* FIX: Better virtual keyboard detect on Android
Firefox has a bug where *sometimes* the visualViewport.height won't be
updated when the keyboard pops up until you scroll, making our composer
stay hidden behind the keyboard. This commit uses both window.innerHeight
and window.visualViewport.height using the minimum of both to check for
height changes.
For Chrome/Edge we feature detect the new VirtualKeyboard API and
opt-into it when the composer opens and use it to detect if a keyboard
is being draw. Opting into the API changes how the viewport is
calculated so we have to also change how the full height composer is
calculated. To minimize breakage we opt-out when the composer component
is destroyed.
This commit also moves the `--composer-ipad-padding` to only happen on
iPads.
Bug report at https://meta.discourse.org/t/-/228382
* DEV: Add test case for syntax highlight of complex HTML
The commit 685e0da upgrade HighlightJS to version 11, which deprecates
syntax highlight of complex HTML elements. See https://github.com/highlightjs/highlight.js/issues/2889
This brought a regression of syntax highlighting of GitHub oneboxes,
which was fixed in 09cec7d. This commit adds a test case to prevent
future regressions like this one.
* fix test and warning
At some point moved from hljs pure source to their CDN assets, but we
did not change the way we created the HLJS bundle. The CDN asset comes
with their "common" languages already included, so we were duplicating
around 35 languagues in the bundle we create.
This patch includes a list of their current common langs so we never
double bundle those.
Changes in size are significant. Numbers before minimization/compression
are 747.53 KB before and 117.57 KB after.
Highlight.js 11 deprecated the feature to highlight HTML blocks while
keeping the HTML structure, which broke our GitHub onebox syntax
highlight.
This patch adds it back by bringing the maintainers code as a plugin.
See https://github.com/highlightjs/highlight.js/issues/2889
If userA has multiple tab/devices on the same topic, and:
1. userA likes a post in tab1
2. userB likes the same post
3. userA post like `acted` attr would desync in tab2
This fix handles this case and also the reverse one when removing likes
interleaved with other users acting on the same post.
Reported in Meta at https://meta.discourse.org/t/-/227239/3
This fixes a corner case of the perf optimization in d4e35f5.
When you have the the same post showing in multiple tab/devices and like
said post in one place, we updated the like count but didn't flip the
`acted` bool in the front-end. This caused a small visual desync.
Co-authored-by: Penar Musaraj <pmusaraj@gmail.com>
Admins won't be able to disable strip_image_metadata if they don't
disable composer_media_optimization_image_enabled first since the later
will strip the same metadata on client during upload, making disabling
the former have no effect.
Bug report at https://meta.discourse.org/t/-/223350
If a user copies a gif from a website into their clipboard and then
tries to paste it into the Discourse composer, we would only paste a
static single frame of the original gif. This happens because the
browser doesn't store the original image in the clipboard, but two
entries:
1. image/png with the frame of the copy moment
2. text/html with the markup of the gif img element
This commit adds an heuristic that detects this and makes us pick the
clipboard content of text/html instead of the image/png when this
happens.
From there our existing HTML paste logic handles and converts the HTML
img tag into markdown, preserving even the alt text.
See https://meta.discourse.org/t/-/218720 for context.
Stop Discourse from prompting for push notification on latest iOS beta
where the navigators exposes a broken Push object.
We had some feature detection functions that where outside our
pre-initializer that it dedicated for this stuff. All feature detection
now lives on sniff-capabilities file.
Also removed some old browser detection from the push notifications
code, and simplified the function signature because of it.
Co-authored-by: Jarek Radosz <jradosz@gmail.com>
* FIX: Remove svg icons from webmanifest shortcuts
While SVGs are valid in the webmanifest, Chromium has not implemented
support for it in this specific manifest member.
Revert when https://bugs.chromium.org/p/chromium/issues/detail?id=1091612
lands.
* fix test
* FEATURE: Export topics to markdown
The route `/raw/TOPIC_ID` will now export whole topics (paginated to 100
posts) in a markdown format.
See https://meta.discourse.org/t/-/152185/12
This is a workaround a behavior change in Chromium v97.
The following text was sent to the blink-dev mailing list:
> This change broke a SingleSignOn login on the FOSS software Discourse. We have a flow like:
>
> 1. User visits forum.siteA.com, click login
> 2. Gets redirected to idp.siteB.com
> 3. Fills login details
> 4. Gets redirected to forum.siteA.com/session/sso_login?parameters
> 5. Gets redirected to forum.siteA.com/homepage
>
> On step 4, the response includes a `set-cookie` header, with proper `HttpOnly; SameSite=Lax; Secure `and set. But if there is an active service worker, the login will fail as that cookie will be rejected by Chromium due to SameSite rules now.
>
> t=2971 [st=258] COOKIE_INCLUSION_STATUS
> --> domain = "forum.siteA.com"
> --> name = "_t"
> --> operation = "store"
> --> path = "/"
> --> status = "EXCLUDE_SAMESITE_LAX, DO_NOT_WARN"
>
> The service worker is a vanilla WorkboxJS service worker that intercepts all GETs with the "Network First" strategy.
>
> Disabling the service worker or using Firefox results in a successful login. There is no warning in either DevTools network tab nor the console that the cookie was rejected.
>
> Chrome 96: login works
> Chrome 97: login does not work
> Chrome 98: login does not work
>
> Is this expected behavior? Even if the request `GET forum.siteA.com` was initiated because of a redirect from a different domain, is it expected that Chrome will silently drop same site cookies from forum.siteA.com?
