We tried to get a F_WRLCK on the open lock; we shouldn't do that for a
read-only tdb.  (TDB1 gets away with it because a read-only open skips
all locking).

We also avoid leaking the fd in two tdb_open() failure paths revealed
by this extra testing.

Allows tests to explicitly avoid continuing when a failure has been
injected.

TDB2 tracks locks using getpid(), and gets upset when we fork behind
its back.

This crept in, it should be the same as the tests in typesafe_cb.h.

More recording of interesting events.  As we don't have an ABI yet, we
don't need to put these at the end.

The original code assumed that unlocking would fail if we didn't have a lock;
this isn't true (at least, on my machine).  So we have to always check the
pid before unlocking.

PAGESIZE used to be defined to getpagesize(); we changed it to a
constant in b556ef1f, which broke the msync() call.

We don't need to copy into a buffer to examine the old data: in the
common case, it's mmaped already.  It's made a bit trickier because
the tdb_access_read() function uses the current I/O methods, so we
need to restore that temporarily.

The difference was in the noise, however (the sync no-doubt
dominates).

Before:
$ time ./growtdb-bench 250000 10 > /dev/null && ls -l /tmp/growtdb.tdb && time ./tdbtorture -s 0 && ls -l torture.tdb && ./speed --transaction 2000000
real	0m45.021s
user	0m16.261s
sys	0m2.432s
-rw------- 1 rusty rusty 364469344 2011-04-27 22:55 /tmp/growtdb.tdb
testing with 3 processes, 5000 loops, seed=0
OK

real	1m10.144s
user	0m0.480s
sys	0m0.460s
-rw------- 1 rusty rusty 391992 2011-04-27 22:56 torture.tdb
Adding 2000000 records:  863 ns (110601144 bytes)
Finding 2000000 records:  565 ns (110601144 bytes)
Missing 2000000 records:  383 ns (110601144 bytes)
Traversing 2000000 records:  409 ns (110601144 bytes)
Deleting 2000000 records:  676 ns (225354680 bytes)
Re-adding 2000000 records:  784 ns (225354680 bytes)
Appending 2000000 records:  1191 ns (247890168 bytes)
Churning 2000000 records:  2166 ns (423133432 bytes)

After:
real	0m47.141s
user	0m16.073s
sys	0m2.460s
-rw------- 1 rusty rusty 364469344 2011-04-27 22:58 /tmp/growtdb.tdb
testing with 3 processes, 5000 loops, seed=0
OK

real	1m4.207s
user	0m0.416s
sys	0m0.504s
-rw------- 1 rusty rusty 313576 2011-04-27 22:59 torture.tdb
Adding 2000000 records:  874 ns (110601144 bytes)
Finding 2000000 records:  565 ns (110601144 bytes)
Missing 2000000 records:  393 ns (110601144 bytes)
Traversing 2000000 records:  404 ns (110601144 bytes)
Deleting 2000000 records:  684 ns (225354680 bytes)
Re-adding 2000000 records:  792 ns (225354680 bytes)
Appending 2000000 records:  1212 ns (247890168 bytes)
Churning 2000000 records:  2191 ns (423133432 bytes)

We don't need to use 4k for our transaction pages; we can use any
value.  For the tools/speed benchmark, any value between about 4k and
64M makes no difference, but that's probably because the entire
database is touched in each transaction.

So instead, I looked at tdbtorture to try to find an optimum value, as
it uses smaller transactions.  4k and 64k were equivalent.  16M was
almost three times slower, 1M was 5-10% slower.  1024 was also 5-10%
slower.

There's a slight advantage of having larger pages, both for allowing
direct access to the database (if it's all in one page we can sometimes
grant direct access even inside a transaction) and for the compactness
of our recovery area (since our code is naive and won't combine one
run across pages).

