Commit graph

722 commits

Author SHA1 Message Date
Sears Russell
4fe5bcaa72 Refactored doubleLinkedList and lruFast. The linked list API is now less insane, and lruFast is simpler. Added asserts to detect double insertion of nodes into the lists / lru 2010-09-24 17:56:51 +00:00
Sears Russell
7815d2f792 add assert to sequential-optimized write path 2010-09-24 17:51:09 +00:00
Sears Russell
a3e7fc0fe2 add hint: stasis_buffer_manager_hint_writes_are_sequential. if set to 1, then only one thread will write back at a time, which leads to greatly improved sequential write throughput 2010-09-23 23:09:43 +00:00
Sears Russell
9a46882ae3 port to new rbtree implementation (still buggy); performance tweaks for sequential throughput (this commit only changes non_blocking which is disabled by default, and in all sane runtime configurations) 2010-09-23 22:51:13 +00:00
Sears Russell
cb31593d1f fix typo in disabled code 2010-09-23 22:48:06 +00:00
Sears Russell
58c909a2ea add implementations of some more rblookup flags to the stl red black tree adapter 2010-09-23 20:38:50 +00:00
Sears Russell
4a5d14f2c6 more non_blocking fixes. now the unit test passes 2010-09-23 19:52:54 +00:00
Sears Russell
774ccfb002 whitespace and comment cleanup 2010-09-23 18:24:03 +00:00
Sears Russell
d3c8c1c003 remove overly-aggressive asserts from last checkin 2010-09-23 00:31:37 +00:00
Sears Russell
546bb4c767 fix rare latch corruption in concurrentBufferManager; add some asserts 2010-09-23 00:21:04 +00:00
Sears Russell
62e7663c27 fix + use macro 2010-09-23 00:19:34 +00:00
Sears Russell
aa8142d708 remove apparent race in concurrentHash.c 2010-09-17 01:38:40 +00:00
Sears Russell
889311a303 un-rollback -r1403. It was not the source of the assert failure. 2010-09-17 01:16:46 +00:00
Sears Russell
bb904e85cf add latch-free slotted pages, and a bug-fix in page.c 2010-09-17 01:10:42 +00:00
Sears Russell
e6067ae60b rollback -r1403, which was leading to tmp == p->TLS assertion failures 2010-09-02 17:59:00 +00:00
Sears Russell
00e474e3fe Fix bug in LinearHashNTA: inserting over an old key, then aborting set deleted the old key instead of restoring it to its original value 2010-09-01 04:01:20 +00:00
Sears Russell
c2cb334506 add comment to explain mysterious policy in concurrentWrapper 2010-08-25 16:37:05 +00:00
Sears Russell
9b0ee8e856 get rid of deadlock in concurrentBufferManager writeBackPage. The deadlock was due to concurrent reads in higher level code 2010-08-19 23:16:02 +00:00
Sears Russell
52055aa20f whitespace cleanup 2010-08-19 17:48:39 +00:00
Sears Russell
e5528ce218 relaxed page latching requirements; page.c, and dirtyPageTable no longer check / rely on holding rwlatch; removed (now) unnecessary latching from unit tests 2010-08-06 00:19:19 +00:00
Sears Russell
0c37a4bccc fix performance-degrading typo 2010-07-26 23:56:05 +00:00
Sears Russell
ad0fe63c2e implement writeBackPage for bufferHash.c 2010-07-26 23:55:43 +00:00
Sears Russell
7a27e39e5b MAX_BUFFER_SIZE is now a runtime (pre-tinit) option; clean up some doxygen warnings 2010-07-16 16:52:08 +00:00
Sears Russell
e2ba421a53 cleaned up buffer manager comments, added some asserts 2010-07-13 18:17:25 +00:00
Sears Russell
ee7aaff9ec add writeback starvation avoidance logic to concurrentBufferManager 2010-07-09 01:55:54 +00:00
Sears Russell
f376d46090 pull multiops outside of the log latch. (how embarassing...) 2010-06-15 23:12:38 +00:00
Sears Russell
97a2fdd600 fix null pointer dereference when recovering loser transaction that makes use of multipage operations 2010-05-19 01:22:40 +00:00
Sears Russell
71292538e8 Amortize bucket initialization costs for linear hash tables. The best case improvement (~ 20%) is seen with hashes with fixed length key/value pairs and insert-only workloads 2010-05-12 17:51:20 +00:00
Sears Russell
dc1f799a37 TarrayListAlloc() and TarrayListExtend() now generate O(1) bytes of log entries; they use the new multipage operation type. 2010-05-11 23:34:13 +00:00
Sears Russell
820f706faf TarrayListExtend now zeros out the contents of the newly allocated records (this is implemented by having fixed.c zero out the records) 2010-05-11 23:27:22 +00:00
Sears Russell
75ae7103fc fix crash in recovery redo that was caused by multipage operations 2010-05-11 23:17:09 +00:00
Sears Russell
2d645beb42 cleaned up multipage operations a bit; added stasis_segments_enabled flag 2010-05-11 22:01:37 +00:00
Sears Russell
07df7c064a add support for multi-page operations 2010-05-11 00:17:07 +00:00
Sears Russell
ffa4f18283 add tryToWriteBackPage. concurrentBufferManager is now compatible with FORCE mode transactions. 2010-04-28 00:05:00 +00:00
Sears Russell
ee6d0768d3 Fix failing assert by replacing trywritelock with writelock in concurrentBufferManager. Revision -r1380 relaxed an invariant, causing the trylock to sometimes contend with other threads 2010-04-26 04:07:38 +00:00
Sears Russell
74803d354e fix latch ordering in concurrent buffer manager. It no longer does a lock(p->loadlatch) while holding a hashtable bucket latch. Instead, it trylocks, then compensates for in progress changes if the latch fails. 2010-04-23 23:51:12 +00:00
Sears Russell
04de939e42 fix null pointer derefence in lruFast.c; add unit tests to check for this bug 2010-04-21 21:34:00 +00:00
Sears Russell
87ba38adb7 fix compilation when CONCURRENT_LRU is defined; make it the default (since it is only used by concurrentBufferManager) 2010-04-20 03:56:47 +00:00
Sears Russell
c7fa929e71 This fixes in Stasis issue 22.
Replacement policies now track the pin count of each page that they interact with.  This behavior used to live in bufferHash, and had never been ported to concurrentBufferManager.