* FEATURE: Optionally send a 'noindex' header in non-canonical responses
This will be used in a SEO experiment.
Co-authored-by: David Taylor <david@taylorhq.com>
PERF: Update like count in visible posts without an extra GET per like
Currently when a user is reading a topic and some post in it receive a
like from another user, the Ember app will be notified via MessageBus
and issue a GET to `/posts/{id}` to get the new like count. This worked
fine for us until today, but it can easily create a self-inflicted DDoS
when a topic with a large number of visitors gets a large number of
likes, since we will issue `visitors * likes` GET requests requests.
This patch optimizes this flow, by sending the new like count down in
the MessageBus notification, removing any need for the extra request.
It shouldn't cause any drift on the count because we send down the full
count instead of the difference too.
Possible follow-ups could include handling like removal.
* FEATURE: Cache CORS preflight requests for 2h
Browsers will cache this for 5 seconds by default. If using MessageBus
in a different domain, Discourse will issue a new long polling, by
default, every 30s or so. This means we would be issuing a new preflight
request **every time**. This can be incredibly wasteful, so let's cache
the authorization in the client for 2h, which is the maximum Chromium
allows us as of today.
* fix tests
DEV: Allow passing cook_method to TopicEmbed.import to override default
This will be used in the rss-polling plugin when we want to have
oneboxes on feed content, like youtube for example.
Allow admins to configure exceptions to our Rails rate limiter.
Configuration happens in the environment variables, and work with both
IPs and CIDR blocks.
Example:
```
env:
DISCOURSE_MAX_REQS_PER_IP_EXCEPTIONS: >-
14.15.16.32/27
216.148.1.2
```
Flips content_security_policy_frame_ancestors default to enabled, and
removes HTTP_REFERER checks on embed requests, as the new referer
privacy options made the check fragile.
There was a UI bug when submitting multiple files in the same batch. We
would remove the disabled status of the submit button after the previous
file was sucesfully uploaded and the next one was still mid
optimization.
Reported at https://meta.discourse.org/t/-/194841/15?u=falco
* FIX: Detect decode failures earlier in image optimization pipeline
Follow up to 9b51b9b but also detects the bug earlier and backs off.
What iOS 15 is doing is returning all zeroes to `ctx.getImageData`,
so we don't have to wait until resize to detect the problem.
* Update app/assets/javascripts/discourse/app/lib/media-optimization-utils.js
Co-authored-by: Jarek Radosz <jradosz@gmail.com>
Co-authored-by: Jarek Radosz <jradosz@gmail.com>
On iOS 15 beta, if you select the camera app when uploading an image
and try to upload a freshly taken picture, from the second picture
onwards the resize WASM operation will return an array filled with
zeroes.
Since every 4th byte is alpha, and at this step we are only dealing with
non-transparent images this a O(1) way to detect that the bug was hit.
(On normal images, all 4th bytes are 255 at this point)
Also adds a "catch-all" when the original image became too small to try
to accomodate other bugs of the same type. By default we only trigger
this whole operation on images over 1MB, so if the end result is <20KB
something weird did happen. Throwing here will let the upload continue
using the original file, so nothing is lost and the user can continue.
There are some hard limits in browser Canvas implementations, that will
throw a runtime exception when crossed. Since those limits are platform
dependent, the best we can do is catch it and back off from trying to
optimize a problematic file.
For example, a 60MB PNG can be processed fine by Chrome but Firefox will
fail trying to extract the ImageData from the CanvasRenderingContext2D
with NS_ERROR_FAILURE.
Also cleans up the media-optimization-utils and add post-resize size logs
Re-lands the change initially proposed on #8359 but without a new nginx
location block, so it has less change surface.
Co-authored-by: Jeff Wong <awole20@gmail.com>
Co-authored-by: Jeff Wong <awole20@gmail.com>
* FEATURE: Small improvements to the topic list embed
- Ability to wrap the list in a custom class so you can styles different
lists using specific CSS
- Adds a topic link to the thumbnail when using the complete template
* FIX: Be more strict about allowed chars in class name
Previosly, if the topic embed request had a missing username parameter,
and SiteSetting.embed_by_username was empty we would fail to create the
new topic and not show any errors.
Now we will fallback using the priority:
1. Username parameter
2. SiteSetting.embed_by_username
3. SiteSetting.site_contact_username
4. system user
`rake emails:test` task was always sending `localhost` as the domain name rather than using `smtp[:domain]` (aka `DISCOURSE_SMTP_DOMAIN`. `discourse-setup` has recently been modified to always set `DISCOURSE_SMTP_DOMAIN`, so it's important that this test rake task actually use the value.
I tested this change on a standard production install, and it's working as expected. Hopefully I managed not to bungle the copy/paste of the single line here in the github edit window.
Co-authored-by: Jay Pfaffman <jay@literatecomputing.com>
* FIX: Subfolder replace should only affect URL prefix
Issue was reported in https://meta.discourse.org/t/-/179504
* DEV: Test subfolder handling in get-url when called twice on the same path
A user browser may rotate a user subscription endpoint/keys
anytime.
Currently, Discourse will receive a 4XX response while trying to
deliver a push notification and silently unsubscribe the device.
With this change, we will gracefully handle desativating the old
subscription and the replacement creation with the need for the user
to resubscribe manually every time it breaks.
https://meta.discourse.org/t/-/125179?u=falco