Before:
$ time ./growtdb-bench 250000 10 > /dev/null && ls -l /tmp/growtdb.tdb && time ./tdbtorture -s 0 && ls -l torture.tdb && ./speed --transaction 2000000
real	0m47.127s
user	0m17.125s
sys	0m2.456s
-rw------- 1 rusty rusty 366680288 2011-04-27 21:34 /tmp/growtdb.tdb
testing with 3 processes, 5000 loops, seed=0
OK

real	1m16.049s
user	0m0.300s
sys	0m0.492s
-rw------- 1 rusty rusty 244472 2011-04-27 21:35 torture.tdb
Adding 2000000 records:  894 ns (110551992 bytes)
Finding 2000000 records:  564 ns (110551992 bytes)
Missing 2000000 records:  398 ns (110551992 bytes)
Traversing 2000000 records:  399 ns (110551992 bytes)
Deleting 2000000 records:  711 ns (225633208 bytes)
Re-adding 2000000 records:  819 ns (225633208 bytes)
Appending 2000000 records:  1252 ns (248196544 bytes)
Churning 2000000 records:  2319 ns (424005056 bytes)

After:
$ time ./growtdb-bench 250000 10 > /dev/null && ls -l /tmp/growtdb.tdb && time ./tdbtorture -s 0 && ls -l torture.tdb && ./speed --transaction 2000000
real	0m45.021s
user	0m16.261s
sys	0m2.432s
-rw------- 1 rusty rusty 364469344 2011-04-27 22:55 /tmp/growtdb.tdb
testing with 3 processes, 5000 loops, seed=0
OK

real	1m10.144s
user	0m0.480s
sys	0m0.460s
-rw------- 1 rusty rusty 391992 2011-04-27 22:56 torture.tdb
Adding 2000000 records:  863 ns (110601144 bytes)
Finding 2000000 records:  565 ns (110601144 bytes)
Missing 2000000 records:  383 ns (110601144 bytes)
Traversing 2000000 records:  409 ns (110601144 bytes)
Deleting 2000000 records:  676 ns (225354680 bytes)
Re-adding 2000000 records:  784 ns (225354680 bytes)
Appending 2000000 records:  1191 ns (247890168 bytes)
Churning 2000000 records:  2166 ns (423133432 bytes)

Currently we use the worst-case-possible size for the recovery area.
Instead, prepare the recovery data, then see whether it's too large.

Note that this currently works out to make the database *larger* on
our speed benchmark, since we happen to need to enlarge the recovery
area at the wrong time now, rather than the old case where its already
hugely oversized.

Before:
$ time ./growtdb-bench 250000 10 > /dev/null && ls -l /tmp/growtdb.tdb && time ./tdbtorture -s 0 && ls -l torture.tdb && ./speed --transaction 2000000
real	0m50.366s
user	0m17.109s
sys	0m2.468s
-rw------- 1 rusty rusty 564215952 2011-04-27 21:31 /tmp/growtdb.tdb
testing with 3 processes, 5000 loops, seed=0
OK

real	1m23.818s
user	0m0.304s
sys	0m0.508s
-rw------- 1 rusty rusty 669856 2011-04-27 21:32 torture.tdb
Adding 2000000 records:  887 ns (110556088 bytes)
Finding 2000000 records:  556 ns (110556088 bytes)
Missing 2000000 records:  385 ns (110556088 bytes)
Traversing 2000000 records:  401 ns (110556088 bytes)
Deleting 2000000 records:  710 ns (244003768 bytes)
Re-adding 2000000 records:  825 ns (244003768 bytes)
Appending 2000000 records:  1255 ns (268404160 bytes)
Churning 2000000 records:  2299 ns (268404160 bytes)

After:
$ time ./growtdb-bench 250000 10 > /dev/null && ls -l /tmp/growtdb.tdb && time ./tdbtorture -s 0 && ls -l torture.tdb && ./speed --transaction 2000000
real	0m47.127s
user	0m17.125s
sys	0m2.456s
-rw------- 1 rusty rusty 366680288 2011-04-27 21:34 /tmp/growtdb.tdb
testing with 3 processes, 5000 loops, seed=0
OK

real	1m16.049s
user	0m0.300s
sys	0m0.492s
-rw------- 1 rusty rusty 244472 2011-04-27 21:35 torture.tdb
Adding 2000000 records:  894 ns (110551992 bytes)
Finding 2000000 records:  564 ns (110551992 bytes)
Missing 2000000 records:  398 ns (110551992 bytes)
Traversing 2000000 records:  399 ns (110551992 bytes)
Deleting 2000000 records:  711 ns (225633208 bytes)
Re-adding 2000000 records:  819 ns (225633208 bytes)
Appending 2000000 records:  1252 ns (248196544 bytes)
Churning 2000000 records:  2319 ns (424005056 bytes)

We don't need to write the whole page to the recovery area if it
hasn't all changed.  Simply skipping the start and end of the pages
which are similar saves us about 20% on growtdb-bench 250000, and 45%
on tdbtorture.  The more thorough examination of page differences
gives us a saving of 90% on growtdb-bench and 98% on tdbtorture!

