Blame


1 b9220ca4 2021-01-11 op # gmid
2 3e4749f7 2020-10-02 op
3 601bc1cc 2021-01-30 op gmid is a Gemini server written with security in mind. I initially
4 601bc1cc 2021-01-30 op wrote it to serve static files, but it has grown into a featureful
5 601bc1cc 2021-01-30 op server that can be used from either the command line to serve local
6 601bc1cc 2021-01-30 op directories
7 3e4749f7 2020-10-02 op
8 601bc1cc 2021-01-30 op gmid docs # serve the directory docs over gemini
9 3e4749f7 2020-10-02 op
10 601bc1cc 2021-01-30 op or as a traditional daemon
11 3e4749f7 2020-10-02 op
12 2f09adb0 2021-01-25 op gmid -c /etc/gmid.conf
13 2f09adb0 2021-01-25 op
14 2f09adb0 2021-01-25 op
15 b9220ca4 2021-01-11 op ## Features
16 3e4749f7 2020-10-02 op
17 601bc1cc 2021-01-30 op (random order)
18 601bc1cc 2021-01-30 op
19 4d2ec6d7 2021-01-13 op - IRI support (RFC3987)
20 286c4f40 2021-01-27 op - punycode support
21 601bc1cc 2021-01-30 op - dual stack (IPv4 and IPv6)
22 601bc1cc 2021-01-30 op - automatic certificate generation for config-less mode
23 b9220ca4 2021-01-11 op - CGI scripts
24 b9220ca4 2021-01-11 op - (very) low memory footprint
25 b9220ca4 2021-01-11 op - small codebase, easily hackable
26 b9220ca4 2021-01-11 op - virtual hosts
27 286c4f40 2021-01-27 op - per-location rules
28 286c4f40 2021-01-27 op - optional directory listings
29 286c4f40 2021-01-27 op - configurable mime types
30 ea58dab1 2021-01-17 op - sandboxed by default on OpenBSD, Linux and FreeBSD
31 0b00962d 2021-01-25 op - chroot support
32 3e4749f7 2020-10-02 op
33 fab952e1 2020-10-03 op
34 b9220ca4 2021-01-11 op ## Drawbacks
35 3e4749f7 2020-10-02 op
36 b9220ca4 2021-01-11 op - not suited for very busy hosts. If you receive an high number of
37 b9220ca4 2021-01-11 op connection per-second you'd probably want to run multiple gmid
38 b9220ca4 2021-01-11 op instances behind relayd/haproxy or a different server.
39 85dff1f9 2021-01-11 op
40 286c4f40 2021-01-27 op
41 286c4f40 2021-01-27 op ## Internationalisation (IRIs, UNICODE, punycode, all that stuff)
42 286c4f40 2021-01-27 op
43 286c4f40 2021-01-27 op Even thought the current Gemini specification doesn't mention anything
44 601bc1cc 2021-01-30 op in this regard, I do think these are important things and so I tried
45 601bc1cc 2021-01-30 op to implement them in the most user-friendly way I could think of.
46 286c4f40 2021-01-27 op
47 601bc1cc 2021-01-30 op For starters, gmid has full support for IRI (RFC3987 —
48 22a742e4 2021-01-29 op Internationalized Resource Identifiers). IRIs are a superset of URIs,
49 286c4f40 2021-01-27 op so there aren't incompatibilities with URI-only clients.
50 286c4f40 2021-01-27 op
51 601bc1cc 2021-01-30 op There is full support also for punycode. In theory, the user doesn't
52 286c4f40 2021-01-27 op even need to know that punycode is a thing. The hostname in the
53 601bc1cc 2021-01-30 op configuration file can (and must be) in the decoded form (e.g. `naïve`
54 601bc1cc 2021-01-30 op and not `xn--nave-6pa`), gmid will do the rest.
55 286c4f40 2021-01-27 op
56 601bc1cc 2021-01-30 op The only missing piece is UNICODE normalisation of the IRI path: gmid
57 601bc1cc 2021-01-30 op doesn't do that (yet).
58 286c4f40 2021-01-27 op
59 286c4f40 2021-01-27 op
60 b9220ca4 2021-01-11 op ## Building
61 3e4749f7 2020-10-02 op
62 42650ade 2021-01-27 op gmid depends on a POSIX libc, OpenSSL/LibreSSL and libtls (provided
63 42650ade 2021-01-27 op either by LibreSSL or libretls). At build time, flex and yacc (or GNU
64 42650ade 2021-01-27 op bison) are also needed.
65 3e4749f7 2020-10-02 op
66 771d8f28 2021-01-17 op The build is as simple as
67 3e4749f7 2020-10-02 op
68 601bc1cc 2021-01-30 op ./configure
69 b9220ca4 2021-01-11 op make
70 3e4749f7 2020-10-02 op
71 8f0da580 2021-01-21 op If the configure scripts fails to pick up something, please open an
72 f980545b 2021-01-21 op issue or notify me via email.
