Menu
▾
▴
[ naviserver-Feature Requests-1151137 ] New driver API and Udp module
|
From: SourceForge.net <no...@so...> - 2005-05-17 19:24:28
|
Feature Requests item #1151137, was opened at 2005-02-24 16:01 Message generated for change (Comment added) made by seryakov You can respond by visiting: https://sourceforge.net/tracker/?func=detail&atid=719009&aid=1151137&group_id=130646 Category: C-API Group: None Status: Open Resolution: None Priority: 5 Submitted By: Vlad Seryakov (seryakov) Assigned to: Vlad Seryakov (seryakov) Summary: New driver API and Udp module Initial Comment: Hi guys, Attached is minor driver extensions which do not change existing drivers but add new functionality. There are some cosmetic changes, like moving some fields in the Ns_Sock/Ns_Driver structres so they can be accessed publically and made some private functions public but functionality preserved as before. I included udp driver as an example of new API, and also added ns_sha1 command in the tclmisc.c, it is just one command and it is uses practically everywhere. ------------------------------------------------------------------------- To test udp driver i use new ns_udp command: ossweb:nscp 8> ns_udp send 127.0.0.1 5060 "GET / HTTP/1.0\n\n" HTTP/1.0 200 OK MIME-Version: 1.0 Date: Thu, 24 Feb 2005 05:39:50 GMT Server: NaviServer/4.0.10 Content-Type: text/html; charset=iso-8859-1 Content-Length: 661 Connection: close <HEAD><TITLE>Seryakov's Family Intranet</TITLE></HEAD> ..... ---------------------------------------------------------------------- >Comment By: Vlad Seryakov (seryakov) Date: 2005-05-17 19:24 Message: Logged In: YES user_id=184124 I am attaching new nsudp driver, it uses existing driver sock, just couple places where NS_DRIVER_UDP flag is checked for initializing and not performing accept calls. ns_udp module is simple driver like nssock. ---------------------------------------------------------------------- Comment By: Vlad Seryakov (seryakov) Date: 2005-05-04 18:01 Message: Logged In: YES user_id=184124 I implemented my SIP driver completely in separate threads and do not use Conn/Queue or jobs at all. When comparing performance with SER(sip proxy), NS still behind because of multi-threading overhead while SER forks many children listening on the same socket, same way as apache 1.x used to do. Also, NS/AS/Tcl memory allocator is slower than non-threaded allocator. My first version was using Conn/SockCallbacks but it was so slow, i could not process more than 500 requests per second. Now i can do 9000 requests per second on the same server. So, having UDP implemented in the core is not that important for me but still, using UDP as signaling/messaging protocol instead of HTTP or implementing other not that performance-needed drivers would be great. ---------------------------------------------------------------------- Comment By: Vlad Seryakov (seryakov) Date: 2005-03-18 18:17 Message: Logged In: YES user_id=184124 >I'm happy to see any of the core changed to add this >support, but it needs to be justified. I don't want to >support multiple different ways of doing the same thing. No, it is not the same thing. To build different drivers, API and tools should be flexible enough, if we have only one way to add a driver, so only drivers which fit will be(if) added, and if something usefull will come up, will we need to change the core again to fit different mehtod? Basically you offered the same thing that i am talking about: if we supply new DriverThread for UDP driver, the only common is how to queue it, public NsQueueConn. The rest is driver/thread specific. Currently, i can live with specific thread proc using sock callbacks, but when driver threads will be available, i can switch to it, but still once i read data from UDP, i need to build Sock/Conn and queue it. This is what i am offering, make NsQueueConn public plus/mines changing some other function but just to only support public NsQueueConn. IT will not conflict with future global driver changes. ---------------------------------------------------------------------- Comment By: Vlad Seryakov (seryakov) Date: 2005-03-18 14:28 Message: Logged In: YES user_id=184124 One good for having NsQueueConn public is to be able to submit connection threads from memory, without actual protocol/driver connection. I have data to spawn connection thread for and i just submit it. This way my server uses thread/connection pooling without implementing my own thread management(for example i do this now in my monitoring, but i could use conn threads instead with better control over the queue). ---------------------------------------------------------------------- Comment By: Vlad Seryakov (seryakov) Date: 2005-03-18 14:23 Message: Logged In: YES user_id=184124 Yes, it will work in the long run. Currently i do not have any problems with having only one DriverThread but having different driver threads with specific ThreadProc would be nice. It is just last time i tried new 4.1 driver, i could not work with it using Mozilla, somehow all connections got stale and mozilla was excepting more and more data from the server while server thought it sent everything already. My point: it needs more testing and even AOL did not start testing it. In the short run i am fine with ability to use sockcallbacks, i just need to be able to submit Sock to the ConnQueue, couple of small functions i added. If we can allow it for now and wait till 4.1 driver will be stable i would be okay. ---------------------------------------------------------------------- Comment