Hello,
I think we need to decide about AVPs. Although not deeply investigated, ser seems to have couple of AVPs lists. In kamailio/openser it is one list.
Can a ser developer describe a bit the AVP implementation as it is now there?
In openser/kamailio we have more or less same architecture as for ser 0.9.6 with couple extensions, but I will detail here so we can have full picture: - avps are kept in shared memory - they are bound to each sip message and moved to transaction if that message create a transaction - avp can have two types of names - integer id, referred as $avp(i:number) - string id, referred as $avp(s:string) - there can be aliases for avp names, defined as "alias=[is]:id", so the script writer can use $avp(alias) - a value of an avp can be either integer or string - not related to avp core part, but important to mention here -- tm, controlled by a parameter, can make the avps from transaction (request) available to onreply_route.
Also, just to have in mind, for the future we got into discussion: - moving from linked list to a structure that allows faster search (e.g., hash table) as number of used avps can be big some times - adding support for structure-like values (the value to be another list of (name,value) pairs). A typical use case is serial forking, where one needs to store data in avps. If for one destination an avp is missing, things get screwed. Example:
$avp(uri) = {A, B, C} $avp(RPID) = {X, not needed, Y}
Internally will look like: $avp(uri) = {A, B, C} $avp(RPID) = {X, Y}
So actually, Y will be used for destination B. Workarounds are possible, or course, but maybe we can do it nicer. so it will be like: $avp(dst) = {{u=A,r=Y}, {u=B},{u=C,r=Y}}
A possibility to specify the access is $avp(d:dst=>u)
Finally, as pointed in another email, it is about making possible to hold "value null" (a mark/flag) and differentiate that from "undefined" (not existing).
Cheers, Daniel
On 12-12 00:25, Daniel-Constantin Mierla wrote:
Hello,
I think we need to decide about AVPs. Although not deeply investigated, ser seems to have couple of AVPs lists. In kamailio/openser it is one list.
Can a ser developer describe a bit the AVP implementation as it is now there?
A long time ago we started using AVPs to store all kinds of configuration information. We needed some sort of general configuration mechanism because real-world scripts were getting too complex and at that time the AVPs were readily available.
We started by loading attribute names and value from a database table into AVPs. This mechanism was/is same as in k. We would always load a set of attributes this way for a particular user (identified by username and domain from his/her SIP URI). With this mechanism in place, we were able to load a generic set of configuration options for a particular user of the server, I guess you have the same.
To make the AVPs easily accessible from the configuration script, we added support for AVP identifiers in the config script, so that you can access AVP with name foo as $foo.
As we migrated more and more configuration options to AVPs, it became clear that having just one set of AVPs was not enough. Although we could load all the config options for the caller (From URI) from the database into a set of AVPs, we could not do the same for the callee (Request-URI) at the same time. We could not do it because they would conflict with the AVPs of the caller as both users could have AVPs with same names but different values.
To get around this issue we added another list of AVPs. The new AVP list works the same way, it can contain AVPs with same names as AVPs in the original list and they do not conflict. All functions that work with lists of AVPs now take the list to work on as parameter.
To make both AVP lists available in the configuration script, we extended the syntax of AVPs identifiers so the script write can choose the AVP list to work with. SO instead of just $foo you can write either
$f.foo or $t.foo
$f refers to the original AVP list which is commonly associated with the caller. The address/uid of the caller is taken from From header, hence the 'f' in the identifier. $t refers to the AVP list which contains configuration settings of the callee. The address/uid of the calle can be taken either from the Request-URI or To headers, hence the 't' in the identifier.
The original syntax without any list identifier is still available, in other words you can still write $foo, this is defined as a shorthand for $f.foo. If you do not specify the AVP list to be used then you are referring to the list containing AVPs of the caller (From).
It also turned out that in some cases we would be having too many attributes in the database table storing AVPs. This can happen in bigger setups, having tens or hundreds of thousands users or serving multiple domains. This can slow the database down and makes SER load too much data. Based on my observations it is common that a large number of users have AVPs with same values. If ten thousands of users have attribute named 'foo' with value 'bar' then the attribute will be stored in the database ten thousand times.
As a remedy for this problem, we introduced the concept of AVP levels. The AVPs described so far and stored in user_attrs table are called user-level attributes and they store configuration specific to particular users. Then we added another two AVP lists to store so called domain-level AVPs.
Domain-level AVPs are used to store configuration information that is shared by a group of users. Domain-level AVPs are stored in a separate database table, the name of the table is domain_attrs and its contents is cached in memory by 'domain' SER module. This is an important difference, while user-level AVPs need to be loaded every time a SIP request is processed, domain-level AVPs are only loaded when the contents of domain_attrs table has changed.
The domain-level AVPs are called 'domain' because they are tied to a particular domain handled by the SIP server. That could be 'iptel.org', 'sip-router.org', and so on. This mainly useful for multi-domain setups. There are two domain-level AVP lists because it is common that SER works with two domains at the same time, the domain from From header (caller's domain) and the domain from Request-URI or To (callee's domain).
Again, we extended the syntax of AVP identifiers in the configuration file. So you can write:
$fu.foo -- this way you are asking for the value of the user-level foo AVP. $fd.foo -- This will return the value of the domain-level foo AVP.
And similarly there is $tu and $td for callee's user-level and domain-level AVPs. If you specify 'u' in the AVP identifiers then SER searches only the list of user-level attributes. If you specify 'd' then SER searches only the list of domain-level attributes.
This behavior changes if you do NOT specify the level to be searched. In that case SER searches the user-level list first and if no match is found then the domain-level list will be searched. Thus if you write:
$f.foo
then you are telling SER to search for the value of 'foo' AVP in the user-level list and continue with the domain-level list if no match is found. In other words, user-level AVPs have higher priority than domain-level AVPs. With this system you can efficiently set a configuration option for all users within a domain by creating a domain-level attribute and yet you can override that option with user-level attributes for particular users within that domain.
