1 #LyX 1.6.7 created this file. For more info see http://www.lyx.org/
6 \use_default_options true
11 \font_typewriter default
12 \font_default_family default
19 \paperfontsize default
27 \paperorientation portrait
30 \paragraph_separation indent
32 \quotes_language english
35 \paperpagestyle default
36 \tracking_changes true
38 \author "Rusty Russell,,,"
45 TDB2: A Redesigning The Trivial DataBase
49 Rusty Russell, IBM Corporation
56 \begin_layout Abstract
57 The Trivial DataBase on-disk format is 32 bits; with usage cases heading
58 towards the 4G limit, that must change.
59 This required breakage provides an opportunity to revisit TDB's other design
60 decisions and reassess them.
67 \begin_layout Standard
68 The Trivial DataBase was originally written by Andrew Tridgell as a simple
69 key/data pair storage system with the same API as dbm, but allowing multiple
70 readers and writers while being small enough (< 1000 lines of C) to include
72 The simple design created in 1999 has proven surprisingly robust and performant
73 , used in Samba versions 3 and 4 as well as numerous other projects.
74 Its useful life was greatly increased by the (backwards-compatible!) addition
75 of transaction support in 2005.
78 \begin_layout Standard
79 The wider variety and greater demands of TDB-using code has lead to some
80 organic growth of the API, as well as some compromises on the implementation.
81 None of these, by themselves, are seen as show-stoppers, but the cumulative
82 effect is to a loss of elegance over the initial, simple TDB implementation.
83 Here is a table of the approximate number of lines of implementation code
84 and number of API functions at the end of each year:
87 \begin_layout Standard
89 <lyxtabular version="3" rows="12" columns="3">
91 <column alignment="center" valignment="top" width="0">
92 <column alignment="center" valignment="top" width="0">
93 <column alignment="center" valignment="top" width="0">
95 <cell alignment="center" valignment="top" topline="true" bottomline="true" leftline="true" usebox="none">
98 \begin_layout Plain Layout
104 <cell alignment="center" valignment="top" topline="true" bottomline="true" leftline="true" usebox="none">
107 \begin_layout Plain Layout
113 <cell alignment="center" valignment="top" topline="true" bottomline="true" leftline="true" rightline="true" usebox="none">
116 \begin_layout Plain Layout
117 Lines of C Code Implementation
124 <cell alignment="center" valignment="top" topline="true" leftline="true" usebox="none">
127 \begin_layout Plain Layout
133 <cell alignment="center" valignment="top" topline="true" leftline="true" usebox="none">
136 \begin_layout Plain Layout
142 <cell alignment="center" valignment="top" topline="true" leftline="true" rightline="true" usebox="none">
145 \begin_layout Plain Layout
153 <cell alignment="center" valignment="top" topline="true" leftline="true" usebox="none">
156 \begin_layout Plain Layout
162 <cell alignment="center" valignment="top" topline="true" leftline="true" usebox="none">
165 \begin_layout Plain Layout
171 <cell alignment="center" valignment="top" topline="true" leftline="true" rightline="true" usebox="none">
174 \begin_layout Plain Layout
182 <cell alignment="center" valignment="top" topline="true" leftline="true" usebox="none">
185 \begin_layout Plain Layout
191 <cell alignment="center" valignment="top" topline="true" leftline="true" usebox="none">
194 \begin_layout Plain Layout
200 <cell alignment="center" valignment="top" topline="true" leftline="true" rightline="true" usebox="none">
203 \begin_layout Plain Layout
211 <cell alignment="center" valignment="top" topline="true" leftline="true" usebox="none">
214 \begin_layout Plain Layout
220 <cell alignment="center" valignment="top" topline="true" leftline="true" usebox="none">
223 \begin_layout Plain Layout
229 <cell alignment="center" valignment="top" topline="true" leftline="true" rightline="true" usebox="none">
232 \begin_layout Plain Layout
240 <cell alignment="center" valignment="top" topline="true" leftline="true" usebox="none">
243 \begin_layout Plain Layout
249 <cell alignment="center" valignment="top" topline="true" leftline="true" usebox="none">
252 \begin_layout Plain Layout
258 <cell alignment="center" valignment="top" topline="true" leftline="true" rightline="true" usebox="none">
261 \begin_layout Plain Layout
269 <cell alignment="center" valignment="top" topline="true" leftline="true" usebox="none">
272 \begin_layout Plain Layout
278 <cell alignment="center" valignment="top" topline="true" leftline="true" usebox="none">
281 \begin_layout Plain Layout
287 <cell alignment="center" valignment="top" topline="true" leftline="true" rightline="true" usebox="none">
290 \begin_layout Plain Layout
298 <cell alignment="center" valignment="top" topline="true" leftline="true" usebox="none">
301 \begin_layout Plain Layout
307 <cell alignment="center" valignment="top" topline="true" leftline="true" usebox="none">
310 \begin_layout Plain Layout
316 <cell alignment="center" valignment="top" topline="true" leftline="true" rightline="true" usebox="none">
319 \begin_layout Plain Layout
327 <cell alignment="center" valignment="top" topline="true" leftline="true" usebox="none">
330 \begin_layout Plain Layout
336 <cell alignment="center" valignment="top" topline="true" leftline="true" usebox="none">
339 \begin_layout Plain Layout
345 <cell alignment="center" valignment="top" topline="true" leftline="true" rightline="true" usebox="none">
348 \begin_layout Plain Layout
356 <cell alignment="center" valignment="top" topline="true" leftline="true" usebox="none">
359 \begin_layout Plain Layout
365 <cell alignment="center" valignment="top" topline="true" leftline="true" usebox="none">
368 \begin_layout Plain Layout
374 <cell alignment="center" valignment="top" topline="true" leftline="true" rightline="true" usebox="none">
377 \begin_layout Plain Layout
385 <cell alignment="center" valignment="top" topline="true" leftline="true" usebox="none">
388 \begin_layout Plain Layout
394 <cell alignment="center" valignment="top" topline="true" leftline="true" usebox="none">
397 \begin_layout Plain Layout
403 <cell alignment="center" valignment="top" topline="true" leftline="true" rightline="true" usebox="none">
406 \begin_layout Plain Layout
414 <cell alignment="center" valignment="top" topline="true" bottomline="true" leftline="true" usebox="none">
417 \begin_layout Plain Layout
423 <cell alignment="center" valignment="top" topline="true" bottomline="true" leftline="true" usebox="none">
426 \begin_layout Plain Layout
432 <cell alignment="center" valignment="top" topline="true" bottomline="true" leftline="true" rightline="true" usebox="none">
435 \begin_layout Plain Layout
449 \begin_layout Standard
450 This review is an attempt to catalog and address all the known issues with
451 TDB and create solutions which address the problems without significantly
452 increasing complexity; all involved are far too aware of the dangers of
453 second system syndrome in rewriting a successful project like this.
456 \begin_layout Section
460 \begin_layout Subsection
461 tdb_open_ex Is Not Expandable
464 \begin_layout Standard
465 The tdb_open() call was expanded to tdb_open_ex(), which added an optional
466 hashing function and an optional logging function argument.
