Re: [SR-Dev] [Serdev] CVS:commitlog: sip_router/modules/tm t_cancel.c t_fwd.c t_reply.c t_reply.h t_suspend.c t_suspend.h tm.c tm_load.c tm_load.h
[crossposting since this is of general interest]
On Nov 12, 2008 at 00:30, Klaus Darilion klaus.mailinglists@pernau.at wrote:
Hi Miklos!
This sounds great. Can you make a sample function which uses the new feature?
We are thinking of doing a dns_prefetch module using this. The sip-router.cfg will look like:
route{ ... dns_prefetch("uri", ROUTE_DNS_OK); # end of script }
route[ROUTE_DNS_OK]{ if (dns_error){ ... } # continue normal processing # the uri was resolved and is in the dns cache ... t_relay() ... }
However don't expect something quickly as everybody is quite busy and we would still need to write a dns resolver process pool (time consuming).
Note that this is not the "final" async. solution since it's not transparent for the script writer (scripts would have to be written in a special way), works only with tm and it's also a little inefficient for dns usage. In the long term we will have completely transparent async. support for dns and maybe db lookups (though it's controversial if we really need it for normal db lookups). This will take much longer and will require among other things reviewing all the global variable use (most of the global variables will either need to be registered or allocated in some known special place). If Miklos doesn't beat me to it, I'll probably start working on it only after the core merging is complete.
Can this also be used for DNS lookups and TCP/TLS connection setup?
Yes for DNS (see above) and in general for any route-level async processing involving tm (e.g. lookup some part of the message in a slow DB and execute automatically another route when the DB responds).
It cannot be used for TCP/TLS, but it's not needed anyway. In ser TCP connection setup and TCP send is already asynchronous, at least if you have tcp_buf_write=yes in ser.cfg (I agree it's not a very well chosen name :-)). TLS does not support it yet, so if you use TLS you should not set this option (that's why it's not yet the default). TLS support for it requires non-trivial changes in the tcp core and it will probably come after the merge (it's started already but far from finished). (BTW: most of the core new options documentation in ser can be found in NEWS)
Andrei
Further, also for Kamailio cooperation better documentation would be needed. Thus, instead of writing long CVS commit messages I think it would be better to write the text into the module documentation, and then copy/paste it into the CVS commit.
regards Klaus
Miklos Tirpak wrote:
tirpi 2008/11/10 13:47:03 CET
SER CVS Repository
Modified files: modules/tm t_cancel.c t_fwd.c t_reply.c t_reply.h tm.c tm_load.c tm_load.h Added files: modules/tm t_suspend.c t_suspend.h Log: t_suspend() and t_continue() functions are introduced.
These fuctions can be used by other modules to implement asynchronous actions: t_suspend() saves the transaction, returns its identifiers, and t_continue() continues the SIP request processing. (The transaction processing does not continue from the same point in the script, a separate route block defined by the parameter of t_continue() is executed instead. The reply lock is held during the route block execution.) FR timer is ticking while the transaction is suspended, and the transaction's failure route is executed if t_continue() is not called in time.
Missing: msg lumps are saved by t_suspend() and are not updated by the subsequent t_relay(). This means that the modifications made between them are lost.
