embarrassing errors ;-)

This allows taking BIO_ERROR into the block layer core and remove unsafe
bio_flags manipulations in the virtio-blk driver.

 1. A lock-free multiple-producer, single-consider, queue
 2. A lock-free mutex using the above queue.

In a new directory include/lockfree/.

The lock-free mutex is based on a technique called "responsibility hand-off"
published by Anders Gidenstam and Marina Papatriantafilou in 2007, and while
I spent the last two days thinking (and nearly losing my mind ;-)) of all the
weird corner cases that happen in this algorithm, I haven't actually tested
this implementation, which I need to do next.

The code is also still very messy - everything is in the .hh (some should
move to a .cc), it's C++ with no C interface, and the comments are various
thoughts I had trying to convince myself that this algorithm is correct.

So I need to write tests and clean up this code.

* 'master' of github.com:cloudius-systems/osv:
  sched: copy the tls image to a new thread, instead of zeroing it

With the current memset(), every thread starts out with zero-initialized
tls variables.

Switch to memcpy(), so it gets the proper static initializer.

Fixes conf-preempt=0.

Integrate virtio-net to the tx network stack. Register the interface into the FreeBsd stack and handle tx packets. At the moment the previous tx test function still exist but it should eventually dropped. Along the way remove few compilation hurdles w/ the rest of the system.

Move osv/bio.h into the .cc implementation and along with it move the virtio_blk_req structure that can be hidden from the header.

Rename all 'log' functions of class log into 'wrt'. The reason is that BSD has a #define log line and there was a clash

SIGSEGV delivery

Conflicts:
	tests/bench/com/cloudiussystems/bench/Bench.java

As an actual benchmark, it's more or less worthless, since the JVM sees
a lot of SEGVs and inserts an inline check eventually, so the test takes
1 ns per iteration.

This only implements SIGSEGv delivery to the thread that triggered it
(i.e., it must be unblocked).

(4K) pages to back this mapping. Now, instead, if the range is large
enough to contain huge pages (2M), we allocate huge pages to back those
up.

This patch improved performance of the sieve test (where a 200MB array
is allocated and skipped inside) by 13%, bringing performance to within
1% of unvirtualized performance (which was also using THP).

aligned and contiguous huge-page of a given size.

length N and runs the simple Sieve of Eratosthenes (of finding all the
prime numbers below N), which jumps around the array at various strides
and therefore exercises, among other things, the TLB, and can be effected
by whether we use huge pages or normal pages for storing the array.

To allow running this benchmark without increasing the VM's memory,
main() uses N=100,000. But to really see the effect of huge pages,
try N=200,000,000 (commented out in main()).

virtio_device is now a simple set of accessors around pci::device.  Since
the only users of virtio_device are virtio_driver and its subclasses, we
can simply fold it into virtio_driver, eliminating the middleman.

Resolves two FIXMEs where we leaked virtio_device objects created during
probing.