\input texinfo @c -*-texinfo-*- @c Copyright 1998 TcX AB, Detron HB and Monty Program KB @c @c %**start of header @setfilename internals.info @c We want the types in the same index @c @synindex tp fn cp @synindex cp fn @iftex @c Well this is normal in Europe. Maybe this should go into the include.texi? @afourpaper @end iftex @c Get version and other info @include include.texi @ifclear tex-debug @c This removes the black squares in the right margin @finalout @end ifclear @c Set background for HTML @set _body_tags BGCOLOR=#FFFFFF TEXT=#000000 LINK=#101090 VLINK=#7030B0 @settitle @strong{MySQL} internals Manual for version @value{mysql_version}. @setchapternewpage off @paragraphindent 0 @c %**end of header @ifinfo @format START-INFO-DIR-ENTRY * mysql-internals: (mysql-internals). @strong{MySQL} internals. END-INFO-DIR-ENTRY @end format @end ifinfo @titlepage @sp 10 @center @titlefont{@strong{MySQL} Internals Manual} @sp 10 @center Copyright @copyright{} 1998 TcX AB, Detron HB and Monty Program KB @end titlepage @node Top, Introduction, (dir), (dir) @ifinfo This is a manual about @strong{MySQL} internals. @end ifinfo @menu @end menu @node caching,,, @chapter How MySQL handles caching @strong{MySQL} has the following caches: (Note that the some of the filename have a wrong spelling of cache. :) @itemize @bullet @item Key cache A shared cache for all B-tree index blocks in the different NISAM files. Uses hashing and reverse linked lists for quick caching of the last used blocks and quick flushing of changed entries for a specific table. (@file{mysys/mf_keycash.c}) @item Record cache This is used for quick scanning of all records in a table. (@file{mysys/mf_iocash.c} and @file{isam/_cash.c}) @item Table cache This holds the last used tables. (@file{sql/sql_base.cc}) @item Hostname cache For quick lookup (with reverse name resolving). Is a must when one has a slow DNS. (@file{sql/hostname.cc}) @item Privilege cache To allow quick change between databases the last used privileges are cached for each user/database combination. (@file{sql/sql_acl.cc}) @item Heap table cache Many use of GROUP BY or DISTINCT caches all found rows in a HEAP table (this is a very quick in-memory table with hash index) @item Join row cache. For every full join in a SELECT statement (a full join here means there were no keys that one could use to find the next table in a list), the found rows are cached in a join cache. One SELECT query can use many join caches in the worst case. @end itemize @node flush tables,,, @chapter How MySQL handles flush tables @itemize @bullet @item Flush tables is handled in @code{sql/sql_base.cc::close_cached_tables()}. @item The idea of flush tables is to force all tables to be closed. This is mainly to ensure that if someone adds a new table outside of @strong{MySQL} (for example with @code{cp}) all threads will start using the new table. This will also ensure that all table changes are flushed to disk (but of course not as optimally as simple calling a sync on all tables)! @item When one does a @code{FLUSH TABLES}, the variable @code{refresh_version} will be incremented. Every time a thread releases a table it checks if the refresh version of the table (updated at open) is the same as the current refresh_version. If not it will close it and broadcast a signal on COND_refresh (to wait any thread that is waiting for all instanses of a table to be closed). @item The current @code{refresh_version} is also compared to the open @code{refresh_version} after a thread gets a lock on a table. If the refresh version is different the thread will free all locks, reopen the table and try to get the locks again; This is just to quickly get all tables to use the newest version. This is handled by @code{sql/lock.cc::mysql_lock_tables()} and @code{sql/sql_base.cc::wait_for_tables()}. @item When all tables has been closed @code{FLUSH TABLES} will return an ok to client. @item If the thread that is doing @code{FLUSH TABLES} has a lock on some tables, it will first close the locked tables, then wait until all other