Revision Changes Path 1.32 +7 -2 sip_router/modules/tm/t_cancel.c http://cvs.berlios.de/cgi-bin/viewcvs.cgi/ser/sip_router/modules/tm/t_cancel... 1.107 +8 -2 sip_router/modules/tm/t_fwd.c http://cvs.berlios.de/cgi-bin/viewcvs.cgi/ser/sip_router/modules/tm/t_fwd.c.... 1.162 +4 -4 sip_router/modules/tm/t_reply.c http://cvs.berlios.de/cgi-bin/viewcvs.cgi/ser/sip_router/modules/tm/t_reply.... 1.31 +7 -1 sip_router/modules/tm/t_reply.h http://cvs.berlios.de/cgi-bin/viewcvs.cgi/ser/sip_router/modules/tm/t_reply.... 1.1 +154 -0 sip_router/modules/tm/t_suspend.c (new) http://cvs.berlios.de/cgi-bin/viewcvs.cgi/ser/sip_router/modules/tm/t_suspen... 1.1 +42 -0 sip_router/modules/tm/t_suspend.h (new) http://cvs.berlios.de/cgi-bin/viewcvs.cgi/ser/sip_router/modules/tm/t_suspen... 1.177 +3 -1 sip_router/modules/tm/tm.c http://cvs.berlios.de/cgi-bin/viewcvs.cgi/ser/sip_router/modules/tm/tm.c.dif... 1.31 +11 -1 sip_router/modules/tm/tm_load.c http://cvs.berlios.de/cgi-bin/viewcvs.cgi/ser/sip_router/modules/tm/tm_load.... 1.31 +4 -1 sip_router/modules/tm/tm_load.h http://cvs.berlios.de/cgi-bin/viewcvs.cgi/ser/sip_router/modules/tm/tm_load.... _______________________________________________ Serdev mailing list Serdev@lists.iptel.org http://lists.iptel.org/mailman/listinfo/serdev
Serdev mailing list Serdev@lists.iptel.org http://lists.iptel.org/mailman/listinfo/serdev
Andrei Pelinescu-Onciul schrieb:
[crossposting since this is of general interest]
On Nov 12, 2008 at 00:30, Klaus Darilion klaus.mailinglists@pernau.at wrote:
Hi Miklos!
This sounds great. Can you make a sample function which uses the new feature?
We are thinking of doing a dns_prefetch module using this. The sip-router.cfg will look like:
route{ ... dns_prefetch("uri", ROUTE_DNS_OK); # end of script }
route[ROUTE_DNS_OK]{ if (dns_error){ ... } # continue normal processing # the uri was resolved and is in the dns cache ... t_relay() ... }
Ok. I see.
However don't expect something quickly as everybody is quite busy and we would still need to write a dns resolver process pool (time consuming).
I still wonder if the DNS resolver in ser is really a good idea. For example if you take a look at bind, it is really mature and still they fix several issues with each release - and doing all the tricks to avoid cache poisoning and handling DNSSEC correctly is not easy. Thus, maybe having a DNS cache in ser can make sense, but having a full resolver IMO not.
Can this also be used for DNS lookups and TCP/TLS connection setup?
Yes for DNS (see above) and in general for any route-level async processing involving tm (e.g. lookup some part of the message in a slow DB and execute automatically another route when the DB responds).
It cannot be used for TCP/TLS, but it's not needed anyway. In ser TCP connection setup and TCP send is already asynchronous, at least if you have tcp_buf_write=yes in ser.cfg (I agree it's not a very well
Interesting. Is it really completely asynchronous with any blocking, or is it handed over to a dedicated "TCP send" process which is then blocked?
(BTW: most of the core new options documentation in ser can be found in NEWS)
Probably this should be merged too.
thanks klaus
On Nov 12, 2008 at 10:33, Klaus Darilion klaus.mailinglists@pernau.at wrote:
Andrei Pelinescu-Onciul schrieb:
[crossposting since this is of general interest]
On Nov 12, 2008 at 00:30, Klaus Darilion klaus.mailinglists@pernau.at wrote:
Hi Miklos!
This sounds great. Can you make a sample function which uses the new feature?
We are thinking of doing a dns_prefetch module using this. The sip-router.cfg will look like:
route{ ... dns_prefetch("uri", ROUTE_DNS_OK); # end of script }
route[ROUTE_DNS_OK]{ if (dns_error){ ... } # continue normal processing # the uri was resolved and is in the dns cache ... t_relay() ... }
Ok. I see.
However don't expect something quickly as everybody is quite busy and we would still need to write a dns resolver process pool (time consuming).
I still wonder if the DNS resolver in ser is really a good idea. For example if you take a look at bind, it is really mature and still they fix several issues with each release - and doing all the tricks to avoid cache poisoning and handling DNSSEC correctly is not easy. Thus, maybe having a DNS cache in ser can make sense, but having a full resolver IMO not.
Sorry for the name, I was not talking about a full resolver. The DNS cache uses in fact a minimalistic resolver. We would just need to replace res_search with a res_mkquery and then "manually" send the request and wait for an answer in a poll() loop (we won't actually use poll, but the io_wait.h optimized stuff that uses epoll(), kqueue() a.s.o depending on the system). We don't need a recursive resolver, we only need something minimal. One should still use a close DNS server since the internal resolver does not support recursive queries, DNSSEC a.s.o (we could add them, but so far there would be little benefit).
