Summary: See PHI425. See T12678. This should be an integer, but may be a string.
Test Plan: Called `differential.revision.edit`, observed integer in result instead of string.
Differential Revision: https://secure.phabricator.com/D19186
Summary: Depends on D19184. Ref T11015. Now that we have a digest index column, we can improve some of the queries a bit.
Test Plan:
- Ran queries from revision pages before and after with and without EXPLAIN.
- Saw the same results with much better EXPLAIN plans.
- Fragment size is now fixed at 12 bytes per fragment, so we can shove more of them in a single query.
Maniphest Tasks: T11015
Differential Revision: https://secure.phabricator.com/D19185
Summary:
Depends on D19183. Ref T11015. Currently, adding a trailing slash works great and omitting it mysteriously doesn't work.
Store a normalized version with an unconditional trailing slash for the lookup logic to operate on, and a separate display version which tracks what the user actually typed.
Test Plan:
- Entered "/src/main.c", "/src/main.c/", saw them de-duplicate.
- Entered "/src/main.c", saw it stay that way in the UI but appear as "/src/main.c/" internally.
- Added a rule for "/src/applications/owners" (no slash), created a revision touching paths in that directory, saw Owners fire for it.
- Changed the display value of a path only ("/src/main.c" to "/src/main.c/"), saw the update reflected in the UI without any beahvioral change.
Maniphest Tasks: T11015
Differential Revision: https://secure.phabricator.com/D19184
Summary:
Depends on D19182. Ref T11015. This changes `path` from `text255` to `longtext` because paths may be arbitrarily long.
It adds `pathDisplay` to prepare for display paths and storage paths having different values. For now, `pathDisplay` is copied from `path` and always has the same value.
Test Plan:
- Ran migration, checked database for sanity (all `pathDisplay` and `path` values identical).
- Added new paths, saw `pathDisplay` and `path` get the same values.
- Added an unreasonably enormous path with far more than 255 characters.
Maniphest Tasks: T11015
Differential Revision: https://secure.phabricator.com/D19183
Summary:
Depends on D19181. Ref T11015. This nukes duplicates from the table if they exist, then adds a unique key.
(Duplicates should not exist and can not be added with any recent version of the web UI.)
Test Plan:
- Tried to add duplicates with web UI, didn't have any luck.
- Explicitly added duplicates with manual `INSERT`s.
- Viewed packages in web UI and saw duplicates.
- Ran migrations, got a clean purge and a nice unique key.
- There's still no way to actually hit a duplicate key error in the UI (unless you can collide hashes, I suppose), this is purely a correctness/robustness change.
Maniphest Tasks: T11015
Differential Revision: https://secure.phabricator.com/D19182
Summary: Ref T11015. This supports making path names arbitrarily long and putting a proper unique key on the table.
Test Plan:
- Migrated, checked database, saw nice digested indexes.
- Edited a package, saw new rows update with digested indexes.
Maniphest Tasks: T11015
Differential Revision: https://secure.phabricator.com/D19181
Summary: Ref T13099. See PHI424. Fixes T11664. Several installs are interested in having these behaviors available in Owners by default and they aren't difficult to provide, it just makes the UI kind of messy. But I think there's enough general interest to justify it, now.
Test Plan: Created a package which owns "/" with a "With Non-Owner Author" review rule which I own. Created a revision, no package reviewer. Changed rule to "All", updated revision, got package reviewer.
Maniphest Tasks: T13099, T11664
Differential Revision: https://secure.phabricator.com/D19180
Summary:
Ref T13099. Ref T12787. See PHI417. Differential has new "irresponsible" warnings in the timeline somewhat recently, but these publish feed stories that don't link to the revision or have other relevant details, so they're confusing on the balance.
These have a high strength so they render on top, but we actually just want to hide them from the feed and let "abraham closed Dxyz by committing rXzzz." be the primary story.
Modularize things more so that we can get this behavior. Also, respect `shouldHideForFeed()` at display time, not just publishing time.
Test Plan: Used `bin/differential attach-commit` on a non-accepted revision to "irresponsibly land" a revision. Verified that feed story now shows "closed by commit" instead of "closed irresponsibly".
Maniphest Tasks: T13099, T12787
Differential Revision: https://secure.phabricator.com/D19179
Summary: Ref T13100. Since rules may begin failing for PRCE configuration reasons soon, provide a more complete explanation of possible causes in the UI.
