WL#3072 Maria Recovery.
Updates to the bitmap flush/pin logic to prepare for when we support multiple writers. storage/maria/ma_bitmap.c: Read lock is less bad than write lock. Changing bitmap->flushable to a counter, to prepare for when multiple writers are allowed on a table. Using bitmap->flush_all_requested instead of share->in_checkpoint; the latter can be true for the time of a whole checkpoint even though the checkpoint is not yet handling our table, or has already handled it, so to decrease the number of broadcasts we use a dedicated my_bool which is true only when checkpoint is handling this table's bitmap. _ma_bitmap_flushable(share,+1) waits for a concurrent _ma_bitmap_flush_all() to finish before incrementing non_flushable; without this, with multiple writers there may always be one thread making the bitmap unflushable and thus checkpoint would stall. storage/maria/ma_blockrec.c: update to new prototype: "flushable is FALSE|TRUE" becomes "add 1|-1 to non_flushable". storage/maria/ma_blockrec.h: new prototype storage/maria/maria_def.h: MARIA_FILE_BITMAP::flushable becomes a counter. New MARIA_FILE_BITMAP::flush_all_requested.
Showing
Please register or sign in to comment