1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
|
<html>
<head>
<title>syslog-protocol support in rsyslog</title>
</head>
<body>
<h1>syslog-protocol support in rsyslog</h1>
<p><b><a href="http://www.rsyslog.com/">Rsyslog</a> provides a trial
implementation of the proposed
<a href="http://www.monitorware.com/Common/en/glossary/syslog-protocol.php">
syslog-protocol</a> standard.</b> The intention of this implementation is to
find out what inside syslog-protocol is causing problems during implementation.
As syslog-protocol is a standard under development, its support in rsyslog is
highly volatile. It may change from release to release. So while it provides
some advantages in the real world, users are cautioned against using it right
now. If you do, be prepared that you will probably need to update all of your
rsyslogds with each new release. If you try it anyhow, please provide feedback
as that would be most benefitial for us.</p>
<h2>Currently supported message format</h2>
<p>Due to recent discussion on syslog-protocol, we do not follow any specific
revision of the draft but rather the candidate ideas. The format supported
currently is:</p>
<p><b><code><PRI>VERSION SP TIMESTAMP SP HOSTNAME SP APP-NAME SP PROCID SP MSGID SP [SD-ID]s
SP MSG</code></b></p>
<p>Field syntax and semantics are as defined in IETF I-D syslog-protocol-15.</p>
<h2>Capabilities Implemented</h2>
<ul>
<li>receiving message in the supported format (see above)</li>
<li>sending messages in the supported format</li>
<li>relaying messages</li>
<li>receiving messages in either legacy or -protocol format and transforming
them into the other one</li>
<li>virtual availability of TAG, PROCID, APP-NAME, MSGID, SD-ID no matter if
the message was received via legacy format, API or syslog-protocol format (non-present
fields are being emulated with great success)</li>
<li>maximum message size is set via preprocessor #define</li>
<li>syslog-protocol messages can be transmitted both over UDP and plain TCP
with some restrictions on compliance in the case of TCP</li>
</ul>
<h2>Findings</h2>
<p>This lists what has been found during implementation:</p>
<ul>
<li>The same receiver must be able to support both legacy and
syslog-protocol syslog messages. Anything else would be a big inconvenience
to users and would make deployment much harder. The detection must be done
automatically (see below on how easy that is).</li>
<li><b>NUL characters inside MSG</b> cause the message to be truncated at
that point. This is probably a major point for many C-based implementations.
No measures have yet been taken against this. Modifying the code to "cleanly"
support NUL characters is non-trivial, even though rsyslogd already has some
byte-counted string library (but this is new and not yet available
everywhere).</li>
<li><b>character encoding in MSG</b>: is is problematic to do the right
UTF-8 encoding. The reason is that we pick up the MSG from the local domain
socket (which got it from the syslog(3) API). The text obtained does not
include any encoding information, but it does include non US-ASCII
characters. It may also include any other encoding. Other than by guessing
based on the provided text, I have no way to find out what it is. In order
to make the syslogd do anything useful, I have now simply taken the message
as is and stuffed it into the MSG part. Please note that I think this will
be a route that other implementors would take, too.</li>
<li>A minimal parser is easy to implement. It took me roughly 2 hours to add
it to rsyslogd. This includes the time for restructering the code to be able
to parse both legacy syslog as well as syslog-protocol. The parser has some
restrictions, though<ul>
<li>STRUCTURED-DATA field is extracted, but not validated. Structured data
"[test ]]" is not caught as an error. Nor are any other errors caught. For
my needs with this syslogd, that level of structued data processing is
probably sufficient. I do not want to parse/validate it in all cases. This
is also a performance issue. I think other implementors could have the same
view. As such, we should not make validation a requirement.</li>
<li>MSG is not further processed (e.g. Unicode not being validated)</li>
<li>the other header fields are also extracted, but no validation is
performed right now. At least some validation should be easy to add (not
done this because it is a proof-of-concept and scheduled to change).</li>
</ul>
</li>
<li>Universal access to all syslog fields (missing ones being emulated) was
also quite easy. It took me around another 2 hours to integrate emulation of
non-present fields into the code base.</li>
<li>The version at the start of the message makes it easy to detect if we
have legacy syslog or syslog-protocol. Do NOT move it to somewhere inside
the middle of the message, that would complicate things. It might not be
totally fail-safe to just rely on "1 " as the "cookie" for a syslog-protocol.
Eventually, it would be good to add some more uniqueness, e.g. "@#1 ".</li>
<li>I have no (easy) way to detect truncation if that happens on the UDP
stack. All I see is that I receive e.g. a 4K message. If the message was e.g.
6K, I received two chunks. The first chunk (4K) is correctly detected as a
syslog-protocol message, the second (2K) as legacy syslog. I do not see what
we could do against this. This questions the usefulness of the TRUNCATE bit.
Eventually, I could look at the UDP headers and see that it is a fragment. I
have looked at a network sniffer log of the conversation. This looks like
two totally-independant messages were sent by the sender stack.</li>
<li>The maximum message size is currently being configured via a
preprocessor #define. It can easily be set to 2K or 4K, but more than 4K is
not possible because of UDP stack limitations. Eventually, this can be
worked around, but I have not done this yet.</li>
<li>rsyslogd can accept syslog-protocol formatted messages but is able to
relay them in legacy format. I find this a must in real-life deployments.