Test Plan: Faked this, hit it via test console, saw explanation in web UI.
Maniphest Tasks: T13100
Differential Revision: https://secure.phabricator.com/D19178
Summary:
Ref T13099. See that task for discussion. Chrome is unhappy with an MFA form submitting to an endpoint which redirects you to an OAuth URI.
Instead, do the redirect entirely on the client.
Chrome's rationale here isn't obvious, so we may be able to revert this at some point.
Test Plan: Went through the OAuth flow locally, was redirected on the client. Will verify in production.
Maniphest Tasks: T13099
Differential Revision: https://secure.phabricator.com/D19177
Summary: Depends on D19175. Ref T13099. This fills in "close" and "update" transactions so that they show which commit(s) caused the action.
Test Plan: Used `transaction.search` to query some revisions, saw commit PHID information.
Maniphest Tasks: T13099
Differential Revision: https://secure.phabricator.com/D19176
Summary: Ref T13099. Move most of the "Update" logic to modular transactions
Test Plan: Created and updated revisions. Flushed the task queue. Grepped for `TYPE_UPDATE`. Reviewed update transactions in the timeline and feed.
Maniphest Tasks: T13099
Differential Revision: https://secure.phabricator.com/D19175
Summary: Depends on D19173. Ref T13096. Adds an optional, disabled-by-default lock log to make it easier to figure out what is acquiring and holding locks.
Test Plan: Ran `bin/lock log --enable`, `--disable`, `--name`, etc. Saw sensible-looking output with log enabled and daemons restarted. Saw no additional output with log disabled and daemons restarted.
Maniphest Tasks: T13096
Differential Revision: https://secure.phabricator.com/D19174
Summary:
Ref T13096. Currently, we do a fair amount of clever digesting and string manipulation to build lock names which are less than 64 characters long while still being reasonably readable.
Instead, do more of this automatically. This will let lock acquisition become simpler and make it more possible to build a useful lock log.
Test Plan: Ran `bin/repository update`, saw a reasonable lock acquire and release.
Maniphest Tasks: T13096
Differential Revision: https://secure.phabricator.com/D19173
Summary:
Ref T4340. See <https://discourse.phabricator-community.org/t/core-exception-during-installation/1193/8>.
If we return a response very early during setup, we may not be able to read from the environment yet. Just decline to build a "Content-Security-Policy" header in these cases.
Test Plan:
- Faked a preflight error (e.g., safe_mode enabled), restarted apache.
- Before patch: environment error while generating CSP.
- After patch: no error.
- Loaded a normal page, observed an normal CSP header.
Maniphest Tasks: T4340
Differential Revision: https://secure.phabricator.com/D19172
Summary:
See PHI416. If you raise a lint message in a deleted file, we don't render any text on the right hand side so the message never displays.
This is occasionally still legitimate/useful, e.g. to display a "don't delete this file" message. At least for now, show these messages on the left.
Test Plan: Posted a lint message on a deleted file via `harbormaster.sendmessage`, viewed revision, saw file expand with synthetic inline for lint.
Differential Revision: https://secure.phabricator.com/D19171
Summary:
Ref T4340. If you don't configure a CDN and visit a custom site (like a Phame blog site, or a CORGI sandbox internally) we serve resources from the main site. This violates the Content-Security-Policy.
When there's no CDN, include the primary domain in the CSP explicitly.
Test Plan: Loaded `local.www.phacility.com`, got resources.
Maniphest Tasks: T4340
Differential Revision: https://secure.phabricator.com/D19170
Summary: See PHI193. Previously, see similar D18763. Skip this legacy-style policy check when creating a project, since we know you can add members, even if the policy doesn't actually resolve in your favor.
Test Plan:
- Created a project with edit policy "Members of project" and myself, plus any other user (so the code goes down this path, not the "join/leave" path) as members.
Differential Revision: https://secure.phabricator.com/D19169
Summary:
See PHI413. You can pre-generate these with `bin/people profileimage --all`, but they're needlessly expensive to generate.
Streamline the workflow and cache some of the cacheable parts to reduce the generation cost.
Test Plan:
- Ran `bin/people profileimage --all` and saw cost drop from {nav 15.801s > 4.839s}.
