This commits updates `FinalDestination#get` to not forward
`Authorization` header on redirects since most HTTP clients I tested like
curl and wget does not it.
This also fixes a recent problem in `DiscourseIpInfo.mmdb_download`
where we will fail to download the databases when both `GlobalSetting.maxmind_account_id` and
`GlobalSetting.maxmind_license_key` has been set. The failure is due to
the bug above where the redirected URL given by MaxMind does not accept
an `Authorization` header.
This commit adds some system specs to test uploads with
direct to S3 single and multipart uploads via uppy. This
is done with minio as a local S3 replacement. We are doing
this to catch regressions when uppy dependencies need to
be upgraded or we change uppy upload code, since before
this there was no way to know outside manual testing whether
these changes would cause regressions.
Minio's server lifecycle and the installed binaries are managed
by the https://github.com/discourse/minio_runner gem, though the
binaries are already installed on the discourse_test image we run
GitHub CI from.
These tests will only run in CI unless you specifically use the
CI=1 or RUN_S3_SYSTEM_SPECS=1 env vars.
For a history of experimentation here see https://github.com/discourse/discourse/pull/22381
Related PRs:
* https://github.com/discourse/minio_runner/pull/1
* https://github.com/discourse/minio_runner/pull/2
* https://github.com/discourse/minio_runner/pull/3
### Background
When SSRF detection fails, the exception bubbles all the way up, causing a log alert. This isn't actionable, and should instead be ignored. The existing `rescue` does already ignore network errors, but fails to account for SSRF exceptions coming from `FinalDestination`.
### What is this change?
This PR does two things.
---
Firstly, it introduces a common root exception class, `FinalDestination::SSRFError` for SSRF errors. This serves two functions: 1) it makes it easier to rescue both errors at once, which is generally what one wants to do and 2) prevents having to dig deep into the class hierarchy for the constant.
This change is fully backwards compatible thanks to how inheritance and exception handling works.
---
Secondly, it rescues this new exception in `UserAvatar.import_url_for_user`, which is causing sporadic errors to be logged in production. After this SSRF errors are handled the same as network errors.
This is closer to git's redirect following behaviour. We prevented git
following redirects when we clone in order to prevent SSRF attacks.
Follow-up-to: 291bbc4fb9
This is a much better description of its function. It performs idempotent normalization of a URL. If consumers truly need to `encode` a URL (including double-encoding of existing encoded entities), they can use the existing `.encode` method.
Currently when generating oneboxes if the connection timeouts and we’re
using the `FinalDestination#get` method, then it raises an exception.
We already catch this exception when using the
`FinalDestination#resolve` method so this patch just applies the same
logic to `FinalDestination#get`.
Meta topic: https://meta.discourse.org/t/prevent-to-linkify-when-there-is-a-redirect/226964/2?u=osama.
This commit adds a new site setting `block_onebox_on_redirect` (default off) for blocking oneboxes (full and inline) of URLs that redirect. Note that an initial http → https redirect is still allowed if the redirect location is identical to the source (minus the scheme of course). For example, if a user includes a link to `http://example.com/page` and the link resolves to `https://example.com/page`, then the link will onebox (assuming it can be oneboxed) even if the setting is enabled. The reason for this is a user may type out a URL (i.e. the URL is short and memorizable) with http and since a lot of sites support TLS with http traffic automatically redirected to https, so we should still allow the URL to onebox.
If the crawled page returned an error, `FinalDestination#safe_get`
yielded `nil` for `uri` and `chunk` arguments. Another problem is that
`get` did not handle the case when `safe_get` failed and did not return
the `location` and `set_cookie` headers.
The `blocked onebox domains` setting lets site owners change what sites
are allowed to be oneboxed. When a link is entered into a post,
Discourse checks the domain of the link against that setting and blocks
the onebox if the domain is blocked. But if there's a chain of
redirects, then only the final destination website is checked against
the site setting.
This commit amends that behavior so that every website in the redirect
chain is checked against the site setting, and if anything is blocked
the original link doesn't onebox at all in the post. The
`Discourse-No-Onebox` header is also checked in every response and the
onebox is blocked if the header is set to "1".
Additionally, Discourse will now include the `Discourse-No-Onebox`
header with every response if the site requires login to access content.
This is done to signal to a Discourse instance that it shouldn't attempt
to onebox other Discourse instances if they're login-only. Non-Discourse
websites can also use include that header if they don't wish to have
Discourse onebox their content.
Internal ticket: t59305.
The default of 1Mb was preventing some valid Onebox requests from successfully completing.
Increasing this to 5Mb should reduce the number of unexpected failures.
In an earlier PR, we decided that we only want to block a domain if
the blocked domain in the SiteSetting is the final destination (/t/59305). That
PR used `FinalDestination#get`. `resolve` however is used several places
but blocks domains along the redirect chain when certain options are provided.
This commit changes the default options for `resolve` to not do that. Existing
users of `FinalDestination#resolve` are
- `Oneboxer#external_onebox`
- our onebox helper `fetch_html_doc`, which is used in amazon, standard embed
and youtube
- these folks already go through `Oneboxer#external_onebox` which already
blocks correctly
FinalDestination's follow_canonical mode used for embedded topics should work when canonical URLs are relative, as specified in [RFC 6596](https://datatracker.ietf.org/doc/html/rfc6596)
If the initial `get`/`head` response within `resolve` returns a status code of `103`, attempt to fetch the same URL with the alternative `small_get` method.