And we do win a bit on timings for transaction commit:

Before:
$ time ./growtdb-bench 250000 10 > /dev/null && ls -l /tmp/growtdb.tdb && time ./tdbtorture -s 0 && ls -l torture.tdb && ./speed --transaction 2000000
real	1m4.844s
user	0m15.537s
sys	0m3.796s
-rw------- 1 rusty rusty 626693096 2011-04-27 21:28 /tmp/growtdb.tdb
testing with 3 processes, 5000 loops, seed=0
OK

real	1m17.021s
user	0m0.272s
sys	0m0.540s
-rw------- 1 rusty rusty 458800 2011-04-27 21:29 torture.tdb
Adding 2000000 records:  894 ns (110556088 bytes)
Finding 2000000 records:  569 ns (110556088 bytes)
Missing 2000000 records:  390 ns (110556088 bytes)
Traversing 2000000 records:  403 ns (110556088 bytes)
Deleting 2000000 records:  710 ns (244003768 bytes)
Re-adding 2000000 records:  825 ns (244003768 bytes)
Appending 2000000 records:  1262 ns (268404160 bytes)
Churning 2000000 records:  2311 ns (268404160 bytes)


After:
$ time ./growtdb-bench 250000 10 > /dev/null && ls -l /tmp/growtdb.tdb && time ./tdbtorture -s 0 && ls -l torture.tdb && ./speed --transaction 2000000
real	0m50.366s
user	0m17.109s
sys	0m2.468s
-rw------- 1 rusty rusty 564215952 2011-04-27 21:31 /tmp/growtdb.tdb
testing with 3 processes, 5000 loops, seed=0
OK

real	1m23.818s
user	0m0.304s
sys	0m0.508s
-rw------- 1 rusty rusty 669856 2011-04-27 21:32 torture.tdb
Adding 2000000 records:  887 ns (110556088 bytes)
Finding 2000000 records:  556 ns (110556088 bytes)
Missing 2000000 records:  385 ns (110556088 bytes)
Traversing 2000000 records:  401 ns (110556088 bytes)
Deleting 2000000 records:  710 ns (244003768 bytes)
Re-adding 2000000 records:  825 ns (244003768 bytes)
Appending 2000000 records:  1255 ns (268404160 bytes)
Churning 2000000 records:  2299 ns (268404160 bytes)

tdb1 always makes the tdb a multiple of the transaction page size,
tdb2 doesn't.  This means that if a transaction hits the exact end of
the file, we might need to save off a partial page.

So that we don't have to rewrite tdb_recovery_size() too, we simply do
a short read and memset the unused section to 0 (to keep valgrind
happy).

We don't have tailers in tdb2, so it's just 8 bytes of confusing wastage.

Instead of walking the entire free list, walk 8 entries, or more if we
are successful: the reward is scaled by the size coalesced.

We also move previously-examined records to the end of the list.

This reduces file size with very little speed penalty.

Before:
$ time ./growtdb-bench 250000 10 > /dev/null && ls -l /tmp/growtdb.tdb && time ./tdbtorture -s 0 && ls -l torture.tdb && ./speed --transaction 2000000
real	1m17.022s
user	0m27.206s
sys	0m3.920s
-rw------- 1 rusty rusty 570130576 2011-04-27 21:17 /tmp/growtdb.tdb
testing with 3 processes, 5000 loops, seed=0
OK

real	1m27.355s
user	0m0.296s
sys	0m0.516s
-rw------- 1 rusty rusty 617352 2011-04-27 21:18 torture.tdb
Adding 2000000 records:  890 ns (110556088 bytes)
Finding 2000000 records:  565 ns (110556088 bytes)
Missing 2000000 records:  390 ns (110556088 bytes)
Traversing 2000000 records:  410 ns (110556088 bytes)
Deleting 2000000 records:  8623 ns (244003768 bytes)
Re-adding 2000000 records:  7089 ns (244003768 bytes)
Appending 2000000 records:  33708 ns (244003768 bytes)
Churning 2000000 records:  2029 ns (268404160 bytes)