467 Additional arguments to open would require the introduction of a tdb_open_ex2
471 \begin_layout Subsubsection
473 \begin_inset CommandInset label
482 \begin_layout Standard
483 tdb_open() will take a linked-list of attributes:
486 \begin_layout LyX-Code
490 \begin_layout LyX-Code
491 TDB_ATTRIBUTE_LOG = 0,
494 \begin_layout LyX-Code
495 TDB_ATTRIBUTE_HASH = 1
498 \begin_layout LyX-Code
502 \begin_layout LyX-Code
503 struct tdb_attribute_base {
506 \begin_layout LyX-Code
507 enum tdb_attribute attr;
510 \begin_layout LyX-Code
511 union tdb_attribute *next;
514 \begin_layout LyX-Code
518 \begin_layout LyX-Code
519 struct tdb_attribute_log {
522 \begin_layout LyX-Code
523 struct tdb_attribute_base base; /* .attr = TDB_ATTRIBUTE_LOG */
526 \begin_layout LyX-Code
530 \begin_layout LyX-Code
534 \begin_layout LyX-Code
538 \begin_layout LyX-Code
539 struct tdb_attribute_hash {
542 \begin_layout LyX-Code
543 struct tdb_attribute_base base; /* .attr = TDB_ATTRIBUTE_HASH */
546 \begin_layout LyX-Code
547 tdb_hash_func hash_fn;
550 \begin_layout LyX-Code
554 \begin_layout LyX-Code
558 \begin_layout LyX-Code
559 union tdb_attribute {
562 \begin_layout LyX-Code
563 struct tdb_attribute_base base;
566 \begin_layout LyX-Code
567 struct tdb_attribute_log log;
570 \begin_layout LyX-Code
571 struct tdb_attribute_hash hash;
574 \begin_layout LyX-Code
578 \begin_layout Standard
579 This allows future attributes to be added, even if this expands the size
583 \begin_layout Subsubsection
587 \begin_layout Standard
591 \begin_layout Subsection
592 tdb_traverse Makes Impossible Guarantees
595 \begin_layout Standard
596 tdb_traverse (and tdb_firstkey/tdb_nextkey) predate transactions, and it
597 was thought that it was important to guarantee that all records which exist
598 at the start and end of the traversal would be included, and no record
599 would be included twice.
602 \begin_layout Standard
603 This adds complexity (see
604 \begin_inset CommandInset ref
606 reference "Reliable-Traversal-Adds"
610 ) and does not work anyway for records which are altered (in particular,
611 those which are expanded may be effectively deleted and re-added behind
615 \begin_layout Subsubsection
616 \begin_inset CommandInset label
618 name "traverse-Proposed-Solution"
625 \begin_layout Standard
626 Abandon the guarantee.
627 You will see every record if no changes occur during your traversal, otherwise
628 you will see some subset.
629 You can prevent changes by using a transaction or the locking API.
632 \begin_layout Subsubsection
636 \begin_layout Standard
638 Delete-during-traverse will still delete every record, too (assuming no
642 \begin_layout Subsection
643 Nesting of Transactions Is Fraught
646 \begin_layout Standard
647 TDB has alternated between allowing nested transactions and not allowing
649 Various paths in the Samba codebase assume that transactions will nest,
650 and in a sense they can: the operation is only committed to disk when the
651 outer transaction is committed.
652 There are two problems, however:
655 \begin_layout Enumerate
656 Canceling the inner transaction will cause the outer transaction commit
657 to fail, and will not undo any operations since the inner transaction began.
658 This problem is soluble with some additional internal code.
661 \begin_layout Enumerate
662 An inner transaction commit can be cancelled by the outer transaction.
663 This is desirable in the way which Samba's database initialization code
664 uses transactions, but could be a surprise to any users expecting a successful
665 transaction commit to expose changes to others.
668 \begin_layout Standard
669 The current solution is to specify the behavior at tdb_open(), with the
670 default currently that nested transactions are allowed.
671 This flag can also be changed at runtime.
674 \begin_layout Subsubsection
678 \begin_layout Standard
679 Given the usage patterns, it seems that the
680 \begin_inset Quotes eld
684 \begin_inset Quotes erd
687 behavior of disallowing nested transactions should become the default.
688 Additionally, it seems the outer transaction is the only code which knows
689 whether inner transactions should be allowed, so a flag to indicate this
690 could be added to tdb_transaction_start.
691 However, this behavior can be simulated with a wrapper which uses tdb_add_flags
692 () and tdb_remove_flags(), so the API should not be expanded for this relatively
696 \begin_layout Subsubsection
700 \begin_layout Standard
701 Incomplete; nesting flag is still defined as per tdb1.
704 \begin_layout Subsection
705 Incorrect Hash Function is Not Detected
708 \begin_layout Standard
709 tdb_open_ex() allows the calling code to specify a different hash function
710 to use, but does not check that all other processes accessing this tdb
711 are using the same hash function.
712 The result is that records are missing from tdb_fetch().
715 \begin_layout Subsubsection
719 \begin_layout Standard
720 The header should contain an example hash result (eg.
721 the hash of 0xdeadbeef), and tdb_open_ex() should check that the given
722 hash function produces the same answer, or fail the tdb_open call.
725 \begin_layout Subsubsection
729 \begin_layout Standard
733 \begin_layout Subsection
734 tdb_set_max_dead/TDB_VOLATILE Expose Implementation
737 \begin_layout Standard
738 In response to scalability issues with the free list (
739 \begin_inset CommandInset ref
741 reference "TDB-Freelist-Is"
745 ) two API workarounds have been incorporated in TDB: tdb_set_max_dead()
746 and the TDB_VOLATILE flag to tdb_open.
747 The latter actually calls the former with an argument of
748 \begin_inset Quotes eld
752 \begin_inset Quotes erd
758 \begin_layout Standard
759 This code allows deleted records to accumulate without putting them in the
761 On delete we iterate through each chain and free them in a batch if there
762 are more than max_dead entries.
763 These are never otherwise recycled except as a side-effect of a tdb_repack.
766 \begin_layout Subsubsection
770 \begin_layout Standard
771 With the scalability problems of the freelist solved, this API can be removed.
772 The TDB_VOLATILE flag may still be useful as a hint that store and delete
773 of records will be at least as common as fetch in order to allow some internal
774 tuning, but initially will become a no-op.
777 \begin_layout Subsubsection
781 \begin_layout Standard
783 TDB_VOLATILE still defined, but implementation should fail on unknown flags
787 \begin_layout Subsection
788 \begin_inset CommandInset label
790 name "TDB-Files-Cannot"
794 TDB Files Cannot Be Opened Multiple Times In The Same Process
797 \begin_layout Standard
798 No process can open the same TDB twice; we check and disallow it.
799 This is an unfortunate side-effect of fcntl locks, which operate on a per-file
800 rather than per-file-descriptor basis, and do not nest.
801 Thus, closing any file descriptor on a file clears all the locks obtained
802 by this process, even if they were placed using a different file descriptor!
805 \begin_layout Standard
806 Note that even if this were solved, deadlock could occur if operations were
807 nested: this is a more manageable programming error in most cases.
