This PR enhances the CORS middleware usage by allowing for the headers
to be configured in `app.ini`.
Fixes#21746
Co-authored-by: KN4CK3R <admin@oldschoolhack.me>
Co-authored-by: John Olheiser <john.olheiser@gmail.com>
Co-authored-by: Lunny Xiao <xiaolunwen@gmail.com>
Attempt clarify the AppWorkPath in the documentation by using different
notation and adding a section to the start of the cheat sheet.
Fix#21523
Signed-off-by: Andrew Thornton <art27@cantab.net>
Signed-off-by: Andrew Thornton <art27@cantab.net>
Related #20471
This PR adds global quota limits for the package registry. Settings for
individual users/orgs can be added in a seperate PR using the settings
table.
Co-authored-by: Lauris BH <lauris@nix.lv>
Co-authored-by: Lunny Xiao <xiaolunwen@gmail.com>
Close https://github.com/go-gitea/gitea/issues/21640
Before: Gitea can create users like ".xxx" or "x..y", which is not
ideal, it's already a consensus that dot filenames have special
meanings, and `a..b` is a confusing name when doing cross repo compare.
After: stricter
Co-authored-by: Jason Song <i@wolfogre.com>
Co-authored-by: Lunny Xiao <xiaolunwen@gmail.com>
Co-authored-by: delvh <dev.lh@web.de>
Fix https://github.com/go-gitea/gitea/issues/21666
Caused by https://github.com/go-gitea/gitea/pull/19663
Before: when install, the INTERNAL_TOKEN was always generated and saved.
But the internal token may be already there by INTERNAL_TOKEN_URI
After: INTERNAL_TOKEN_URI file must be non-empty. When install, skip
internal token generation if the token exists.
_This is a different approach to #20267, I took the liberty of adapting
some parts, see below_
## Context
In some cases, a weebhook endpoint requires some kind of authentication.
The usual way is by sending a static `Authorization` header, with a
given token. For instance:
- Matrix expects a `Bearer <token>` (already implemented, by storing the
header cleartext in the metadata - which is buggy on retry #19872)
- TeamCity #18667
- Gitea instances #20267
- SourceHut https://man.sr.ht/graphql.md#authentication-strategies (this
is my actual personal need :)
## Proposed solution
Add a dedicated encrypt column to the webhook table (instead of storing
it as meta as proposed in #20267), so that it gets available for all
present and future hook types (especially the custom ones #19307).
This would also solve the buggy matrix retry #19872.
As a first step, I would recommend focusing on the backend logic and
improve the frontend at a later stage. For now the UI is a simple
`Authorization` field (which could be later customized with `Bearer` and
`Basic` switches):
The header name is hard-coded, since I couldn't fine any usecase
justifying otherwise.
## Questions
- What do you think of this approach? @justusbunsi @Gusted @silverwind
- ~~How are the migrations generated? Do I have to manually create a new
file, or is there a command for that?~~
- ~~I started adding it to the API: should I complete it or should I
drop it? (I don't know how much the API is actually used)~~
## Done as well:
- add a migration for the existing matrix webhooks and remove the
`Authorization` logic there
_Closes #19872_
Co-authored-by: Lunny Xiao <xiaolunwen@gmail.com>
Co-authored-by: Gusted <williamzijl7@hotmail.com>
Co-authored-by: delvh <dev.lh@web.de>
I found myself wondering whether a PR I scheduled for automerge was
actually merged. It was, but I didn't receive a mail notification for it
- that makes sense considering I am the doer and usually don't want to
receive such notifications. But ideally I want to receive a notification
when a PR was merged because I scheduled it for automerge.
This PR implements exactly that.
The implementation works, but I wonder if there's a way to avoid passing
the "This PR was automerged" state down so much. I tried solving this
via the database (checking if there's an automerge scheduled for this PR
when sending the notification) but that did not work reliably, probably
because sending the notification happens async and the entry might have
already been deleted. My implementation might be the most
straightforward but maybe not the most elegant.
Signed-off-by: Andrew Thornton <art27@cantab.net>
Co-authored-by: Lauris BH <lauris@nix.lv>
Co-authored-by: Andrew Thornton <art27@cantab.net>
Co-authored-by: Lunny Xiao <xiaolunwen@gmail.com>
A patch to #17335.
