IF YOU WOULD LIKE TO GET AN ACCOUNT, please write an
email to Administrator. User accounts are meant only to access repo
and report issues and/or generate pull requests.
This is a purpose-specific Git hosting for
BaseALT
projects. Thank you for your understanding!
Только зарегистрированные пользователи имеют доступ к сервису!
Для получения аккаунта, обратитесь к администратору.
more agressively coalesce entries in the linked list of the undo
log. The second is to ensure that writes during a transaction into the
hash table don't cause the size of the undo log linked list to grow.
These optimisations don't affect Samba much, but they make a huge
difference to the use of ldb in kde
to msync/mmap not changing the mtime of the file. This patch ensures
that for successfully completed transactions we update the mtime.
I don't do this on all tdb writes as its too expensive, but doing it
just on transactions is bearable, as those cost quite a lot anyway.
the transaction lock in tdb_traverse_read(). This prevents a pattern
of locks which triggers the deadlock detection code in solaris10. I
suspect solaris10 is trying to prevent lock starvation by granting
locks in the order they were requested, which makes it much easier to
produce deadlocks.
allowing searches to proceed while another process is in a
transaction, then only upgrading the transaction lock to a write lock
on commit.
The solution is:
- split tdb_traverse() into two calls, called tdb_traverse() and
tdb_traverse_read(). The _read() version only gets read locks, and
will fail any write operations made in the callback from the
traverse.
- the normal tdb_traverse() call allows for read or write operations
in the callback, but gets the transaction lock, preventing
transastions from starting inside the traverse
In addition we enforce the following rule that you may not start a
transaction within a traverse callback, although you can start a
traverse within a transaction
With these rules in place I believe all the deadlock possibilities are
removed, and we can now allow for searches to happen in parallel with
transactions
samba-technical posting for more details on the transactions design.
This also adds a number of command line arguments to tdbtorture,
making it more flexible, and fixes some lock deadlock conditions in
the tdbtorture code.
just involves splitting up the core tdb.c code into separate files on
logical boundaries, but there are some minor functional changes as well:
- move the 'struct tdb_context' into tdb_private.h, hiding it from
users. This was done to allow the structure to change without
breaking code that uses tdb.
- added accessor functions tdb_fd(), tdb_name(), and tdb_log_fn() to
access the elements of struct tdb_context that were used by
external code but are no longer visible
- simplied tdb_append() to use tdb_fetch()/tdb_store(), which is just
as good due to the way tdb locks work
- changed some of the types (such as tdb_off to tdb_off_t) to make
syntax highlighting work better
- removed the old optional spinlock code. It was a bad idea.
- fixed a bug in tdb_reopen_all() that caused tdbtorture to sometimes
fail or report nasty looking errors. This is the only real bug
fixed in this commit. Jeremy/Jerry, you might like to pickup this
change for Samba3, as that could definately affect smbd in
Samba3.
The aim of all of these changes is to make the tdb
transactions/journaling code I am working on easier to write. I
started to write it on top of the existing tdb.c code and it got very
messy. Splitting up the code makes it much easier to follow.
There are more cleanups we could do in tdb, such as using uint32_t
instead of u32 (suggested by metze). I'll leave those for another day.
optimise this case a bit more. The total speedup using non-indexed
ldbtest is now around a factor of 80x. The code is ugly as hell, but
I think this speed is worth it.
Of course, if we only ever do indexed searches in ldb then this
doesn't help, but it seems all too common that we get unindexable
searches, so the optimisation is worthwhile
added a tdb optimisation that speeds up non-indexed ldb by a large
margin (often 10x or more). I'd be interested in any comments on the
safety of this optimisation. See the comment in the code for an
explanation.
called from multiple backends. (ldb_sqlite3 needs it too.) Added parameter
for a callback function that determines whether an attribute needs case
folding.
- begin to prepare for sqlite3 in build process
- work-in-progress updates, on ldb_sqlite3
returning NULL/0, which is the same as we used for a failure. Having
to look at tdb->ecode (which we never do) is too error prone.
Instead, tdb_fetch() should behave like malloc() and talloc(), where
zero length is not special and malloc(0) returns a valid pointer.
- similarly in data_blob(), asking for data_blob(NULL, 0) should
return a zero blob, but asking for data_blob(ptr, 0) should return a
zero length blob with a valid pointer, just like talloc() and malloc()
This change fixes the SummaryInformation stream stored in the tdb
backend when manipulated from w2k. The w2k client was using
SET_EOF_INFORMATION to create a zero-length stream, which we return
STATUS_NOT_FOUND on, as the tdb_fetch() gave us back a NULL/0 blob,
which we returned as not-found
less likely that anyone will use pstring for new code
- got rid of winbind_client.h from includes.h. This one triggered a
huge change, as winbind_client.h was including system/filesys.h and
defining the old uint32 and uint16 types, as well as its own
pstring and fstring.
- removed the u32 hack in events.c as I think this was only needed as
tdb.h defines u32. Metze, can you check that this hack is indeed no
longer needed on your suse system?
convert a ... varargs function to a va_list by just a cast!!)
also mark the tdb log function with PRINTF_ATTRIBUTE() and fixed some
bad format errors in tdb.c that jim found.