Then there are the global AVPs. Global AVPs can be used to store settings applicable to the whole SIP server and all domains on that server. Global AVPs are stored in global_attrs table and the contents of that table is cached in memory by gflags module. There is only one list of global AVPs. Global AVPs can be accessed with:
$g.foo
(note that there is no 'f' or 't' in the identifier). The list of global AVPs is searched after the list of domain-level AVPs. So if you write:
$f.foo
Then SER searches user-level AVPs first, then domain-level and then the global AVPs.
And finally there are also so called URI-level AVPs. They work just like user-level AVPs, but a single user can have multiple sets of uri-level AVPs. They are tied to SIP URIs of that user. Uri-level AVPs are denoted by 'r' in the AVP identifier, for example:
$fr.foo
In openser/kamailio we have more or less same architecture as for ser 0.9.6 with couple extensions, but I will detail here so we can have full picture:
- avps are kept in shared memory
In our case global and domain level AVPs are stored in shared memory.
- they are bound to each sip message and moved to transaction if that
message create a transaction
This is same.
- avp can have two types of names
- integer id, referred as $avp(i:number)
- string id, referred as $avp(s:string)
- there can be aliases for avp names, defined as "alias=[is]:id", so the
script writer can use $avp(alias)
We've kind of dropped support for integer avps, we've been using string names pretty much everywhere, although the original code still exists.
- a value of an avp can be either integer or string
- not related to avp core part, but important to mention here -- tm,
controlled by a parameter, can make the avps from transaction (request) available to onreply_route.
In SER tm restores all AVPs stored in a transaction in all route blocks called from tm, such as onreply_route, failure_route. At least I hope I remember it correctly.
Jan.
Jan Janak writes:
A long time ago we started using AVPs to store all kinds of configuration information. We needed some sort of general configuration mechanism because real-world scripts were getting too complex and at that time the AVPs were readily available.
jan,
i have been doing in kamailio exactly what you describe, but using radius.
a function that loads caller's attributes does not do anything to the avp name (integer) in db, but a function that loads callee's attributes adds a constant (say 100) to the name. in cfg file, i then define m4 macros, like this:
define(`CALLER_LANGUAGE_AVP', i:102) define(`CALLEE_LANGUAGE_AVP', i:202)
i too have domain level avps in a separate table and when avp's of a user (caller or callee) are loaded, domain level table is access first, then user level table. this means that reply attributes returned from radius are loaded into k avps in the same order. if there is only a domain level attribute, that gets used, but if there are both, the user level one that was loaded later overrides the domain one.
in addition to user and domain level, i have uri level, i.e., if an uri of a user has uri level attribute, it overrides the corresponding user level avp value.
all this has worked fine for a long time and for this purpose, i don't need any additions whatsoever to k's avp implementation. for other purposes (as daniel mentioned), it would be nice if avp value could be a structure.
-- juha
Hello,
On 12/12/08 17:22, Jan Janak wrote:
On 12-12 00:25, Daniel-Constantin Mierla wrote:
Hello,
I think we need to decide about AVPs. Although not deeply investigated, ser seems to have couple of AVPs lists. In kamailio/openser it is one list.
Can a ser developer describe a bit the AVP implementation as it is now there?
A long time ago we started using AVPs to store all kinds of configuration information. We needed some sort of general configuration mechanism because real-world scripts were getting too complex and at that time the AVPs were readily available.
We started by loading attribute names and value from a database table into AVPs. This mechanism was/is same as in k. We would always load a set of attributes this way for a particular user (identified by username and domain from his/her SIP URI). With this mechanism in place, we were able to load a generic set of configuration options for a particular user of the server, I guess you have the same.
To make the AVPs easily accessible from the configuration script, we added support for AVP identifiers in the config script, so that you can access AVP with name foo as $foo.
As we migrated more and more configuration options to AVPs, it became clear that having just one set of AVPs was not enough. Although we could load all the config options for the caller (From URI) from the database into a set of AVPs, we could not do the same for the callee (Request-URI) at the same time. We could not do it because they would conflict with the AVPs of the caller as both users could have AVPs with same names but different values.
To get around this issue we added another list of AVPs. The new AVP list works the same way, it can contain AVPs with same names as AVPs in the original list and they do not conflict. All functions that work with lists of AVPs now take the list to work on as parameter.
To make both AVP lists available in the configuration script, we extended the syntax of AVPs identifiers so the script write can choose the AVP list to work with. SO instead of just $foo you can write either
$f.foo or $t.foo
$f refers to the original AVP list which is commonly associated with the caller. The address/uid of the caller is taken from From header, hence the 'f' in the identifier. $t refers to the AVP list which contains configuration settings of the callee. The address/uid of the calle can be taken either from the Request-URI or To headers, hence the 't' in the identifier.
The original syntax without any list identifier is still available, in other words you can still write $foo, this is defined as a shorthand for $f.foo. If you do not specify the AVP list to be used then you are referring to the list containing AVPs of the caller (From).
It also turned out that in some cases we would be having too many attributes in the database table storing AVPs. This can happen in bigger setups, having tens or hundreds of thousands users or serving multiple domains. This can slow the database down and makes SER load too much data. Based on my observations it is common that a large number of users have AVPs with same values. If ten thousands of users have attribute named 'foo' with value 'bar' then the attribute will be stored in the database ten thousand times.
As a remedy for this problem, we introduced the concept of AVP levels. The AVPs described so far and stored in user_attrs table are called user-level attributes and they store configuration specific to particular users. Then we added another two AVP lists to store so called domain-level AVPs.