Just like AppPath, Gitea writes its own CustomConf into git hook scripts
too. If Gitea's CustomConf changes, then the git push may fail.
Co-authored-by: techknowlogick <techknowlogick@gitea.io>
This patch provide a mechanism to disable
[sitemap](https://github.com/go-gitea/gitea/pull/18407).
Signed-off-by: Xinyu Zhou <i@sourcehut.net>
Co-authored-by: 6543 <6543@obermui.de>
Follow-up to #21352: when testing this out, I found out that the syntax
for the `symbolic-ref` command was different than originally assumed in
that PR. Never got around to submitting a PR until now.
Co-authored-by: Lunny Xiao <xiaolunwen@gmail.com>
Our implementation of `IsZero` can't work. An "empty" timestamp (= 0)
calls `time.Unix(int64(ts), 0).IsZero()` which is always `false`. Only
`time.Time{}.IsZero()` is `true`.
We call this method ~~only at one place~~ and there the value
(`UpdatedUnix`) should be always != 0 so this PR may not have
consequences.
Co-authored-by: Lunny Xiao <xiaolunwen@gmail.com>
I noticed an admin is not allowed to upload packages for other users
because `ctx.IsSigned` was not set.
I added a check for `user.IsActive` and `user.ProhibitLogin` too because
both was not checked. Tests enforce this now.
Co-authored-by: Lauris BH <lauris@nix.lv>
The OAuth spec [defines two types of
client](https://datatracker.ietf.org/doc/html/rfc6749#section-2.1),
confidential and public. Previously Gitea assumed all clients to be
confidential.
> OAuth defines two client types, based on their ability to authenticate
securely with the authorization server (i.e., ability to
> maintain the confidentiality of their client credentials):
>
> confidential
> Clients capable of maintaining the confidentiality of their
credentials (e.g., client implemented on a secure server with
> restricted access to the client credentials), or capable of secure
client authentication using other means.
>
> **public
> Clients incapable of maintaining the confidentiality of their
credentials (e.g., clients executing on the device used by the resource
owner, such as an installed native application or a web browser-based
application), and incapable of secure client authentication via any
other means.**
>
> The client type designation is based on the authorization server's
definition of secure authentication and its acceptable exposure levels
of client credentials. The authorization server SHOULD NOT make
assumptions about the client type.
https://datatracker.ietf.org/doc/html/rfc8252#section-8.4
> Authorization servers MUST record the client type in the client
registration details in order to identify and process requests
accordingly.
Require PKCE for public clients:
https://datatracker.ietf.org/doc/html/rfc8252#section-8.1
> Authorization servers SHOULD reject authorization requests from native
apps that don't use PKCE by returning an error message
Fixes#21299
Co-authored-by: wxiaoguang <wxiaoguang@gmail.com>
Co-authored-by: Lunny Xiao <xiaolunwen@gmail.com>
Previously mentioning a user would link to its profile, regardless of
whether the user existed. This change tests if the user exists and only
if it does - a link to its profile is added.
* Fixes#3444
Signed-off-by: Yarden Shoham <hrsi88@gmail.com>
Co-authored-by: wxiaoguang <wxiaoguang@gmail.com>
Co-authored-by: Lunny Xiao <xiaolunwen@gmail.com>
It distractingly shows up on unit tests
* Looks like a leftover from #20571
Signed-off-by: Yarden Shoham <hrsi88@gmail.com>
Co-authored-by: Lunny Xiao <xiaolunwen@gmail.com>
At the moment a repository reference is needed for webhooks. With the
upcoming package PR we need to send webhooks without a repository
reference. For example a package is uploaded to an organization. In
theory this enables the usage of webhooks for future user actions.
This PR removes the repository id from `HookTask` and changes how the
hooks are processed (see `services/webhook/deliver.go`). In a follow up
PR I want to remove the usage of the `UniqueQueue´ and replace it with a
normal queue because there is no reason to be unique.
Co-authored-by: 6543 <6543@obermui.de>
A lot of our code is repeatedly testing if individual errors are
specific types of Not Exist errors. This is repetitative and unnecesary.
`Unwrap() error` provides a common way of labelling an error as a
NotExist error and we can/should use this.
This PR has chosen to use the common `io/fs` errors e.g.