threads have also closed them, and then reopen them and get the locks. After this it will give other threads a chance to open the same tables. @end itemize @node Filesort,,, @chapter How MySQL does sorting (filesort) @itemize @bullet @item Read all rows according to key or by table scanning. @item Store the sort-key in a buffer (@code{sort_buffer}). @item When the buffer gets full, run a qsort on it and store the result in a temporary file. Save a pointer to the sorted block. @item Repeat the above until all rows have been read. @item Repeat the following until there is less than @code{MERGEBUFF2} (15) blocks left. @item Do a multi-merge of up to @code{MERGEBUFF} (7) regions to one block in another temporary file. Repeat until all blocks from the first file are in the second file. @item On the last multi-merge, only the pointer to the row (last part of the sort-key) is written to a result file. @item Now the code in @file{sql/records.cc} will be used to read through them in sorted order by using the row pointers in the result file. To optimize this, we read in a big block of row pointers, sort these and then we read the rows in the sorted order into a row buffer (@code{record_buffer}) . @end itemize @node Coding guidelines,,, @chapter Coding guidelines @itemize @bullet @item We are using @uref{http://www.bitkeeper.com/, BitKeeper} for source management. @item You should use the @strong{MySQL} 4.0 source for all developments. @item If you have any questions about the @strong{MySQL} source, you can post these to @email{dev-public@@mysql.com} and we will answer them. Please remember to not use this internal email list in public! @item Try to write code in a lot of black boxes that can be reused or use at least a clean, easy to change interface. @item Reuse code; There is already a lot of algorithms in MySQL for list handling, queues, dynamic and hashed arrays, sorting, etc. that can be reused. @item Use the @code{my_*} functions like @code{my_read()}/@code{my_write()}/ @code{my_malloc()} that you can find in the @code{mysys} library instead of the direct system calls; This will make your code easier to debug and more portable. @item Try to always write optimized code, so that you don't have to go back and rewrite it a couple of months later. It's better to spend 3 times as much time designing and writing an optimal function than having to do it all over again later on. @item Avoid CPU wasteful code, even where it does not matter, so that you will not develop sloppy coding habits. @item If you can write it in fewer lines, do it (as long as the code will not be slower or much harder to read). @item Don't use two commands on the same line. @item Do not check the same pointer for @code{NULL} more than once. @item Use long function and variable names in English; This makes your code easier to read. @item Use my_var as opposed to myVar or MyVar (@samp{_} rather than dancing SHIFT to seperate words in identifiers). @item Think assembly - make it easier for the compiler to optimize your code. @item Comment your code when you do something that someone else may think is not ''trivial''. @item Use @code{libstring} functions (in the strings directory) instead of standard libc string functions whenever possible. @item Avoid using @code{malloc()} (its REAL slow); For memory allocations that only need to live for the lifetime of one thread, one should use @code{sql_alloc()} instead. @item Before making big design decisions, please first post a summary of what you want to do, why you want to do it, and how you plan to do it. This way we can easily provide you with feedback and also easily discuss it thoroughly if some other developer thinks there is better way to do the same thing! @item Class names start with a capital letter. @item Structure types are @code{typedef}'ed to an all-caps identifier. @item Any @code{#define}'s are in all-caps. @item Matching @samp{@{} are in the same column. @item Put the @samp{@{} after a 'switch' on the same line @example switch (arg) { @end example Because this gives better overall