Domain-level AVPs are used to store configuration information that is shared by a group of users. Domain-level AVPs are stored in a separate database table, the name of the table is domain_attrs and its contents is cached in memory by 'domain' SER module. This is an important difference, while user-level AVPs need to be loaded every time a SIP request is processed, domain-level AVPs are only loaded when the contents of domain_attrs table has changed.
The domain-level AVPs are called 'domain' because they are tied to a particular domain handled by the SIP server. That could be 'iptel.org', 'sip-router.org', and so on. This mainly useful for multi-domain setups. There are two domain-level AVP lists because it is common that SER works with two domains at the same time, the domain from From header (caller's domain) and the domain from Request-URI or To (callee's domain).
Again, we extended the syntax of AVP identifiers in the configuration file. So you can write:
$fu.foo -- this way you are asking for the value of the user-level foo AVP. $fd.foo -- This will return the value of the domain-level foo AVP.
And similarly there is $tu and $td for callee's user-level and domain-level AVPs. If you specify 'u' in the AVP identifiers then SER searches only the list of user-level attributes. If you specify 'd' then SER searches only the list of domain-level attributes.
This behavior changes if you do NOT specify the level to be searched. In that case SER searches the user-level list first and if no match is found then the domain-level list will be searched. Thus if you write:
$f.foo
then you are telling SER to search for the value of 'foo' AVP in the user-level list and continue with the domain-level list if no match is found. In other words, user-level AVPs have higher priority than domain-level AVPs. With this system you can efficiently set a configuration option for all users within a domain by creating a domain-level attribute and yet you can override that option with user-level attributes for particular users within that domain.
Then there are the global AVPs. Global AVPs can be used to store settings applicable to the whole SIP server and all domains on that server. Global AVPs are stored in global_attrs table and the contents of that table is cached in memory by gflags module. There is only one list of global AVPs. Global AVPs can be accessed with:
$g.foo
(note that there is no 'f' or 't' in the identifier). The list of global AVPs is searched after the list of domain-level AVPs. So if you write:
$f.foo
Then SER searches user-level AVPs first, then domain-level and then the global AVPs.
And finally there are also so called URI-level AVPs. They work just like user-level AVPs, but a single user can have multiple sets of uri-level AVPs. They are tied to SIP URIs of that user. Uri-level AVPs are denoted by 'r' in the AVP identifier, for example:
$fr.foo
so, if I understood correctly, there are couple of combinations between the dot, with one or two letters, right?
Like:
$xy.name
where x = f|t|g and y may be missing or u|d|r
Then we can get rid of overlapping in the namespace if we enforce usage of $f.foo for $foo . In K there are couple of PV that use dot in classname (a PV can be $classname or $classname(innername)), but they are proposed to be removed.
In the same idea as in ser, thenK AVPs will map to From list.
Cheers, Daniel
In openser/kamailio we have more or less same architecture as for ser 0.9.6 with couple extensions, but I will detail here so we can have full picture:
- avps are kept in shared memory
In our case global and domain level AVPs are stored in shared memory.
- they are bound to each sip message and moved to transaction if that
message create a transaction
This is same.
- avp can have two types of names
- integer id, referred as $avp(i:number)
- string id, referred as $avp(s:string)
- there can be aliases for avp names, defined as "alias=[is]:id", so the
script writer can use $avp(alias)
We've kind of dropped support for integer avps, we've been using string names pretty much everywhere, although the original code still exists.
- a value of an avp can be either integer or string
- not related to avp core part, but important to mention here -- tm,
controlled by a parameter, can make the avps from transaction (request) available to onreply_route.
In SER tm restores all AVPs stored in a transaction in all route blocks called from tm, such as onreply_route, failure_route. At least I hope I remember it correctly.
Jan.
sr-dev mailing list sr-dev@lists.sip-router.org http://lists.sip-router.org/cgi-bin/mailman/listinfo/sr-dev
On Dec 16, 2008 at 00:36, Daniel-Constantin Mierla miconda@gmail.com wrote:
Hello,
On 12/12/08 17:22, Jan Janak wrote:
On 12-12 00:25, Daniel-Constantin Mierla wrote:
Hello,
I think we need to decide about AVPs. Although not deeply investigated, ser seems to have couple of AVPs lists. In kamailio/openser it is one list.
Can a ser developer describe a bit the AVP implementation as it is now there?
A long time ago we started using AVPs to store all kinds of configuration information. We needed some sort of general configuration mechanism because real-world scripts were getting too complex and at that time the AVPs were readily available.
We started by loading attribute names and value from a database table into AVPs. This mechanism was/is same as in k. We would always load a set of attributes this way for a particular user (identified by username and domain from his/her SIP URI). With this mechanism in place, we were able to load a generic set of configuration options for a particular user of the server, I guess you have the same.
To make the AVPs easily accessible from the configuration script, we added support for AVP identifiers in the config script, so that you can access AVP with name foo as $foo.
As we migrated more and more configuration options to AVPs, it became clear that having just one set of AVPs was not enough. Although we could load all the config options for the caller (From URI) from the database into a set of AVPs, we could not do the same for the callee (Request-URI) at the same time. We could not do it because they would conflict with the AVPs of the caller as both users could have AVPs with same names but different values.
To get around this issue we added another list of AVPs. The new AVP list works the same way, it can contain AVPs with same names as AVPs in the original list and they do not conflict. All functions that work with lists of AVPs now take the list to work on as parameter.
To make both AVP lists available in the configuration script, we extended the syntax of AVPs identifiers so the script write can choose the AVP list to work with. SO instead of just $foo you can write either
$f.foo or $t.foo
$f refers to the original AVP list which is commonly associated with the caller. The address/uid of the caller is taken from From header, hence the 'f' in the identifier. $t refers to the AVP list which contains configuration settings of the callee. The address/uid of the calle can be taken either from the Request-URI or To headers, hence the 't' in the identifier.
