wessels [Sat, 28 Mar 1998 05:44:18 +0000 (05:44 +0000)]
Changed comm connect routines so that the timeout value and handler
are CLEARED when connection succeeds. This means all the connect
callback (CNCB) functions need to re-set the timeout handlers.
kostas [Wed, 25 Mar 1998 16:21:39 +0000 (16:21 +0000)]
- renovated mib and added descriptions and comments
- added hit and byte counters to client_db , for cacheClientTable
- cacheClientTable, netdbTable, cachePeerTable, cacheConnTable and
fqdnCacheTable are now indexed by ip address. hash_lookup was enhanced
to allow for subsequent hash_next's similar to hash_first-hash_next,
to speed up getnext's in tables which refer to hash-table structures.
- added generic (well, sorf of) table indexing functionality
- added makefile dependencies for snmplib and cache_snmp.h
- WaisHost, WaisPort, Timeouts removed
- FdTable split into FdTable and ConnTable. FdTable simplified
- PeerTable and PeerStat merged and put into new cacheMesh group
- cacheClientTable added for client statistics and accounting (cacheMesh 2)
- cacheSec and cacheAccounting groups removed
- fixed acl bug when communities not defined
- acl config now survives bad configuration (public is defined "implicitly")
wessels [Wed, 25 Mar 1998 12:30:54 +0000 (12:30 +0000)]
From: "Michael O'Reilly" <michael@metal.iinet.net.au>
#2. More importantly, comm_select_incoming is getting called WAY too
often on machines with lot of file handles. It's getting called when a
filehandle is tested, rather than when a handler is actually called
(no point in doing it per check, the check is _very_ cheap. The
handler may not be).
wessels [Wed, 25 Mar 1998 12:21:47 +0000 (12:21 +0000)]
From: "Michael O'Reilly" <michael@metal.iinet.net.au>
#1. close() called on a network socket on linux doesn't block, so
there's no point in paying the cost an aioClose.
wessels [Tue, 24 Mar 1998 05:10:48 +0000 (05:10 +0000)]
Fixes for multiple clients on one StoreEntry. storeAbort() was
incorrectly being called when one of multiple clients would
disconnect.
One problem seemed to be that storePendingNClients()
counted the number of store_client's with callback functions registered
(i.e. assumed all clients had recently called storeClientCopy).
storePendingNClients() has been changed to return mem->nclients.
This assumes that all clients properly use storeUnregister(), which
may not have been the case in the past.
Also, clientWriteComplete() would call CheckQuickAbort() if comm_write
returned an error. Then CheckQuickAbort() would get called again
in httpRequestFree(). The first call seems to be unnecessary, so
it is #ifdef'd out for now.
rousskov [Sun, 22 Mar 1998 14:38:59 +0000 (14:38 +0000)]
- Fixed a bug when we were printing a non-existent "->" link as a part of the
ftp dir entry. For example, the following should not lead to a coredump now:
rousskov [Sat, 21 Mar 1998 01:06:37 +0000 (01:06 +0000)]
- Separated raw HTTP headers from their "compiled" values. Squid is now
forwarding headers even it cannot "compile" them. Virtually no changes to
common format headers are done. Compiled values are stored in Reply/Request
objects.
wessels [Thu, 19 Mar 1998 14:13:29 +0000 (14:13 +0000)]
reinstall ICP reqnum high-byte hack for encoding request method
call neighborsUdpAck() even when we don't have the cache key.
i.e. look up the StoreEntry in neighborsUdpAck() instead of
icpHandleIcpV2()
wessels [Tue, 17 Mar 1998 12:12:30 +0000 (12:12 +0000)]
SNMP changes/fixes
- changed some snmplib debugging messages
- simplifed and fixed snmplib_debug_hook interface. It
wasn't properly passing all the 'varargs' args. Now
it just vsnprintf's to a buf and passes that.
- struct _snmpconf was never used.
- made 'snmp_agent_conf' a stringlist type
- moved 'snmp_agent_conf' processing to snmpInit().
- moved snmpInit() call to before 'ready to serve requests'
message.