3
<title>syslog-protocol support in rsyslog</title>
6
<h1>syslog-protocol support in rsyslog</h1>
7
<p><b><a href="http://www.rsyslog.com/">Rsyslog</a> provides a trial
8
implementation of the proposed
9
<a href="http://www.monitorware.com/Common/en/glossary/syslog-protocol.php">
10
syslog-protocol</a> standard.</b> The intention of this implementation is to
11
find out what inside syslog-protocol is causing problems during implementation.
12
As syslog-protocol is a standard under development, its support in rsyslog is
13
highly volatile. It may change from release to release. So while it provides
14
some advantages in the real world, users are cautioned against using it right
15
now. If you do, be prepared that you will probably need to update all of your
16
rsyslogds with each new release. If you try it anyhow, please provide feedback
17
as that would be most beneficial for us.</p>
18
<h2>Currently supported message format</h2>
19
<p>Due to recent discussion on syslog-protocol, we do not follow any specific
20
revision of the draft but rather the candidate ideas. The format supported
22
<p><b><code><PRI>VERSION SP TIMESTAMP SP HOSTNAME SP APP-NAME SP PROCID SP MSGID SP [SD-ID]s
24
<p>Field syntax and semantics are as defined in IETF I-D syslog-protocol-15.</p>
25
<h2>Capabilities Implemented</h2>
27
<li>receiving message in the supported format (see above)</li>
28
<li>sending messages in the supported format</li>
29
<li>relaying messages</li>
30
<li>receiving messages in either legacy or -protocol format and transforming
31
them into the other one</li>
32
<li>virtual availability of TAG, PROCID, APP-NAME, MSGID, SD-ID no matter if
33
the message was received via legacy format, API or syslog-protocol format (non-present
34
fields are being emulated with great success)</li>
35
<li>maximum message size is set via preprocessor #define</li>
36
<li>syslog-protocol messages can be transmitted both over UDP and plain TCP
37
with some restrictions on compliance in the case of TCP</li>
40
<p>This lists what has been found during implementation:</p>
42
<li>The same receiver must be able to support both legacy and
43
syslog-protocol syslog messages. Anything else would be a big inconvenience
44
to users and would make deployment much harder. The detection must be done
45
automatically (see below on how easy that is).</li>
46
<li><b>NUL characters inside MSG</b> cause the message to be truncated at
47
that point. This is probably a major point for many C-based implementations.
48
No measures have yet been taken against this. Modifying the code to "cleanly"
49
support NUL characters is non-trivial, even though rsyslogd already has some
50
byte-counted string library (but this is new and not yet available
52
<li><b>character encoding in MSG</b>: is is problematic to do the right
53
UTF-8 encoding. The reason is that we pick up the MSG from the local domain
54
socket (which got it from the syslog(3) API). The text obtained does not
55
include any encoding information, but it does include non US-ASCII
56
characters. It may also include any other encoding. Other than by guessing
57
based on the provided text, I have no way to find out what it is. In order
58
to make the syslogd do anything useful, I have now simply taken the message
59
as is and stuffed it into the MSG part. Please note that I think this will
60
be a route that other implementors would take, too.</li>
61
<li>A minimal parser is easy to implement. It took me roughly 2 hours to add
62
it to rsyslogd. This includes the time for restructuring the code to be able
63
to parse both legacy syslog as well as syslog-protocol. The parser has some
64
restrictions, though<ul>
65
<li>STRUCTURED-DATA field is extracted, but not validated. Structured data
66
"[test ]]" is not caught as an error. Nor are any other errors caught. For
67
my needs with this syslogd, that level of structured data processing is
68
probably sufficient. I do not want to parse/validate it in all cases. This
69
is also a performance issue. I think other implementors could have the same
70
view. As such, we should not make validation a requirement.</li>
71
<li>MSG is not further processed (e.g. Unicode not being validated)</li>
72
<li>the other header fields are also extracted, but no validation is
73
performed right now. At least some validation should be easy to add (not
74
done this because it is a proof-of-concept and scheduled to change).</li>
77
<li>Universal access to all syslog fields (missing ones being emulated) was
78
also quite easy. It took me around another 2 hours to integrate emulation of
79
non-present fields into the code base.</li>
80
<li>The version at the start of the message makes it easy to detect if we
81
have legacy syslog or syslog-protocol. Do NOT move it to somewhere inside
82
the middle of the message, that would complicate things. It might not be
83
totally fail-safe to just rely on "1 " as the "cookie" for a syslog-protocol.
84
Eventually, it would be good to add some more uniqueness, e.g. "@#1 ".</li>
85
<li>I have no (easy) way to detect truncation if that happens on the UDP
86
stack. All I see is that I receive e.g. a 4K message. If the message was e.g.
87
6K, I received two chunks. The first chunk (4K) is correctly detected as a
88
syslog-protocol message, the second (2K) as legacy syslog. I do not see what
89
we could do against this. This questions the usefulness of the TRUNCATE bit.
90
Eventually, I could look at the UDP headers and see that it is a fragment. I
91
have looked at a network sniffer log of the conversation. This looks like
92
two totally-independent messages were sent by the sender stack.</li>
93
<li>The maximum message size is currently being configured via a
94
preprocessor #define. It can easily be set to 2K or 4K, but more than 4K is
95
not possible because of UDP stack limitations. Eventually, this can be
96
worked around, but I have not done this yet.</li>
97
<li>rsyslogd can accept syslog-protocol formatted messages but is able to
98
relay them in legacy format. I find this a must in real-life deployments.