`fs.ErrNotExist` for our errors. This is in some ways not completely
correct as these are not filesystem errors but it seems like a
reasonable thing to do and would allow us to simplify a lot of our code
to `errors.Is(err, fs.ErrNotExist)` instead of
`package.IsErr...NotExist(err)`
I am open to suggestions to use a different base error - perhaps
`models/db.ErrNotExist` if that would be felt to be better.
Signed-off-by: Andrew Thornton <art27@cantab.net>
Co-authored-by: delvh <dev.lh@web.de>
Fixes#21379
The commits are capped by `setting.UI.FeedMaxCommitNum` so
`len(commits)` is not the correct number. So this PR adds a new
`TotalCommits` field.
Co-authored-by: wxiaoguang <wxiaoguang@gmail.com>
This changes the rendering logic of issue titles. If a substring in an
issue title is enclosed with a pair of backticks, it'll be rendered with
a monospace font (HTML `code` tag).
* Closes#20887
Signed-off-by: Yarden Shoham <hrsi88@gmail.com>
Co-authored-by: Gusted <williamzijl7@hotmail.com>
Co-authored-by: wxiaoguang <wxiaoguang@gmail.com>
Co-authored-by: 6543 <6543@obermui.de>
After some discussion, introduce a new slice `brokenArgs` to make
`gitCmd.Run()` return errors if any dynamic argument is invalid.
Co-authored-by: delvh <dev.lh@web.de>
Due to a bug in presumably an older version of Gitea, multiple of my
repositories still have their HEADs pointing to a `master` branch while
the default branch on the UI is listed as `main`. This adds a `gitea
doctor` command that will fix all of the HEAD references for repos when
they're not synchronized with the default branch in the DB.
This will help with cloning to ensure that git automatically checks out
the right branch, instead of a nonexistent one.
Note: I'm not sure if I actually need to do more other than add a file
here. Will try testing this out on my server soon.
Co-authored-by: zeripath <art27@cantab.net>
Co-authored-by: wxiaoguang <wxiaoguang@gmail.com>
Fixes#17655
If you rename `user1` to `user2` and visit `/user1` you get redirected
to `/user2`. But if you visit `/user1/repo` you just get a 404 error.
With this PR the user is redirected to `/user2/repo`.
Partially fix#21213.
This PR will get client IP address from SSH_CONNECTION env which should
be the first field of that. And deliver it to the internal API so Gitea
routers could record the real IP from SSH requests.
Co-authored-by: wxiaoguang <wxiaoguang@gmail.com>
Co-authored-by: 6543 <6543@obermui.de>
Co-authored-by: zeripath <art27@cantab.net>
We should only log CheckPath errors if they are not simply due to
context cancellation - and we should add a little more context to the
error message.
Fix#20709
Signed-off-by: Andrew Thornton <art27@cantab.net>
There are repeated issues reported whereby users are unable to interpret
the template errors. This PR adds some (somewhat complex) error handling
to the panic recovery for template renderering but hopefully makes the
interpretation of the error easier.
Reference #21344
Signed-off-by: Andrew Thornton <art27@cantab.net>
Close#20315 (fix the panic when parsing invalid input), Speed up #20231 (use ls-tree without size field)
Introduce ListEntriesRecursiveFast (ls-tree without size) and ListEntriesRecursiveWithSize (ls-tree with size)
Calls to ToCommit are very slow due to fetching diffs, analyzing files.
This patch lets us supply `stat` as false to speed fetching a commit
when we don't need the diff.
/v1/repo/commits has a default `stat` set as true now. Set to false to
experience fetching thousands of commits per second instead of 2-5 per
second.
There is a small bug in #20571 whereby `$a a$b b$` will not be correctly
detected as a math inline block of `a a$b b`. This PR fixes this.
Also reenable test cases as per #21340
Signed-off-by: Andrew Thornton <art27@cantab.net>
Co-authored-by: wxiaoguang <wxiaoguang@gmail.com>
Only load SECRET_KEY and INTERNAL_TOKEN if they exist.
