Tuesday, January 17, 2006
Real-time LaBrea data processing program
I finished writing the real-time LaBrea data processing program (most of it last night at The Hospital) and the final few bugs were real doozies.
The program works by getting data from LaBrea, then it looks up the connection and updates the information accordingly. The code pretty much looks like:
void start_tarpit(time_t stamp,char *line)
{
struct tprecord *exists;
struct tprecord rec;
size_t index;
read_record(&rec,line);
exists = tpr_search(&rec,&index);
if (exists == NULL)
exists = pull_free_record();
exists->src = rec.src;
exists->sport = rec.sport;
/* ... */
record_add(index,exists);
}
/*******************************/
struct tprecord *pull_free_record(void)
{
if (g_poolnum == g_poolmax)
do_forced_garbage_collection();
return(&g_pool[g_poolnum++]);
}
/*****************************/
void record_add(size_t index,struct tprecord *rec)
{
if (g_recnum == g_recmax)
do_forced_garbage_collection();
memmove(
&g_rec[index + 1],
&g_rec[index],
(g_recnum - index + 1) * sizeof(struct tprecord *)
);
g_rec[index] = rec;
g_recnum++;
}
tpr_search() is the binary search routine I wrote the other day, and as
you can see, it returns the index to where the record is in the array, or to
where it should be. pull_free_record() just returns
the next free slot in the structure array, and if there are no slots
available, it does a removes some older records according to some criteria.
And record_add will add the record to the pointer array,
also removing older records if there is no space left.
Some records are deleted. All remaining records move about. The pointer array is resorted. So between the time
exists = tpr_search(&rec,&index);
and
record_add(index,exists);
index may not be a valid index anymore!
Oops.
(Never mind the fact that one of the two calls to
do_forced_garbage_collection() is redundant)
Simple enough to fix once I knew what was going on.
Another bug dealt with named pipes. Instead of directly piping the data
from LaBrea to ltpstat (what I call the read-time LaBrea data
processing program), I decided to go through a named pipe, which would allow
me to start and stop either one independantly from the other.
Now, I'm testing my program, running cat sample >
labrea.pipe in one window, and ltpstat --input
labrea.pipe in another. It's working, until the data in
sample runs out and cat closes its side of the
named pipe.
Now, the code that reads in the data is in a library I wrote, and it just
assumes when a read() returns 0, that it's the end of the file,
and marks it as being closed. ltpstat ignores the “end of
file” status, and keeps trying to read a now-closed file. We get into a
busy loop and the system load shoots up. Also, if I now try to pump more
data through the named pipe, ltpstat ignores the data.
Even when I modify the library code to not mark “end of file” when there's
nothing to read does nothing, as from that point on, read()
just returns nothing anyway and seems to be a “feature” of Linux
(or Unix—I didn't have Advanced
Programming in the Unix® Environment with me to look this up), so I
restructure the main loop:
while(1)
{
in = openinput();
while(!StreamEOF(in))
{
process_labrea_output();
}
StreamFree(in);
}
That was all fine and good, until I threw signals into the mix.
I use signals to tell ltpstat (what I call the read-time
LaBrea data processing program) to dump various
information—SIGHUP to print the number of connections and
unique IPs being tracked, SIGUSR1 to do a raw dump of all the
data accumulated so far and SIGUSR2 to generate a more or less
human readable dump.
But signals are basically interrupts. And operating system theory states that one process should never find another process with its pants down (so to speak—and I should warn you—the paper I linked to is way more technical than I've gotten here). Signals and system calls interact (that is, if a process is signaled while making a system call into the kernel) in one of two ways—the system call will simply fail, or it will be restarted automatically. And one can select which method to use.
If I elected to have the system call fail, and ltpstat would
fail, either to open (a system call) the named pipe, or in reading (another
system call) the named pipe after I signal the program.
If I elected to have the system call restarted, and the signal handlers I set up would never get called (due to the way I handle the signals).
I ended up reimplementing two routines from my library (which is used in more than just this program) for just this program. I select for the “system call fail” method, check to see if the system called failed due to a signal, and if so, check for the signals, and try again.
Again, this took a few hours to track down.
But now the program works, and I can finally get real time statistics from LaBrea.
