Filter: lots of documentation

This commit is contained in:
Maria Matejka 2019-07-15 13:19:01 +02:00
parent 1b9db6d4a7
commit 0da06b7103
3 changed files with 156 additions and 49 deletions

View file

@ -7,7 +7,42 @@
* *
* Can be freely distributed and used under the terms of the GNU GPL. * Can be freely distributed and used under the terms of the GNU GPL.
* *
* Filter instructions. You shall define your instruction only here * The filter code goes through several phases:
*
* 1 Parsing
* Flex- and Bison-generated parser decodes the human-readable data into
* a struct f_inst tree. This is an infix tree that was interpreted by
* depth-first search execution in previous versions of the interpreter.
* All instructions have their constructor: f_new_inst(FI_EXAMPLE, ...)
* translates into f_new_inst_FI_EXAMPLE(...) and the types are checked in
* compile time. If the result of the instruction is always the same,
* it's reduced to FI_CONSTANT directly in constructor. This phase also
* counts how many instructions are underlying in means of f_line_item
* fields to know how much we have to allocate in the next phase.
*
* 2 Linearize before interpreting
* The infix tree is always interpreted in the same order. Therefore we
* sort the instructions one after another into struct f_line. Results
* and arguments of these instructions are implicitly put on a value
* stack; e.g. the + operation just takes two arguments from the value
* stack and puts the result on there.
*
* 3 Interpret
* The given line is put on a custom execution stack. If needed (FI_CALL,
* FI_SWITCH, FI_AND, FI_OR, FI_CONDITION, ...), another line is put on top
* of the stack; when that line finishes, the execution continues on the
* older lines on the stack where it stopped before.
*
* 4 Same
* On config reload, the filters have to be compared whether channel
* reload is needed or not. The comparison is done by comparing the
* struct f_line's recursively.
*
* The main purpose of this rework was to improve filter performance
* by making the interpreter non-recursive.
*
* The other outcome is concentration of instruction definitions to
* one place -- right here. You shall define your instruction only here
* and nowhere else. * and nowhere else.
* *
* Beware. This file is interpreted by M4 macros. These macros * Beware. This file is interpreted by M4 macros. These macros
@ -48,11 +83,122 @@
* m4_dnl RESULT_VOID; return undef * m4_dnl RESULT_VOID; return undef
* m4_dnl } * m4_dnl }
* *
* Also note that the { ... } blocks are not respected by M4 at all.
* If you get weird unmatched-brace-pair errors, check what it generated and why.
* What is really considered as one instruction is not the { ... } block
* after m4_dnl INST() but all the code between them.
*
* Other code is just copied into the interpreter part. * Other code is just copied into the interpreter part.
* *
* If you want to write something really special, see FI_CALL * If you are satisfied with this, you don't need to read the following
* or FI_CONSTANT or whatever else to see how to use the FID_* * detailed description of what is really done with the instruction definitions.
* macros. *
* m4_dnl Now let's look under the cover. The code between each INST()
* m4_dnl is copied to several places, namely these (numbered by the M4 diversions
* m4_dnl used in filter/decl.m4):
*
* m4_dnl (102) struct f_inst *f_new_inst(FI_EXAMPLE [[ put it here ]])
* m4_dnl {
* m4_dnl ... (common code)
* m4_dnl (103) [[ put it here ]]
* m4_dnl ...
* m4_dnl if (all arguments are constant)
* m4_dnl (108) [[ put it here ]]
* m4_dnl }
* m4_dnl For writing directly to constructor argument list, use FID_NEW_ARGS.
* m4_dnl For computing something in constructor (103), use FID_NEW_BODY.
* m4_dnl For constant pre-interpretation (108), see below at FID_INTERPRET_BODY.
*
* m4_dnl struct f_inst {
* m4_dnl ... (common fields)
* m4_dnl union {
* m4_dnl struct {
* m4_dnl (101) [[ put it here ]]
* m4_dnl } i_FI_EXAMPLE;
* m4_dnl ...
* m4_dnl };
* m4_dnl };
* m4_dnl This structure is returned from constructor.
* m4_dnl For writing directly to this structure, use FID_STRUCT_IN.
*
* m4_dnl linearize(struct f_line *dest, const struct f_inst *what, uint pos) {
* m4_dnl ...
* m4_dnl switch (what->fi_code) {
* m4_dnl case FI_EXAMPLE:
* m4_dnl (105) [[ put it here ]]
* m4_dnl break;
* m4_dnl }
* m4_dnl }
* m4_dnl This is called when translating from struct f_inst to struct f_line_item.
* m4_dnl For accessing your custom instruction data, use following macros:
* m4_dnl whati -> for accessing (struct f_inst).i_FI_EXAMPLE
* m4_dnl item -> for accessing (struct f_line)[pos].i_FI_EXAMPLE
* m4_dnl For writing directly here, use FID_LINEARIZE_BODY.
*
* m4_dnl (107) struct f_line_item {
* m4_dnl ... (common fields)
* m4_dnl union {
* m4_dnl struct {
* m4_dnl (101) [[ put it here ]]
* m4_dnl } i_FI_EXAMPLE;
* m4_dnl ...
* m4_dnl };
* m4_dnl };
* m4_dnl The same as FID_STRUCT_IN (101) but for the other structure.
* m4_dnl This structure is returned from the linearizer (105).
* m4_dnl For writing directly to this structure, use FID_LINE_IN.
*
* m4_dnl f_dump_line_item_FI_EXAMPLE(const struct f_line_item *item, const int indent)
* m4_dnl {
* m4_dnl (104) [[ put it here ]]
* m4_dnl }
* m4_dnl This code dumps the instruction on debug. Note that the argument
* m4_dnl is the linearized instruction; if the instruction has arguments,
* m4_dnl their code has already been linearized and their value is taken
* m4_dnl from the value stack.
* m4_dnl For writing directly here, use FID_DUMP_BODY.
*
* m4_dnl f_same(...)
* m4_dnl {
* m4_dnl switch (f1_->fi_code) {
* m4_dnl case FI_EXAMPLE:
* m4_dnl (106) [[ put it here ]]
* m4_dnl break;
* m4_dnl }
* m4_dnl }
* m4_dnl This code compares the two given instrucions (f1_ and f2_)
* m4_dnl on reconfigure. For accessing your custom instruction data,
* m4_dnl use macros f1 and f2.
* m4_dnl For writing directly here, use FID_SAME_BODY.
*
* m4_dnl interpret(...)
* m4_dnl {
* m4_dnl switch (what->fi_code) {
* m4_dnl case FI_EXAMPLE:
* m4_dnl (108) [[ put it here ]]
* m4_dnl break;
* m4_dnl }
* m4_dnl }
* m4_dnl This code executes the instruction. Every pre-defined macro
* m4_dnl resets the output here. For setting it explicitly,
* m4_dnl use FID_INTERPRET_BODY.
* m4_dnl This code is put on two places; one is the interpreter, the other
* m4_dnl is instruction constructor. If you need to distinguish between
* m4_dnl these two, use FID_INTERPRET_EXEC or FID_INTERPRET_NEW respectively.
* m4_dnl To address the difference between interpreter and constructor
* m4_dnl environments, there are several convenience macros defined:
* m4_dnl runtime() -> for spitting out runtime error like division by zero
* m4_dnl RESULT(...) -> declare result; may overwrite arguments
* m4_dnl v1, v2, v3 -> positional arguments, may be overwritten by RESULT()
* m4_dnl falloc(size) -> allocate memory from the appropriate linpool
* m4_dnl fpool -> the current linpool
* m4_dnl NEVER_CONSTANT-> don't generate pre-interpretation code at all
* m4_dnl ACCESS_RTE -> check that route is available, also NEVER_CONSTANT
* m4_dnl ACCESS_EATTRS -> pre-cache the eattrs; use only with ACCESS_RTE
* m4_dnl f_rta_cow(fs) -> function to call before any change to route should be done
*
* m4_dnl If you are stymied, see FI_CALL or FI_CONSTANT or just search for
* m4_dnl the mentioned macros in this file to see what is happening there in wild.
*/ */
/* Binary operators */ /* Binary operators */