810 \begin_layout Subsubsection
814 \begin_layout Standard
815 We could lobby POSIX to fix the perverse rules, or at least lobby Linux
816 to violate them so that the most common implementation does not have this
818 This would be a generally good idea for other fcntl lock users.
821 \begin_layout Standard
822 Samba uses a wrapper which hands out the same tdb_context to multiple callers
823 if this happens, and does simple reference counting.
824 We should do this inside the tdb library, which already emulates lock nesting
825 internally; it would need to recognize when deadlock occurs within a single
827 This would create a new failure mode for tdb operations (while we currently
828 handle locking failures, they are impossible in normal use and a process
829 encountering them can do little but give up).
832 \begin_layout Standard
833 I do not see benefit in an additional tdb_open flag to indicate whether
834 re-opening is allowed, as though there may be some benefit to adding a
835 call to detect when a tdb_context is shared, to allow other to create such
839 \begin_layout Subsubsection
843 \begin_layout Standard
847 \begin_layout Subsection
848 TDB API Is Not POSIX Thread-safe
851 \begin_layout Standard
852 The TDB API uses an error code which can be queried after an operation to
853 determine what went wrong.
854 This programming model does not work with threads, unless specific additional
855 guarantees are given by the implementation.
856 In addition, even otherwise-independent threads cannot open the same TDB
858 \begin_inset CommandInset ref
860 reference "TDB-Files-Cannot"
867 \begin_layout Subsubsection
871 \begin_layout Standard
872 Reachitecting the API to include a tdb_errcode pointer would be a great
873 deal of churn; we are better to guarantee that the tdb_errcode is per-thread
874 so the current programming model can be maintained.
877 \begin_layout Standard
878 This requires dynamic per-thread allocations, which is awkward with POSIX
879 threads (pthread_key_create space is limited and we cannot simply allocate
880 a key for every TDB).
883 \begin_layout Standard
884 Internal locking is required to make sure that fcntl locks do not overlap
885 between threads, and also that the global list of tdbs is maintained.
888 \begin_layout Standard
889 The aim is that building tdb with -DTDB_PTHREAD will result in a pthread-safe
890 version of the library, and otherwise no overhead will exist.
891 Alternatively, a hooking mechanism similar to that proposed for
892 \begin_inset CommandInset ref
894 reference "Proposed-Solution-locking-hook"
898 could be used to enable pthread locking at runtime.
901 \begin_layout Subsubsection
905 \begin_layout Standard
909 \begin_layout Subsection
910 *_nonblock Functions And *_mark Functions Expose Implementation
913 \begin_layout Standard
918 \begin_layout Plain Layout
919 Clustered TDB, see http://ctdb.samba.org
924 wishes to operate on TDB in a non-blocking manner.
925 This is currently done as follows:
928 \begin_layout Enumerate
929 Call the _nonblock variant of an API function (eg.
930 tdb_lockall_nonblock).
934 \begin_layout Enumerate
935 Fork a child process, and wait for it to call the normal variant (eg.
939 \begin_layout Enumerate
940 If the child succeeds, call the _mark variant to indicate we already have
945 \begin_layout Enumerate
946 Upon completion, tell the child to release the locks (eg.
950 \begin_layout Enumerate
951 Indicate to tdb that it should consider the locks removed (eg.
955 \begin_layout Standard
956 There are several issues with this approach.
957 Firstly, adding two new variants of each function clutters the API for
958 an obscure use, and so not all functions have three variants.
959 Secondly, it assumes that all paths of the functions ask for the same locks,
960 otherwise the parent process will have to get a lock which the child doesn't
961 have under some circumstances.
962 I don't believe this is currently the case, but it constrains the implementatio
967 \begin_layout Subsubsection
968 \begin_inset CommandInset label
970 name "Proposed-Solution-locking-hook"
977 \begin_layout Standard
978 Implement a hook for locking methods, so that the caller can control the
979 calls to create and remove fcntl locks.
980 In this scenario, ctdbd would operate as follows:
983 \begin_layout Enumerate
984 Call the normal API function, eg tdb_lockall().
987 \begin_layout Enumerate
988 When the lock callback comes in, check if the child has the lock.
989 Initially, this is always false.
991 Otherwise, try to obtain it in non-blocking mode.
992 If that fails, return EWOULDBLOCK.
995 \begin_layout Enumerate
996 Release locks in the unlock callback as normal.
999 \begin_layout Enumerate
1000 If tdb_lockall() fails, see if we recorded a lock failure; if so, call the
1001 child to repeat the operation.
1004 \begin_layout Enumerate
1005 The child records what locks it obtains, and returns that information to
1009 \begin_layout Enumerate
1010 When the child has succeeded, goto 1.
1013 \begin_layout Standard
1014 This is flexible enough to handle any potential locking scenario, even when
1015 lock requirements change.
1016 It can be optimized so that the parent does not release locks, just tells
1017 the child which locks it doesn't need to obtain.
1020 \begin_layout Standard
1021 It also keeps the complexity out of the API, and in ctdbd where it is needed.
1024 \begin_layout Subsubsection
1028 \begin_layout Standard
1032 \begin_layout Subsection
1033 tdb_chainlock Functions Expose Implementation
1036 \begin_layout Standard
1037 tdb_chainlock locks some number of records, including the record indicated
1039 This gave atomicity guarantees; no-one can start a transaction, alter,
1040 read or delete that key while the lock is held.
1043 \begin_layout Standard
1044 It also makes the same guarantee for any other key in the chain, which is
1045 an internal implementation detail and potentially a cause for deadlock.
1048 \begin_layout Subsubsection
1052 \begin_layout Standard
1054 It would be nice to have an explicit single entry lock which effected no
1056 Unfortunately, this won't work for an entry which doesn't exist.
1057 Thus while chainlock may be implemented more efficiently for the existing
1058 case, it will still have overlap issues with the non-existing case.
1059 So it is best to keep the current (lack of) guarantee about which records
1060 will be effected to avoid constraining our implementation.
1063 \begin_layout Subsection
1064 Signal Handling is Not Race-Free
1067 \begin_layout Standard
1068 The tdb_setalarm_sigptr() call allows the caller's signal handler to indicate
1069 that the tdb locking code should return with a failure, rather than trying
1070 again when a signal is received (and errno == EAGAIN).
1071 This is usually used to implement timeouts.
1074 \begin_layout Standard
1075 Unfortunately, this does not work in the case where the signal is received
1076 before the tdb code enters the fcntl() call to place the lock: the code
1077 will sleep within the fcntl() code, unaware that the signal wants it to
1079 In the case of long timeouts, this does not happen in practice.
1082 \begin_layout Subsubsection
1086 \begin_layout Standard
1087 The locking hooks proposed in
1088 \begin_inset CommandInset ref
1090 reference "Proposed-Solution-locking-hook"
1094 would allow the user to decide on whether to fail the lock acquisition
1096 This allows the caller to choose their own compromise: they could narrow
1097 the race by checking immediately before the fcntl call.
1101 \begin_layout Plain Layout
1102 It may be possible to make this race-free in some implementations by having
1103 the signal handler alter the struct flock to make it invalid.
1104 This will cause the fcntl() lock call to fail with EINVAL if the signal
1105 occurs before the kernel is entered, otherwise EAGAIN.