By: Stephen Deasey (sdeasey) Date: 2005-03-18 13:03 Message: Logged In: YES user_id=87254 I'm not against making NsQueueConn() public per-se, I am against publishing it as a work-around to creating a good API for pluggable protocols. I'm keen to see great support for text based and binary protocols, packet based and stream based protocols, implementations which want to use a listen thread and multiple worker threads, and those which want to do all work in a single thread. I'm happy to see any of the core changed to add this support, but it needs to be justified. I don't want to support multiple different ways of doing the same thing. Maybe we should backport the changes in 4.1 to our tree before we start hacking into it? We could leave out the adp changes etc. as I believe someone said they were broken. We should make a point release before doing this. What do you think? ---------------------------------------------------------------------- Comment By: Stephen Deasey (sdeasey) Date: 2005-03-18 12:46 Message: Logged In: YES user_id=87254 Sorry, I don't have a patch to show how UDP/Unix sockets could be integrated into nssock. The 4.1 code base would seem to be a better starting point as it creates a new thread for each loaded comm driver. Here's how it could work: Unix domain socket support is easy. Use the same syntax as for the binder for the address param of the driver module in the config file. Call the correct socket listen routine depending on type within NsStartDrivers(). The nssock module should require no changes to work with this socket type. UDP sockets are harder. The 4.1 code base has changed NsStartDrivers() to spin off a thread for each loaded comm module. As the requirements for stream and packet sockets are different, a different routine would be passed to Ns_ThreadCreate() for each type. I would rename DriverThread() StreamThread() and add a new routine, PacketThread(). Much of the contents of DriverThread() can be junked for handling packet based protocols, but not all should. The new QueueWait callbacks and pre-queue filters are just as applicable here as for stream based protocols. The advantage of this is that there is a clear distinction between network transport, protocol, and application. Network transports such as SSL/TLS, SSH, etc. can be swapped in below different protocol implementations. Many applications can be built on a protocol implementation if it provides the standard hooks (registered proc, filters, QueueWait callbacks, pre-queue filteres...). Is this enough to give you an idea? It seems feasable to me, what do you think? ---------------------------------------------------------------------- Comment By: Vlad Seryakov (seryakov) Date: 2005-03-18 04:31 Message: Logged In: YES user_id=184124 Do we have any conclusions about this? Soon i will have to develop some application with support for SIP over UDP, i would like to do it on Naviserver and not on my hacked AS. ---------------------------------------------------------------------- Comment By: Vlad Seryakov (seryakov) Date: 2005-03-04 16:44 Message: Logged In: YES user_id=184124 And again, i have nothing against it, i said it before, what i am offering is a littl ebit different, for network based protocols driver stuff is kay, i just want to be able to submit conn threads from within the server, not through the main driver. If i have already request data, through shared memory, file or other way, i want to submit wroker thread directly, not by connecting to myself and submitting data using HTTP or other protocol i will have to implement just to queue te request. This is for other applications, not HTTP/Web based, more like messaging/authentication servers and alike. ---------------------------------------------------------------------- Comment By: Vlad Seryakov (seryakov) Date: 2005-03-04 14:41 Message: Logged In: YES user_id=184124 Can you provide what you have to modify in driver/nssock, i think that will not be easy and will require too many modifications. ---------------------------------------------------------------------- Comment By: Stephen Deasey (sdeasey) Date: 2005-03-04 10:08 Message: Logged In: YES user_id=87254 What do you think of my suggestion that the nssock module and nsd/driver.c could be easily modified to handle udp and unix domain sockets when specified in the config file? A change like that combined with the work you've already done to the binder would mean no development is required to handle all socket types, and would be transparent to the upper layers. ---------------------------------------------------------------------- Comment By: Vlad Seryakov (seryakov) Date: 2005-02-26 17:10 Message: Logged In: YES user_id=184124 Stephan, I am suggesting a compromise, let's put ParseProc and DriverSockQueu patches, they can coexists easily but will offer two way to add new drivers. Once we have new drivers working we will see what can be modified/adjusted, for now we do not have anything except my several drivers i use in the production and i use my own loop for smtp driver to keep processing and state machine in one place. Again, i have nothing against your method, i used to do similar in my previous versions but now i need to be completely free to implement aditional driver and i think we should have this ability. Zoran, what do you think? ---------------------------------------------------------------------- Comment By: Vlad Seryakov (seryakov) Date: 2005-02-25 22:57 Message: Logged In: YES user_id=184124 Try to implement HTTP over UDP as i provided in the patch, you