After:
$ time ./growtdb-bench 250000 10 > /dev/null && ls -l /tmp/growtdb.tdb && time ./tdbtorture -s 0 && ls -l torture.tdb && ./speed --transaction 2000000
real	1m7.096s
user	0m15.637s
sys	0m3.812s
-rw------- 1 rusty rusty 561270928 2011-04-27 21:22 /tmp/growtdb.tdb
testing with 3 processes, 5000 loops, seed=0
OK

real	1m13.850s
user	0m0.268s
sys	0m0.492s
-rw------- 1 rusty rusty 429768 2011-04-27 21:23 torture.tdb
Adding 2000000 records:  892 ns (110556088 bytes)
Finding 2000000 records:  570 ns (110556088 bytes)
Missing 2000000 records:  390 ns (110556088 bytes)
Traversing 2000000 records:  407 ns (110556088 bytes)
Deleting 2000000 records:  706 ns (244003768 bytes)
Re-adding 2000000 records:  822 ns (244003768 bytes)
Appending 2000000 records:  1262 ns (268404160 bytes)
Churning 2000000 records:  2320 ns (268404160 bytes)

This simply uses a 7 bit counter which gets incremented on each addition
to the list (but not decremented on removals).  When it wraps, we walk the
entire list looking for things to coalesce.

This causes performance problems, especially when appending records, so
we limit it in the next patch:

Before:
$ time ./growtdb-bench 250000 10 > /dev/null && ls -l /tmp/growtdb.tdb && time ./tdbtorture -s 0 && ls -l torture.tdb && ./speed --transaction 2000000
real	0m59.687s
user	0m11.593s
sys	0m4.100s
-rw------- 1 rusty rusty 752004064 2011-04-27 21:14 /tmp/growtdb.tdb
testing with 3 processes, 5000 loops, seed=0
OK

real	1m17.738s
user	0m0.348s
sys	0m0.580s
-rw------- 1 rusty rusty 663360 2011-04-27 21:15 torture.tdb
Adding 2000000 records:  926 ns (110556088 bytes)
Finding 2000000 records:  592 ns (110556088 bytes)
Missing 2000000 records:  416 ns (110556088 bytes)
Traversing 2000000 records:  422 ns (110556088 bytes)
Deleting 2000000 records:  741 ns (244003768 bytes)
Re-adding 2000000 records:  799 ns (244003768 bytes)
Appending 2000000 records:  1147 ns (295244592 bytes)
Churning 2000000 records:  1827 ns (568411440 bytes)

After:
$ time ./growtdb-bench 250000 10 > /dev/null && ls -l /tmp/growtdb.tdb && time ./tdbtorture -s 0 && ls -l torture.tdb && ./speed --transaction 2000000
real	1m17.022s
user	0m27.206s
sys	0m3.920s
-rw------- 1 rusty rusty 570130576 2011-04-27 21:17 /tmp/growtdb.tdb
testing with 3 processes, 5000 loops, seed=0
OK

real	1m27.355s
user	0m0.296s
sys	0m0.516s
-rw------- 1 rusty rusty 617352 2011-04-27 21:18 torture.tdb
Adding 2000000 records:  890 ns (110556088 bytes)
Finding 2000000 records:  565 ns (110556088 bytes)
Missing 2000000 records:  390 ns (110556088 bytes)
Traversing 2000000 records:  410 ns (110556088 bytes)
Deleting 2000000 records:  8623 ns (244003768 bytes)
Re-adding 2000000 records:  7089 ns (244003768 bytes)
Appending 2000000 records:  33708 ns (244003768 bytes)
Churning 2000000 records:  2029 ns (268404160 bytes)

I noticed a counter-intuitive phenomenon as I tweaked the coalescing
code: the more coalescing we did, the larger the tdb grew!  This was
measured using "growtdb-bench 250000 10".

The cause: more coalescing means larger transactions, and every time
we do a larger transaction, we need to allocate a larger recovery
area.  The only way to do this is to append to the file, so the file
keeps growing, even though it's mainly unused!

Overallocating by 25% seems reasonable, and gives better results in
such benchmarks.

The real fix is to reduce the transaction to a run-length based format
rather then the naive block system used now.