1113 \begin_layout Subsubsection
1117 \begin_layout Standard
1121 \begin_layout Subsection
1122 The API Uses Gratuitous Typedefs, Capitals
1125 \begin_layout Standard
1126 typedefs are useful for providing source compatibility when types can differ
1127 across implementations, or arguably in the case of function pointer definitions
1128 which are hard for humans to parse.
1129 Otherwise it is simply obfuscation and pollutes the namespace.
1132 \begin_layout Standard
1133 Capitalization is usually reserved for compile-time constants and macros.
1136 \begin_layout Description
1137 TDB_CONTEXT There is no reason to use this over 'struct tdb_context'; the
1138 definition isn't visible to the API user anyway.
1141 \begin_layout Description
1142 TDB_DATA There is no reason to use this over struct TDB_DATA; the struct
1143 needs to be understood by the API user.
1146 \begin_layout Description
1148 \begin_inset space ~
1151 TDB_DATA This would normally be called 'struct tdb_data'.
1154 \begin_layout Description
1156 \begin_inset space ~
1159 TDB_ERROR Similarly, this would normally be enum tdb_error.
1162 \begin_layout Subsubsection
1166 \begin_layout Standard
1168 Introducing lower case variants would please pedants like myself, but if
1169 it were done the existing ones should be kept.
1170 There is little point forcing a purely cosmetic change upon tdb users.
1173 \begin_layout Subsection
1174 \begin_inset CommandInset label
1176 name "tdb_log_func-Doesnt-Take"
1180 tdb_log_func Doesn't Take The Private Pointer
1183 \begin_layout Standard
1184 For API compatibility reasons, the logging function needs to call tdb_get_loggin
1185 g_private() to retrieve the pointer registered by the tdb_open_ex for logging.
1188 \begin_layout Subsubsection
1192 \begin_layout Standard
1193 It should simply take an extra argument, since we are prepared to break
1197 \begin_layout Subsubsection
1201 \begin_layout Standard
1205 \begin_layout Subsection
1206 Various Callback Functions Are Not Typesafe
1209 \begin_layout Standard
1210 The callback functions in tdb_set_logging_function (after
1211 \begin_inset CommandInset ref
1213 reference "tdb_log_func-Doesnt-Take"
1217 is resolved), tdb_parse_record, tdb_traverse, tdb_traverse_read and tdb_check
1218 all take void * and must internally convert it to the argument type they
1222 \begin_layout Standard
1223 If this type changes, the compiler will not produce warnings on the callers,
1224 since it only sees void *.
1227 \begin_layout Subsubsection
1231 \begin_layout Standard
1232 With careful use of macros, we can create callback functions which give
1233 a warning when used on gcc and the types of the callback and its private
1235 Unsupported compilers will not give a warning, which is no worse than now.
1236 In addition, the callbacks become clearer, as they need not use void *
1237 for their parameter.
1240 \begin_layout Standard
1241 See CCAN's typesafe_cb module at http://ccan.ozlabs.org/info/typesafe_cb.html
1244 \begin_layout Subsubsection
1248 \begin_layout Standard
1252 \begin_layout Subsection
1253 TDB_CLEAR_IF_FIRST Must Be Specified On All Opens, tdb_reopen_all Problematic
1256 \begin_layout Standard
1257 The TDB_CLEAR_IF_FIRST flag to tdb_open indicates that the TDB file should
1258 be cleared if the caller discovers it is the only process with the TDB
1260 However, if any caller does not specify TDB_CLEAR_IF_FIRST it will not
1261 be detected, so will have the TDB erased underneath them (usually resulting
1265 \begin_layout Standard
1266 There is a similar issue on fork(); if the parent exits (or otherwise closes
1267 the tdb) before the child calls tdb_reopen_all() to establish the lock
1268 used to indicate the TDB is opened by someone, a TDB_CLEAR_IF_FIRST opener
1269 at that moment will believe it alone has opened the TDB and will erase
1273 \begin_layout Subsubsection
1277 \begin_layout Standard
1278 Remove TDB_CLEAR_IF_FIRST.
1279 Other workarounds are possible, but see
1280 \begin_inset CommandInset ref
1282 reference "TDB_CLEAR_IF_FIRST-Imposes-Performance"
1289 \begin_layout Subsubsection
1293 \begin_layout Standard
1294 Incomplete, TDB_CLEAR_IF_FIRST still defined, but not implemented.
1297 \begin_layout Subsection
1298 Extending The Header Is Difficult
1301 \begin_layout Standard
1302 We have reserved (zeroed) words in the TDB header, which can be used for
1304 If the future features are compulsory, the version number must be updated
1305 to prevent old code from accessing the database.
1306 But if the future feature is optional, we have no way of telling if older
1307 code is accessing the database or not.
1310 \begin_layout Subsubsection
1314 \begin_layout Standard
1315 The header should contain a
1316 \begin_inset Quotes eld
1320 \begin_inset Quotes erd
1324 This is divided into two 32-bit parts:
1327 \begin_layout Enumerate
1328 The lower part reflects the format variant understood by code accessing
1332 \begin_layout Enumerate
1333 The upper part reflects the format variant you must understand to write
1334 to the database (otherwise you can only open for reading).
1337 \begin_layout Standard
1338 The latter field can only be written at creation time, the former should
1339 be written under the OPEN_LOCK when opening the database for writing, if
1340 the variant of the code is lower than the current lowest variant.
1343 \begin_layout Standard
1344 This should allow backwards-compatible features to be added, and detection
1345 if older code (which doesn't understand the feature) writes to the database.
1348 \begin_layout Subsubsection
1352 \begin_layout Standard
1356 \begin_layout Subsection
1357 Record Headers Are Not Expandible
1360 \begin_layout Standard
1361 If we later want to add (say) checksums on keys and data, it would require
1362 another format change, which we'd like to avoid.
1365 \begin_layout Subsubsection
1369 \begin_layout Standard
1370 We often have extra padding at the tail of a record.
1371 If we ensure that the first byte (if any) of this padding is zero, we will
1372 have a way for future changes to detect code which doesn't understand a
1373 new format: the new code would write (say) a 1 at the tail, and thus if
1374 there is no tail or the first byte is 0, we would know the extension is
1375 not present on that record.
1378 \begin_layout Subsubsection
1382 \begin_layout Standard
1386 \begin_layout Subsection
1387 TDB Does Not Use Talloc
1390 \begin_layout Standard
1391 Many users of TDB (particularly Samba) use the talloc allocator, and thus
1392 have to wrap TDB in a talloc context to use it conveniently.
1395 \begin_layout Subsubsection
1399 \begin_layout Standard
1400 The allocation within TDB is not complicated enough to justify the use of
1401 talloc, and I am reluctant to force another (excellent) library on TDB
1403 Nonetheless a compromise is possible.
1405 \begin_inset CommandInset ref
1407 reference "attributes"
1411 ) can be added later to tdb_open() to provide an alternate allocation mechanism,
1412 specifically for talloc but usable by any other allocator (which would
1414 \begin_inset Quotes eld
1418 \begin_inset Quotes erd
1424 \begin_layout Standard
1425 This would form a talloc heirarchy as expected, but the caller would still
1426 have to attach a destructor to the tdb context returned from tdb_open to
1428 All TDB_DATA fields would be children of the tdb_context, and the caller
1429 would still have to manage them (using talloc_free() or talloc_steal()).
