasn't able to handle "complete" loading information, and neither was `PDFDataTransportStream._onProgress` since that method would only ever attempt to report it through a RangeReader (which won't exist when `disableRange=true` is set).
Note how at https://mozilla.github.io/pdf.js/api/ it's being described as API docs, however `src/core/annotation.js` is not part of the public API.
Furthermore, given that the code residing in the `src/core/` folder is run in a worker-thread, it's not even accessible on the main-thread (since `postMessage` is being used to transfer the data).
Hence the different API methods simply returns a "proxy" to the underlying data, but not actually the same objects and data structures as in the worker-thread itself; thus it doesn't make a whole lot of sense to expose this in API docs as far as I'm concerned.
Finally, the patch fixes a small JSDoc related typo in `src/display/api.js` when referring to the `TextStyle` typedef.
Currently if trying to set `disableRange=true` in the built-in PDF Viewer in Firefox, either through `about:config` or via the URL hash, the PDF document will never load. It appears that this has been broken for a couple of years, without anyone noticing.
Obviously it's not a good idea to set `disableRange=true`, however it seems that this bug affects the PDF Viewer in Firefox even with default settings:
- In the case where `initialData` already contains the *entire* file, we're forced to dispatch a range request to re-fetch already available data just so that file loading may complete.
- (In the case where the data arrives, via streaming, before being specifically requested through `requestDataRange`, we're also forced to re-fetch data unnecessarily.) *This part was removed, to reduce the scope/risk of the patch somewhat.*
In the cases outlined above, we're having to re-fetch already available data thus potentially delaying loading/rendering of PDF files in Firefox (and wasting resources in the process).
Natively supported JPEG images are sent as-is, using a `blob:` or possibly a `data` URL, to the main-thread for loading/decoding.
However there's currently no attempt at releasing these resources, which are held alive by `blob:` URLs, which seems unfortunately given that images can be arbitrarily large.
As mentioned in https://developer.mozilla.org/en-US/docs/Web/API/URL/createObjectURL the lifetime of these URLs are tied to the document, hence they are not being removed when a page is cleaned-up/destroyed (e.g. when being removed from the `PDFPageViewBuffer` in the viewer).
This is easy to test with the help of `about:memory` (in Firefox), which clearly shows the number of `blob:` URLs becomming arbitrarily large *without* this patch. With this patch however the `blob:` URLs are immediately release upon clean-up as expected, and the memory consumption should thus be considerably reduced for long documents with (simple) JPEG images.
Note how `PDFDocumentProxy.destroy` is a nothing more than an alias for `PDFDocumentLoadingTask.destroy`. While removing the latter method would be a breaking API change, there's still room for at least some clean-up here.
The main changes in this patch are:
- Stop providing a `PDFDocumentLoadingTask` instance *separately* when creating a `PDFDocumentProxy`, since the loadingTask is already available through the `WorkerTransport` instance.
- Stop tracking the `PDFDocumentProxy` instance on the `WorkerTransport`, since that property is completely unused.
- Simplify the 'Multiple `getDocument` instances' unit-tests by only destroying *once*, rather than twice, for each document.
Given that the function is (purposely) independent of the verbosity level and that its message is worded to only apply on the main-thread, there's no reason to duplicate this across the built `pdf.js`/`pdf.worker.js` files.
Currently for every single parsed/rendered page there's no less than *four* `Date.now()` calls being made on the worker-side. This seems totally unnecessary, since the result of these calls are, by default, not used for anything *unless* the verbosity level is set to `INFO`.
This file (currently) contains not only DOM-specific helper functions/classes, but is used generally for various helper code relevant for main-thread functionality.
After PR 9340 all glyphs are now re-mapped to a Private Use Area (PUA) which means that if a font fails to load, for whatever reason[1], all glyphs in the font will now render as Unicode glyph outlines.
This obviously doesn't look good, to say the least, and might be seen as a "regression" since previously many glyphs were left in their original positions which provided a slightly better fallback[2].
Hence this patch, which implements a *general* fallback to the PDF.js built-in font renderer for fonts that fail to load (i.e. are rejected by the sanitizer). One caveat here is that this only works for the Font Loading API, since it's easy to handle errors in that case[3].
