mod_h[ttp]2

HTTP/2 for Apache httpd

Bucket Beams, Baby!

Copyright (C) 2016 greenbytes GmbH

So, Apache 2.4.20 is out and I had a little time in between to work on the data transport inside mod_http2. The results are Bucket Beams! And they are awesome, if I may say so myself.

(Sorry, this is a post about the very internals of Apache httpd and I can understand that it might not be of interest to a user of said server. If you are not into programming servers in C, you are excused. Do something else, smell the roses!)

Copy Copy the Problem

Apache has a very nice runtime, names APR, that provides good abstractions and utilities. One feature, very central to data transport in Apache, are the apr_buckets and corresponding apr_bucket_brigades.

A bucket is just a chunk of data which one can, eventually, read. It can also be of 0 length and contain meta information like "flush everything before" or "the data stream ends here". Buckets can hold allocated data on the heap, from memory pools, static data, file handles and even sockets. They can split themselves when needed and also morph from one form to another.

For example: when you read a file bucket, it might happen that the bucket reads 8000 bytes from the file, splits itself into two (one 8000 bytes bucket and one for the rest) and then transforms itself (the first) into a heap bucket. The second one is a new file bucket with a different offset and length and when you read that, the same might happen again.

And a row of buckets can be kept in a apr_bucket_brigade. Those can tell you the combined length of all their buckets, you can write to them in various ways and they try to find effecient ways to manage the buckets or allocate new ones. Lifetimes are handled using APR memory pools, everything cleans up after itself, re-uses mempory in free lists etc. etc. Very nice.

But.

It does not work across threads.

I tried.

There is really nothing wrong with buckets and brigades and their allocators. It is just that HTTP/1.1 does not need to work across threads. Because a server can process a HTTP/1.1 connection using a single thread only. It might use a different thread in the end as the one it started with, but there is only ever one single thread using it at a time.

But HTTP/2 processing, at least in mod_http2, needs at least two threads. And the response from the thread processing a request needs to be transported to the thread managing the connection. And "the response" often means a chain of buckets. Sometimes it's a small number, but it can be a lot. Which requires the connection thread to work on the first buckets as soon as they arrive from the request thread. While that one produces new buckets.

So, a bucket can be handed over from one thread to another without problem. But when you read the data inside, it might allocate new memory or new buckets. Which invoke the memory pools and allocators it was created with...which are also being used by the other thread...which - segmentation fault.

And even if you read the data before passing the bucket, discarding the bucket once it was sent will again invoke the allocator which will manage its free lists...which is also invoked from the other thread... which...boom.

So, when passing a bucket acrosse threads, the receiving thread may neither read nor discard this bucket. Nor do anything else that changes it, like splitting.

So, before version 1.5.0 mod_http2 was copying bucket data. Reading it in the sending thread and writing it into new buckets allocated by the receiving thread. It had some special handling for file buckets so that it did not need to read them when passing the data. But everything else involved copying.

Beam Me Up

The idea for the bucket beams is inspired (or so I prefer to think) from the good, old transporter and star gate technology. It's a tested and proven method to keep things you beam across the galaxy in your local memory buffers and erase them only when the transport was successful. Numerous protagonists own their very screen life to this simple fact.

For ease of explanation, think of bucket beams as having a 'red' sending side where the 'red' thread puts its 'red' buckets into the beam. And a 'green' receiving side where the 'green' thread gets 'green' buckets out of the beam. The methods called by the red thread may operate on red buckets only and vice versa for the green thread.

How to convert red buckets into green ones then?

When sending, so in a red call, red buckets are read. This gives a pointer to the data and the number of data bytes. This data will continue to exist as long as the red bucket is not changed or destroyed. The red bucket is then placed into a to-be-transferred list inside the beam.

When the green thread calls receive() the to-be-transferred list is inspected and a corresponding green bucket is created. There might be no corresponding green bucket available, for example unknown metadata buckets cannot be duplicated. The red bucket is then put into the hold bucket list. For red data buckets, a special beam bucket is created. This refers to the red's data buffer as determined during the sending as well as the red bucket itself. The beam bucket is handed over to the green receiving side.

The green thread then works on the received buckets just as with any other, eventually reading the data and sending it out on the connection after which the bucket gets destroyed.

Destroying a beam bucket will notify the beam it came from: its red bucket is no longer needed. The beam bucket makes sure that, even though it might get itself split into a number of new buckets, it will only call in to the beam when the last one gets destroyed.

The beam will then transfer the red bucket from the hold list to the purge list. Since the notify call comes from the green thread, it cannot destroy the red bucket immediately.

And last, whenver the red thread calls in, the purge list is cleared.

This uses less memory than the copying done before. The memory in the hold was already allocated and is also needed as long as the former copy would have been there. So, only the data in the purge area occupies unwanted space. But that only as long as the red side does not call.

That is basically the bucket beam implementation. There are more details of course on how life times of red and green pools need to be treated and how locking is done. How it allows to limit the amount of data buffered and can block using condition variables, etc. But if you want to know that kind of detail, it's probably best to dive right into the source.

Enjoy Apache. And the sun outside.

M√ľnster, 19.04.2016,

Stefan Eissing, greenbytes GmbH

Copying and distribution of this file, with or without modification, are permitted in any medium without royalty provided the copyright notice and this notice are preserved. This file is offered as-is, without warranty of any kind. See LICENSE for details.