Can this also be used for DNS lookups and TCP/TLS connection setup?
Yes for DNS (see above) and in general for any route-level async processing involving tm (e.g. lookup some part of the message in a slow DB and execute automatically another route when the DB responds).
It cannot be used for TCP/TLS, but it's not needed anyway. In ser TCP connection setup and TCP send is already asynchronous, at least if you have tcp_buf_write=yes in ser.cfg (I agree it's not a very well
Interesting. Is it really completely asynchronous with any blocking, or is it handed over to a dedicated "TCP send" process which is then blocked?
Yes, it's completely asynchronous, no process is blocked. A brief description on how it's working for send on an already opened connection: - check if the connection has data buffered. If it does it means that either is still connecting or a previous send failed to be executed in-place and queued the data => add to the queued data, return. - if no data buffered try non-blocking send. If it fails queue data to the ser internal tcp connection structure and enable waiting for POLLOUT events (socket send buffer has enough space). return.
For connect is similar: connect() non blocking, queue data, return.
When a socket can be written, the main tcp process will write the queued data (or as much as possible from it). -
(BTW: most of the core new options documentation in ser can be found in NEWS)
Probably this should be merged too.
Yes.
Andrei
Hi Klaus,
regarding your original question, you are right, I have not documented the functions yet. I will add the doc to the devel section of the readme as soon as I will have some time.
I also have a small module that I was testing with, I can also commit this module if you wish as a sample. There is no real functionality in the module, it only suspends the transaction for some seconds, and when the timer fires, the transaction continues. It is not even good for performance testing, because there is a single timer process that calls t_continue() for all the transactions. As Andrei wrote, writing a real async function would take more time, this was only a baby step.
Regards, Miklos
On 11/12/2008 11:16 AM, Andrei Pelinescu-Onciul wrote:
On Nov 12, 2008 at 10:33, Klaus Darilion klaus.mailinglists@pernau.at wrote:
Andrei Pelinescu-Onciul schrieb:
[crossposting since this is of general interest]
On Nov 12, 2008 at 00:30, Klaus Darilion klaus.mailinglists@pernau.at wrote:
Hi Miklos!
This sounds great. Can you make a sample function which uses the new feature?
We are thinking of doing a dns_prefetch module using this. The sip-router.cfg will look like:
route{ ... dns_prefetch("uri", ROUTE_DNS_OK); # end of script }
route[ROUTE_DNS_OK]{ if (dns_error){ ... } # continue normal processing # the uri was resolved and is in the dns cache ... t_relay() ... }
Ok. I see.
However don't expect something quickly as everybody is quite busy and we would still need to write a dns resolver process pool (time consuming).
I still wonder if the DNS resolver in ser is really a good idea. For example if you take a look at bind, it is really mature and still they fix several issues with each release - and doing all the tricks to avoid cache poisoning and handling DNSSEC correctly is not easy. Thus, maybe having a DNS cache in ser can make sense, but having a full resolver IMO not.
Sorry for the name, I was not talking about a full resolver. The DNS cache uses in fact a minimalistic resolver. We would just need to replace res_search with a res_mkquery and then "manually" send the request and wait for an answer in a poll() loop (we won't actually use poll, but the io_wait.h optimized stuff that uses epoll(), kqueue() a.s.o depending on the system). We don't need a recursive resolver, we only need something minimal. One should still use a close DNS server since the internal resolver does not support recursive queries, DNSSEC a.s.o (we could add them, but so far there would be little benefit).
Can this also be used for DNS lookups and TCP/TLS connection setup?
Yes for DNS (see above) and in general for any route-level async processing involving tm (e.g. lookup some part of the message in a slow DB and execute automatically another route when the DB responds).
It cannot be used for TCP/TLS, but it's not needed anyway. In ser TCP connection setup and TCP send is already asynchronous, at least if you have tcp_buf_write=yes in ser.cfg (I agree it's not a very well
Interesting. Is it really completely asynchronous with any blocking, or is it handed over to a dedicated "TCP send" process which is then blocked?