Never write the config file if the keys do not exist, which was only a fallback for Gitea upgraded from < 1.5
Co-authored-by: wxiaoguang <wxiaoguang@gmail.com>
This adds an api endpoint `/files` to PRs that allows to get a list of changed files.
built upon #18228, reviews there are included
closes https://github.com/go-gitea/gitea/issues/654
Co-authored-by: Anton Bracke <anton@ju60.de>
Co-authored-by: 6543 <6543@obermui.de>
Co-authored-by: wxiaoguang <wxiaoguang@gmail.com>
This PR would presumably
Fix#20522Fix#18773Fix#19069Fix#21077Fix#13622
-----
1. Check whether unit type is currently enabled
2. Check if it _will_ be enabled via opt
3. Allow modification as necessary
Signed-off-by: jolheiser <john.olheiser@gmail.com>
Co-authored-by: wxiaoguang <wxiaoguang@gmail.com>
Co-authored-by: 6543 <6543@obermui.de>
The behaviour of `PreventSurroundingPre` has changed in
https://github.com/alecthomas/chroma/pull/618 so that apparently it now
causes line wrapper tags to be no longer emitted, but we need some form
of indication to split the HTML into lines, so I did what
https://github.com/yuin/goldmark-highlighting/pull/33 did and added the
`nopWrapper`.
Maybe there are more elegant solutions but for some reason, just
splitting the HTML string on `\n` did not work.
Co-authored-by: wxiaoguang <wxiaoguang@gmail.com>
Go creates temporary files as 600, but sometimes we want the group to be able to read them (for example,
for another user to back up the storage)
This PR applies the umask to the renamed tmp files in local storage.
Co-authored-by: wxiaoguang <wxiaoguang@gmail.com>
Close#20098, in the NPM registry API, implemented to match what's described by https://github.com/npm/registry/blob/master/docs/REGISTRY-API.md#get-v1search
Currently have only implemented the bare minimum to work with the [Unity Package Manager](https://docs.unity3d.com/Manual/upm-ui.html).
Co-authored-by: Jack Vine <jackv@jack-lemur-suse.cat-prometheus.ts.net>
Co-authored-by: KN4CK3R <admin@oldschoolhack.me>
Co-authored-by: Lunny Xiao <xiaolunwen@gmail.com>
Co-authored-by: wxiaoguang <wxiaoguang@gmail.com>
Fix#20863
When REQUIRE_SIGNIN_VIEW = true, even with public repositories, you can only see them after you login. The packages should not be accessed without login.
Co-authored-by: Lauris BH <lauris@nix.lv>
Co-authored-by: wxiaoguang <wxiaoguang@gmail.com>
Follows #21119
The manual length check doesn't make sense nowadays:
1. The length check is already done by form's `binding:MaxSize` (then the manual check is unnecessary)
2. The CreateRepository doesn't have such check (then the manual check is inconsistent)
So this PR removes these manual length checks.
Both allow only limited characters. If you input more, you will get a error
message. So it make sense to limit the characters of the input fields.
Slightly relax the MaxSize of repo's Description and Website
Using `append(args, strings.Fields(arg)...)` is dangerous, it may
generate incorrect results.
For example: `arg1 "the dangerous"` will be splitted to 3 arguments:
`arg1`, `"the`, `dangerous"`. In some cases the incorrect arguments may
lead to security problems.
This fixes#5709 and #17316 by changing the order of listed branches
and tags to show the ones with latest commits atop.
It's achieved with changing underlying "show-ref" git command with
"for-each-ref" as suggested in https://stackoverflow.com/a/5188364
Also, it's passing format string so the output matches "show-ref"
command output.
close#5709close#17316
This PR adds mathematical rendering with KaTeX.
The first step is to add a Goldmark extension that detects the latex
(and tex) mathematics delimiters.
The second step to make this extension only run if math support is
enabled.
The second step is to then add KaTeX CSS and JS to the head which will
load after the dom is rendered.
Fix#3445
Signed-off-by: Andrew Thornton <art27@cantab.net>
Signed-off-by: Andrew Thornton <art27@cantab.net>
Co-authored-by: silverwind <me@silverwind.io>
Co-authored-by: Lunny Xiao <xiaolunwen@gmail.com>
This commit adds the possibility to click the labels in the message
"User added/removed the X label", which will lead to the issues page,
filtered by the clicked label.
This pull requests aims to fix#19768
_I've tried to follow the contribution guidelines as carefully as
possible, but in case I made a mistake: Please correct me._
<!--
Please check the following:
1. Make sure you are targeting the `main` branch, pull requests on
release branches are only allowed for bug fixes.