Before:
$ time ./growtdb-bench 250000 10 > /dev/null && ls -l /tmp/growtdb.tdb && time ./tdbtorture -s 0 && ls -l torture.tdb && ./speed --transaction 2000000
real	0m57.403s
user	0m11.361s
sys	0m4.056s
-rw------- 1 rusty rusty 689536976 2011-04-27 21:10 /tmp/growtdb.tdb
testing with 3 processes, 5000 loops, seed=0
OK

real	1m24.901s
user	0m0.380s
sys	0m0.512s
-rw------- 1 rusty rusty 655368 2011-04-27 21:12 torture.tdb
Adding 2000000 records:  941 ns (110551992 bytes)
Finding 2000000 records:  603 ns (110551992 bytes)
Missing 2000000 records:  428 ns (110551992 bytes)
Traversing 2000000 records:  416 ns (110551992 bytes)
Deleting 2000000 records:  741 ns (199517112 bytes)
Re-adding 2000000 records:  819 ns (199517112 bytes)
Appending 2000000 records:  1228 ns (376542552 bytes)
Churning 2000000 records:  2042 ns (553641304 bytes)

After:
$ time ./growtdb-bench 250000 10 > /dev/null && ls -l /tmp/growtdb.tdb && time ./tdbtorture -s 0 && ls -l torture.tdb && ./speed --transaction 2000000
real	0m59.687s
user	0m11.593s
sys	0m4.100s
-rw------- 1 rusty rusty 752004064 2011-04-27 21:14 /tmp/growtdb.tdb
testing with 3 processes, 5000 loops, seed=0
OK

real	1m17.738s
user	0m0.348s
sys	0m0.580s
-rw------- 1 rusty rusty 663360 2011-04-27 21:15 torture.tdb
Adding 2000000 records:  926 ns (110556088 bytes)
Finding 2000000 records:  592 ns (110556088 bytes)
Missing 2000000 records:  416 ns (110556088 bytes)
Traversing 2000000 records:  422 ns (110556088 bytes)
Deleting 2000000 records:  741 ns (244003768 bytes)
Re-adding 2000000 records:  799 ns (244003768 bytes)
Appending 2000000 records:  1147 ns (295244592 bytes)
Churning 2000000 records:  1827 ns (568411440 bytes)

We currently start walking the free list again when we coalesce any record;
this is overzealous, as we only care about the next record being blatted,
or the record we currently consider "best".

We can also opportunistically try to add the coalesced record into the
new free list: if it fails, we go back to the old "mark record,
unlock, re-lock" code.

Before:
$ time ./growtdb-bench 250000 10 > /dev/null && ls -l /tmp/growtdb.tdb && time ./tdbtorture -s 0 && ls -l torture.tdb && ./speed --transaction 2000000
real	1m0.243s
user	0m13.677s
sys	0m4.336s
-rw------- 1 rusty rusty 683302864 2011-04-27 21:03 /tmp/growtdb.tdb
testing with 3 processes, 5000 loops, seed=0
OK

real	1m24.074s
user	0m0.344s
sys	0m0.468s
-rw------- 1 rusty rusty 836040 2011-04-27 21:04 torture.tdb
Adding 2000000 records:  1015 ns (110551992 bytes)
Finding 2000000 records:  641 ns (110551992 bytes)
Missing 2000000 records:  445 ns (110551992 bytes)
Traversing 2000000 records:  439 ns (110551992 bytes)
Deleting 2000000 records:  807 ns (199517112 bytes)
Re-adding 2000000 records:  851 ns (199517112 bytes)
Appending 2000000 records:  1301 ns (376542552 bytes)
Churning 2000000 records:  2423 ns (553641304 bytes)

After:
$ time ./growtdb-bench 250000 10 > /dev/null && ls -l /tmp/growtdb.tdb && time ./tdbtorture -s 0 && ls -l torture.tdb && ./speed --transaction 2000000
real	0m57.403s
user	0m11.361s
sys	0m4.056s
-rw------- 1 rusty rusty 689536976 2011-04-27 21:10 /tmp/growtdb.tdb
testing with 3 processes, 5000 loops, seed=0
OK

real	1m24.901s
user	0m0.380s
sys	0m0.512s
-rw------- 1 rusty rusty 655368 2011-04-27 21:12 torture.tdb
Adding 2000000 records:  941 ns (110551992 bytes)
Finding 2000000 records:  603 ns (110551992 bytes)
Missing 2000000 records:  428 ns (110551992 bytes)
Traversing 2000000 records:  416 ns (110551992 bytes)
Deleting 2000000 records:  741 ns (199517112 bytes)
Re-adding 2000000 records:  819 ns (199517112 bytes)
Appending 2000000 records:  1228 ns (376542552 bytes)
Churning 2000000 records:  2042 ns (553641304 bytes)

This makes life easier for the next patch.