The solution implemented in this patch does *not* in any way delay the loading of valid fonts, which was the problem with my previous attempt at a solution, and will only require a bit of extra work/waiting for those fonts that actually fail to load.
*Please note:* This patch doesn't fix any of the underlying PDF.js font conversion bugs that's responsible for creating corrupt font files, however it does *improve* rendering in a number of cases; refer to this possibly incomplete list:
[Bug 1524888](https://bugzilla.mozilla.org/show_bug.cgi?id=1524888)
Issue 10175
Issue 10232
---
[1] Usually because the PDF.js font conversion code wasn't able to parse the font file correctly.
[2] Glyphs fell back to some default font, which while not accurate was more useful than the current state.
[3] Furthermore I'm not sure how to implement this generally, assuming that's even possible, and don't really have time/interest to look into it either.
- The only existing call-site, of this method, is never passing more than *one* font at a time anyway.
- As far as I can remember, this functionality has never actually been used (caveat: I didn't check the git history).
- This allows simplification of the method, especially by making use of the fact that it's now asynchronous.
- It should be just as easy to call `BaseFontLoader.bind` from within a loop, rather than having the loop in the method itself.
Currently all fonts are using the `_queueLoadingCallback` method to determine when they have been loaded[1]. However in most cases this is just adding unnecessary overhead, especially with `BaseFontLoader.bind` now being asynchronous, given how fonts are loaded:
- For fonts loaded using the Font Loading API, it's already possible to easily tell when a font has been loaded simply by checking the `loaded` promise on the FontFace object itself.
- For browsers, e.g. Firefox, which support synchronous font loading it's already assumed that fonts are immediately available.
Hence the `_queueLoadingCallback` method is moved into the `GenericFontLoader`, such that it's only utilized for fonts which are loaded using CSS.
---
[1] In the "fonts loaded using CSS" case, this is already a hack anyway as outlined in the comments.
This polyfill is currently used in only *one* file, i.e. `src/display/api.js`, and only when trying to build a *fallback* `workerSrc` path.
Given that the global `workerSrc` should *always* be set[1] when using the PDF.js library[2], and that the fallback `workerSrc` should only be regarded as a best-effort solution anyway, there isn't a particularily strong reason to keep the compatibility code in my opinion.
---
[1] Other supported options include setting the global `workerPort`, or passing in a `PDFWorker` instance as part of the `getDocument` call.
[2] Which is clearly mentioned in the JSDocs in `src/display/worker_options.js`.
This piggybacks of the existing `cancel` functionality, to ensure that any pending operations are closed *and* that any temporary canvases are actually being removed.
Also simplifies `finishPaintTask` in `PDFPageView.draw` slightly, by converting it to an async function.
If, as PR 10368 suggests, more parameters should be added to `getViewport` I think that it would be a mistake to not change the signature *first* to avoid needlessly unwieldy call-sites.
To not break any existing code and third-party use-cases, this is obviously implemented with a deprecation warning *and* with a working fallback[1] for the old method signature.
---
[1] This is limited to `GENERIC` builds, which should be sufficient.
Note that the OpenAction dictionary may contain other information besides just a destination array, e.g. instructions for auto-printing[1].
Given first of all that an arbitrary `Dict` cannot be sent from the Worker (since cloning would fail), and second of all that the data obviously needs to be validated, this patch purposely only adds support for fetching a destination from the OpenAction entry[2].
---
[1] This information is, currently in PDF.js, being included through the `getJavaScript` API method.
[2] This significantly reduces the complexity of the implementation, which seems fine for now. If there's ever need for other kinds of OpenAction to be fetched, additional API methods could/should be implemented as necessary (could e.g. follow the `getOpenActionWhatever` naming scheme).
This changes all occurrences of `var` to `let`/`const` in this code, and updates the signature of the constructor to use object destructuring for better readability (and self documentation).
Also, `useRequestAnimationFrame` is changed to a parameter and the `typeof window` check is now done *once* rather than at every `_scheduleNext` call.