2. Read contributing guidelines:
https://github.com/go-gitea/gitea/blob/main/CONTRIBUTING.md
3. Describe what your pull request does and which issue you're targeting
(if any)
-->
Co-authored-by: wxiaoguang <wxiaoguang@gmail.com>
Co-authored-by: 6543 <6543@obermui.de>
Co-authored-by: techknowlogick <techknowlogick@gitea.io>
Remove this small, but unnecessary
[module](https://fomantic-ui.com/elements/image.html) and use `img`
selector over previous `.image`. Did a few tests, could not notice any
visual regression.
Co-authored-by: 6543 <6543@obermui.de>
Co-authored-by: Lauris BH <lauris@nix.lv>
The `go-licenses` make task introduced in #21034 is being run on make vendor
and occasionally causes an empty go-licenses file if the vendors need to
change. This should be moved to the generate task as it is a generated file.
Now because of this change we also need to split generation into two separate
steps:
1. `generate-backend`
2. `generate-frontend`
In the future it would probably be useful to make `generate-swagger` part of `generate-frontend` but it's not tolerated with our .drone.yml
Ref #21034
Signed-off-by: Andrew Thornton <art27@cantab.net>
Signed-off-by: Andrew Thornton <art27@cantab.net>
Co-authored-by: delvh <dev.lh@web.de>
Add support for triggering webhook notifications on wiki changes.
This PR contains frontend and backend for webhook notifications on wiki actions (create a new page, rename a page, edit a page and delete a page). The frontend got a new checkbox under the Custom Event -> Repository Events section. There is only one checkbox for create/edit/rename/delete actions, because it makes no sense to separate it and others like releases or packages follow the same schema.
The actions itself are separated, so that different notifications will be executed (with the "action" field). All the webhook receivers implement the new interface method (Wiki) and the corresponding tests.
When implementing this, I encounter a little bug on editing a wiki page. Creating and editing a wiki page is technically the same action and will be handled by the ```updateWikiPage``` function. But the function need to know if it is a new wiki page or just a change. This distinction is done by the ```action``` parameter, but this will not be sent by the frontend (on form submit). This PR will fix this by adding the ```action``` parameter with the values ```_new``` or ```_edit```, which will be used by the ```updateWikiPage``` function.
I've done integration tests with matrix and gitea (http).
Fix#16457
Signed-off-by: Aaron Fischer <mail@aaron-fischer.net>
A testing cleanup.
This pull request replaces `os.MkdirTemp` with `t.TempDir`. We can use the `T.TempDir` function from the `testing` package to create temporary directory. The directory created by `T.TempDir` is automatically removed when the test and all its subtests complete.
This saves us at least 2 lines (error check, and cleanup) on every instance, or in some cases adds cleanup that we forgot.
Reference: https://pkg.go.dev/testing#T.TempDir
```go
func TestFoo(t *testing.T) {
// before
tmpDir, err := os.MkdirTemp("", "")
require.NoError(t, err)
defer os.RemoveAll(tmpDir)
// now
tmpDir := t.TempDir()
}
```
Signed-off-by: Eng Zer Jun <engzerjun@gmail.com>
When setting.Git.DisablePartialClone is set to false then the web server will add filter support to web http. It does this by using`-c` command arguments but this will not work on gitea serv as the upload-pack and receive-pack commands do not support this.
Instead we move these options into the .gitconfig instead.
Fix#20400
Signed-off-by: Andrew Thornton <art27@cantab.net>
When migrating add several more important sanity checks:
* SHAs must be SHAs
* Refs must be valid Refs
* URLs must be reasonable
Signed-off-by: Andrew Thornton <art27@cantab.net>
Signed-off-by: Andrew Thornton <art27@cantab.net>
Co-authored-by: techknowlogick <matti@mdranta.net>
* feat: extend issue template for yaml
* feat: support yaml template
* feat: render form to markdown
* feat: support yaml template for pr
* chore: rename to Fields
* feat: template unmarshal
* feat: split template
* feat: render to markdown
* feat: use full name as template file name
* chore: remove useless file
* feat: use dropdown of fomantic ui
* feat: update input style
* docs: more comments
* fix: render text without render
* chore: fix lint error
* fix: support use description as about in markdown
* fix: add field class in form
* chore: generate swagger
* feat: validate template
* feat: support is_nummber and regex
* test: fix broken unit tests
* fix: ignore empty body of md template
* fix: make multiple easymde editors work in one page
* feat: better UI
* fix: js error in pr form
* chore: generate swagger
* feat: support regex validation
* chore: generate swagger
* fix: refresh each markdown editor
* chore: give up required validation
* fix: correct issue template candidates
* fix: correct checkboxes style
* chore: ignore .hugo_build.lock in docs
* docs: separate out a new doc for merge templates
* docs: introduce syntax of yaml template
* feat: show a alert for invalid templates
* test: add case for a valid template
* fix: correct attributes of required checkbox
* fix: add class not-under-easymde for dropzone
* fix: use more back-quotes
* chore: remove translation in zh-CN
* fix EasyMDE statusbar margin
* fix: remove repeated blocks
* fix: reuse regex for quotes
Co-authored-by: wxiaoguang <wxiaoguang@gmail.com>
The recovery, API, Web and package frameworks all create their own HTML
Renderers. This increases the memory requirements of Gitea
unnecessarily with duplicate templates being kept in memory.