indentation for the switch statement. @item In all other cases, @{ and @} should be on their own line, except if there is nothing inside @{ @}. @item Have a space after 'if' @item Put a space after ',' for function arguments @item Functions return 0 on success, and non-zero on error, so you can do: @example if(a() || b() || c()) { error("something went wrong"); } @end example @item Using @code{goto} is okay if not abused. @item Avoid default variable initalizations, use @code{LINT_INIT()} if the compiler complains after making sure that there is really no way the variable can be used uninitialized. @item Do not instantiate a class if you do not have to. @item Use pointers rather than array indexing when operating on strings. @end itemize Suggested mode in emacs: @example (load "cc-mode") (setq c-mode-common-hook '(lambda () (turn-on-font-lock) (setq comment-column 48))) (setq c-style-alist (cons '("MY" (c-basic-offset . 2) (c-comment-only-line-offset . 0) (c-offsets-alist . ((statement-block-intro . +) (knr-argdecl-intro . 0) (substatement-open . 0) (label . -) (statement-cont . +) (arglist-intro . c-lineup-arglist-intro-after-paren) (arglist-close . c-lineup-arglist) )) ) c-style-alist)) (c-set-style "MY") (setq c-default-style "MY") @end example @node mysys functions,,, @chapter mysys functions Functions i mysys: (For flags se my_sys.h) int my_copy _A((const char *from,const char *to,myf MyFlags)); - Copy file int my_delete _A((const char *name,myf MyFlags)); - Delete file int my_getwd _A((string buf,uint size,myf MyFlags)); int my_setwd _A((const char *dir,myf MyFlags)); - Get and set working directory string my_tempnam _A((const char *pfx,myf MyFlags)); - Make a uniq temp file name by using dir and adding something after pfx to make name uniq. Name is made by adding a uniq 6 length-string and TMP_EXT after pfx. Returns pointer to malloced area for filename. Should be freed by free(). File my_open _A((const char *FileName,int Flags,myf MyFlags)); File my_create _A((const char *FileName,int CreateFlags, int AccsesFlags, myf MyFlags)); int my_close _A((File Filedes,myf MyFlags)); uint my_read _A((File Filedes,byte *Buffer,uint Count,myf MyFlags)); uint my_write _A((File Filedes,const byte *Buffer,uint Count, myf MyFlags)); ulong my_seek _A((File fd,ulong pos,int whence,myf MyFlags)); ulong my_tell _A((File fd,myf MyFlags)); - Use instead of open,open-with-create-flag, close read and write to get automatic error-messages (flag: MYF_WME) and only have to test for != 0 if error (flag: MY_NABP). int my_rename _A((const char *from,const char *to,myf MyFlags)); - Rename file FILE *my_fopen _A((const char *FileName,int Flags,myf MyFlags)); FILE *my_fdopen _A((File Filedes,int Flags,myf MyFlags)); int my_fclose _A((FILE *fd,myf MyFlags)); uint my_fread _A((FILE *stream,byte *Buffer,uint Count,myf MyFlags)); uint my_fwrite _A((FILE *stream,const byte *Buffer,uint Count, myf MyFlags)); ulong my_fseek _A((FILE *stream,ulong pos,int whence,myf MyFlags)); ulong my_ftell _A((FILE *stream,myf MyFlags)); - Same read-interface for streams as for files gptr _mymalloc _A((uint uSize,const char *sFile, uint uLine, myf MyFlag)); gptr _myrealloc _A((string pPtr,uint uSize,const char *sFile, uint uLine, myf MyFlag)); void _myfree _A((gptr pPtr,const char *sFile,uint uLine)); int _sanity _A((const char *sFile,unsigned int uLine)); gptr _myget_copy_of_memory _A((const byte *from,uint length, const char *sFile, uint uLine, myf MyFlag)); - malloc(size,myflag) is mapped to this functions if not compiled with -DSAFEMALLOC void TERMINATE _A((void)); - Writes malloc-info on stdout if compiled with -DSAFEMALLOC. int my_chsize _A((File fd,ulong newlength,myf MyFlags)); - Change size of file void my_error _D((int nr,myf MyFlags, ...)); - Writes message using error number (se mysys/errors.h) on stdout or curses if MYSYS_PROGRAM_USES_CURSES() is called. void my_message _A((const char *str,myf MyFlags)); - Writes message-string on stdout or curses if MYSYS_PROGRAM_USES_CURSES() is called. void my_init _A((void )); - Start each program (in main) with this. void my_end _A((int infoflag)); - Gives info about program. - If infoflag & MY_CHECK_ERROR prints if some files are left open - If infoflag & MY_GIVE_INFO prints timing info and malloc info about prog. int my_redel _A((const char *from, const char *to, int MyFlags)); - Delete from before rename of to to from. Copyes state from old file to new file. If MY_COPY_TIME is set sets old time. int my_copystat _A((const char *from, const char *to, int MyFlags)); - Copye state from old file to new file. If MY_COPY_TIME is set sets copy also time. string my_filename _A((File fd)); - Give filename of open file. int dirname _A((string to,const char *name)); - Copy name of directory from filename. int test_if_hard_path _A((const char *dir_name)); - Test if dirname is a hard path (Starts from root) void convert_dirname _A((string name)); - Convert dirname acording to system. - In MSDOS changes all caracters to capitals and changes '/' to '\' string fn_ext _A((const char *name)); - Returns pointer to extension in filename string fn_format _A((string to,const char *name,const char *dsk, const char *form,int flag)); format a filename with replace of library and extension and converts between different systems. params to and name may be identicall function dosn't change name if name != to Flag may be: 1 force replace filnames library with 'dsk' 2 force replace extension with 'form' */ 4 force Unpack filename (replace ~ with home) 8 Pack filename as short as possibly for output to user. All open requests should allways use at least: "open(fn_format(temp_buffe,name,"","",4),...)" to unpack home and convert filename to system-form. string fn_same _A((string toname,const char *name,int flag)); - Copys directory and extension from name to toname if neaded. copy can be forced by same flags that in fn_format. int wild_compare _A((const char *str,const char *wildstr)); - Compare if str matches wildstr. Wildstr can contain "*" and "?" as match-characters. Returns 0 if match. void get_date _A((string to,int timeflag)); - Get current date in a form ready for printing. void soundex _A((string out_pntr, string in_pntr)) - Makes in_pntr to a 5 chars long string. All words that sounds alike have the same string. int init_key_cache _A((ulong use_mem,ulong leave_this_much_mem)); - Use cacheing of keys in MISAM, PISAM, and ISAM. KEY_CACHE_SIZE is a good size. - Remember to lock databases for optimal cacheing void end_key_cache _A((void)); - End key-cacheing. @node DBUG,,, @chapter The DBUG tags to use: Here is some of the tags we now use: (We should probably add a couple of new ones) "enter" Arguments to the function. "exit" Results from the function. "info" is something that may be interesting. "warning" is when something doesn't go the usual route or may be wrong. "error" when something went wrong. "loop" write in a loop, that is probably only useful when debugging the loop. These should normally be deleted when on is satisfied with the code and it has been in real use for a while. Some specific to mysqld, because we want to watch these carefully: "trans" Starting/stopping transactions. "quit" 'info' when mysqld is preparing to die. "query" Print query @node protocol,,, @chapter MySQL client/server protocol Raw packet without compression ============================== ------------------------------------------------- | Packet Length | Packet no | Data | | 3 Bytes | 1 Byte | n Bytes | ------------------------------------------------- 3 Byte packet length The length is calculated with int3store See include/global.h for details. The max packetsize can be 16 MB. 1 Byte packet no If no compression is used the first 4 bytes of each paket is the header of the paket. The packet number is incremented for each sent packet. The first packet starts with 0 n Byte data The packet length can be recalculated with: length = byte1 + (256 * byte2) + (256 * 256 * byte3) Raw packet with compression =========================== ----------------------------------------------------- | Packet Length | Packet no | Uncomp. Packet Length | | 3 Bytes | 1 Byte | 3 Bytes | ----------------------------------------------------- 3 Byte packet length The length is calculated with int3store See include/global.h for details. The max packetsize can be 16 MB. 1 Byte packet no 3 Byte uncompressed packet length If compression is used the first 7 bytes of each paket is the header of the paket. Basic packets ============== OK-packet For details see sql/net_pkg.cc function send_ok ------------------------------------------------- | Header | No of Rows | Affected Rows | | | 1 Byte | 1-8 Byte | ------------------------------------------------- | ID (last_insert_id) | Status | Length | | 1-8 Byte | 2 Byte | 1-8 Byte | ------------------------------------------------- | Messagetext | | n Byte | ------------------------------------------------- Header 1 byte number of rows ? (always 0 ?) 1-8 bytes affected rows 1-8 byte id (last_insert_id) 2 byte Status (usually 0) If the OK-packege includes a message: 1-8 bytes length of message n bytes messagetext Error-packet ------------------------------------------------- | Header | Statuscode | Error no | | | 1 Byte | 2 Byte | ------------------------------------------------- | Messagetext | 0x00 | | n Byte | 1 Byte | ------------------------------------------------- Header 1 byte status code (0xFF = ERROR) 2 byte error number (is only sent to new 3.23 clients. n byte errortext 1 byte 0x00 The communication ================= > Packet from server to client < Paket from client tor server Login ------ > 1. packet Header 1 byte protocolversion n byte serverversion 1 byte 0x00 4 byte threadnumber 8 byte crypt seed 1 byte 0x00 2 byte CLIENT_xxx options (see include/mysql_com.h that is supported by the server 1 byte number of current server charset 2 byte server status variables (SERVER_STATUS_xxx flags) 13 byte 0x00 (not used yet). < 2. packet Header 2 byte CLIENT_xxx options 3 byte max_allowed_packet for the client n byte username 1 byte 0x00 8 byte crypted password 1 byte 0x00 n byte databasename 1 byte 0x00 > 3. packet OK-packet Command -------- < 1. packet Header 1 byte command type (e.g.0x03 = query) n byte query Result set (after command) -------------------------- > 2. packet Header 1-8 byte field_count (packed with net_store_length()) If field_count == 0 (command): 1-8 byte affected rows 1-8 byte insert id 2 bytes server_status (SERVER_STATUS_xx) If field_count == NULL_LENGTH (251) LOAD DATA LOCAL INFILE If field_count > 0 Result Set: > n packets Header Info Column description: 5 data object /column (See code in unpack_fields()) Columninfo for each column: 1 data block table_name 1 byte length of block n byte data 1 data block field_name 1 byte length of block... n byte data 1 data block display length of field 1 byte length of block 3 bytes display length of filed 1 data block type field of type (enum_field_types) 1 byte length of block 1 bytexs field of type 1 data block flags 1 byte length of block 2 byte flags for the columns (NOT_NULL_FLAG, ZEROFILL_FLAG....) 1 byte decimals if table definition: 1 data block default value Actual result (one packet per row): 4 byte header 1-8 byte length of data n data Fieldtype Codes: ================ display_length |enum_field_type |flags ---------------------------------------------------- Blob 03 FF FF 00 |01 FC |03 90 00 00 Mediumblob 03 FF FF FF |01 FC |03 90 00 00 Tinyblob 03 FF 00 00 |01 FC |03 90 00 00 Text 03 FF FF 00 |01 FC |03 10 00 00 Mediumtext 03 FF FF FF |01 FC |03 10 00 00 Tinytext 03 FF 00 00 |01 FC |03 10 00 00 Integer 03 0B 00 00 |01 03 |03 03 42 00 Mediumint 03 09 00 00 |01 09 |03 00 00 00 Smallint 03 06 00 00 |01 02 |03 00 00 00 Tinyint 03 04 00 00 |01 01 |03 00 00 00 Varchar 03 XX 00 00 |01 FD |03 00 00 00 Enum 03 05 00 00 |01 FE |03 00 01 00 Datetime 03 13 00 00 |01 0C |03 00 00 00 Timestamp 03 0E 00 00 |01 07 |03 61 04 00 Time 03 08 00 00 |01 0B |03 00 00 00 Date 03 0A 00 00 |01 0A |03 00 00 00 @c The Index was empty, and ugly, so I removed it. (jcole, Sep 7, 2000) @c @node Index @c @unnumbered Index @c @printindex fn @summarycontents @contents @bye