For this, I needed to do some field mapping so that APP-NAME/PROCID are
mapped into a TAG.</li>
<li>rsyslogd can also accept legacy syslog message and relay them in
syslog-protocol format. For this, I needed to apply some sub-parsing of the
TAG, which on most occasions provides correct results. There might be some
misinterpretations but I consider these to be mostly non-intrusive. </li>
<li>Messages received from the syslog API (the normal case under *nix) also
do not have APP-NAME and PROCID and I must parse them out of TAG as
described directly above. As such, this algorithm is absolutely vital to
make things work on *nix.</li>
<li>I have an issue with messages received via the syslog(3) API (or, to be
more precise, via the local domain socket this API writes to): These
messages contain a timestamp, but that timestamp does neither have the year
nor the high-resolution time. The year is no real issue, I just take the
year of the reception of that message. There is a very small window of
exposure for messages read from the log immediately after midnight Jan 1st.
The message in the domain socket might have been written immediately before
midnight in the old year. I think this is acceptable. However, I can not
assign a high-precision timestamp, at least it is somewhat off if I take the
timestamp from message reception on the local socket. An alternative might
be to ígnore the timestamp present and instead use that one when the message
is pulled from the local socket (I am talking about IPC, not the network -
just a reminder...). This is doable, but eventually not advisable. It looks
like this needs to be resolved via a configuration option.</li>
<li>rsyslogd already advertised its origin information on application
startup (in a syslog-protocol-14 compatible format). It is fairly easy to
include that with any message if desired (not currently done).</li>
<li>A big problem I noticed are malformed messages. In -syslog-protocol, we
recommend/require to discard malformed messages. However, in practice users
would like to see everything that the syslogd receives, even if it is in
error. For the first version, I have not included any error handling at all.
However, I think I would deliberately ignore any "discard" requirement. My
current point of view is that in my code I would eventually flag a message
as being invalid and allow the user to filter on this invalidness. So these
invalid messages could be redirected into special bins.</li>
<li>The error logging recommendations (those I insisted on;)) are not really
practical. My application has its own error logging philosophy and I will
not change this to follow a draft.</li>
<li>Relevance of support for leap seconds and senders without knowledge of
time is questionable. I have not made any specific provisions in the code
nor would I know how to handle that differently. I could, however, pull the
local reception timestamp in this case, so it might be useful to have this
feature. I do not think any more about this for the initial proof-of-concept.
Note it as a potential problem area, especially when logging to databases.</li>
<li>The HOSTNAME field for internally generated messages currently contains
the hostname part only, not the FQDN. This can be changed inside the code
base, but it requires some thinking so that thinks are kept compatible with
legacy syslog. I have not done this for the proof-of-concept, but I think it
is not really bad. Maybe an hour or half a day of thinking.</li>
<li>It is possible that I did not receive a TAG with legacy syslog or via
the syslog API. In this case, I can not generate the APP-NAME. For
consistency, I have used "-" in such cases (just like in PROCID, MSGID and
STRUCTURED-DATA).</li>
<li>As an architectural side-effect, syslog-protocol formatted messages can
also be transmitted over non-standard syslog/raw tcp. This implementation
uses the industry-standard LF termination of tcp syslog records. As such,
syslog-protocol messages containing a LF will be broken invalidly. There is
nothing that can be done against this without specifying a TCP transport.
This issue might be more important than one thinks on first thought. The
reason is the wide deployment of syslog/tcp via industry standard.</li>
</ul>
<p><b>Some notes on syslog-transport-udp-06</b></p>
<ul>
<li>I did not make any low-level modifications to the UDP code and think I
am still basically covered with this I-D.</li>
<li>I deliberately violate section 3.3 insofar as that I do not necessarily
accept messages destined to port 514. This feature is user-required and a
must. The same applies to the destination port. I am not sure if the "MUST"
in section 3.3 was meant that this MUST be an option, but not necessarily be
active. The wording should be clarified.</li>
<li>section 3.6: I do not check checksums. See the issue with discarding
messages above. The same solution will probably be applied in my code.</li>
</ul>
<p> </p>
<h2>Conlusions/Suggestions</h2>
<p>These are my personal conclusions and suggestions. Obviously, they must be
discussed ;)</p>
<ul>
<li>NUL should be disallowd in MSG</li>
<li>As it is not possible to definitely know the character encoding of the
application-provided message, MSG should <b>not</b> be specified to use UTF-8
exclusively. Instead, it is suggested that any encoding may be used but
UTF-8 is preferred. To detect UTF-8, the MSG should start with the UTF-8
byte order mask of "EF BB BF" if it is UTF-8 encoded (see section 155.9 of
<a href="http://www.unicode.org/versions/Unicode4.0.0/ch15.pdf">
http://www.unicode.org/versions/Unicode4.0.0/ch15.pdf</a>) </li>
<li>Requirements to drop messages should be reconsidered. I guess I would
not be the only implementor ignoring them.</li>
<li>Logging requirements should be reconsidered and probably be removed.</li>
<li>It would be advisable to specify "-" for APP-NAME is the name is not
known to the sender.</li>
<li>The implications of the current syslog/tcp industry standard on
syslog-protocol should be further evaluated and be fully understood</li>
</ul>
<p> </p>
</body>
</html>
|