The original syntax without any list identifier is still available, in other words you can still write $foo, this is defined as a shorthand for $f.foo. If you do not specify the AVP list to be used then you are referring to the list containing AVPs of the caller (From).
It also turned out that in some cases we would be having too many attributes in the database table storing AVPs. This can happen in bigger setups, having tens or hundreds of thousands users or serving multiple domains. This can slow the database down and makes SER load too much data. Based on my observations it is common that a large number of users have AVPs with same values. If ten thousands of users have attribute named 'foo' with value 'bar' then the attribute will be stored in the database ten thousand times.
As a remedy for this problem, we introduced the concept of AVP levels. The AVPs described so far and stored in user_attrs table are called user-level attributes and they store configuration specific to particular users. Then we added another two AVP lists to store so called domain-level AVPs.
Domain-level AVPs are used to store configuration information that is shared by a group of users. Domain-level AVPs are stored in a separate database table, the name of the table is domain_attrs and its contents is cached in memory by 'domain' SER module. This is an important difference, while user-level AVPs need to be loaded every time a SIP request is processed, domain-level AVPs are only loaded when the contents of domain_attrs table has changed.
The domain-level AVPs are called 'domain' because they are tied to a particular domain handled by the SIP server. That could be 'iptel.org', 'sip-router.org', and so on. This mainly useful for multi-domain setups. There are two domain-level AVP lists because it is common that SER works with two domains at the same time, the domain from From header (caller's domain) and the domain from Request-URI or To (callee's domain).
Again, we extended the syntax of AVP identifiers in the configuration file. So you can write:
$fu.foo -- this way you are asking for the value of the user-level foo AVP. $fd.foo -- This will return the value of the domain-level foo AVP.
And similarly there is $tu and $td for callee's user-level and domain-level AVPs. If you specify 'u' in the AVP identifiers then SER searches only the list of user-level attributes. If you specify 'd' then SER searches only the list of domain-level attributes.
This behavior changes if you do NOT specify the level to be searched. In that case SER searches the user-level list first and if no match is found then the domain-level list will be searched. Thus if you write:
$f.foo
then you are telling SER to search for the value of 'foo' AVP in the user-level list and continue with the domain-level list if no match is found. In other words, user-level AVPs have higher priority than domain-level AVPs. With this system you can efficiently set a configuration option for all users within a domain by creating a domain-level attribute and yet you can override that option with user-level attributes for particular users within that domain.
Then there are the global AVPs. Global AVPs can be used to store settings applicable to the whole SIP server and all domains on that server. Global AVPs are stored in global_attrs table and the contents of that table is cached in memory by gflags module. There is only one list of global AVPs. Global AVPs can be accessed with:
$g.foo
(note that there is no 'f' or 't' in the identifier). The list of global AVPs is searched after the list of domain-level AVPs. So if you write:
$f.foo
Then SER searches user-level AVPs first, then domain-level and then the global AVPs.
And finally there are also so called URI-level AVPs. They work just like user-level AVPs, but a single user can have multiple sets of uri-level AVPs. They are tied to SIP URIs of that user. Uri-level AVPs are denoted by 'r' in the AVP identifier, for example:
$fr.foo
so, if I understood correctly, there are couple of combinations between the dot, with one or two letters, right?
Like:
$xy.name
where x = f|t|g and y may be missing or u|d|r
x = f|t ("track" in ser avp terminology y= r|u|d|g ("class")
g is a special case (it cannot be combined with f or t).
Here is the parse_avp_ident() usage comment:
/** parse an avp indentifier. * * Parses the following avp indentifier forms: * - "i:<number>" - old form, deprecated (e.g. i:42) * - "s:<string>" - old form, deprecated (e.g. s:foo) * - "<track>.<name>" (e.g.: f.bar) * - "<track>.<name>[<index>]" (e.g.: f.bar[1]) * - "<track><class>.<name>" (e.g: tu.bar) * - "<track><class>.<name>[<index>]" (e.g: fd.bar[2]) * - "<string>" (e.g.: foo) * Where: * <string> = ascii string * <id> = ascii string w/o '[', ']', '.' and '/' * <name> = <id> | '/' regex '/' * (Note: regex use is deprecated) * <track> = 'f' | 't' * (from or to) * <class> = 'r' | 'u' | 'd' | 'g' * (uri, user, domain or global) * <index> = <number> | '-' <number> | '' * (the avp index, if missing it means AVP_INDEX_ALL, but * it's use is deprecated) * More examples: * "fr.bar[1]" - from track, uri class, avp "bar", the value 1. * "tu./^foo/" - to track, user class, all avps for which the name * starts with foo (note RE in avp names are deprecated). * "t.did" - to track, "did" avp * * @param name - avp identifier * @param *attr - the result will be stored here * @return 0 on success, -1 on error */
Then we can get rid of overlapping in the namespace if we enforce usage of $f.foo for $foo . In K there are couple of PV that use dot in classname (a PV can be $classname or $classname(innername)), but they are proposed to be removed.
For the time being in sip-router this is handled in the following way:
$(...) -> pvar $foo(...) -> pvar $foo[...] -> avp $bar.foo -> avp (if it has a dot in it -> avp) $foo (no dot, [] or ()) -> try first to resolve it as pvar and if it fails consider it an avp.
For the current version I think this is the most compatible way. I'll use the same approach for ser module fixups (so all ser module that use core fixup functions will be able to use also pvars). One could also force kamailio or ser behaviour, by starting the script with #!KAMAILIO or #!SER.
For future versions we should rethink how everything is accessed. For example I would use $foo for script variables and some other form for avps or the current pvar $foo. Also we would have to unify selects with the select like part of the pvars. (Note for ser users: pvars in kamailio are used to access script vars, avps, message headers and something which is kind of similar to ser selects.).