will need to modify core driver again, this is one example. Next, if i want to implement small driver that do not interact with web part of aolserver, no filters, it just need to do one small thing, it should be fast but will handle very many requests. I do not need filters, callbacks, i need my own main loop. I do not understand why you are insisting that all drivers should go through http driver's main loop. If i implement dns server and web interface to it, why they should compete in the same driver's main loop and go through registered filters if they share only backend database structure? ---------------------------------------------------------------------- Comment By: Stephen Deasey (sdeasey) Date: 2005-02-25 22:22 Message: Logged In: YES user_id=87254 You don't have to use registered filters or procs. It's an option for protocol implementations where the initial accept and parsing take place in the driver thread, and processing takes place in a conn thread. Once you submit to the queue however, either a filter, a registered proc, or a registered proxy proc must handle the request. It makes no difference whether you submit explicitly with Ns_DriverSockQueue or implicitly after your Ns_ParseProc completes. How much of the request structure you fill in is up to you, it's all optional. If you choose not to use it there is no overhead. It doesn't have to be a text based protocol for this to make sense. You might fill in the request structure if you wanted to enable people to handle different request types via C or Tcl. Or, you could store binary data in Ns_Cls storage and access it from registered procs etc. In the case where you want to handle connection accept/parse/reponse processing in the same thread, nothing needs to be added. You can do that today with Ns_SockCallback (or by placing all your code in an Ns_ParseProc and returning the response from there) Can you be more specific about the overhead of using an Ns_ParseProc? Exactly what memory is allocated, what code is run that should not..? What specifically can you not do with the Ns_ParseProc interface that you need the Ns_Driver* routines for? I'm looking for concrete examples. As far as I can tell, with Ns_ParseProc you write less code, get more options, and it takes advantage of infrastructure to give you more speed. I must be missing something, but you'll have to explain it to me in more detail because I don't get it. ---------------------------------------------------------------------- Comment By: Vlad Seryakov (seryakov) Date: 2005-02-25 20:43 Message: Logged In: YES user_id=184124 Okay, the biggest problem with your way i see that i am enforced to go through http driver, even with my parse proc, i need to know how it works to implement my driver so correct hooks/filters will be called. If i want skip some parts i am not able to do it, http driver works only one way. Another issue is if something will not fit into current driver, additional hook needs to be introduced and core driver needs to be modified again. If my driver fits http-like/text mode paradigm, reusing http driver is the easiest way, but if i need something specific, all extra step to mimic the connection as http request, so all other parts of the http driver will not fail is just unnecesssary extra efforts. And i you mentioned, using Ns_DriverSockQueue is low level function and requires from the developer full attention for bulding the driver, but if this is what i want, why not. I can spawn my own thread if some callbacks are long running Tcl scripts and queue connection from my own loop. If i omly want to received packets, decode them and submit connection, i do not need all http driver infrusturcture, why to enforce using it? So we can have at least 3 ways for supporting multiprotocols: - standard callbacks, completely different threads, no connection pooling (Exists now) - Ns_DriverSockQueue, new thread or callback thread, reusing connection pooling. All filters/traces are reused as well - driver Parse proc in the http driver to reuse drivers thread and connection pooling with all filters/traces All 3 methiods can co-exists and do not interfere with each other, that's my point, and all they add just couple extra functions. I would implement all methods, i have 3 drivers i need to run: smtp, dns, snmp, can add imap in the future ---------------------------------------------------------------------- Comment By: Stephen Deasey (sdeasey) Date: 2005-02-25 20:31 Message: Logged In: YES user_id=87254 "Each way has its own drawbacks..." That's what I'm asking. What are the drawbacks for each approach we've come up with so far? ---------------------------------------------------------------------- Comment By: Vlad Seryakov (seryakov) Date: 2005-02-25 15:23 Message: Logged In: YES user_id=184124 I am not arguing with you about your ParseProc driver extension, it is very usefull and makes driver more flexible. I just want to add couple of new API functions that will allow developer to queue connections from any place, that's it. How drivers will be written and how developer will decide to handle it is up to each particular developer. But nobody will be locked up in only one way of doing it. Each way has its own drawbacks so we have multiple choices and developer will decide for itself how it should be. Both our ways do not mutually exclusive. ---------------------------------------------------------------------- Comment By: Stephen Deasey (sdeasey) Date: 2005-02-25 08:28 Message: Logged In: YES user_id=87254 Ah I see you're right, an extra thread is not created for each incoming request. 