- Set `defaultProfileImagePHID` to `NULL` in `phabricator_user.user` and purged caches with `bin/cache purge --all`.
- Loaded user directory.
- Saw default images regenerate relatively quickly.
Differential Revision: https://secure.phabricator.com/D19168
Summary:
Depends on D19166. Ref T13088. When the user scrolls away from a followed log, break the focus lock.
Let users stop following a live log.
Show when lines are added more clearly.
Don't refresh quite as quickly give users a better shot at clicking the stop button.
These behaviors can probably be refined but are at least more plausible and less actively user-hostile than the first version of this behavior was.
Test Plan: Used `write-log --rate` to write a large log slowly. Clicked "Follow Log", followed for a bit. Scrolled away, still got live updates but no more scroll lock. Clicked stop, no more updates.
Subscribers: PHID-OPKG-gm6ozazyms6q6i22gyam
Maniphest Tasks: T13088
Differential Revision: https://secure.phabricator.com/D19167
Summary:
Depends on D19165. Ref T13088. Currently, in other applications, we use Zero Width Spaces and Javascript "copy" listeners to prevent line numbers from being copied. This isn't terribly elegant.
Modern browsers support a second approach: using psuedo-elements with `content`. Try this in Harbormaster since it's conceptually cleaner, at least. One immediate drawback is that Command-F can't find this text either.
Test Plan: In Safari, Chrome and Firefox, highlighted ranges of lines and copy/pasted text. Got just text (no line numbers) in all cases.
Subscribers: PHID-OPKG-gm6ozazyms6q6i22gyam
Maniphest Tasks: T13088
Differential Revision: https://secure.phabricator.com/D19166
Summary: Depends on D19164. Ref T13088. Now that the JS behaviors are generic, use them on the Harbormaster standalone page.
Test Plan: Clicked lines and dragged across line ranges. Reloaded pages. Saw expected highlighting behavior in the client and on the server across reloads.
Subscribers: PHID-OPKG-gm6ozazyms6q6i22gyam
Maniphest Tasks: T13088
Differential Revision: https://secure.phabricator.com/D19165
Summary: Depends on D19163. Ref T13088. Increase the generality of this code so it can be shared with Harbormaster.
Test Plan: Clicked individual lines, clicked-and-dragged, etc., in Paste. Got sensible URI and highlight behaviors.
Maniphest Tasks: T13088
Differential Revision: https://secure.phabricator.com/D19164
Summary:
Depends on D19162. Ref T13088. When a user links to `$1234`, we need to render a default view of the log with a piece at the head, a piece at the end, and a piece in the middle.
We also need to figure out the offset for line 1234, or multiple offsets for "1234-2345".
Since the logic views/reads mostly anticipated this it isn't too much of a mess, although there are a couple of bugs this exposes with view specifications that use combinations of parameters which were previously impossible.
Test Plan: Viewed a large log with no line marker. Viewed `$1`. Viewed `$end`. Viewed `$35-40`, etc. Expanded context around logs.
Subscribers: PHID-OPKG-gm6ozazyms6q6i22gyam
Maniphest Tasks: T13088
Differential Revision: https://secure.phabricator.com/D19163
Summary: Ref T13088. This lifts the code for parsing "$x-y" line ranges in URIs into AphrontRequest so Diffusion, Paste, Harbormaster, etc., can share it.
Test Plan: Viewed lines, line ranges, no lines, negative line ranges, line ranges with 0, and extremely long line ranges in Paste.
Maniphest Tasks: T13088
Differential Revision: https://secure.phabricator.com/D19162
Summary:
Ref T4340. Some "Register/Login" and "Link External Account" buttons are forms which submit to third-party sites. Whitelist these targets when pages render an OAuth form.
Safari, at least, also prevents a redirect to a third-party domain after a form submission to the local domain, so when we first redirect locally (as with Twitter and other OAuth1 providers) we need to authorize an additional URI.
Test Plan: Clicked all my registration buttons locally without hitting CSP issues.
Maniphest Tasks: T4340
Differential Revision: https://secure.phabricator.com/D19159
Summary: Ref T4340. We don't use "<base />" so we can safely block it.
Test Plan: Injected "<base />" into a page, saw an error in the console showing that the browser had blocked it.