In the same idea as in ser, thenK AVPs will map to From list.
[...]
Andrei
On 12/16/08 10:15, Andrei Pelinescu-Onciul wrote:
On Dec 16, 2008 at 00:36, Daniel-Constantin Mierla miconda@gmail.com wrote:
Hello,
On 12/12/08 17:22, Jan Janak wrote:
On 12-12 00:25, Daniel-Constantin Mierla wrote:
Hello,
I think we need to decide about AVPs. Although not deeply investigated, ser seems to have couple of AVPs lists. In kamailio/openser it is one list.
Can a ser developer describe a bit the AVP implementation as it is now there?
A long time ago we started using AVPs to store all kinds of configuration information. We needed some sort of general configuration mechanism because real-world scripts were getting too complex and at that time the AVPs were readily available.
We started by loading attribute names and value from a database table into AVPs. This mechanism was/is same as in k. We would always load a set of attributes this way for a particular user (identified by username and domain from his/her SIP URI). With this mechanism in place, we were able to load a generic set of configuration options for a particular user of the server, I guess you have the same.
To make the AVPs easily accessible from the configuration script, we added support for AVP identifiers in the config script, so that you can access AVP with name foo as $foo.
As we migrated more and more configuration options to AVPs, it became clear that having just one set of AVPs was not enough. Although we could load all the config options for the caller (From URI) from the database into a set of AVPs, we could not do the same for the callee (Request-URI) at the same time. We could not do it because they would conflict with the AVPs of the caller as both users could have AVPs with same names but different values.
To get around this issue we added another list of AVPs. The new AVP list works the same way, it can contain AVPs with same names as AVPs in the original list and they do not conflict. All functions that work with lists of AVPs now take the list to work on as parameter.
To make both AVP lists available in the configuration script, we extended the syntax of AVPs identifiers so the script write can choose the AVP list to work with. SO instead of just $foo you can write either
$f.foo or $t.foo
$f refers to the original AVP list which is commonly associated with the caller. The address/uid of the caller is taken from From header, hence the 'f' in the identifier. $t refers to the AVP list which contains configuration settings of the callee. The address/uid of the calle can be taken either from the Request-URI or To headers, hence the 't' in the identifier.
The original syntax without any list identifier is still available, in other words you can still write $foo, this is defined as a shorthand for $f.foo. If you do not specify the AVP list to be used then you are referring to the list containing AVPs of the caller (From).
It also turned out that in some cases we would be having too many attributes in the database table storing AVPs. This can happen in bigger setups, having tens or hundreds of thousands users or serving multiple domains. This can slow the database down and makes SER load too much data. Based on my observations it is common that a large number of users have AVPs with same values. If ten thousands of users have attribute named 'foo' with value 'bar' then the attribute will be stored in the database ten thousand times.
As a remedy for this problem, we introduced the concept of AVP levels. The AVPs described so far and stored in user_attrs table are called user-level attributes and they store configuration specific to particular users. Then we added another two AVP lists to store so called domain-level AVPs.
Domain-level AVPs are used to store configuration information that is shared by a group of users. Domain-level AVPs are stored in a separate database table, the name of the table is domain_attrs and its contents is cached in memory by 'domain' SER module. This is an important difference, while user-level AVPs need to be loaded every time a SIP request is processed, domain-level AVPs are only loaded when the contents of domain_attrs table has changed.
The domain-level AVPs are called 'domain' because they are tied to a particular domain handled by the SIP server. That could be 'iptel.org', 'sip-router.org', and so on. This mainly useful for multi-domain setups. There are two domain-level AVP lists because it is common that SER works with two domains at the same time, the domain from From header (caller's domain) and the domain from Request-URI or To (callee's domain).
Again, we extended the syntax of AVP identifiers in the configuration file. So you can write:
$fu.foo -- this way you are asking for the value of the user-level foo AVP. $fd.foo -- This will return the value of the domain-level foo AVP.
And similarly there is $tu and $td for callee's user-level and domain-level AVPs. If you specify 'u' in the AVP identifiers then SER searches only the list of user-level attributes. If you specify 'd' then SER searches only the list of domain-level attributes.
This behavior changes if you do NOT specify the level to be searched. In that case SER searches the user-level list first and if no match is found then the domain-level list will be searched. Thus if you write:
$f.foo
then you are telling SER to search for the value of 'foo' AVP in the user-level list and continue with the domain-level list if no match is found. In other words, user-level AVPs have higher priority than domain-level AVPs. With this system you can efficiently set a configuration option for all users within a domain by creating a domain-level attribute and yet you can override that option with user-level attributes for particular users within that domain.
Then there are the global AVPs. Global AVPs can be used to store settings applicable to the whole SIP server and all domains on that server. Global AVPs are stored in global_attrs table and the contents of that table is cached in memory by gflags module. There is only one list of global AVPs. Global AVPs can be accessed with:
$g.foo
(note that there is no 'f' or 't' in the identifier). The list of global AVPs is searched after the list of domain-level AVPs. So if you write:
$f.foo
Then SER searches user-level AVPs first, then domain-level and then the global AVPs.
And finally there are also so called URI-level AVPs. They work just like user-level AVPs, but a single user can have multiple sets of uri-level AVPs. They are tied to SIP URIs of that user. Uri-level AVPs are denoted by 'r' in the AVP identifier, for example:
$fr.foo
so, if I understood correctly, there are couple of combinations between the dot, with one or two letters, right?
Like:
$xy.name
where x = f|t|g and y may be missing or u|d|r
x = f|t ("track" in ser avp terminology y= r|u|d|g ("class")
g is a special case (it cannot be combined with f or t).