1432 \begin_layout Subsubsection
1436 \begin_layout Standard
1440 \begin_layout Section
1441 Performance And Scalability Issues
1444 \begin_layout Subsection
1445 \begin_inset CommandInset label
1447 name "TDB_CLEAR_IF_FIRST-Imposes-Performance"
1451 TDB_CLEAR_IF_FIRST Imposes Performance Penalty
1454 \begin_layout Standard
1455 When TDB_CLEAR_IF_FIRST is specified, a 1-byte read lock is placed at offset
1458 While these locks never conflict in normal tdb usage, they do add substantial
1459 overhead for most fcntl lock implementations when the kernel scans to detect
1460 if a lock conflict exists.
1461 This is often a single linked list, making the time to acquire and release
1462 a fcntl lock O(N) where N is the number of processes with the TDB open,
1463 not the number actually doing work.
1466 \begin_layout Standard
1467 In a Samba server it is common to have huge numbers of clients sitting idle,
1468 and thus they have weaned themselves off the TDB_CLEAR_IF_FIRST flag.
1472 \begin_layout Plain Layout
1473 There is a flag to tdb_reopen_all() which is used for this optimization:
1474 if the parent process will outlive the child, the child does not need the
1476 This is a workaround for this very performance issue.
1484 \begin_layout Subsubsection
1488 \begin_layout Standard
1490 It was a neat idea, but even trivial servers tend to know when they are
1491 initializing for the first time and can simply unlink the old tdb at that
1495 \begin_layout Subsubsection
1499 \begin_layout Standard
1500 Incomplete; TDB_CLEAR_IF_FIRST still defined, but does nothing.
1503 \begin_layout Subsection
1504 TDB Files Have a 4G Limit
1507 \begin_layout Standard
1508 This seems to be becoming an issue (so much for
1509 \begin_inset Quotes eld
1513 \begin_inset Quotes erd
1516 !), particularly for ldb.
1519 \begin_layout Subsubsection
1523 \begin_layout Standard
1524 A new, incompatible TDB format which uses 64 bit offsets internally rather
1526 For simplicity of endian conversion (which TDB does on the fly if required),
1527 all values will be 64 bit on disk.
1528 In practice, some upper bits may be used for other purposes, but at least
1529 56 bits will be available for file offsets.
1532 \begin_layout Standard
1533 tdb_open() will automatically detect the old version, and even create them
1534 if TDB_VERSION6 is specified to tdb_open.
1537 \begin_layout Standard
1538 32 bit processes will still be able to access TDBs larger than 4G (assuming
1539 that their off_t allows them to seek to 64 bits), they will gracefully
1540 fall back as they fail to mmap.
1541 This can happen already with large TDBs.
1544 \begin_layout Standard
1545 Old versions of tdb will fail to open the new TDB files (since 28 August
1546 2009, commit 398d0c29290: prior to that any unrecognized file format would
1547 be erased and initialized as a fresh tdb!)
1550 \begin_layout Subsubsection
1554 \begin_layout Standard
1558 \begin_layout Subsection
1559 TDB Records Have a 4G Limit
1562 \begin_layout Standard
1563 This has not been a reported problem, and the API uses size_t which can
1564 be 64 bit on 64 bit platforms.
1565 However, other limits may have made such an issue moot.
1568 \begin_layout Subsubsection
1572 \begin_layout Standard
1573 Record sizes will be 64 bit, with an error returned on 32 bit platforms
1574 which try to access such records (the current implementation would return
1575 TDB_ERR_OOM in a similar case).
1576 It seems unlikely that 32 bit keys will be a limitation, so the implementation
1577 may not support this (see
1578 \begin_inset CommandInset ref
1580 reference "sub:Records-Incur-A"
1587 \begin_layout Subsubsection
1591 \begin_layout Standard
1595 \begin_layout Subsection
1596 Hash Size Is Determined At TDB Creation Time
1599 \begin_layout Standard
1600 TDB contains a number of hash chains in the header; the number is specified
1601 at creation time, and defaults to 131.
1602 This is such a bottleneck on large databases (as each hash chain gets quite
1603 long), that LDB uses 10,000 for this hash.
1604 In general it is impossible to know what the 'right' answer is at database
1608 \begin_layout Subsubsection
1609 \begin_inset CommandInset label
1611 name "sub:Hash-Size-Solution"
1618 \begin_layout Standard
1619 After comprehensive performance testing on various scalable hash variants
1623 \begin_layout Plain Layout
1624 http://rusty.ozlabs.org/?p=89 and http://rusty.ozlabs.org/?p=94 This was annoying
1625 because I was previously convinced that an expanding tree of hashes would
1626 be very close to optimal.
1631 , it became clear that it is hard to beat a straight linear hash table which
1632 doubles in size when it reaches saturation.
1633 Unfortunately, altering the hash table introduces serious locking complications
1634 : the entire hash table needs to be locked to enlarge the hash table, and
1635 others might be holding locks.
1636 Particularly insidious are insertions done under tdb_chainlock.
1639 \begin_layout Standard
1640 Thus an expanding layered hash will be used: an array of hash groups, with
1641 each hash group exploding into pointers to lower hash groups once it fills,
1642 turning into a hash tree.
1643 This has implications for locking: we must lock the entire group in case
1644 we need to expand it, yet we don't know how deep the tree is at that point.
1647 \begin_layout Standard
1648 Note that bits from the hash table entries should be stolen to hold more
1649 hash bits to reduce the penalty of collisions.
1650 We can use the otherwise-unused lower 3 bits.
1651 If we limit the size of the database to 64 exabytes, we can use the top
1652 8 bits of the hash entry as well.
1653 These 11 bits would reduce false positives down to 1 in 2000 which is more
1654 than we need: we can use one of the bits to indicate that the extra hash
1656 This means we can choose not to re-hash all entries when we expand a hash
1657 group; simply use the next bits we need and mark them invalid.
1660 \begin_layout Subsubsection
1664 \begin_layout Standard
1668 \begin_layout Subsection
1669 \begin_inset CommandInset label
1671 name "TDB-Freelist-Is"
1675 TDB Freelist Is Highly Contended
1678 \begin_layout Standard
1679 TDB uses a single linked list for the free list.
1680 Allocation occurs as follows, using heuristics which have evolved over
1684 \begin_layout Enumerate
1685 Get the free list lock for this whole operation.
1688 \begin_layout Enumerate
1689 Multiply length by 1.25, so we always over-allocate by 25%.
1692 \begin_layout Enumerate
1693 Set the slack multiplier to 1.
1696 \begin_layout Enumerate
1697 Examine the current freelist entry: if it is > length but < the current
1698 best case, remember it as the best case.
1701 \begin_layout Enumerate
1702 Multiply the slack multiplier by 1.05.
1705 \begin_layout Enumerate
1706 If our best fit so far is less than length * slack multiplier, return it.
1707 The slack will be turned into a new free record if it's large enough.
1710 \begin_layout Enumerate
1711 Otherwise, go onto the next freelist entry.
1714 \begin_layout Standard
1715 Deleting a record occurs as follows:
1718 \begin_layout Enumerate
1719 Lock the hash chain for this whole operation.
1722 \begin_layout Enumerate
1723 Walk the chain to find the record, keeping the prev pointer offset.
1726 \begin_layout Enumerate
1727 If max_dead is non-zero:
1731 \begin_layout Enumerate
1732 Walk the hash chain again and count the dead records.
1735 \begin_layout Enumerate
1736 If it's more than max_dead, bulk free all the dead ones (similar to steps
1737 4 and below, but the lock is only obtained once).