Further the reloading framework in dev mode for these involves locking
and recompiling all of the templates on each load. This will potentially
hide concurrency issues and it is inefficient.
This PR stores the templates renderer in the context and stores this
context in the NormalRoutes, it then creates a fsnotify.Watcher
framework to watch files.
The watching framework is then extended to the mailer templates which
were previously not being reloaded in dev.
Then the locales are simplified to a similar structure.
Fix#20210Fix#20211Fix#20217
Signed-off-by: Andrew Thornton <art27@cantab.net>
Adds a new option to only show relevant repo's on the explore page, for bigger Gitea instances like Codeberg this is a nice option to enable to make the explore page more populated with unique and "high" quality repo's. A note is shown that the results are filtered and have the possibility to see the unfiltered results.
Co-authored-by: vednoc <vednoc@protonmail.com>
Co-authored-by: delvh <dev.lh@web.de>
Co-authored-by: 6543 <6543@obermui.de>
* Set no-tags in git fetch on compare
In the compare endpoint the git fetch is restricted to a certain branch however,
this does not completely prevent tag acquisition/pollution as git fetch will collect
any tags on that branch.
This causes pollution of the tag namespace and could cause confusion by users.
This PR adds `--no-tags` to the `git fetch` call.
Signed-off-by: Andrew Thornton <art27@cantab.net>
* Update modules/git/repo_compare.go
* Update modules/git/repo_compare.go
Signed-off-by: Andrew Thornton <art27@cantab.net>
This should solve the main problem of dynamic assets getting stale after
a version upgrade. Everything not affected will use query-string based
cache busting, which includes files loaded via HTML or worker scripts.
This PR adds functionality to allow Gitea to sit behind an
HAProxy and HAProxy protocolled connections directly.
Fix#7508
Signed-off-by: Andrew Thornton <art27@cantab.net>
The graceful manager waits for 4 listeners to be created or to be told that they are not needed. If it is not told about them it will indefinitely and timeout.
This leads to SVC hosts not being told of being in the readyState but on Unix would lead to the termination of the process.
There was an unfortunate regression in #20299 which missed this subtly and in the case whereby SSH is disabled the `builtinUnused()` is not called.
This PR adds a call to `builtinUnused()` when not using the builtin ssh to allow `createServerWaitGroup.Done()` to be called.
In addition it was noted that the if/else clauses for timeout informing of the SVC host were in the wrong order. These have been swapped.
Fix#20609
* Disable doctor logging on panic
If permissions are incorrect for writing to the doctor log simply disable the log file
instead of panicing.
Related #20570
Signed-off-by: Andrew Thornton <art27@cantab.net>
* Update cmd/doctor.go
* Update cmd/doctor.go
Co-authored-by: delvh <dev.lh@web.de>
Signed-off-by: Andrew Thornton <art27@cantab.net>
Co-authored-by: delvh <dev.lh@web.de>
Co-authored-by: techknowlogick <techknowlogick@gitea.io>
This adds support for getting the user's full name from the reverse
proxy in addition to username and email.
Tested locally with caddy serving as reverse proxy with Tailscale
authentication.
Signed-off-by: Will Norris <will@tailscale.com>
Signed-off-by: Will Norris <will@tailscale.com>
Co-authored-by: Lunny Xiao <xiaolunwen@gmail.com>