Maniphest Tasks: T4340
Differential Revision: https://secure.phabricator.com/D19158
Summary: Depends on D19156. Ref T13094. This replaces the remaining forms in the file embed view and lightbox with normal download links.
Test Plan: Clicked "Download" and lightbox -> download for embedded files.
Maniphest Tasks: T13094
Differential Revision: https://secure.phabricator.com/D19157
Summary:
Depends on D19155. Ref T13094. Ref T4340.
We can't currently implement a strict `form-action 'self'` content security policy because some file downloads rely on a `<form />` which sometimes POSTs to the CDN domain.
Broadly, stop generating these forms. We just redirect instead, and show an interstitial confirm dialog if no CDN domain is configured. This makes the UX for installs with no CDN domain a little worse and the UX for everyone else better.
Then, implement the stricter Content-Security-Policy.
This also removes extra confirm dialogs for downloading Harbormaster build logs and data exports.
Test Plan:
- Went through the plain data export, data export with bulk jobs, ssh key generation, calendar ICS download, Diffusion data, Paste data, Harbormaster log data, and normal file data download workflows with a CDN domain.
- Went through all those workflows again without a CDN domain.
- Grepped for affected symbols (`getCDNURI()`, `getDownloadURI()`).
- Added an evil form to a page, tried to submit it, was rejected.
- Went through the ReCaptcha and Stripe flows again to see if they're submitting any forms.
Subscribers: PHID-OPKG-gm6ozazyms6q6i22gyam
Maniphest Tasks: T13094, T4340
Differential Revision: https://secure.phabricator.com/D19156
Summary:
Depends on D19154. Ref T13094. This controller was removed at some point and this route no longer works.
I plan to add a new `download/` route to let us tighten the `form-action` Content Security Policy.
Test Plan: Grepped for the route and controller, no hits.
Maniphest Tasks: T13094
Differential Revision: https://secure.phabricator.com/D19155
Summary: See PHI399. Ref T4340. We don't require Flash/Java anywhere and can safely block them unconditionally in the Content-Security-Policy header.
Test Plan: Added a `<object ... />` tag to a page, saw "Blocked Plug-In" and a CSP warning in the browser console.
Maniphest Tasks: T4340
Differential Revision: https://secure.phabricator.com/D19154
Summary: Depends on D19152. Ref T13088. This adds live log tailing. It is probably not the final version of this feature because it prevents escape once you begin tailing a log.
Test Plan: Used `bin/harbormaster write-log --rate ...` to write a log slowly. Viewed it in the web UI. Clicked "Follow Log". Followed the log until the write finished, a lifetime later.
Subscribers: PHID-OPKG-gm6ozazyms6q6i22gyam
Maniphest Tasks: T13088
Differential Revision: https://secure.phabricator.com/D19153
Summary: Depends on D19151. Ref T13088. While dramatically less exciting than using `lolcat` and less general than `pv`, this should do the job adequately.
Test Plan: Piped a sizable log into `bin/harbormaster write-log` with `--rate 2048`, saw a progress bar. Loaded the log in the web UI and saw it grow as the page reloaded.
Reviewers: yelirekim
Reviewed By: yelirekim
Subscribers: yelirekim, PHID-OPKG-gm6ozazyms6q6i22gyam
Maniphest Tasks: T13088
Differential Revision: https://secure.phabricator.com/D19152
Summary:
Depends on D19150. Ref T13088. Allow clients to retrieve information about build logs, including log data, over the API.
(To fetch log data, take the `filePHID` to `file.search`, then issue a normal GET against the URI. Use a `Content-Range` header to get part of the log.)
Test Plan: Ran `harbormaster.log.search`, got sensible-looking results.
Subscribers: PHID-OPKG-gm6ozazyms6q6i22gyam
Maniphest Tasks: T13088
Differential Revision: https://secure.phabricator.com/D19151
Summary: Depends on D19149. Ref T13088. Since the new log requires a bunch of log reprocessing, the cutover is going to require at least some time for installs to run migrations. Add a link in the UI to ease the transition, smooth over some behaviors a little, and fix a fetch issue where we'd request past the end of the log (since this is now enforced).
Test Plan: Viewed a traditional Harbormaster build, saw links to the new standalone log pages.