1740 \begin_layout Enumerate
1741 Simply mark this record as dead and return.
1746 \begin_layout Enumerate
1747 Get the free list lock for the remainder of this operation.
1750 \begin_layout Enumerate
1751 \begin_inset CommandInset label
1753 name "right-merging"
1757 Examine the following block to see if it is free; if so, enlarge the current
1758 block and remove that block from the free list.
1759 This was disabled, as removal from the free list was O(entries-in-free-list).
1762 \begin_layout Enumerate
1763 Examine the preceeding block to see if it is free: for this reason, each
1764 block has a 32-bit tailer which indicates its length.
1765 If it is free, expand it to cover our new block and return.
1768 \begin_layout Enumerate
1769 Otherwise, prepend ourselves to the free list.
1772 \begin_layout Standard
1773 Disabling right-merging (step
1774 \begin_inset CommandInset ref
1776 reference "right-merging"
1780 ) causes fragmentation; the other heuristics proved insufficient to address
1781 this, so the final answer to this was that when we expand the TDB file
1782 inside a transaction commit, we repack the entire tdb.
1785 \begin_layout Standard
1786 The single list lock limits our allocation rate; due to the other issues
1787 this is not currently seen as a bottleneck.
1790 \begin_layout Subsubsection
1794 \begin_layout Standard
1795 The first step is to remove all the current heuristics, as they obviously
1796 interact, then examine them once the lock contention is addressed.
1799 \begin_layout Standard
1800 The free list must be split to reduce contention.
1801 Assuming perfect free merging, we can at most have 1 free list entry for
1803 This implies that the number of free lists is related to the size of the
1804 hash table, but as it is rare to walk a large number of free list entries
1805 we can use far fewer, say 1/32 of the number of hash buckets.
1808 \begin_layout Standard
1809 It seems tempting to try to reuse the hash implementation which we use for
1810 records here, but we have two ways of searching for free entries: for allocatio
1811 n we search by size (and possibly zone) which produces too many clashes
1812 for our hash table to handle well, and for coalescing we search by address.
1813 Thus an array of doubly-linked free lists seems preferable.
1816 \begin_layout Standard
1817 There are various benefits in using per-size free lists (see
1818 \begin_inset CommandInset ref
1820 reference "sub:TDB-Becomes-Fragmented"
1824 ) but it's not clear this would reduce contention in the common case where
1825 all processes are allocating/freeing the same size.
1826 Thus we almost certainly need to divide in other ways: the most obvious
1827 is to divide the file into zones, and using a free list (or table of free
1829 This approximates address ordering.
1832 \begin_layout Standard
1833 Unfortunately it is difficult to know what heuristics should be used to
1834 determine zone sizes, and our transaction code relies on being able to
1836 \begin_inset Quotes eld
1840 \begin_inset Quotes erd
1843 by simply appending to the file (difficult if it would need to create a
1845 Thus we use a linked-list of free tables; currently we only ever create
1846 one, but if there is more than one we choose one at random to use.
1847 In future we may use heuristics to add new free tables on contention.
1848 We only expand the file when all free tables are exhausted.
1851 \begin_layout Standard
1852 The basic algorithm is as follows.
1856 \begin_layout Enumerate
1857 Identify the correct free list.
1860 \begin_layout Enumerate
1861 Lock the corresponding list.
1864 \begin_layout Enumerate
1865 Re-check the list (we didn't have a lock, sizes could have changed): relock
1869 \begin_layout Enumerate
1870 Place the freed entry in the list.
1873 \begin_layout Standard
1874 Allocation is a little more complicated, as we perform delayed coalescing
1878 \begin_layout Enumerate
1879 Pick a free table; usually the previous one.
1882 \begin_layout Enumerate
1883 Lock the corresponding list.
1886 \begin_layout Enumerate
1887 If the top entry is -large enough, remove it from the list and return it.
1890 \begin_layout Enumerate
1891 Otherwise, coalesce entries in the list.If there was no entry large enough,
1892 unlock the list and try the next largest list
1895 \begin_layout Enumerate
1896 If no list has an entry which meets our needs, try the next free table.
1899 \begin_layout Enumerate
1900 If no zone satisfies, expand the file.
1903 \begin_layout Standard
1904 This optimizes rapid insert/delete of free list entries by not coalescing
1906 First-fit address ordering ordering seems to be fairly good for keeping
1907 fragmentation low (see
1908 \begin_inset CommandInset ref
1910 reference "sub:TDB-Becomes-Fragmented"
1915 Note that address ordering does not need a tailer to coalesce, though if
1916 we needed one we could have one cheaply: see
1917 \begin_inset CommandInset ref
1919 reference "sub:Records-Incur-A"
1927 \begin_layout Standard
1928 Each free entry has the free table number in the header: less than 255.
1929 It also contains a doubly-linked list for easy deletion.
1932 \begin_layout Subsection
1933 \begin_inset CommandInset label
1935 name "sub:TDB-Becomes-Fragmented"
1939 TDB Becomes Fragmented
1942 \begin_layout Standard
1943 Much of this is a result of allocation strategy
1947 \begin_layout Plain Layout
1948 The Memory Fragmentation Problem: Solved? Johnstone & Wilson 1995 ftp://ftp.cs.ute
1949 xas.edu/pub/garbage/malloc/ismm98.ps
1954 and deliberate hobbling of coalescing; internal fragmentation (aka overallocati
1955 on) is deliberately set at 25%, and external fragmentation is only cured
1956 by the decision to repack the entire db when a transaction commit needs
1957 to enlarge the file.
1960 \begin_layout Subsubsection
1964 \begin_layout Standard
1965 The 25% overhead on allocation works in practice for ldb because indexes
1966 tend to expand by one record at a time.
1967 This internal fragmentation can be resolved by having an
1968 \begin_inset Quotes eld
1972 \begin_inset Quotes erd
1975 bit in the header to note entries that have previously expanded, and allocating
1976 more space for them.
1979 \begin_layout Standard
1980 There are is a spectrum of possible solutions for external fragmentation:
1981 one is to use a fragmentation-avoiding allocation strategy such as best-fit
1982 address-order allocator.
1983 The other end of the spectrum would be to use a bump allocator (very fast
1984 and simple) and simply repack the file when we reach the end.
1987 \begin_layout Standard
1988 There are three problems with efficient fragmentation-avoiding allocators:
1989 they are non-trivial, they tend to use a single free list for each size,
1990 and there's no evidence that tdb allocation patterns will match those recorded
1991 for general allocators (though it seems likely).
1994 \begin_layout Standard
1995 Thus we don't spend too much effort on external fragmentation; we will be
1996 no worse than the current code if we need to repack on occasion.
