Emil Lerch
7e2a3b04fa
This was the hard part. Under abnormal termination (kill -9, seg faults, whatever) socket files do not get cleaned up properly. As long as this is the IPC mechanism, there is very little that can be done about this. So, with this change we are leaving them to linger. Fixing this we can do one of two things. First, we can look for EADDRINUSE and just delete the file and retry. However, that prevents someone from running two binaries at the same time (maybe to run two servers on different ports). The second option is to make the files unique, so this is the path chosen. Here we append a system timestamp to the file name. Part two of this eventually will need to be locating the TEMP directory and placing the files there. |
||
---|---|---|
.gitea/workflows | ||
docker | ||
src | ||
.gitignore | ||
build.zig | ||
FlexiLib.svg | ||
LICENSE | ||
proxy.ini | ||
README.md |
FlexiLib
This is a web server written with the following goals:
- Low memory consumption
- Low latency
- Flexible "reverse proxy" capabilities
- Ability to simulate FAAS capabilities of various online providers (AWS, CloudFlare, etc)
This last point is indirectly supported through the ability of the server to load, at run time, dynamic libraries to support requests. It will also reload these libraries after any in flight requests have completed, to support the experience of developing new libaries.
Libraries can be written in any programming language that supports a standard Linux C-Based calling convention, which is to say, nearly every programming language.
This project provides slightly better development and performance characteristics if the library used is written in zig. An example zig-based library can be found in src/main-lib.zig.
Deployment
Gitea actions are configured to build, sign, and deploy the source code. Each successful build will generate an artifact, which can be found at https://git.lerch.org/lobo/FlexiLib/actions.
Two artifacts will be available:
- The flexilib binary, compiled for linux x86_64 with GNU libc. GNU libc is necessary due to the dynamic loading involved, but otherwise it should be possible to use the binary as is on any glibc-based Linux distribution.
- A signature file generated from the HSM-based signing process. This can be verified for authenticity against sigstore public transparency log with rekor.
Additionally, a docker container image will be build and uploaded using the
tag git.lerch.org/lobo/flexilib:<shortsha>
. For example, docker pull git.lerch.org/lobo/flexilib:c02cd20
will get the docker container with flexilib from git commit c02cd20
.
Signature Validation
To verify the build artifacts, you will need the rekor CLI and four additional things:
- The signature file stored as a build artifact
- The flexilib executable, also from the build
- A downloaded version of the server public key. Theoretically rekor can take the URL at the command line, but this doesn't seem to work for me
- The sigstore entry URL from the Sign job in the Gitea action
Once those four things are assembled, the following command will verify the executable matches the output from the build run at the time of the run:
rekor verify --artifact flexilib --entry <entry url> --signature signature --pki-format x509 --public-key serverpublic.pem
As an example, using output from run 8:
rekor verify \
--artifact flexilib \
--entry https://rekor.sigstore.dev/api/v1/log/entries/73a64ca9cc712f9645bfe79ae104b101e3ef7022172f0bfc3aa34d4f45ca2af8 \
--signature signature \
--pki-format x509 \
--public-key serverpublic.pem
Architecture
This library assumes the use of Linux as a host. While the primary engine is not tied to Linux, the file watcher module uses inotify and friends and will not work outside that OS. PRs are welcome.
The system is zig version 0.11.
To achieve the lowest latency possible, this server loads dynamic libraries
using dlopen(3) based on a configuration
file in the current working directory called proxy.ini
. An example of the
configuration is in this directory, and it is relatively simple string prefix
matching, again, for speed.
On startup, a thread pool will be created. Request paths and header matching is loaded from the configuration file, and file watches are initiated on all libraries mentioned in the configuration file. Libraries are loaded on demand when a request arrives that needs the library. When a library changes for a new version, the file watcher will take note and unload the previous version.
Changes to the configuration file are not watched, relying instead on a HUP signal to force a reload. At that point, all libraries ("executors") are unloaded, and configuration is re-read.
As libraries are loaded directly into main process space, bugs in the libraries can and will crash the engine. As such, some supervisory process (dockerd, systemd, etc) should monitor and restart if necessary.
Security
There is little attempt to secure libraries from interfering with the current thread or even the main process. As such, the libraries should be fully trusted. However, libraries themselves may be hardened to run other non-trusted code. For example: A "I run WASM code" library may be written to create a WASM VM and run user-supplied WASM code. In that case, the "I run WASM code" library is trusted, although the code it runs may not be.
Configuration
Very little has been done so far in terms of configuration. By default, the
number of threads created to serve requests is equal to the number of CPUs
reported by the system (although thread count is limited to 4 threads when
compiled in debug mode). This can be controlled with the environment variable
SERVER_THREAD_COUNT
.
The port by default is 8069, although this can be set with the PORT
environment variable. Future plans include an environment variable for IP
address as well as the amount of pre-allocated memory for response data (currently
hardcoded to 8k/thread). Pre-allocated memory reduces the number of system
calls required for memory allocation, and pre-allocation/allocation statistics
per request are reported in the logs. The current pre-allocation provides
approximately 4k per request without requiring system calls.
Logs
Request logs are sent to standard out, and are likely to change. Here is a sample:
127.0.0.1:59940 - - "GET / HTTP/1.1" 200 ttfb 2000.420ms 11 ttlb 2000.568ms (pre-alloc: 1569, alloc: 4350)
The first part mirrors common logs from Apache/nginx.
ttfb: Time to first byte. This represents the number of ms of processing within the library ttlb: Time to last byte. This includes processing as well a transmission of data pre-alloc: The amount of memory actually pre-allocated (1k is just a minimum and the system may allocate more) alloc: The amount of memory actually allocated during the request