Here is the parse_avp_ident() usage comment:
/** parse an avp indentifier.
- Parses the following avp indentifier forms:
- "i:<number>" - old form, deprecated (e.g. i:42)- "s:<string>" - old form, deprecated (e.g. s:foo)- "<track>.<name>" (e.g.: f.bar)- "<track>.<name>[<index>]" (e.g.: f.bar[1])- "<track><class>.<name>" (e.g: tu.bar)- "<track><class>.<name>[<index>]" (e.g: fd.bar[2])- "<string>" (e.g.: foo)- Where:
<string> = ascii string<id> = ascii string w/o '[', ']', '.' and '/'<name> = <id> | '/' regex '/'(Note: regex use is deprecated)<track> = 'f' | 't'(from or to)<class> = 'r' | 'u' | 'd' | 'g'(uri, user, domain or global)<index> = <number> | '-' <number> | ''(the avp index, if missing it means AVP_INDEX_ALL, butit's use is deprecated)- More examples:
"fr.bar[1]" - from track, uri class, avp "bar", the value 1."tu./^foo/" - to track, user class, all avps for which the namestarts with foo (note RE in avp names are deprecated)."t.did" - to track, "did" avp- @param name - avp identifier
- @param *attr - the result will be stored here
- @return 0 on success, -1 on error
*/
Then we can get rid of overlapping in the namespace if we enforce usage of $f.foo for $foo . In K there are couple of PV that use dot in classname (a PV can be $classname or $classname(innername)), but they are proposed to be removed.
For the time being in sip-router this is handled in the following way:
$(...) -> pvar $foo(...) -> pvar $foo[...] -> avp $bar.foo -> avp (if it has a dot in it -> avp) $foo (no dot, [] or ()) -> try first to resolve it as pvar and if it fails consider it an avp.
For the current version I think this is the most compatible way.
yes, indeed, thanks!
I'll use the same approach for ser module fixups (so all ser module that use core fixup functions will be able to use also pvars). One could also force kamailio or ser behaviour, by starting the script with #!KAMAILIO or #!SER.
For future versions we should rethink how everything is accessed. For example I would use $foo for script variables and some other form for avps or the current pvar $foo. Also we would have to unify selects with the select like part of the pvars. (Note for ser users: pvars in kamailio are used to access script vars, avps, message headers and something which is kind of similar to ser selects.).
Perhaps select implementations can be moved in modules, as well. Integration has to be done at some point anyhow.
Cheers, Daniel
On 16-12 10:52, Daniel-Constantin Mierla wrote:
On 12/16/08 10:15, Andrei Pelinescu-Onciul wrote:
On Dec 16, 2008 at 00:36, Daniel-Constantin Mierla miconda@gmail.com wrote:
Hello,
On 12/12/08 17:22, Jan Janak wrote:
On 12-12 00:25, Daniel-Constantin Mierla wrote:
Hello,
I think we need to decide about AVPs. Although not deeply investigated, ser seems to have couple of AVPs lists. In kamailio/openser it is one list.
Can a ser developer describe a bit the AVP implementation as it is now there?
A long time ago we started using AVPs to store all kinds of configuration information. We needed some sort of general configuration mechanism because real-world scripts were getting too complex and at that time the AVPs were readily available.
We started by loading attribute names and value from a database table into AVPs. This mechanism was/is same as in k. We would always load a set of attributes this way for a particular user (identified by username and domain from his/her SIP URI). With this mechanism in place, we were able to load a generic set of configuration options for a particular user of the server, I guess you have the same.
To make the AVPs easily accessible from the configuration script, we added support for AVP identifiers in the config script, so that you can access AVP with name foo as $foo.
As we migrated more and more configuration options to AVPs, it became clear that having just one set of AVPs was not enough. Although we could load all the config options for the caller (From URI) from the database into a set of AVPs, we could not do the same for the callee (Request-URI) at the same time. We could not do it because they would conflict with the AVPs of the caller as both users could have AVPs with same names but different values.
To get around this issue we added another list of AVPs. The new AVP list works the same way, it can contain AVPs with same names as AVPs in the original list and they do not conflict. All functions that work with lists of AVPs now take the list to work on as parameter.
To make both AVP lists available in the configuration script, we extended the syntax of AVPs identifiers so the script write can choose the AVP list to work with. SO instead of just $foo you can write either
$f.foo or $t.foo
$f refers to the original AVP list which is commonly associated with the caller. The address/uid of the caller is taken from From header, hence the 'f' in the identifier. $t refers to the AVP list which contains configuration settings of the callee. The address/uid of the calle can be taken either from the Request-URI or To headers, hence the 't' in the identifier.
The original syntax without any list identifier is still available, in other words you can still write $foo, this is defined as a shorthand for $f.foo. If you do not specify the AVP list to be used then you are referring to the list containing AVPs of the caller (From).
It also turned out that in some cases we would be having too many attributes in the database table storing AVPs. This can happen in bigger setups, having tens or hundreds of thousands users or serving multiple domains. This can slow the database down and makes SER load too much data. Based on my observations it is common that a large number of users have AVPs with same values. If ten thousands of users have attribute named 'foo' with value 'bar' then the attribute will be stored in the database ten thousand times.
As a remedy for this problem, we introduced the concept of AVP levels. The AVPs described so far and stored in user_attrs table are called user-level attributes and they store configuration specific to particular users. Then we added another two AVP lists to store so called domain-level AVPs.
Domain-level AVPs are used to store configuration information that is shared by a group of users. Domain-level AVPs are stored in a separate database table, the name of the table is domain_attrs and its contents is cached in memory by 'domain' SER module. This is an important difference, while user-level AVPs need to be loaded every time a SIP request is processed, domain-level AVPs are only loaded when the contents of domain_attrs table has changed.