1997 More effort is spent on reducing freelist contention, and reducing overhead.
2000 \begin_layout Subsection
2001 \begin_inset CommandInset label
2003 name "sub:Records-Incur-A"
2007 Records Incur A 28-Byte Overhead
2010 \begin_layout Standard
2011 Each TDB record has a header as follows:
2014 \begin_layout LyX-Code
2018 \begin_layout LyX-Code
2019 tdb_off_t next; /* offset of the next record in the list */
2022 \begin_layout LyX-Code
2023 tdb_len_t rec_len; /* total byte length of record */
2026 \begin_layout LyX-Code
2027 tdb_len_t key_len; /* byte length of key */
2030 \begin_layout LyX-Code
2031 tdb_len_t data_len; /* byte length of data */
2034 \begin_layout LyX-Code
2035 uint32_t full_hash; /* the full 32 bit hash of the key */
2038 \begin_layout LyX-Code
2039 uint32_t magic; /* try to catch errors */
2042 \begin_layout LyX-Code
2043 /* the following union is implied:
2046 \begin_layout LyX-Code
2050 \begin_layout LyX-Code
2051 char record[rec_len];
2054 \begin_layout LyX-Code
2058 \begin_layout LyX-Code
2062 \begin_layout LyX-Code
2063 char data[data_len];
2066 \begin_layout LyX-Code
2070 \begin_layout LyX-Code
2071 uint32_t totalsize; (tailer)
2074 \begin_layout LyX-Code
2078 \begin_layout LyX-Code
2082 \begin_layout LyX-Code
2086 \begin_layout Standard
2087 Naively, this would double to a 56-byte overhead on a 64 bit implementation.
2090 \begin_layout Subsubsection
2094 \begin_layout Standard
2095 We can use various techniques to reduce this for an allocated block:
2098 \begin_layout Enumerate
2099 The 'next' pointer is not required, as we are using a flat hash table.
2102 \begin_layout Enumerate
2103 'rec_len' can instead be expressed as an addition to key_len and data_len
2104 (it accounts for wasted or overallocated length in the record).
2105 Since the record length is always a multiple of 8, we can conveniently
2106 fit it in 32 bits (representing up to 35 bits).
2109 \begin_layout Enumerate
2110 'key_len' and 'data_len' can be reduced.
2111 I'm unwilling to restrict 'data_len' to 32 bits, but instead we can combine
2112 the two into one 64-bit field and using a 5 bit value which indicates at
2113 what bit to divide the two.
2114 Keys are unlikely to scale as fast as data, so I'm assuming a maximum key
2118 \begin_layout Enumerate
2119 'full_hash' is used to avoid a memcmp on the
2120 \begin_inset Quotes eld
2124 \begin_inset Quotes erd
2127 case, but this is diminishing returns after a handful of bits (at 10 bits,
2128 it reduces 99.9% of false memcmp).
2129 As an aside, as the lower bits are already incorporated in the hash table
2130 resolution, the upper bits should be used here.
2131 Note that it's not clear that these bits will be a win, given the extra
2132 bits in the hash table itself (see
2133 \begin_inset CommandInset ref
2135 reference "sub:Hash-Size-Solution"
2142 \begin_layout Enumerate
2143 'magic' does not need to be enlarged: it currently reflects one of 5 values
2144 (used, free, dead, recovery, and unused_recovery).
2145 It is useful for quick sanity checking however, and should not be eliminated.
2148 \begin_layout Enumerate
2149 'tailer' is only used to coalesce free blocks (so a block to the right can
2150 find the header to check if this block is free).
2151 This can be replaced by a single 'free' bit in the header of the following
2152 block (and the tailer only exists in free blocks).
2156 \begin_layout Plain Layout
2157 This technique from Thomas Standish.
2158 Data Structure Techniques.
2159 Addison-Wesley, Reading, Massachusetts, 1980.
2164 The current proposed coalescing algorithm doesn't need this, however.
2167 \begin_layout Standard
2168 This produces a 16 byte used header like this:
2171 \begin_layout LyX-Code
2172 struct tdb_used_record {
2175 \begin_layout LyX-Code
2176 uint32_t used_magic : 16,
2179 \begin_layout LyX-Code
2183 \begin_layout LyX-Code
2187 \begin_layout LyX-Code
2191 \begin_layout LyX-Code
2192 uint32_t extra_octets;
2195 \begin_layout LyX-Code
2196 uint64_t key_and_data_len;
2199 \begin_layout LyX-Code
2203 \begin_layout Standard
2204 And a free record like this:
2207 \begin_layout LyX-Code
2208 struct tdb_free_record {
2211 \begin_layout LyX-Code
2212 uint64_t free_magic: 8,
2215 \begin_layout LyX-Code
2219 \begin_layout LyX-Code
2223 \begin_layout LyX-Code
2224 uint64_t free_table: 8,
2227 \begin_layout LyX-Code
2231 \begin_layout LyX-Code
2235 \begin_layout LyX-Code
2239 \begin_layout Standard
2241 \change_deleted 0 1291206079
2244 Note that by limiting valid offsets to 56 bits, we can pack everything we
2245 need into 3 64-byte words, meaning our minimum record size is 8 bytes.
2248 \begin_layout Subsubsection
2252 \begin_layout Standard
2256 \begin_layout Subsection
2257 Transaction Commit Requires 4 fdatasync
2260 \begin_layout Standard
2261 The current transaction algorithm is:
2264 \begin_layout Enumerate
2265 write_recovery_data();
2268 \begin_layout Enumerate
2272 \begin_layout Enumerate
2273 write_recovery_header();
2276 \begin_layout Enumerate
2280 \begin_layout Enumerate
2281 overwrite_with_new_data();
2284 \begin_layout Enumerate
2288 \begin_layout Enumerate
2289 remove_recovery_header();
2292 \begin_layout Enumerate
2296 \begin_layout Standard
2297 On current ext3, each sync flushes all data to disk, so the next 3 syncs
2298 are relatively expensive.
2299 But this could become a performance bottleneck on other filesystems such
2303 \begin_layout Subsubsection
2307 \begin_layout Standard
2308 Neil Brown points out that this is overzealous, and only one sync is needed:
2311 \begin_layout Enumerate
2312 Bundle the recovery data, a transaction counter and a strong checksum of
2316 \begin_layout Enumerate
2317 Strong checksum that whole bundle.
2320 \begin_layout Enumerate
2321 Store the bundle in the database.
2324 \begin_layout Enumerate
2325 Overwrite the oldest of the two recovery pointers in the header (identified
2326 using the transaction counter) with the offset of this bundle.
2329 \begin_layout Enumerate
2333 \begin_layout Enumerate
2334 Write the new data to the file.
2337 \begin_layout Standard
2338 Checking for recovery means identifying the latest bundle with a valid checksum
2339 and using the new data checksum to ensure that it has been applied.
2340 This is more expensive than the current check, but need only be done at
2342 For running databases, a separate header field can be used to indicate
2343 a transaction in progress; we need only check for recovery if this is set.