99
For this, I needed to do some field mapping so that APP-NAME/PROCID are
100
mapped into a TAG.</li>
101
<li>rsyslogd can also accept legacy syslog message and relay them in
102
syslog-protocol format. For this, I needed to apply some sub-parsing of the
103
TAG, which on most occasions provides correct results. There might be some
104
misinterpretations but I consider these to be mostly non-intrusive. </li>
105
<li>Messages received from the syslog API (the normal case under *nix) also
106
do not have APP-NAME and PROCID and I must parse them out of TAG as
107
described directly above. As such, this algorithm is absolutely vital to
108
make things work on *nix.</li>
109
<li>I have an issue with messages received via the syslog(3) API (or, to be
110
more precise, via the local domain socket this API writes to): These
111
messages contain a timestamp, but that timestamp does neither have the year
112
nor the high-resolution time. The year is no real issue, I just take the
113
year of the reception of that message. There is a very small window of
114
exposure for messages read from the log immediately after midnight Jan 1st.
115
The message in the domain socket might have been written immediately before
116
midnight in the old year. I think this is acceptable. However, I can not
117
assign a high-precision timestamp, at least it is somewhat off if I take the
118
timestamp from message reception on the local socket. An alternative might
119
be to ignore the timestamp present and instead use that one when the message
120
is pulled from the local socket (I am talking about IPC, not the network -
121
just a reminder...). This is doable, but eventually not advisable. It looks
122
like this needs to be resolved via a configuration option.</li>
123
<li>rsyslogd already advertised its origin information on application
124
startup (in a syslog-protocol-14 compatible format). It is fairly easy to
125
include that with any message if desired (not currently done).</li>
126
<li>A big problem I noticed are malformed messages. In -syslog-protocol, we
127
recommend/require to discard malformed messages. However, in practice users
128
would like to see everything that the syslogd receives, even if it is in
129
error. For the first version, I have not included any error handling at all.
130
However, I think I would deliberately ignore any "discard" requirement. My
131
current point of view is that in my code I would eventually flag a message
132
as being invalid and allow the user to filter on this invalidness. So these
133
invalid messages could be redirected into special bins.</li>
134
<li>The error logging recommendations (those I insisted on;)) are not really
135
practical. My application has its own error logging philosophy and I will
136
not change this to follow a draft.</li>
137
<li>Relevance of support for leap seconds and senders without knowledge of
138
time is questionable. I have not made any specific provisions in the code
139
nor would I know how to handle that differently. I could, however, pull the
140
local reception timestamp in this case, so it might be useful to have this
141
feature. I do not think any more about this for the initial proof-of-concept.
142
Note it as a potential problem area, especially when logging to databases.</li>
143
<li>The HOSTNAME field for internally generated messages currently contains
144
the hostname part only, not the FQDN. This can be changed inside the code
145
base, but it requires some thinking so that thinks are kept compatible with
146
legacy syslog. I have not done this for the proof-of-concept, but I think it
147
is not really bad. Maybe an hour or half a day of thinking.</li>
148
<li>It is possible that I did not receive a TAG with legacy syslog or via
149
the syslog API. In this case, I can not generate the APP-NAME. For
150
consistency, I have used "-" in such cases (just like in PROCID, MSGID and
151
STRUCTURED-DATA).</li>
152
<li>As an architectural side-effect, syslog-protocol formatted messages can
153
also be transmitted over non-standard syslog/raw tcp. This implementation
154
uses the industry-standard LF termination of tcp syslog records. As such,
155
syslog-protocol messages containing a LF will be broken invalidly. There is
156
nothing that can be done against this without specifying a TCP transport.
157
This issue might be more important than one thinks on first thought. The
158
reason is the wide deployment of syslog/tcp via industry standard.</li>
160
<p><b>Some notes on syslog-transport-udp-06</b></p>
162
<li>I did not make any low-level modifications to the UDP code and think I
163
am still basically covered with this I-D.</li>
164
<li>I deliberately violate section 3.3 insofar as that I do not necessarily
165
accept messages destined to port 514. This feature is user-required and a
166
must. The same applies to the destination port. I am not sure if the "MUST"
167
in section 3.3 was meant that this MUST be an option, but not necessarily be
168
active. The wording should be clarified.</li>
169
<li>section 3.6: I do not check checksums. See the issue with discarding
170
messages above. The same solution will probably be applied in my code.</li>
173
<h2>Conlusions/Suggestions</h2>
174
<p>These are my personal conclusions and suggestions. Obviously, they must be
177
<li>NUL should be disallowed in MSG</li>
178
<li>As it is not possible to definitely know the character encoding of the
179
application-provided message, MSG should <b>not</b> be specified to use UTF-8
180
exclusively. Instead, it is suggested that any encoding may be used but
181
UTF-8 is preferred. To detect UTF-8, the MSG should start with the UTF-8
182
byte order mask of "EF BB BF" if it is UTF-8 encoded (see section 155.9 of
183
<a href="http://www.unicode.org/versions/Unicode4.0.0/ch15.pdf">
184
http://www.unicode.org/versions/Unicode4.0.0/ch15.pdf</a>) </li>
185
<li>Requirements to drop messages should be reconsidered. I guess I would
186
not be the only implementor ignoring them.</li>
187
<li>Logging requirements should be reconsidered and probably be removed.</li>
188
<li>It would be advisable to specify "-" for APP-NAME is the name is not
189
known to the sender.</li>
190
<li>The implications of the current syslog/tcp industry standard on
191
syslog-protocol should be further evaluated and be fully understood</li>