Summary: Depends on D19796. Simplify some timing code by using phutil_microseconds_since() instead of duplicate casting and arithmetic.
Test Plan: Grepped for `1000000` to find these. Pulled, pushed, made a conduit call. This isn't exhaustive but it should be hard for these to break in a bad way since they're all just diagnostic.
Reviewers: amckinley
Reviewed By: amckinley
Differential Revision: https://secure.phabricator.com/D19797
Summary:
Depends on D18987. See PHI343. Fixes T13060. See also T7339.
When the main process starts up with `LANG=POSIX` (this is the default on Ubuntu) and we later try to run a subprocess with a UTF8 character in the argument list (like `git cat-file blob 🐑.txt`), the argument is not passed to the subprocess correctly.
We already set `LANG=en_US.UTF-8` in the //subprocess// environment, but this only controls behavior for the subprocess itself. It appears that the argument list encoding before the actual subprocess starts depends on the parent process's locale setting, which makes some degree of sense.
Setting `putenv('LANG=en_US.UTF-8')` has no effect on this, but my guess is that the parent process's locale setting is read at startup (rather than read anew from `LANG` every time) and not changed by further modifications of `LANG`.
Using `setlocale(...)` does appear to fix this.
Ideally, installs would probably set some UTF-8-compatible LANG setting as the default. However, this makes setup harder and I couldn't figure out how to do it on our production Ubuntu AMI after spending a reasonable amount of time at it (see T13060).
Since it's very rare that this setting matters, try to just do the right thing. This may fail if "en_US.UTF-8" isn't available, but I think warnings/remedies to this are in the scope of T7339, since we want this locale to exist for other legitimate reasons anyway.
Test Plan:
- Applied this fix in production, processed the failing worker task from PHI343 after kicking Apache hard enough.
- Ran locally with `setlocale(LC_ALL, 'duck.quack')` to make sure a bad/invalid/unavailable setting didn't break anything, didn't hit any issues.
Reviewers: amckinley
Reviewed By: amckinley
Maniphest Tasks: T13060
Differential Revision: https://secure.phabricator.com/D18988
Summary:
Ref T13008. We haven't hit any issues with this, but I can imagine we might in the future.
When one host makes an intracluster request to another host, the `$viewer` ends up as the omnipotent viewer. This viewer isn't logged in, so they'll currently accumulate rate limit points at a high rate.
Instead, don't give them any points. These requests are always legitimate, and if they originated from a user request, that request should be the one getting rate limited.
Test Plan: Browsed around.
Reviewers: amckinley
Reviewed By: amckinley
Maniphest Tasks: T13008
Differential Revision: https://secure.phabricator.com/D18708
Summary:
Depends on D18702. Ref T13008. This replaces the old hard-coded single rate limit with multiple flexible limits, and defines two types of limits:
- Rate: reject requests if a client has completed too many requests recently.
- Connection: reject requests if a client has too many more connections than disconnections recently.
The connection limit adds +1 to the score for each connection, then adds -1 for each disconnection. So the overall number is how many open connections they have, at least approximately.
Supporting multiple limits will let us do limiting by Hostname and by remote address (e.g., a specific IP can't exceed a low limit, and all requests to a hostname can't exceed a higher limit).
Configuring the new limits looks something like this:
```
PhabricatorStartup::addRateLimit(new PhabricatorClientRateLimit())
->setLimitKey('rate')
->setClientKey($_SERVER['REMOTE_ADDR'])
->setLimit(5);
PhabricatorStartup::addRateLimit(new PhabricatorClientConnectionLimit())
->setLimitKey('conn')
->setClientKey($_SERVER['REMOTE_ADDR'])
->setLimit(2);
```
Test Plan:
- Configured limits as above.
- Made a lot of requests, got cut off by the rate limit.
- Used `curl --limit-rate -F 'data=@the_letter_m.txt' ...` to upload files really slowly. Got cut off by the connection limit. With `enable_post_data_reading` off, this correctly killed the connections //before// the uploads finished.
- I'll send this stuff to `secure` before production to give it more of a chance.
Reviewers: amckinley
Reviewed By: amckinley
Maniphest Tasks: T13008
Differential Revision: https://secure.phabricator.com/D18703
