WL#3071 - Maria checkpoint
- serializing calls to flush_pagecache_blocks_int() on the same file to avoid known concurrency bugs - having that, we can now enable the background thread, as the flushes it does are now supposedly safe in concurrent situations. - new type of flush FLUSH_KEEP_LAZY: when the background checkpoint thread is flushing a packet of dirty pages between two checkpoints, it uses this flush type, indeed if a file is already being flushed by another thread it's smarter to move on to the next file than wait. - maria_checkpoint_frequency renamed to maria_checkpoint_interval. include/my_sys.h: new type of flushing for the page cache: FLUSH_KEEP_LAZY mysql-test/r/maria.result: result update mysys/mf_keycache.c: indentation. No FLUSH_KEEP_LAZY support in key cache. storage/maria/ha_maria.cc: maria_checkpoint_frequency was somehow a hidden part of the Checkpoint API and that was not good. Now we have checkpoint_interval, local to ha_maria.cc, which serves as container for the user-visible maria_checkpoint_interval global variable; setting it calls update_checkpoint_interval which passes the new value to ma_checkpoint_init(). There is no hiding anymore. By default, enable background thread which does checkpoints every 30 seconds, and dirty page flush in between. That thread takes a checkpoint when it ends, so no need for maria_hton_panic to take one. The | is | and not ||, because maria_panic() must always be called. frequency->interval. storage/maria/ma_checkpoint.c: Use FLUSH_KEEP_LAZY for background thread when it flushes packets of dirty pages between two checkpoints: it is smarter to move on to the next file than wait for it to have been completely flushed, which may take long. Comments about flush concurrency bugs moved from ma_pagecache.c. Removing out-of-date comment. frequency->interval. create_background_thread -> (interval>0). In ma_checkpoint_background(), some variables need to be preserved between iterations. storage/maria/ma_checkpoint.h: new prototype storage/maria/ma_pagecache.c: - concurrent calls of flush_pagecache_blocks_int() on the same file cause bugs (see @note in that function); we fix them by serializing in this situation. For that we use a global hash of (file, wqueue). When flush_pagecache_blocks_int() starts it looks into the hash, using the file as key. If not found, it inserts (file,wqueue) into the hash, flushes the file, and finally removes itself from the hash and wakes up any waiter in the queue. If found, it adds itself to the wqueue and waits. - As a by-product, we can remove changed_blocks_is_incomplete and replace it by scanning the hash, replace the sleep() by a queue wait. - new type of flush FLUSH_KEEP_LAZY: when flushing a file, if it's already being flushed by another thread (even partially), return immediately. storage/maria/ma_pagecache.h: In pagecache, a hash of files currently being flushed (i.e. there is a call to flush_pagecache_blocks_int() for them). storage/maria/ma_recovery.c: new prototype storage/maria/ma_test1.c: new prototype storage/maria/ma_test2.c: new prototype
Showing
This diff is collapsed.
This diff is collapsed.
Please register or sign in to comment