hashtable: Updated docs.

[libucw.git] / ucw / doc / log.txt

diff --git a/ucw/doc/log.txt b/ucw/doc/log.txt
index 94e1e82a87bd10d58c84932fcc0c66faab367dc9..28265f7290af6c4795cb5d85e8736f75df062435 100644 (file)
--- a/ucw/doc/log.txt
+++ b/ucw/doc/log.txt
@@ -18,9 +18,8 @@ A newline character is automatically appended; the message should not contain
 any control characters.
 
 The first argument of `msg` can be OR'ed with additional flags. Most notably, you can
-add `L_SIGHANDLER` if you wish to log a message from a signal handler (calling
-time-related functions in libc from signal handlers is generally unsafe, so
-`msg` does not log a timestamp in such cases).
+add `L_SIGHANDLER` if you wish to log a message from a signal handler (see below
+for discussion on signals and reentrancy in general).
 
 By default, all messages are logged to stderr. If you wish to use a log file,
 call `log_file(@name)`. All subsequent logging will use this file and stderr
@@ -83,7 +82,7 @@ Example
        int main(int argc, char **argv)
        {
          log_init(argv[0]);
-         struct log_stream *ls = log_new_file("/var/log/utterances");
+         struct log_stream *ls = log_new_file("/var/log/utterances", 0);
          msg(L_INFO | ls->regnum, "Aye captain, we have a log file");
          msg(L_INFO, "Alas, stderr still works");
          return 0;
@@ -118,6 +117,38 @@ LibUCW therefore offers only limited logging services in such situations and
 you must use the `L_SIGHANDLER` flag to request it. Otherwise, deadlocks get
 ready to happen.
 
+Messages logged with `L_SIGHANDLER` set are written directly to stderr (which
+is usually an alias for the main log file, at least if you use <<log_file()>>)
+and they do not carry a timestamp. Logging of sighandler messages to general
+log streams or to syslog is therefore not supported.
+
 ucw/log.h
 ---------
 !!ucw/log.h
+
+Limiting rate: ucw/tbf.h
+------------------------
+
+LibUCW also offers simple means of limiting the rate of log messages (or of any other
+events) by means of a so called 'Token Bucket Filter.' The idea behind this filter is
+simple: To log a message, we need a token. The available tokens are accumulated in
+a bucket which has a fixed 'filling rate' (the number of tokens arriving in the bucket
+per second, which may be a fractional number) and fixed 'maximum capacity.' The
+bucket receives the tokens continuously with the given rate and when it reaches
+the maximum capacity, the extra tokens are dropped on the floor. When a message
+has to be sent, we take a single token from the bucket and if there wasn't any,
+we drop the message.
+
+The filling rate therefore describes the maximum sustained rate of messages,
+while the bucket capacity tells the filter the maximum length of a short burst,
+which can temporarily exceed the rate.
+
+A general bucket filter is available in `ucw/tbf.h`. The usual way of using it
+to limit logging is to set up a filter hook of a stream which asks the TBF for
+every message. (Remember, though, that if your program is multithreaded, the
+filter hook can be run in multiple threads in parallel, so it has to guard the
+TBF by a lock.) The configuration interface for log streams described above
+is able to attach rate limiters to streams per user's request, so you usually
+need not take any extra care.
+
+!!ucw/tbf.h