Henrik Nordström wrote:
tis 2010-05-18 klockan 23:34 +0000 skrev Amos Jeffries:
I've discovered the VC connections in DNS will need a re-working to handle
the new TCP connection setup handling. I've left that for now since it
appears that you are working on redesigning that area anyway. The new setup
routines will play VERY nicely with persistent TCP links to the
nameservers.
I have not started on the DNS rewrite yet.
I took some extra time last night and broke the back of the selection and
forwarding rewrite. I'm now down to the fine detail build errors. When
those are done I'll push the branch to LP for you to do the DNS fixes on
top of.
Ok.
Pushed to launchpad: lp:~yadi/squid/cleanup-comm
This builds, but has not yet been run tested.
What has changed:
ConnectionDetails objects have been renamed Comm::Connection and been
extended to hold the FD and Squids' socket flags.
Peer selection has been extended to do DNS lookups on the peers chosen
for forwarding to and produce a vector<> of possible connection
endpoints (squid local IP via tcp_outgoing_address or tproxy) and remote
server.
Various connection openers have been converted to use the new
ConnectStateData API and CommCalls (function based so far).
ConnectStateData has been moved into src/comm/ (not yet namespaced) and
had all its DNS lookup operations dropped. To be replaced by a looping
process of attempting to open a socket and join a link as described by
some Comm::Connection or vector<> of same.
ConnectStateData::connect() will go away and do some async work. Will
come back at some point by calling the handler with COMM_OK,
COMM_ERR_CONNECT, COMM_TIMEOUT and ptrs to the Comm::Connection or
vector (whichever were passed in).
On COMM_OK the Comm::Connection pointer or the first entry of the
vector will be an open conn which we can now use.
On COMM_ERR_CONNECT the vector will be empty (all tried and
discarded), the single ptr will be closed if not NULL.
On COMM_TIMEOUT their content is as per COMM_ERR_CONNECT but the
vector may have untried paths still present but closed.
FD opening, FD problems, connection errors, timeouts, early remote
TCP_RST or NACK closure during the setup are all now wrapped out of
sight inside ConnectStateData.
The main-level component may set FD handlers as needed for read/write
and closure of the link in their connection-done handler where the FD
first becomes visible to them.
Besides the testing there is some work to:
* make it obey squid.conf limits on retries and paths looked up.
* make DNS TCP links ('VC') work again.
* make the CommCalls proper AsynCalls and not function handler based.
* make Comm::Connection ref-counted so we can have them stored
in the peer details and further reduce the DNS steps.
* make ICAP do DNS lookups to set its server Comm::Connection properly.
For now it's stuck with the gethostbyname() blocking lookup.
Future work once this is stable is to:
a) push the IDENT, NAT, EUI and TLS operations down into the Comm
layer with simple flags for other layers to turn them on/off as desired.
b) make the general code pass Comm::Connection around so everything
like ACLs can access the client and server conn when they need to.
Amos
--
Please be using
Current Stable Squid 2.7.STABLE9 or 3.1.3