Subscribers: PHID-OPKG-gm6ozazyms6q6i22gyam
Maniphest Tasks: T13088
Differential Revision: https://secure.phabricator.com/D19150
Summary: Depends on D19148. Ref T13088. The new rendering always executes range requests for data it needs, and we can satisfy these requests by loading the smallest number of chunks which span that range.
Test Plan: Piped 50,000 lines of Apache log into Harbormaster, viewed it in the new UI, got sensible rendering times and a reasonable amount of data actually going over the wire.
Subscribers: PHID-OPKG-gm6ozazyms6q6i22gyam
Maniphest Tasks: T13088
Differential Revision: https://secure.phabricator.com/D19149
Summary: Ref T13088. This variable bled through from an earlier loop and caused us to drop some of the lines in the middle.
Test Plan: Clicked "Show More", got an equal number of header and footer lines.
Subscribers: PHID-OPKG-gm6ozazyms6q6i22gyam
Maniphest Tasks: T13088
Differential Revision: https://secure.phabricator.com/D19148
Summary: Ref T13093. Depends on D19145. See PHI398. Previously, see D18933. This provides the current viewer to `ConduitCall` so that we don't try to use device credentials from unprivileged web hosts.
Test Plan: Evaluated the "Branches" field locally, saw an appropriate field value.
Maniphest Tasks: T13093
Differential Revision: https://secure.phabricator.com/D19146
Summary:
Ref T13093. See PHI396. These are possibly somewhat niche, but reasonable to support and consistent with the existing "Pusher's projects".
Also relabel "Pusher's projects" and "Project tags" for consistency and, hopefully, clarity.
Test Plan:
- Created new "commit" and "hook: commit content" Herald rules which run against "Author's projects" and "Committer's projects".
- Test console'd the "Commit" rules.
- Pushed through the "Hook" rule.
- In all cases, saw fields populate appropriately.
Maniphest Tasks: T13093
Differential Revision: https://secure.phabricator.com/D19145
Summary:
Ref T4340. Some browsers respect this header and referrers are a plague upon the earth.
Also, upgrade "never" to the more modern value "no-referrer".
Test Plan:
In Safari, Firefox and Chrome, disabled `rel="noreferrer"` on links and generated a normal link to an external site. Then clicked it and checked if a referrer was sent.
- Safari respects meta only, but "no-referrer" is fine.
- Firefox respects both (either the header or meta tag are individually sufficient to stop referrers).
- Chrome respects both (same as Firefox).
Maniphest Tasks: T4340
Differential Revision: https://secure.phabricator.com/D19144
Summary:
See PHI399. Ref T4340. This header provides an additional layer of protection against various attacks, including XSS attacks which embed inline `<script ...>` or `onhover="..."` content into the document.
**style-src**: The "unsafe-inline" directive affects both `style="..."` and `<style>`. We use a lot of `style="..."`, some very legitimately, so we can't realistically get away from this any time soon. We only use one `<style>` (for monospaced font preferences) but can't disable `<style>` without disabling `style="..."`.
**img-src**: We use "data:" URIs to inline small images into CSS, and there's a significant performance benefit from doing this. There doesn't seem to be a way to allow "data" URIs in CSS without allowing them in the document itself.
**script-src** and **frame-src**: For a small number of flows (Recaptcha, Stripe) we embed external javascript, some of which embeds child elements (or additional resources) into the document. We now whitelist these narrowly on the respective pages.
This won't work with Quicksand, so I've blacklisted it for now.
**connect-src**: We need to include `'self'` for AJAX to work, and any websocket URIs.
**Clickjacking**: We now have three layers of protection:
- X-Frame-Options: works in older browsers.
- `frame-ancestors 'none'`: does the same thing.
- Explicit framebust in JX.Stratcom after initialization: works in ancient IE.
We could probably drop the explicit framebust but it wasn't difficult to retain.
**script tags**: We previously used an inline `<script>` tag to start Javelin. I've moved this to `<data data-javelin-init ...>` tags, which seems to work properly.
**`__DEV__`**: We previously used an inline `<script>` tag to set the `__DEV__` mode flag. I tried using the "initialization" tags for this, but they fire too late. I moved it to `<html data-developer-mode="1">`, which seems OK everywhere.
**CSP Scope**: Only the CSP header on the original request appears to matter -- you can't refine the scope by emitting headers on CSS/JS. To reduce confusion, I disabled the headers on those response types. More headers could be disabled, although we're likely already deep in the land of diminishing returns.