2346 \begin_layout Subsubsection
2350 \begin_layout Standard
2354 \begin_layout Subsection
2355 \begin_inset CommandInset label
2357 name "sub:TDB-Does-Not"
2361 TDB Does Not Have Snapshot Support
2364 \begin_layout Subsubsection
2365 Proposed SolutionNone.
2366 At some point you say
2367 \begin_inset Quotes eld
2371 \begin_inset Quotes erd
2375 \begin_inset CommandInset ref
2377 reference "replay-attribute"
2384 \begin_layout Standard
2385 But as a thought experiment, if we implemented transactions to only overwrite
2386 free entries (this is tricky: there must not be a header in each entry
2387 which indicates whether it is free, but use of presence in metadata elsewhere),
2388 and a pointer to the hash table, we could create an entirely new commit
2389 without destroying existing data.
2390 Then it would be easy to implement snapshots in a similar way.
2393 \begin_layout Standard
2394 This would not allow arbitrary changes to the database, such as tdb_repack
2395 does, and would require more space (since we have to preserve the current
2396 and future entries at once).
2397 If we used hash trees rather than one big hash table, we might only have
2398 to rewrite some sections of the hash, too.
2401 \begin_layout Standard
2402 We could then implement snapshots using a similar method, using multiple
2403 different hash tables/free tables.
2406 \begin_layout Subsubsection
2410 \begin_layout Standard
2414 \begin_layout Subsection
2415 Transactions Cannot Operate in Parallel
2418 \begin_layout Standard
2419 This would be useless for ldb, as it hits the index records with just about
2421 It would add significant complexity in resolving clashes, and cause the
2422 all transaction callers to write their code to loop in the case where the
2423 transactions spuriously failed.
2426 \begin_layout Subsubsection
2430 \begin_layout Standard
2432 \begin_inset CommandInset ref
2434 reference "replay-attribute"
2439 We could solve a small part of the problem by providing read-only transactions.
2440 These would allow one write transaction to begin, but it could not commit
2441 until all r/o transactions are done.
2442 This would require a new RO_TRANSACTION_LOCK, which would be upgraded on
2446 \begin_layout Subsubsection
2450 \begin_layout Standard
2454 \begin_layout Subsection
2455 Default Hash Function Is Suboptimal
2458 \begin_layout Standard
2459 The Knuth-inspired multiplicative hash used by tdb is fairly slow (especially
2460 if we expand it to 64 bits), and works best when the hash bucket size is
2461 a prime number (which also means a slow modulus).
2462 In addition, it is highly predictable which could potentially lead to a
2463 Denial of Service attack in some TDB uses.
2466 \begin_layout Subsubsection
2470 \begin_layout Standard
2471 The Jenkins lookup3 hash
2475 \begin_layout Plain Layout
2476 http://burtleburtle.net/bob/c/lookup3.c
2481 is a fast and superbly-mixing hash.
2482 It's used by the Linux kernel and almost everything else.
2483 This has the particular properties that it takes an initial seed, and produces
2484 two 32 bit hash numbers, which we can combine into a 64-bit hash.
2487 \begin_layout Standard
2488 The seed should be created at tdb-creation time from some random source,
2489 and placed in the header.
2490 This is far from foolproof, but adds a little bit of protection against
2494 \begin_layout Subsubsection
2498 \begin_layout Standard
2502 \begin_layout Subsection
2503 \begin_inset CommandInset label
2505 name "Reliable-Traversal-Adds"
2509 Reliable Traversal Adds Complexity
2512 \begin_layout Standard
2513 We lock a record during traversal iteration, and try to grab that lock in
2515 If that grab on delete fails, we simply mark it deleted and continue onwards;
2516 traversal checks for this condition and does the delete when it moves off
2520 \begin_layout Standard
2521 If traversal terminates, the dead record may be left indefinitely.
2524 \begin_layout Subsubsection
2528 \begin_layout Standard
2529 Remove reliability guarantees; see
2530 \begin_inset CommandInset ref
2532 reference "traverse-Proposed-Solution"
2539 \begin_layout Subsubsection
2543 \begin_layout Standard
2547 \begin_layout Subsection
2548 Fcntl Locking Adds Overhead
2551 \begin_layout Standard
2552 Placing a fcntl lock means a system call, as does removing one.
2553 This is actually one reason why transactions can be faster (everything
2554 is locked once at transaction start).
2555 In the uncontended case, this overhead can theoretically be eliminated.
2558 \begin_layout Subsubsection
2562 \begin_layout Standard
2566 \begin_layout Standard
2567 We tried this before with spinlock support, in the early days of TDB, and
2568 it didn't make much difference except in manufactured benchmarks.
2571 \begin_layout Standard
2572 We could use spinlocks (with futex kernel support under Linux), but it means
2573 that we lose automatic cleanup when a process dies with a lock.
2574 There is a method of auto-cleanup under Linux, but it's not supported by
2575 other operating systems.
2576 We could reintroduce a clear-if-first-style lock and sweep for dead futexes
2577 on open, but that wouldn't help the normal case of one concurrent opener
2579 Increasingly elaborate repair schemes could be considered, but they require
2580 an ABI change (everyone must use them) anyway, so there's no need to do
2581 this at the same time as everything else.
2584 \begin_layout Subsection
2585 Some Transactions Don't Require Durability
2588 \begin_layout Standard
2589 Volker points out that gencache uses a CLEAR_IF_FIRST tdb for normal (fast)
2590 usage, and occasionally empties the results into a transactional TDB.
2591 This kind of usage prioritizes performance over durability: as long as
2592 we are consistent, data can be lost.
2595 \begin_layout Standard
2596 This would be more neatly implemented inside tdb: a
2597 \begin_inset Quotes eld
2601 \begin_inset Quotes erd
2604 transaction commit (ie.
2605 syncless) which meant that data may be reverted on a crash.
2608 \begin_layout Subsubsection
2612 \begin_layout Standard
2616 \begin_layout Standard
2617 Unfortunately any transaction scheme which overwrites old data requires
2618 a sync before that overwrite to avoid the possibility of corruption.
2621 \begin_layout Standard
2622 It seems possible to use a scheme similar to that described in
2623 \begin_inset CommandInset ref
2625 reference "sub:TDB-Does-Not"
2629 ,where transactions are committed without overwriting existing data, and
2630 an array of top-level pointers were available in the header.
2631 If the transaction is
2632 \begin_inset Quotes eld
2636 \begin_inset Quotes erd
2639 then we would not need a sync at all: existing processes would pick up
2640 the new hash table and free list and work with that.
2643 \begin_layout Standard
2644 At some later point, a sync would allow recovery of the old data into the
2645 free lists (perhaps when the array of top-level pointers filled).
2646 On crash, tdb_open() would examine the array of top levels, and apply the
2647 transactions until it encountered an invalid checksum.
2650 \begin_layout Subsection
2651 Tracing Is Fragile, Replay Is External
2654 \begin_layout Standard
2655 The current TDB has compile-time-enabled tracing code, but it often breaks
2656 as it is not enabled by default.
2657 In a similar way, the ctdb code has an external wrapper which does replay
2658 tracing so it can coordinate cluster-wide transactions.
2661 \begin_layout Subsubsection
2663 \begin_inset CommandInset label
2665 name "replay-attribute"
2672 \begin_layout Standard
2673 Tridge points out that an attribute can be later added to tdb_open (see
2675 \begin_inset CommandInset ref
2677 reference "attributes"
2681 ) to provide replay/trace hooks, which could become the basis for this and
2682 future parallel transactions and snapshot support.
2685 \begin_layout Subsubsection
2689 \begin_layout Standard