1. pthread_join should allow retval=NULL, in which case the return value is
ignored. We therefore need to use void**, not void*&, otherwise passing NULL
causes pthread_join to crash in a strange way.

2. pthread_join forgot to delete the object allocated in pthread_create

mmap/munmap and its cousins mmu::map_*() used to leak the 48-byte "vma"
object, twice. The idea to separate reserve() and map() was a good one,
but it was implemented incorrectly causing a vma object to be allocated
twice per mmap, and both of them were leaked when evacuate() didn't free
them.

So switched to a simpler implementation, where the "vma" object is internal
to mmu.cc, and not used by any of the callers; The struct vma is now properly
allocated only once per mmap(), and freed on munmap().

We used rbtree::insert_unique(), which only inserts if the value if unique.
But uniqueness is defined in terms of the ordering function, which compares
vruntime, not object identity.  The result is that if another thread was
queued with exactly the same vruntime, then the new thread is not queued.

Fix by using rbtree:insert_equal(), which inserts unconditionally.

Fixes the callout unit test.

creates 2 threads, a client and a server that perform the tcp echo service,
complete 400 cycles of {accept(),read(),write(),close()} on the server
and {socket(),connect(),write(),read(),close()} in the client.

1. fix many races
2. use a mutex to protect access to callout data structure
3. delete callout->c_state and use the original callout->c-flags
   also update CALLOUT_PENDING and CALLOUT_ACTIVE accordingly.
4. support correct behaviour for callout_pending(), callout_active() and
   callout_deactivate()
5. implement a mutex version: callout_init_mtx() used by tcp syncache
6. add debug prints

these are used by TCP.

connect() fails due to these missing lines.

log lines weren't reaching the console in a synchronized way

previousely, the len argument weren't handled correctly, it should be returned
to the user and the sockaddr salen should be initialized with the api when
possible.

the rwlock stub were written before our mutex were recursive, now
we just don't need any extra recursive layer in rwlock

move have more work mark before the handler, during the handler execution
another thread may signal the netisr worker thread, avoids a race.

SOCK_RAW should be used with sosend() and not sosend_dgram(),
same for soreceive_dgram()

The bufring was never used by any driver in OSv so it's
not useful

tst-leak.so now enables leak detection, runs a few things of interest,
and disables leak detection. So "osv leak show" after running this test
will show any leaks. I'm using this to confirm fixing the leaks I'm now
trying to fix.

We want to reschedule if we woke a local thread, since it may have higher
priority than us.  However commit 31572c85 broke this by changing the
preemption counter.

Fix by adjusting the check to take this into account.

Noted by Nadav.

Saves ~17MB.

handle_incoming_wakeups() may not be called from preemptible context,
since it manipulates per-cpu variables.

Fix by eliminating these calls.  In one call site, the call is removed,
since it will be called immediately afterwards with interrupts disabled.
In the other call site, push the call into an existing irq disabled region.

Fixes livelocks where a thread is placed into an incoming_wakeups queue
but has the wrong bit set in incoming_wakeups_mask.

Since wake() manipulates per-cpu variables, we need to disable preemption
so the cpu pointers aren't invalidated by migration.

The recording mechanism aligns trace messages; replay needs to follow suit.