Previously, a two file descriptors were used for the database: file & metafile. The "file" file
descriptor was used for async writes while the "metafile" file descriptor was used with O_SYNC
writes. This commit changes that so that there's only one file descriptor and it uses fdatasync()
to synchronize writes.
The count and overflow columns are meaningless for freed pages since
there could be random overflow data in there. This commit removes those
columns for free pages.
Well, this is embarassing. Somehow the freelist was never getting written after each commit.
This commit fixes that and fixes a small reporting issue with "bolt pages".
The bucket page is allocated separately from the rest of the pages but the old bucket pages were
not being added to the freelist. This change fixes that and adds a simple check for database
consistency. More advanced consistency checks can be added in the future.
Fixes#82.
I consolidated the DB.Tx() and DB.RWTx() calls into a single
DB.Begin(writable bool) call. This is more consistent with the
database/sql library.
I also changed the DB.Do() and DB.With() call to DB.Update() and
DB.View(), respectively. This is more intuitive and more inline with
other database verbiage.
Commit d2173f5f0ecbf4ed93c768e975435b04df3186ec removed the complete
os & syscall mocking layer as overly complex. This commit adds back
the simplest possible thing: hooks to control the database file
writes.
Missing tests: TestDBOpenMetaFileError, TestDBMmapStatError.
These are harder to test without more extensive mocking.
Conflicts:
db_test.go
Write errors are often delayed and reported only by the close.
The extra close in defer on success is harmless, (*os.File).Close
protects itself against multiple closes, and this way it's immediately
obvious there is no code path that would leak open files.
Commit/Rollback and mutable calls on Tx and Bucket now return ErrTxClosed
if the transaction has already been committed or rolled back. Non-mutable
calls have added an assertion to check if the transaction is closed which
will cause a panic. I don't want to introduce an error return for accessor
methods that are being used improperly so I think the panic is appropriate.
Transaction created from Do() and With() are now considered "managed".
Managed transactions cannot be manually committed or rolled back since
the Do() and With() functions provide that functionally automatically.
Previously, a Tx could be manually committed and then any changes after
that would be lost.
This avoids a case where writes can be reordered so meta page is
written before a page it refers to, potentially causing a corrupt
database after a power loss or kernel crash.