We took the original expansion heuristic from TDB1, and they just
fixed theirs, so copy that.

Before:

After:
time ./growtdb-bench 250000 10 > /dev/null && ls -l /tmp/growtdb.tdb && time ./tdbtorture -s 0 && ls -l torture.tdb && ./speed --transaction 2000000
growtdb-bench.c: In function ‘main’:
growtdb-bench.c:74:8: warning: ignoring return value of ‘system’, declared with attribute warn_unused_result
growtdb-bench.c:108:9: warning: ignoring return value of ‘system’, declared with attribute warn_unused_result

real	1m0.243s
user	0m13.677s
sys	0m4.336s
-rw------- 1 rusty rusty 683302864 2011-04-27 21:03 /tmp/growtdb.tdb
testing with 3 processes, 5000 loops, seed=0
OK

real	1m24.074s
user	0m0.344s
sys	0m0.468s
-rw------- 1 rusty rusty 836040 2011-04-27 21:04 torture.tdb
Adding 2000000 records:  1015 ns (110551992 bytes)
Finding 2000000 records:  641 ns (110551992 bytes)
Missing 2000000 records:  445 ns (110551992 bytes)
Traversing 2000000 records:  439 ns (110551992 bytes)
Deleting 2000000 records:  807 ns (199517112 bytes)
Re-adding 2000000 records:  851 ns (199517112 bytes)
Appending 2000000 records:  1301 ns (376542552 bytes)
Churning 2000000 records:  2423 ns (553641304 bytes)

And testing reveals a latent bug on 32 bit systems.

This is definitely a bad idea in general, but SAMBA uses nested transactions
in many and varied ways (some of them probably reflect real bugs) and it's
far easier to support them inside tdb2 with a flag.

We already have part of the TDB1 infrastructure in place, so this patch
just completes it and fixes one place where I'd messed it up.

It's probably not a good idea, because it's a recipe for deadlocks if
anyone else grabs any *other* two chainlocks, or the allrecord lock,
but SAMBA definitely does it, so allow it as TDB1 does.

Now we have tdb_get_attribute, it makes sense to make that the method
of accessing statistics.  That way they are always available, and it's
probably cheaper doing the direct increment than even the unlikely()
branch.

Otherwise tdb_name() can be NULL in log functions.  And we might as
well allocate it with the tdb, as well.

It makes sense for some attributes to be manipulated after tdb_open,
so allow that.

This allows overriding of low-level locking functions.  This allows
special effects such as non-blocking operations, and lock proxying.

We rename some local lock vars to l to avoid -Wshadow warnings.

And don't double-log.  Both of these cause problems if we want to do
tdb_transaction_prepare_commit non-blocking (and have it fail so we can
try again).

This allows the caller to implement clear-if-first semantics as per
TDB1.  The flag was removed for good reasons: performance and
unreliability, but SAMBA3 still uses it widely, so this allows them to
reimplement it themselves.

(There is no way to do it without help like this from tdb2, since it has
 to be done under the open lock).

1) The logging function needs to append a \n.
2) The transaction start code should be after the comment and print.
3) We should run tdb_check to make sure the database is OK after each op.

Also, rename private logfn to log_fn for consistency with other members.

We use underscores everywhere else, so be consistent.

It's redundant calling hash.hash_fn for example.  Standardize on fn
and data as names (private conflicts with C++).

This gives us more locks for future use, plus allows a clear-if-first-style
hack to be implemented.

1) Fix the bogus reporting of uncoalesced runs: there has to be more than 1
   free record to make a "run", and any other record interrups the run.
2) Fix the bogus data count in the top line (which was number of records,
   not bytes).
3) Remove the count of free buckets: it's now a constant.

Douglas Bagnell noted that it didn't specify.