FinalDestination now supports the `follow_canonical` option, which will perform an initial GET request, parse the canonical link if present, and perform a HEAD request to it.
We use this mode during embeds to avoid treating URLs with different query parameters as different topics.
* FEATURE: Onebox can match engines based on the content_type
`FinalDestination` now returns the `content_type` of a resolved URL.
`Oneboxer` passes this value to `Onebox` itself. Onebox engines can now specify a `matches_content_type` regex of content_types that the engine can handle, regardless of the URL.
`ImageOnebox` will match URLs with a content type of `image/png`, `jpg`, `gif`, `bmp`, `tif`, etc.
This will allow images that exist at a URL without a file type extension to be correctly rendered, assuming a valid `content_type` is returned.
Not specifying an `Accept-Language` should be equivalent to specifying an `Accept-Language` of `*`, however some webservers seem to prefer it if we are explicit about being able to handle a response of content in any language.
* DEV: Allow wildcards in Oneboxer optional domain Site Settings
Allows a wildcard to be used as a subdomain on Oneboxer-related SiteSettings, e.g.:
- `force_get_hosts`
- `cache_onebox_response_body_domains`
- `force_custom_user_agent_hosts`
* DEV: fix typos
* FIX: Try doing a GET after receiving a 500 error from a HEAD
By default we try to do a `HEAD` requests. If this results in a 500 error response, we should try to do a `GET`
* DEV: `force_get_hosts` should be a hidden setting
* DEV: Oneboxer Strategies
Have an alternative oneboxing ‘strategy’ (i.e., set of options) to use when an attempt to generate a Onebox fails. Keep track of any non-default strategies that were used on a particular host, and use that strategy for that host in the future.
Initially, the alternate strategy (`force_get_and_ua`) forces the FinalDestination step of Oneboxing to do a `GET` rather than `HEAD`, and forces a custom user agent.
* DEV: change stubbed return code
The stubbed status code needs to be a value not recognized by FinalDestination
Previous refactors have lost usage of read_timeout in `FileHelper.download` and `FinalDestination` was incorrectly using `Net::HTTP.start` by setting `open_timeout` in the block instead of directly during the invocation.
Couldn't figure how to write a good test for this without slowing the spec.
* FEATURE: Cache successful HTTP GET requests during Oneboxing
Some oneboxes may fail if when making excessive and/or odd requests against the target domains. This change provides a simple mechanism to cache the results of succesful GET requests as part of the oneboxing process, with the goal of reducing repeated requests and ultimately improving the rate of successful oneboxing.
To enable:
Set `SiteSetting.cache_onebox_response_body` to `true`
Add the domains you’re interesting in caching to `SiteSetting. cache_onebox_response_body_domains` e.g. `example.com|example.org|example.net`
Optionally set `SiteSetting.cache_onebox_user_agent` to a user agent string of your choice to use when making requests against domains in the above list.
* FIX: Swap order of duration and value in redis call
The correct order for `setex` arguments is `key`, `duration`, and `value`.
Duration and value had been flipped, however the code would not have thrown an error because we were caching the value of `1.day.to_i` for a period of 1 seconds… The intention appears to be to set a value of 1 (purely as a flag) for a period of 1 day.
* FEATURE: display error if Oneboxing fails due to HTTP error
- display warning if onebox URL is unresolvable
- display warning if attributes are missing
* FEATURE: Use new Instagram oEmbed endpoint if access token is configured
Instagram requires an Access Token to access their oEmbed endpoint. The requirements (from https://developers.facebook.com/docs/instagram/oembed/) are as follows:
- a Facebook Developer account, which you can create at developers.facebook.com
- a registered Facebook app
- the oEmbed Product added to the app
- an Access Token
- The Facebook app must be in Live Mode
The generated Access Token, once added to SiteSetting.facebook_app_access_token, will be passed to onebox. Onebox can then use this token to access the oEmbed endpoint to generate a onebox for Instagram.
* DEV: update user agent string
* DEV: don’t do HEAD requests against news.yahoo.com
* DEV: Bump onebox version from 2.1.5 to 2.1.6
* DEV: Avoid re-reading templates
* DEV: Tweaks to onebox mustache templates
* DEV: simplified error message for missing onebox data
* Apply suggestions from code review
Co-authored-by: Gerhard Schlager <mail@gerhard-schlager.at>
The following methods have long been deprecated in ruby due to flaws in their implementation per http://blade.nagaokaut.ac.jp/cgi-bin/vframe.rb/ruby/ruby-core/29293?29179-31097:
URI.escape
URI.unescape
URI.encode
URI.unencode
escape/encode are just aliases for one another. This PR uses the Addressable gem to replace these methods with its own encode, unencode, and encode_component methods where appropriate.
I have put all references to Addressable::URI here into the UrlHelper to keep them corralled in one place to make changes to this implementation easier.
Addressable is now also an explicit gem dependency.
This also corrects FileHelper.download so it supports "follow_redirect"
correctly (it used to always follow 1 redirect) and adds a `validate_url`
param that will bypass all uri validation if set to false (default is true)
This reduces chances of errors where consumers of strings mutate inputs
and reduces memory usage of the app.
Test suite passes now, but there may be some stuff left, so we will run
a few sites on a branch prior to merging
* `rescue nil` is a really bad pattern to use in our code base.
We should rescue errors that we expect the code to throw and
not rescue everything because we're unsure of what errors the
code would throw. This would reduce the amount of pain we face
when debugging why something isn't working as expexted. I've
been bitten countless of times by errors being swallowed as a
result during debugging sessions.