'udpThread' is perhaps not the best name for the socket callback though... :-) I still don't see the advantage of using Ns_SockCalback. A single thread is created by the server to handle all registered callbacks, including those from ns_sockcallback. At runtime, while this single thread is running tcl code to handle one callback, your dirver callback cannot run, no new connections will be accepted, etc. Why is it better to use the Ns_SockCallback thread rather than a driver thread? Maybe I'm reading this wrong, but how do you handle the case where the request arrives as more than one packet? Ns_DriverSockRead() is called from the 'udpThread' socket callback. The only return value checked is NS_OK, but couldn't this also be SOCK_MORE? How would you handle things like keepalive? I think you'd have to reimplement that in the Ns_SockCallback thread. Re the proxy stuff, Ns_RegisterRequest() and Ns_RegisterProxyRequest() seem very simillar. With Ns_RegisterRequest(), filters are run and you get the choice of using C or Tcl. Ns_RegisterProxyRequest() offers no advantage that I can see -- you still need a complete Request structure, even if you just ignore it. The comm driver initialization should probably be changed to automatically handle unix domain sockets. You add /foo rather that 127.0.0.1 in the config file and it knows to create the correct type of socket. nssock and nsopenssl wouldn't have to be modified at all. UDP is different. There are two types of protocols: single packet, such as DNS or RADIUS; multi packet, as used in some streaming media and p2p protocols. I think for the single packet case we again might want to modify the driver code to automatically handle it. No read-ahead is necessary, there's only one packet per request, so it's placed in the request buffer and passed on to the next stage, which is parsing. Here, a return code of SOCK_MORE would be illegal. Every multi-packet UDP protocol will require custom framing/sequencing and the developer will have to create a new socket driver. Taking RADIUS as an example, which is a single packet UDP protocol, you'd create a very simple Ns_ParseProc who's only job is to check that the number of bytes specified in the header actually arrived, and return SOCK_READY. A default request structure is created for you so the only other thing you have to do is Ns_RegisterRequest() for the '/' URL. Within your request proc, call Ns_ConnContent() and parse the buffer. Now, you do have a number of other options to make this more flexible. You could parse the request in your Ns_ParseProc and then fill out the request structure. e.g. the different RADIUS message types could be expressed as HTTP verbs. This buys you flexibility. Now you can Ns_RegisterRequest() a different routine for each RADIUS message type, and someone reusing your code can override your default implementation. You can also handle some message types in C, and others in Tcl. You might decide to put some useful information about the RADIUS request in the URL line. Now you get logging for free. You might decide to parse the key/value pairs from the RADIUS request into query vars or headers. Now you don't have to write a bunch of RADIUS specific Tcl commands to access the data, should you want to handle that from Tcl. ns_queryget or ns_conn headers will work. How far you want to go is up to you, the writer of the protocol module. But at the low end all you need is an Ns_ParseProc and usually a single registered request proc, which is a lot less code to write than the Ns_Driver* stuff, I think. I'm sure it's not perfect! So what's wrong with it? Why would it suck to write a RADIUS protocol module as I've just described, to take one example? ---------------------------------------------------------------------- Comment By: Vlad Seryakov (seryakov) Date: 2005-02-25 05:51 Message: Logged In: YES user_id=184124 I think we need to collect all solutions and then see where are going, i am holding to port all my old/running modules because i do not know how naviserver will handle foreign protocols. ---------------------------------------------------------------------- Comment By: Vlad Seryakov (seryakov) Date: 2005-02-25 05:48 Message: Logged In: YES user_id=184124 yes, using proxy can solve immediate requirements without hacking NsConnProc by adding hooks to call driver specific C functions. If i need my smtp server and main loop is in C, i need somehow call it in the connection thread. Using registered proxy function i can do it now, i do not need filter/traces. This is for completely new modules. I can implement main loop in the module as Tcl command and then call it in the connection filter, it is possible, it will just require many different parts to be in place and still filter should be registered as Tcl proc which will call another driver main loop. sometimes low level stuff makes things easier and simpler:-))) ---------------------------------------------------------------------- Comment By: Vlad Seryakov (seryakov) Date: 2005-02-25 05:45 Message: Logged In: YES user_id=184124 In all my installations i need sha module but it wasn't in the distribution, so i need to download it or repackage aolserver to include nssha1. That is my point, it is just one simple function use more often than something like ns_jpegsize or so. ---------------------------------------------------------------------- Comment By: Stephen Deasey (sdeasey) Date: 2005-02-25 05:42 Message: Logged In: YES user_id=87254 Well I don't know about the