Yes, it's completely asynchronous, no process is blocked. A brief description on how it's working for send on an already opened connection:
- check if the connection has data buffered. If it does it means that
either is still connecting or a previous send failed to be executed in-place and queued the data => add to the queued data, return.
- if no data buffered try non-blocking send. If it fails queue data
to the ser internal tcp connection structure and enable waiting for POLLOUT events (socket send buffer has enough space). return.
For connect is similar: connect() non blocking, queue data, return.
When a socket can be written, the main tcp process will write the queued data (or as much as possible from it).
(BTW: most of the core new options documentation in ser can be found in NEWS)
Probably this should be merged too.
Yes.
Andrei
Andrei Pelinescu-Onciul schrieb:
[crossposting since this is of general interest]
On Nov 12, 2008 at 00:30, Klaus Darilion klaus.mailinglists@pernau.at wrote:
Can this also be used for DNS lookups and TCP/TLS connection setup?
Yes for DNS (see above) and in general for any route-level async processing involving tm (e.g. lookup some part of the message in a slow DB and execute automatically another route when the DB responds).
It cannot be used for TCP/TLS, but it's not needed anyway. In ser TCP connection setup and TCP send is already asynchronous, at least if you have tcp_buf_write=yes in ser.cfg (I agree it's not a very well chosen name :-)). TLS does not support it yet, so if you use TLS
Hi Andrei!
I wonder what happen if the asynchronous sending fails (e.g. TCP connection is lost). Will it jump to failure route (for requests)? What happens if the sending of the response (e.g. 200 ok) fails? Will acc module account the failed 200 ok?
regards klaus
On Dec 09, 2008 at 12:07, Klaus Darilion klaus.mailinglists@pernau.at wrote:
Andrei Pelinescu-Onciul schrieb:
[crossposting since this is of general interest]
On Nov 12, 2008 at 00:30, Klaus Darilion klaus.mailinglists@pernau.at wrote:
Can this also be used for DNS lookups and TCP/TLS connection setup?
Yes for DNS (see above) and in general for any route-level async processing involving tm (e.g. lookup some part of the message in a slow DB and execute automatically another route when the DB responds).
It cannot be used for TCP/TLS, but it's not needed anyway. In ser TCP connection setup and TCP send is already asynchronous, at least if you have tcp_buf_write=yes in ser.cfg (I agree it's not a very well chosen name :-)). TLS does not support it yet, so if you use TLS
Hi Andrei!
I wonder what happen if the asynchronous sending fails (e.g. TCP connection is lost). Will it jump to failure route (for requests)?
It will time out and then jump into the failure route (it doesn't jump directly on async send failure, it will wait for the transaction timeout).
What happens if the sending of the response (e.g. 200 ok) fails? Will acc module account the failed 200 ok?
No, not directly (it has to look for the ACK). On tcp in general is not possible to say what packet made it to the other side when a connection is suddenly closed, while still having data in the socket send buffers. With the async send we could make it a little better and detect _some_ packets that we know for sure we did not send (and callback tm and fallback or execute the failure route for them), but we cannot tell anything about all the stuff in the socket send buffers (64k+). I think this would help only for the async. connect part (where we don't have any kernel buffering), but for already opened connection that close due to an error it won't help much.
Andrei
Andrei Pelinescu-Onciul writes:
I wonder what happen if the asynchronous sending fails (e.g. TCP connection is lost). Will it jump to failure route (for requests)?
It will time out and then jump into the failure route (it doesn't jump directly on async send failure, it will wait for the transaction timeout).
if this is the same as what happens when udp destination is not responding, then it is fine.
-- juha
On Dec 12, 2008 at 08:28, Juha Heinanen jh@tutpro.com wrote:
Andrei Pelinescu-Onciul writes:
I wonder what happen if the asynchronous sending fails (e.g. TCP connection is lost). Will it jump to failure route (for requests)?
It will time out and then jump into the failure route (it doesn't jump directly on async send failure, it will wait for the transaction timeout).
if this is the same as what happens when udp destination is not responding, then it is fine.
Yes, it's the same.
Andrei
-
Andrei Pelinescu-Onciul -
jh＠tutpro.com -
Klaus Darilion -
Miklos Tirpak