For various reasons, the pin counts have to be updated atomically with the LRU state.  This would be very difficult to do inside of concurrentBufferManager, which sometimes needs to add and remove pages to replacement policies without holding any other latches.
2010-04-20 01:09:32 +00:00
Sears Russell
887832582c expose stasis_handle_t up through the buffer manager; add sequential I/O hinting; fix stupid busy-wait in prefetch 2010-04-12 20:54:10 +00:00
Sears Russell
7c5eb2a7e9 added dup method to io handles 2010-04-08 23:50:01 +00:00
Sears Russell
ff2f2e9e1e add TregionPrefetch() 2010-03-31 22:41:33 +00:00
Sears Russell
358f037441 fix concurrency bug that cropped up when the concurrent buffer manager was used with the concurrent wrapper replacement policy. Concurrent wrapper hashes on pageid, but the buffer manager was changing the pageid out from under it 2010-03-31 22:39:07 +00:00
Sears Russell
d281282533 fix concurrentWrapper so that it implements the same api as the other replacement policies (it used to livelock when no pages were available. It now returns null). 2010-03-31 22:37:42 +00:00
Sears Russell
3c27ed566e add prefetch primitives 2010-03-29 22:29:30 +00:00
Sears Russell
e17e948697 add missing return statement 2010-03-29 22:29:14 +00:00
Sears Russell
6deb34f5b2 fix handling of loadPageForOperation during recovery; initialize rid.size in slottedLast 2010-03-22 19:56:55 +00:00
Sears Russell
85abeb9d77 Fix ArrayList regressions 2010-02-18 19:50:06 +00:00
Sears Russell
5ffd1110e9 Added concept of "transaction fingerprints"
The region allocator marks freed regions "Condemned" before actually freeing them.  Condemned regions are associated 
with a transaction; before they can be reused / coalesced the transaction that condemned them must complete.

The region allocator used "TisActiveTransaction(xid)" to determine whether a given transaction had completed.  The 
problem is that Stasis now reuses transaction id's, and condemned regions were never reclaimed in practice.  A transaction
fingerprint is a unique identifier (currently transaction id, rec_lsn) that can be used to distinguish transactions that
did not execute concurrently.

I considered an alternative design based upon end-of-transaction actions.  This would have worked as well, but would 
have been a more invasive change.
2010-02-18 19:12:46 +00:00
Sears Russell
6454d55a42 fix bug in loadUninitializedPage(). It was passing INVALID_XID into the transaction table, which is not allowed 2010-02-18 19:04:39 +00:00