The domain-level AVPs are called 'domain' because they are tied to a particular domain handled by the SIP server. That could be 'iptel.org', 'sip-router.org', and so on. This mainly useful for multi-domain setups. There are two domain-level AVP lists because it is common that SER works with two domains at the same time, the domain from From header (caller's domain) and the domain from Request-URI or To (callee's domain).
Again, we extended the syntax of AVP identifiers in the configuration file. So you can write:
$fu.foo -- this way you are asking for the value of the user-level foo AVP. $fd.foo -- This will return the value of the domain-level foo AVP.
And similarly there is $tu and $td for callee's user-level and domain-level AVPs. If you specify 'u' in the AVP identifiers then SER searches only the list of user-level attributes. If you specify 'd' then SER searches only the list of domain-level attributes.
This behavior changes if you do NOT specify the level to be searched. In that case SER searches the user-level list first and if no match is found then the domain-level list will be searched. Thus if you write:
$f.foo
then you are telling SER to search for the value of 'foo' AVP in the user-level list and continue with the domain-level list if no match is found. In other words, user-level AVPs have higher priority than domain-level AVPs. With this system you can efficiently set a configuration option for all users within a domain by creating a domain-level attribute and yet you can override that option with user-level attributes for particular users within that domain.
Then there are the global AVPs. Global AVPs can be used to store settings applicable to the whole SIP server and all domains on that server. Global AVPs are stored in global_attrs table and the contents of that table is cached in memory by gflags module. There is only one list of global AVPs. Global AVPs can be accessed with:
$g.foo
(note that there is no 'f' or 't' in the identifier). The list of global AVPs is searched after the list of domain-level AVPs. So if you write:
$f.foo
Then SER searches user-level AVPs first, then domain-level and then the global AVPs.
And finally there are also so called URI-level AVPs. They work just like user-level AVPs, but a single user can have multiple sets of uri-level AVPs. They are tied to SIP URIs of that user. Uri-level AVPs are denoted by 'r' in the AVP identifier, for example:
$fr.foo
so, if I understood correctly, there are couple of combinations between the dot, with one or two letters, right?
Like:
$xy.name
where x = f|t|g and y may be missing or u|d|r
x = f|t ("track" in ser avp terminology y= r|u|d|g ("class")
g is a special case (it cannot be combined with f or t).
Here is the parse_avp_ident() usage comment:
/** parse an avp indentifier.
- Parses the following avp indentifier forms:
- "i:<number>" - old form, deprecated (e.g. i:42)- "s:<string>" - old form, deprecated (e.g. s:foo)- "<track>.<name>" (e.g.: f.bar)- "<track>.<name>[<index>]" (e.g.: f.bar[1])- "<track><class>.<name>" (e.g: tu.bar)- "<track><class>.<name>[<index>]" (e.g: fd.bar[2])- "<string>" (e.g.: foo)- Where:
<string> = ascii string<id> = ascii string w/o '[', ']', '.' and '/'<name> = <id> | '/' regex '/'(Note: regex use is deprecated)<track> = 'f' | 't'(from or to)<class> = 'r' | 'u' | 'd' | 'g'(uri, user, domain or global)<index> = <number> | '-' <number> | ''(the avp index, if missing it means AVP_INDEX_ALL, butit's use is deprecated)- More examples:
"fr.bar[1]" - from track, uri class, avp "bar", the value 1."tu./^foo/" - to track, user class, all avps for which the namestarts with foo (note RE in avp names are deprecated)."t.did" - to track, "did" avp- @param name - avp identifier
- @param *attr - the result will be stored here
- @return 0 on success, -1 on error
*/
Then we can get rid of overlapping in the namespace if we enforce usage of $f.foo for $foo . In K there are couple of PV that use dot in classname (a PV can be $classname or $classname(innername)), but they are proposed to be removed.
For the time being in sip-router this is handled in the following way:
$(...) -> pvar $foo(...) -> pvar $foo[...] -> avp $bar.foo -> avp (if it has a dot in it -> avp) $foo (no dot, [] or ()) -> try first to resolve it as pvar and if it fails consider it an avp.
For the current version I think this is the most compatible way.
yes, indeed, thanks!
I'll use the same approach for ser module fixups (so all ser module that use core fixup functions will be able to use also pvars). One could also force kamailio or ser behaviour, by starting the script with #!KAMAILIO or #!SER.
For future versions we should rethink how everything is accessed. For example I would use $foo for script variables and some other form for avps or the current pvar $foo. Also we would have to unify selects with the select like part of the pvars. (Note for ser users: pvars in kamailio are used to access script vars, avps, message headers and something which is kind of similar to ser selects.).
Perhaps select implementations can be moved in modules, as well. Integration has to be done at some point anyhow.
It already is, the core contains the core selects and shared parts.
Jan.
On Dec 16, 2008 at 14:39, Jan Janak jan@ryngle.com wrote: [...]
Perhaps select implementations can be moved in modules, as well. Integration has to be done at some point anyhow.
It already is, the core contains the core selects and shared parts.
I think Daniel meant moving also the core selects into a select module and leave only the shared part in core (like he did for pvars in k).
Andrei
On 16-12 14:49, Andrei Pelinescu-Onciul wrote:
On Dec 16, 2008 at 14:39, Jan Janak jan@ryngle.com wrote: [...]
Perhaps select implementations can be moved in modules, as well. Integration has to be done at some point anyhow.
It already is, the core contains the core selects and shared parts.
I think Daniel meant moving also the core selects into a select module and leave only the shared part in core (like he did for pvars in k).
Would this be a smart thing to do? Lots of modules use or depend on selects. Maybe we could turn it into a library if there is a need to move it away from the core, but I personally would leave it where it is now.
Jan.
