ref: 3ab9f80448b5d80352bed6251b98f41aeee7af09
parent: d95c2f24d7ff9f1eaacb11ac2ffb7ca69f1d3197
author: Noam Preil <noam@pixelhero.dev>
date: Thu Jul 4 18:19:28 EDT 2024
server: give each client its own Proc
--- a/notebook
+++ b/notebook
@@ -1229,4 +1229,173 @@
Good news is that it's still broken after reboot. Did something get accidentally added to a commit and break it? :/
+8e31146d45d0612b82a5438012e6ed03be3995a9 works. cool, let's bisect this.
+cdc922649bc14a2e25abd4593b244f36fffff1d0 does not. Okay, did ca929 break it?
+Nope, that one works. Um. cdc922649bc14a2e25abd4593b244f36fffff1d0 broke it??
+
+git/export cdc922649bc14a2e25abd4593b244f36fffff1d0
+From: Noam Preil <noam@pixelhero.dev>
+Date: Thu, 04 Jul 2024 21:07:18 +0000
+Subject: [PATCH] server: handle each client in a separate proc
+
+---
+diff ca929c334fedc59fc1240b2c64b887082ef331f2 cdc922649bc14a2e25abd4593b244f36fffff1d0
+--- a/notebook
++++ b/notebook
+@@ -1155,3 +1155,46 @@
+ % dd -if /dev/jasnah/arenas -bs 1024 -count 8 -skip 248
+
+ that works. lol.
++
++Okay, config parsing is good. Let's fix multiclient support.
++
++With a foreground proc for handling, I - oh. This is why testing is important! neoventi shuts down on disconnect >_<
++
++Ohhhhh on vthangup we're bouncing() with 2 as the parameter, which triggers exits(). That was easy lmao.
++
++annnnd with that fixed it crashes on hangup with one thread, too!
++
++7.neoventi 55793: suicide: sys: trap: misaligned pc pc=0x5
++mk: ./7.neoventi -c /dev/jasnah/arenas : exit status=7.neoventi 55793:
++
++... huh.
++
++waiting for packet...Received 0 bytes!
++hanging up: unknown control request
++
++unknown control request is not a neoventi error, that's a driver error. Preeeetty sure it's just from the connection dropping, though. Probably. Maybe not.
++
++The trampoline appears to be going to the wrong place... the heck?
++
++inittrampoline calls setjmp(). The longjmp is correctly returning to that setjmp, but the return in inittrampoline is not returning to handleproc a second time??
++
++ Longjmp restores the environment saved by the last call of
++ setjmp. It then causes execution to continue as if the call
++ of setjmp had just returned with value val. The invoker of
++ setjmp must not itself have returned in the interim. All
++ accessible data have values as of the time longjmp was
++ called.
++The invoker of
++ setjmp must not itself have returned in the interim.
++
++ohhhh dumb. duuumb.
++
++manually inlining initrampoline() fixed that. Multithread??
++
++Yep!!! :D
++
++Technically works. Ish. So long as only one actually does anything at a time >_<
++
++Cache not locking right, maybe?
++
++Yep!
+--- a/server.c
++++ b/server.c
+@@ -17,6 +17,7 @@
+ msg = vsmprint(msg, args);
+ werrstr(msg);
+ free(msg);
++ fprint(2, "error: %r\n");
+ va_end(args);
+ if(tbuf != nil){+ len = snprint(tbuf+6, 0x10000, "neoventi: %r");
+@@ -171,38 +172,31 @@
+ }
+ }
+
+-static int
+-inittrampoline(VtConn *conn)
++static void
++handleproc(void *fd)
+ {+- switch(setjmp(conn->bounce)){++ char buf[MaxPacketSize];
++ VtConn conn;
++ conn.fd = (int)(usize)fd;
++ switch(setjmp(conn.bounce)){+ case 0:
+- return 1;
++ vthello(conn, buf);
++ vtloop(conn, buf);
++ abort();
+ case 1:
+ fprint(2, "abandoning client: %r\n");
++ break;
+ case 2:
+- exits(nil);
+- return 0;
++ fprint(2, "hanging up: %r\n");
++ break;
+ default:
+ fprint(2, "internal error: unexpected bounce code\n");
+- return 0;
++ break;
+ }
++ close(conn.fd);
+ }
+
+ static void
+-handleproc(void *fd)
+-{+- char buf[MaxPacketSize];
+- VtConn conn;
+- conn.fd = (int)(usize)fd;
+- if(!inittrampoline(&conn)){+- close(conn.fd);
+- return;
+- }
+- vthello(conn, buf);
+- vtloop(conn, buf);
+-}
+-
+-static void
+ handle(int ctl, char *dir)
+ {+ int fd = accept(ctl, dir);
+@@ -210,8 +204,7 @@
+ fprint(2, "failed to accept connection\n");
+ return;
+ }
+- handleproc((void*)fd);
+-// proccreate(handleproc, (void*)fd, mainstacksize);
++ proccreate(handleproc, (void*)fd, mainstacksize);
+ }
+
+ void
+
+
+but, but that's just what it's supposed to be?!?!
+
+Dammit. Isolate this into smaller chunks :(
+
+% git/branch front
+% git/revert -c ca929c334fedc59fc1240b2c64b887082ef331f2 server.c
+% mk all; mk fg
+% mk vac
+
+Works. Okay. Um.
+
+% git/revert server.c
+% mk all; mk fg
+% mk vac
+
+Hangs. Yep. That confirms it.
+
+- handleproc((void*)fd);
+-// proccreate(handleproc, (void*)fd, mainstacksize);
++// handleproc((void*)fd);
++ proccreate(handleproc, (void*)fd, mainstacksize);
+ }
+
+That does it??? really?? but it worked fine!!
+
+Is it because I'm passing the conn - and thus the jmp_buf - around on the stack?
+
+Ohhh my gosh it's the stack size >_<
+
+I had it bumped to see if that fixed a different issue, then fixed that when committing >_<
+
+It needs to be bumped, though. Which makes sense; we've got two packet-sized buffers, and those are 64K each, so 128K isn't big enough. Let's bump mainstacksize to, I dunno, 256K should be plenty.
+
+Yep, that's good. Cool!
--- a/server.c
+++ b/server.c
@@ -202,8 +202,7 @@
fprint(2, "failed to accept connection\n");
return;
}
- handleproc((void*)fd);
-// proccreate(handleproc, (void*)fd, mainstacksize);
+ proccreate(handleproc, (void*)fd, 256*1024);
}
void