wisdom of adding sha1 to the core at this stage :-) I see your point though, encryption and hashing functions are almost universally required for systems/server work. Maybe we need to consider adding a new encryption module to the core distribution. Like nsdb it would export a C API via libnscypher.so (or whatever) as well as the Tcl module nscypher.so. Times have changed and things like the openssl libraries are common on all platforms so it's not the big deal it used to be to add such dependiencies. Such a module should make implementing SSL in nsopenssl easier. ---------------------------------------------------------------------- Comment By: Vlad Seryakov (seryakov) Date: 2005-02-25 05:41 Message: Logged In: YES user_id=184124 No, i do not spawn new thread, i use callback feature of the server, when socket is ready, server calls provided callback ad that callback just submits the socket to connection pool. if pool is full, Ns_DriverSockQueue will return NS_TIMEOUT, you can retry. Yes, it is low-level, for high level, HTTP driver provides a lot of functionality, it could be extended like you did, but still it is HTTP driver hacked. If i want completely new driver, like RADIUS server, http driver will not help me, i need low level stuff, and it is there already. To reuse resource limiting, i added Ns_DriverSockQueue function, so new conections can be queue instead of creating new threads. I do not have anything against your patch for extending current driver, it is very usefull. i am offering new API for low level drivers. ---------------------------------------------------------------------- Comment By: Stephen Deasey (sdeasey) Date: 2005-02-25 05:36 Message: Logged In: YES user_id=87254 The proxy stuff is too late in the cycle to do much. The request has to be fully parsed by then (for read-ahead). If a proxy function is available, then all filters, the auth phase, registered procs, cleanup procs etc. are bypassed. A lot of that stuff can be very useful for non-HTTP protocols. I don't think handling stuff via a proxy function buys you much. ---------------------------------------------------------------------- Comment By: Stephen Deasey (sdeasey) Date: 2005-02-25 05:32 Message: Logged In: YES user_id=87254 I don't understand what you're trying to acheive here (well, apart from multi-protocol... :-). The newly exposed Ns_Driver* entry points are quite low level, and so the implementor of the new protocol is left to do a lot of the heavy lifting. For example, the way I read it you have to create your own listen socket and register a callback. Every time a new reaquest comes in a thread is spawned to handle it. From that thread you then submit the parsed request to one of the conn threads. Excessive thread creation and message passing between threads is not going to perform well. And it seems you have to write more code than e.g. the example POP3 driver I posted some time ago. You're also not taking advantage of the other facilities that the server offers. What happens if 1000 connections arrive, do you spawn 1000 threads? You could of course code up some limit checks, but this already exists. What if a client sends you a continuous stream of data, 2GB... etc. By using the driver hooks to provide the new protocol parser, you deny yourself the opportunity to use something like the nsopenssl module. This should work just fine for protocols like SMTP, IMAP, POP3 and probably others. Anyway, I think one of the most carefully coded aspects of the server is it's attention to resource usage. That goes for IO, context switching, memory, etc. It's espescialy nice that most of the time you're not even aware that all this work is being done for you. I'd like new protocol drivers to be able to transparently take advantage of that. Could you take a look at my old POP3 demo driver? It's the attachment nspopd-0.3.tar.bz2 over here: http://sourceforge.net/tracker/index.php?func=detail&aid=973010&group_id=3152&atid=353152 It's not obvious that anything interesting is going on, so it's not much to look at. But actually, conn socket read-ahead is happening eficiently in one thread with async IO, the conn threads are treated as a precious resource (heavy-weight Tcl interps) and are allocated at the last minute, there's an easy API in C and Tcl to implement the actual reading of data from the INBOX (could be from the file system, db etc.). You've got a lot of experience writing servers, what do you think is wrong with this model? What can it not do, or what could it do better? ---------------------------------------------------------------------- Comment By: Vlad Seryakov (seryakov) Date: 2005-02-24 21:29 Message: Logged In: YES user_id=184124 Another thing, once we can submit connections from any place, no need to build any drivers, even in C, i can register new proxy proc and set protocol field in my request, so when submitted, connection will run registered proxy proc. for example: in my smtp driver/module, i create driver, register proxy for smtp: protocol, register callback for the socket. Once connection accepted, in my module i submit that connection to the queue with request-protocl set to smtp:. queue.c will call my proxy handler, which is C function. No need to add anything else. This way even standard aolserver can be extended without touching precious http driver thread. Sorry for sarcasm. ---------------------------------------------------------------------- You can respond by visiting: https://sourceforge.net/tracker/?func=detail&atid=719009&aid=1151137&group_id=130646 |