**Initialization**: The initialization sequence has changed slightly. Previously, we waited for the <script> in bottom of the document to evaluate. Now, we go fishing for tags when domcontentready fires.
Test Plan:
- Browsed around in Firefox, Safari and Chrome looking for console warnings. Interacted with various Javascript behaviors. Enabled Quicksand.
- Disabled all the framebusting, launched a clickjacking attack, verified that each layer of protection is individually effective.
- Verified that the XHProf iframe in Darkconsole and the PHPAST frame layout work properly.
- Enabled notifications, verified no complaints about connecting to Aphlict.
- Hit `__DEV__` mode warnings based on the new data attribute.
- Tried to do sketchy stuff with `data:` URIs and SVGs. This works but doesn't seem to be able to do anything dangerous.
- Went through the Stripe and Recaptcha workflows.
- Dumped and examined the CSP headers with `curl`, etc.
- Added a raw <script> tag to a page (as though I'd found an XSS attack), verified it was no longer executed.
Maniphest Tasks: T4340
Differential Revision: https://secure.phabricator.com/D19143
Summary:
Depends on D19141. Ref T13088. Some of the fundamental log behaviors like "loading the correct rows" are now a bit better behaved.
The UI is a little less garbage, too.
Test Plan: Viewed some logs and loaded more context by clicking the buttons.
Subscribers: PHID-OPKG-gm6ozazyms6q6i22gyam
Maniphest Tasks: T13088
Differential Revision: https://secure.phabricator.com/D19142
Summary: Depends on D19139. Ref T13088. This doesn't actually work, but is close enough that a skilled attacker might be able to briefly deceive a small child.
Test Plan:
- Viewed some very small logs under very controlled conditions, saw content.
- Larger logs vaguely do something resembling working correctly.
Subscribers: PHID-OPKG-gm6ozazyms6q6i22gyam
Maniphest Tasks: T13088
Differential Revision: https://secure.phabricator.com/D19141
Summary:
Depends on D19138. Ref T13088. When we want to read the last part of a logfile //and show accurate line numbers//, we need to be able to get from byte offsets to line numbers somehow.
Our fundamental unit must remain byte offsets, because a test can emit an arbitrarily long line, and we should accommodate it cleanly if a test emits 2GB of the letter "A".
To support going from byte offsets to line numbers, compute a map with periodic line markers throughout the offsets of the file. From here, we can figure out the line numbers for arbitrary positions in the file with only a constant amount of work.
Test Plan: Added unit tests; ran unit tests.
Subscribers: PHID-OPKG-gm6ozazyms6q6i22gyam
Maniphest Tasks: T13088
Differential Revision: https://secure.phabricator.com/D19139
Summary: Depends on D19137. Ref T13088. This allows `rebuild-log` to skip work if the chunks are already compressed. It also prepares for a future GC which is looking for "text" or "gzip" chunks to throw away in favor of archival into Files; such a GC can use this column to find collectable logs and then write "file" to it, meaning "chunks are gone, this data is only available in Files".
Test Plan: Ran migration, saw logs populate as "text". Ran `rebuild-log`, saw logs rebuild as "gzip".
Subscribers: PHID-OPKG-gm6ozazyms6q6i22gyam
Maniphest Tasks: T13088
Differential Revision: https://secure.phabricator.com/D19138
Summary: Depends on D19136. Ref T13088. Since it's probably impractical to do all the migrations these changes imply during `bin/storage upgrade`, provide some support for performing them online.
Test Plan: Ran `bin/harbormaster rebuild-log` with `--all`, `--id`, and with and without `--force`.
Subscribers: PHID-OPKG-gm6ozazyms6q6i22gyam
Maniphest Tasks: T13088
Differential Revision: https://secure.phabricator.com/D19137
Summary: Depends on D19135. Ref T13088. Denormalize the total log size onto the log itself. This makes reasoning about the log at display time easier, and we don't need to fish around in the database as much to figure out what we're dealing with.
Test Plan: Ran `bin/harbormaster rebuild-log`, saw an existing log populate. Ran `bin/harbormaster write-log`, saw new log write with proper length information.
Subscribers: PHID-OPKG-gm6ozazyms6q6i22gyam
Maniphest Tasks: T13088
Differential Revision: https://secure.phabricator.com/D19136
