[proposal] protocol extension to support loadable stream filters

От: Brent Verner
Тема: [proposal] protocol extension to support loadable stream filters
Дата: ,
Msg-id: 20050425173230.GA4347@rcfile.org
(см: обсуждение, исходный текст)
Ответы: Re: [proposal] protocol extension to support loadable stream filters  (Tom Lane)
Список: pgsql-hackers

 I'd like to introduce the concept of (dynamically loaded) stream
filters that would be used to wrap calls to send/recv by the FE/BE 
protocol.  The initial "StreamFilter" will be a zlib compression 
filter.  Yeah, I know it could just be added along-side (in the
same way as) the SSL code, in fact, having done just that is 
precisely why I think a generic filter API is a good idea.  The 
SSL code could then be moved out into a loadable filter.  At 
first blush, this looks like it will entail the following:

 1) Define a StreamFilter struct (see below) and a base implementation    that simply wraps send/recv calls.  This base
StreamFiltersource    file will also contain a handful of utility functions to deal    with the loading/managing
 2) Add StreamFilter member to Port and PGconn, initialized with the    base StreamFilter.
 3) Add support for a new ProtocolVersion in ProcessStartupPacket      NEGOTIATE_FILTER_CODE PG_PROTOCOL(1234,5680)
Inaddition, the client would send the name of the filter    for which it is requesting support along with some filter-
 specific options...maybe a string like this,      "zlib:required=1;level=7"    where everything up to the ':' is the
filtername, and    the remainder would be optional config options.
    Q: Would this /require/ a bump of the protocol version?
    Q: Is there some chunk of existing code that I could easily       use to parse the options bit of the filter
request      string above?  I looked at PQconninfoOption, but I'm       not sure that's what I want.  All I want is for
that      options string to be parsed into a simple structure       from which I can request values by name.
 4) Add support to PQconnectPoll (and throughout client...) for     requesting and initializing a StreamFilter.
 5) Create a pg_dlsym_ptr() function that returns a void*, instead    of a PGFunction like pg_dlsym.
 6) The server will (attempt to) load the filter by name using       pg_dlopen( expand_dynamic_library_name( "sf_" +
filterName) )    or something along those lines. That library's create()    function will return a StreamFilter...
 7) add support for StreamFilter(s) to pqsecure_read/write and     secure_read/write.  If SSL were factored out as a
streamFilter,   fe-secure.c and be-secure.c could really be factored away, and    the calls to [pq]secure_read/write
wouldbe replaced by direct    calls to StreamFilter->read/write.
 Ok, I think that is a fair high-level description of what I'd
like to do.  Comments?  Questions?  I don't find much time to work
on fun stuff during the week (or even to keep up with fun stuff...),
but I do expect to have time this weekend to wrap up a working 
prototype of this.  Hopefully, I'll have time knock out the jdbc
zlib client as well.


typedef struct _StreamFilter { struct _StreamFilter* next; /* next StreamFilter.  NULL iff last filter. */ Port* port;
PGconnconn; int sock; int version; char* name; void* filterOptions;  /* type other than void* ??? */ /* filter-specific
data...ptrto some-struct-or-another */ void* filterStateData;
 // ... PostgresPollingStatusType (*connectPoll)(struct _StreamFilter*); struct _StreamFilter* (*create)(void); int
(*initialize)(struct_StreamFilter*); int (*destroy)(struct _StreamFilter*); int (*openServer)(struct _StreamFilter*
filter,Port* port); int (*openClient)(struct _StreamFilter* filter, PGconn* conn); ssize_t (*read)(struct
_StreamFilter*filter, void* ptr, size_t len); ssize_t (*write)(struct _StreamFilter* filter, void* ptr, size_t len);
int(*close)(struct _StreamFilter* filter); const char* (*getName)(struct _StreamFilter* filter); char*
(*getInfo)(struct_StreamFilter* filter);

} StreamFilter;

В списке pgsql-hackers по дате сообщения:

От: "Dave Held"
Сообщение: Re: [PERFORM] Bad n_distinct estimation; hacks suggested?
От: Christopher Kings-Lynne
Сообщение: Re: [PATCHES] Continue transactions after errors in psql