- Completely rewritten aiops.c that creates and manages a pool of
threads so thread creation overhead is eliminated
- Lots of extra debugging
- Lots of mods to store.c to detect and cancel outstanding ASYNC ops.
Code is not proven exhaustive and there are definately still cases
to be found where outstanding disk ops aren't cancelled properly.
- Changes to call interface to a few routines to support disk op
`tagging', so operations can be cleanly cancelled on store_abort()s.
- Implementation of swap.state files as transaction logs. Removed
objects are now noted with a negative object size. This allows
reliatively clean rebuilds from non-clean shutdowns.
- Now that the swap.state files are transaction logs, there's now no
need to validate by stat()ing. All the validation procedure does
is now just set the valid bit AFTER all the swap.state files have
been read, because by that time, only valid objects can be left.
Object still need to be marked invalid when reading the swap.state
file because there's no guarantee the file has been retaken or
deleted.
- An fstat() call is now added after every storeSwapInFileOpened() so
object sizes can be checked. Added code to storeRelease() the
object if the sizes don't match.
- #defining USE_ASYNC_IO now uses the async unlink() rather than
unlinkd()
- #defining MONOTONIC_STORE will support the creation of disk objects
clustered into directories. This GREATLY improves disk performance
(factor of 3) over old `write-over-old-object' method. If using the
MONOTONIC_STORE, the {get/put}_unusedFileno stack stuff is disabled.
This is actually a good thing and greatly reduces the risk of serving
up bad objects.
- Fixed unlink() in storeWriteCleanLogs to be real unlink() rather
than ASYNC/unlinkd unlinks. swap.state.new files were being removed
just after they were created due to delayed unlinks.
- Disabled various assertions and made these into debug warning
messages to make the code more stable until the bugs can be tracked
down.
- Added most of Michael O'Reilly's patches which included many bug
fixes. Ask him for full details.
- Moved aio_check_callbacks in comm_{poll|select}(). It was called
after the fdset had been built which was wrong because the callbacks
were changing the state of the read/write handlers prior to the
poll/select() calls.