On 12/16/08 16:33, Jan Janak wrote:
On 16-12 14:49, Andrei Pelinescu-Onciul wrote:
On Dec 16, 2008 at 14:39, Jan Janak jan@ryngle.com wrote: [...]
Perhaps select implementations can be moved in modules, as well. Integration has to be done at some point anyhow.
It already is, the core contains the core selects and shared parts.
I think Daniel meant moving also the core selects into a select module and leave only the shared part in core (like he did for pvars in k).
Would this be a smart thing to do? Lots of modules use or depend on selects. Maybe we could turn it into a library if there is a need to move it away from the core, but I personally would leave it where it is now.
same was in K with PVs, most of modules depend on them. But the core gets pretty fat over the time will all there. While couple of them are pretty used, some are not at all. Because there will be pseudo-variables and selects, with lot of overlapping, keeping in core just the shared part is an optimal solution.
This step can be done along with integration with PVs (removal of duplicates), perhaps second release of sip-router.
Daniel
On 16-12 16:44, Daniel-Constantin Mierla wrote:
On 12/16/08 16:33, Jan Janak wrote:
On 16-12 14:49, Andrei Pelinescu-Onciul wrote:
On Dec 16, 2008 at 14:39, Jan Janak jan@ryngle.com wrote: [...]
Perhaps select implementations can be moved in modules, as well. Integration has to be done at some point anyhow.
It already is, the core contains the core selects and shared parts.
I think Daniel meant moving also the core selects into a select module and leave only the shared part in core (like he did for pvars in k).
Would this be a smart thing to do? Lots of modules use or depend on selects. Maybe we could turn it into a library if there is a need to move it away from the core, but I personally would leave it where it is now.
same was in K with PVs, most of modules depend on them. But the core gets pretty fat over the time will all there. While couple of them are pretty used, some are not at all. Because there will be pseudo-variables and selects, with lot of overlapping, keeping in core just the shared part is an optimal solution.
This step can be done along with integration with PVs (removal of duplicates), perhaps second release of sip-router.
I don't understand. Are you suggesting to move selects somewhere else or do you want to keep it the way it is now? Note that by "selects" I was refering to the select framework, i.e. the code that is used to parse select identifiers, lookup functions to be called and calling those functions.
That code is in ser core now, along with a set of select implementation functions that didn't fit anywhere else.
Jan.
On Dec 16, 2008 at 15:58, Jan Janak jan@ryngle.com wrote:
On 16-12 16:44, Daniel-Constantin Mierla wrote:
On 12/16/08 16:33, Jan Janak wrote:
On 16-12 14:49, Andrei Pelinescu-Onciul wrote:
On Dec 16, 2008 at 14:39, Jan Janak jan@ryngle.com wrote: [...]
> Perhaps select implementations can be moved in modules, as well. Integration has to be done at some point anyhow.
It already is, the core contains the core selects and shared parts.
I think Daniel meant moving also the core selects into a select module and leave only the shared part in core (like he did for pvars in k).
Would this be a smart thing to do? Lots of modules use or depend on selects. Maybe we could turn it into a library if there is a need to move it away from the core, but I personally would leave it where it is now.
same was in K with PVs, most of modules depend on them. But the core gets pretty fat over the time will all there. While couple of them are pretty used, some are not at all. Because there will be pseudo-variables and selects, with lot of overlapping, keeping in core just the shared part is an optimal solution.
This step can be done along with integration with PVs (removal of duplicates), perhaps second release of sip-router.
I don't understand. Are you suggesting to move selects somewhere else or do you want to keep it the way it is now? Note that by "selects" I was refering to the select framework, i.e. the code that is used to parse select identifiers, lookup functions to be called and calling those functions.
That code is in ser core now, along with a set of select implementation functions that didn't fit anywhere else.
No. it's not about the framework, is about the "functions that didn't fit anywhere else" being moved to some select module in the future (from my point of view this has very very low priority).
Andrei
On 16-12 15:59, Andrei Pelinescu-Onciul wrote:
On Dec 16, 2008 at 15:58, Jan Janak jan@ryngle.com wrote:
On 16-12 16:44, Daniel-Constantin Mierla wrote:
On 12/16/08 16:33, Jan Janak wrote:
On 16-12 14:49, Andrei Pelinescu-Onciul wrote:
On Dec 16, 2008 at 14:39, Jan Janak jan@ryngle.com wrote: [...]
>> > Perhaps select implementations can be moved in modules, as well. > Integration has to be done at some point anyhow. > It already is, the core contains the core selects and shared parts.
I think Daniel meant moving also the core selects into a select module and leave only the shared part in core (like he did for pvars in k).
Would this be a smart thing to do? Lots of modules use or depend on selects. Maybe we could turn it into a library if there is a need to move it away from the core, but I personally would leave it where it is now.
same was in K with PVs, most of modules depend on them. But the core gets pretty fat over the time will all there. While couple of them are pretty used, some are not at all. Because there will be pseudo-variables and selects, with lot of overlapping, keeping in core just the shared part is an optimal solution.
This step can be done along with integration with PVs (removal of duplicates), perhaps second release of sip-router.
I don't understand. Are you suggesting to move selects somewhere else or do you want to keep it the way it is now? Note that by "selects" I was refering to the select framework, i.e. the code that is used to parse select identifiers, lookup functions to be called and calling those functions.
That code is in ser core now, along with a set of select implementation functions that didn't fit anywhere else.
No. it's not about the framework, is about the "functions that didn't fit anywhere else" being moved to some select module in the future (from my point of view this has very very low priority).
Ah, I understand now, thanks. Yes, moving those functions away from the core makes sense to me.
Jan.
-
Andrei Pelinescu-Onciul -
Daniel-Constantin Mierla -
Jan Janak -
jh＠tutpro.com