View file

@ -7,39 +7,7 @@
* Can be freely distributed and used under the terms of the GNU GPL. * Can be freely distributed and used under the terms of the GNU GPL.
* *
* Filter interpreter data structures and internal API. * Filter interpreter data structures and internal API.
* The filter code goes through several phases: * See filter/f-inst.c for documentation.
*
* 1 Parsing
* Flex- and Bison-generated parser decodes the human-readable data into
* a struct f_inst tree. This is an infix tree that was interpreted by
* depth-first search execution in previous versions of the interpreter.
* All instructions have their constructor: f_new_inst(FI_code, ...)
* translates into f_new_inst_FI_code(...) and the types are checked in
* compile time.
*
* 2 Linearize before interpreting
* The infix tree is always interpreted in the same order. Therefore we
* sort the instructions one after another into struct f_line. Results
* and arguments of these instructions are implicitly put on a value
* stack; e.g. the + operation just takes two arguments from the value
* stack and puts the result on there.
*
* 3 Interpret
* The given line is put on a custom execution stack. If needed (FI_CALL,
* FI_SWITCH, FI_AND, FI_OR, FI_CONDITION, ...), another line is put on top
* of the stack; when that line finishes, the execution continues on the
* older lines on the stack where it stopped before.
*
* 4 Same
* On config reload, the filters have to be compared whether channel
* reload is needed or not. The comparison is done by comparing the
* struct f_line's recursively.
*
* The main purpose of this rework was to improve filter performance
* by making the interpreter non-recursive.
*
* The other outcome is concentration of instruction definitions to
* one place -- filter/f-inst.c
*/ */
#ifndef _BIRD_F_INST_H_ #ifndef _BIRD_F_INST_H_

View file

@ -15,20 +15,13 @@
* the source from user into a tree of &f_inst structures. These trees are * the source from user into a tree of &f_inst structures. These trees are
* later interpreted using code in |filter/filter.c|. * later interpreted using code in |filter/filter.c|.
* *
* A filter is represented by a tree of &f_inst structures, one structure per * A filter is represented by a tree of &f_inst structures, later translated
* "instruction". Each &f_inst contains @code, @aux value which is * into lists called &f_line. All the instructions are defined and documented
* usually the data type this instruction operates on and two generic * in |filter/f-inst.c| definition file.
* arguments (@a[0], @a[1]). Some instructions contain pointer(s) to other
* instructions in their (@a[0], @a[1]) fields.
* *
* Filters use a &f_val structure for their data. Each &f_val * Filters use a &f_val structure for their data. Each &f_val
* contains type and value (types are constants prefixed with %T_). Few * contains type and value (types are constants prefixed with %T_).
* of the types are special; %T_RETURN can be or-ed with a type to indicate * Look into |filter/data.h| for more information and appropriate calls.
* that return from a function or from the whole filter should be
* forced. Important thing about &f_val's is that they may be copied
* with a simple |=|. That's fine for all currently defined types: strings
* are read-only (and therefore okay), paths are copied for each
* operation (okay too).
*/ */
#undef LOCAL_DEBUG #undef LOCAL_DEBUG