[manual index][section index]


styxflush - handler for 9P (Styx) flush protocol


include "sys.m";
include "styx.m";
include "styxflush.m";

styxflush := load Styxflush Styxflush->PATH;
init: fn();
tmsg: fn(m: ref Styx->Tmsg,
	flushc: chan of (int, chan of int),
	reply: chan of ref Styx->Rmsg): (int, ref Styx->Rmsg);
rmsg: fn(m: ref Styx->Rmsg): int;
Einterrupted: con "interrupted";


Getting the semantics of the 9P flush(5) protocol correct when handling requests concurrently is surprisingly hard to do. Styxflush is designed to help get it right. It deals with 9P messages for a single 9P session - if a server needs to deal with multiple sessions, then multiple instances of styxflush should be loaded. It assumes there is a loop in a central process that both reads T-messages and sends their R-message replies. Styxflush handles the flushing of requests that are being run outside the central process.

Init must be called before anything else in styxflush to intialise its internal data structures.

When a T-message request arrives that will be dealt with concurrently, tmsg(mflushcreply) should be called to inform styxflush of the new request. M gives the T-message; flushc gives a channel that will be used if the request is flushed (see below), and reply should hold an unbuffered channel that can be used to send a reply to the central loop. Flushc will usually be a fresh channel for each request, but several requests may share the same flushc if, for instance, one process is managing several requests. Tmsg returns a tuple (handled, rm), where handled is non-zero if styxflush has dealt with the request itself. If it has, then the caller must not handle the request; it must send rm as a reply if it is not nil.

Rmsg should be called when a reply message arrives at the central process (the same process that has called tmsg). It returns non-zero if the reply message should actually be sent to the client - otherwise it should be discarded.

Flush Channel
Styxflush notifies a request that it has been flushed by sending a tuple, say (tagrc) on its flush channel. Tag gives the tag of the message that has been flushed, and rc is a channel that should be replied on when the request has been dealt with. There is no requirement that a request read on its flush channel - if it does not, then the replies to any flushes of that request will be delayed until the request is replied to. If it does read a flush request, however, it must reply to the original request before sending on rc. If it has succeeded in aborting the request, it should send an error(5) R-message with the message interrupted (defined as Einterrupted); otherwise it should send its reply as usual.




This is a skeleton of a prototypical structure of a program that uses styxflush.
replyc: chan of ref Rmsg;
centralloop(tm: chan of ref Tmsg, fd: ref Sys->FD)
	replyc = chan of Rmsg;
	m := <-tm =>
		if(m == nil || tagof m == tagof Tmsg.Readerror){
			cleanup();		# kill outstanding processes, etc.
		flushc := chan of (int, chan of int);
		(handled, rm) := styxflush->tmsg(m, flushc, replyc);
			spawn request(m, flushc);
		else if(rm != nil)
	rm := <- replyc =>
sendreply(fd: ref Sys->FD, rm: ref Rmsg)
	d := rm.pack();
	sys->write(fd, d, len d);

request(tm: ref Tmsg, flushc: chan of (int, chan of int))
	pick m := tm {
	Open =>
		replyc <-= ref Rmsg.Open(m.tag, ...);
	Read =>
		x := <-readc =>
			# read from data produced on readc
			replyc <-= ref Rmsg.Read(m.tag, ...);
		(nil, rc) := <-flushc =>
			# read request has been flushed.
			replyc <-= ref Rmsg.Error(m.tag, Einterrupted);
			rc <-= 1;
	etc ...

STYXFLUSH(2 ) Rev:  Tue Mar 31 02:42:39 GMT 2015