73 8f0da580 2021-01-21 op
74 771d8f28 2021-01-17 op To install execute:
75 881a9dd9 2021-01-16 op
76 771d8f28 2021-01-17 op make install
77 771d8f28 2021-01-17 op
78 601bc1cc 2021-01-30 op ### Docker
79 771d8f28 2021-01-17 op
80 601bc1cc 2021-01-30 op If you have trouble installing LibreSSL or libretls, you can use
81 601bc1cc 2021-01-30 op Docker to build a `gmid` image with:
82 601bc1cc 2021-01-30 op
83 17b09e3c 2021-01-18 op docker build -t gmid .
84 17b09e3c 2021-01-18 op
85 17b09e3c 2021-01-18 op and then run it with something along the lines of
86 17b09e3c 2021-01-18 op
87 17b09e3c 2021-01-18 op docker run --rm -it -p 1965:1965 \
88 6c117838 2021-01-25 op -v /path/to/gmid.conf:...:ro \
89 17b09e3c 2021-01-18 op -v /path/to/docs:/var/gemini \
90 6c117838 2021-01-25 op gmid -c .../gmid.conf
91 17b09e3c 2021-01-18 op
92 601bc1cc 2021-01-30 op (ellipses used for brevity)
93 17b09e3c 2021-01-18 op
94 dd8cc7d3 2021-01-22 op ### Local libretls
95 17b09e3c 2021-01-18 op
96 8f0da580 2021-01-21 op This is **NOT** recommended, please try to port LibreSSL/LibreTLS to
97 8f0da580 2021-01-21 op your distribution of choice or use docker instead.
98 8f0da580 2021-01-21 op
99 601bc1cc 2021-01-30 op However, it's possible to statically-link `gmid` to locally-installed
100 601bc1cc 2021-01-30 op libretls quite easily. (It's how I test gmid on Fedora, for instance)
101 8f0da580 2021-01-21 op
102 04397b32 2021-01-21 op Let's say you have compiled and installed libretls in `$LIBRETLS`,
103 8f0da580 2021-01-21 op then you can build `gmid` with
104 8f0da580 2021-01-21 op
105 04397b32 2021-01-21 op ./configure CFLAGS="-I$LIBRETLS/include" \
106 45b4aa6e 2021-01-23 op LDFLAGS="$LIBRETLS/lib/libtls.a -lssl -lcrypto -lpthread"
107 8f0da580 2021-01-21 op make
108 8f0da580 2021-01-21 op
109 dd8cc7d3 2021-01-22 op ### Testing
110 dd8cc7d3 2021-01-22 op
111 31a4993a 2021-01-23 op Execute
112 dd8cc7d3 2021-01-22 op
113 dd8cc7d3 2021-01-22 op make regress
114 dd8cc7d3 2021-01-22 op
115 601bc1cc 2021-01-30 op to start the suite. Keep in mind that the regression tests will
116 601bc1cc 2021-01-30 op create files inside the `regress` directory and bind the 10965 port.
117 dd8cc7d3 2021-01-22 op
118 dd8cc7d3 2021-01-22 op
119 881a9dd9 2021-01-16 op ## Architecture/Security considerations
120 881a9dd9 2021-01-16 op
121 881a9dd9 2021-01-16 op gmid is composed by two processes: a listener and an executor. The
122 881a9dd9 2021-01-16 op listener process is the only one that needs internet access and is
123 881a9dd9 2021-01-16 op sandboxed. When a CGI script needs to be executed, the executor
124 881a9dd9 2021-01-16 op (outside of the sandbox) sets up a pipe and gives one end to the
125 881a9dd9 2021-01-16 op listener, while the other is bound to the CGI script standard output.
126 909ea500 2021-01-27 op This way, is still possible to execute CGI scripts without
127 601bc1cc 2021-01-30 op restrictions even in the presence of a sandboxed network process.
128 881a9dd9 2021-01-16 op
129 601bc1cc 2021-01-30 op On OpenBSD, the listener runs with the `stdio recvfd rpath inet`
130 601bc1cc 2021-01-30 op pledges, while the executor has `stdio sendfd proc exec`; both have
131 601bc1cc 2021-01-30 op unveiled only the served directories.
132 881a9dd9 2021-01-16 op
133 ea58dab1 2021-01-17 op On FreeBSD, the executor process is sandboxed with `capsicum(4)`.
134 d939d0f0 2021-01-17 op
135 601bc1cc 2021-01-30 op On Linux, a `seccomp(2)` filter is installed in the listener to allow
136 601bc1cc 2021-01-30 op only certain syscalls, see [sandbox.c](sandbox.c) for more information
137 601bc1cc 2021-01-30 op on the BPF program.
138 ea58dab1 2021-01-17 op
139 ea58dab1 2021-01-17 op In any case, you are invited to run gmid inside some sort of
140 0b00962d 2021-01-25 op container/jail/chroot.