drivers/libtsm includes a bunch of Doxygen comments, and worst of all, a
"@mainpage" tag controlling Doxygen's main page. Since libtsm isn't really
part of the API we intend to advertise, let's exclude this directory from
being scanned by doxygen.

Signed-off-by: Nadav Har'El <nyh@cloudius-systems.com>
Signed-off-by: Pekka Enberg <penberg@cloudius-systems.com>

Don Martin writes:

  I'm experimenting with Capstan, trying to rule out some things for
  getting a bigger application working. I have a simple program to
  recursively list directory contents, here...

    https://github.com/dmarti/mini-ls



  It works on the host, but when I build the .so, the "capstan build" step
  works, and then when I do "capstan run"...

  $ capstan run
  OSv v0.05-501-gdbdf4f7
  Failed looking up symbol alphasort

This is missing libc functionality in OSv. To fix that, add the
alphasort function from musl.

Signed-off-by: Pekka Enberg <penberg@cloudius-systems.com>

When we discard a bad buffer, we don't advance to the next one, so we'll free
it again on the next pass, resulting in a double-free.

Tested-by: Asias He <asias@cloudius-systems.com>
Signed-off-by: Avi Kivity <avi@cloudius-systems.com>
Signed-off-by: Pekka Enberg <penberg@cloudius-systems.com>

The API is listening to port 8000, not 8080.

Reviewed-by: Nadav Har'El <nyh@cloudius-systems.com>
Signed-off-by: Pekka Enberg <penberg@cloudius-systems.com>

Reviewed-by: Nadav Har'El <nyh@cloudius-systems.com>
Signed-off-by: Tomasz Grabiec <tgrabiec@cloudius-systems.com>
Signed-off-by: Pekka Enberg <penberg@cloudius-systems.com>

Add similar function of the following route command
in Freebsd to OSv.

   $ route add -host $host_ip -interface $interface_name
   e.g.
   $ route add -host 10.240.0.1 -interface eth0

This is useful when the gateway is not reachable within the network
specified by the interface. For instance, GCE's DHCP server provides:

   ip=10.240.10.15, netmask=255.255.255.255, gw=10.240.10.1

A host route entry is needed before we can add the default route entry.

With this patch, the route table looks like:

   Destination         Gateway             Flags  Netif
   default             10.240.0.1          UGS    eth0
   10.240.0.1          42:01:0a:f0:e0:c5   UHS    eth0
   10.240.10.15        link#2              U      eth0
   10.240.10.15        link#2              UHS    lo0
   127.0.0.1           link#1              UH     lo0

Signed-off-by: Asias He <asias@cloudius-systems.com>
Signed-off-by: Avi Kivity <avi@cloudius-systems.com>

Fixes #198.

Signed-off-by: Sanidhya Kashyap <sanidhya.k@gmail.com>
Signed-off-by: Pekka Enberg <penberg@cloudius-systems.com>

Fixes #198.

Signed-off-by: Sanidhya Kashyap <sanidhya.k@gmail.com>
Signed-off-by: Pekka Enberg <penberg@cloudius-systems.com>

tcp_close() related bug fix
switch virtio to page buffers

Signed-off-by: Avi Kivity <avi@cloudius-systems.com>

Python3 no longer has "long" type, and what was "long" on python2 is
now "int". To make the code portable commit
eb5eb910 switched long() conversions
to int() conversion, which should be equivalent. It turns out that
GDB's implementation of gdb.Value, when compiled against python 2, has
a slight difference between int() and long() conversions. The long
conversion can be used on pointers whereas int conversion cannot. I
don't about any reason it should be like this, maybe it is a GDB bug:

This is int() handler:

      if (!is_intlike (type, 0))
	error (_("Cannot convert value to int."));
      l = value_as_long (value);

This is long() handler:

      if (!is_intlike (type, 1))
	error (_("Cannot convert value to long."));
      l = value_as_long (value);

Note the second argument to is_intlike() differs between variants, "1"
means "pointers allowed". Since on python3 int() conversion calls the
long handler, it works as expected. As a workaround for this I defined
to_int() function which performs the conversion in a portable way. This
needs to be done only when the value may be a pointer.

Places which check:

 if int(x) != 0:

can be simplified to:

 if x:

which works with both pointers and numbers in both versions.

Reported-by: Raphael S. Carvalho <raphaelsc@cloudius-systems.com>
Signed-off-by: Tomasz Grabiec <tgrabiec@cloudius-systems.com>
Signed-off-by: Avi Kivity <avi@cloudius-systems.com>

"
System V shared memory implementation

It looks like fio needs it although with single address space any memory
is shared anyway.
"

Signed-off-by: Avi Kivity <avi@cloudius-systems.com>

Signed-off-by: Gleb Natapov <gleb@cloudius-systems.com>

Implement shmget/shmat/shmdt/shmctl function required by System V shared
memory interface.

Signed-off-by: Gleb Natapov <gleb@cloudius-systems.com>

shmdt() is defined to get const pointer. Next patch implements shmdt
using unmap(), so const qualifier either need to be casted off or
unmap() and friends need to get const pointer as a parameter too. This
patch implement the later.

Signed-off-by: Gleb Natapov <gleb@cloudius-systems.com>

shmfd is a pseudo file that allows sharing of anonymous memory. Each
mapping for this file points to the same memory. Only huge pages are
supported for now.

Signed-off-by: Gleb Natapov <gleb@cloudius-systems.com>

Implement page_allocator specialization that delegates page
allocation/deallocation to mmaped file. It will be used to mmap files
that support get_page/put_page calls. Others fall back to old way of
reading file.

Signed-off-by: Gleb Natapov <gleb@cloudius-systems.com>

map_page_ops is used by page mapper to managed individual page
allocations and freeing, but the name is not reflects that.
page_allocator describes better the purpose of the structure.

Signed-off-by: Gleb Natapov <gleb@cloudius-systems.com>

Some kind of files may want to handle mmap differently. By delegating
file_vma creation to the file class it is possible for each file to
create customized vma.

Signed-off-by: Gleb Natapov <gleb@cloudius-systems.com>

"Scsi fix and improvements".

Signed-off-by: Avi Kivity <avi@cloudius-systems.com>

It will be used outsize of mmu code.

Signed-off-by: Gleb Natapov <gleb@cloudius-systems.com>

Curently virtio uses MCLBYTES (2KB) buffers for virtio.  This hits a weakness
in our UMA and memory allocation implementations:

- UMA appends a 4-byte reference count to the allocation, resulting in an
  2052 byte allocation request
- Our allocator only services requests up to 2048 bytes from the mempool
  allocator; larger requests are serviced by the large allocator, which is
  significately slower, and is also serialized by a mutex.

Switch to an EXT_EXTREF mbuf with the buffer and refcount allocated
separately.  This also allows us to switch to 4K allocations serviced by
the page allocator (memory::alloc_page()), which is our fastest allocator,
and halves the lengths of the mbuf chains on streaming workloads.

Increases iperf stream (host->guest) results from ~12Gbps to ~34Gbps on my
desktop.

Signed-off-by: Avi Kivity <avi@cloudius-systems.com>

tcp_do_segment() has a nasty habit of sometimes closing the tcp connection.
This plays havoc with the locking rules (as updated for net channels), since
tcp_do_segment() is supposed to return with the lock held, yet tcp_close()
will unlock if the connection is disassociated from a socket.

Fix by having tcp_do_segment() not close the connection, and instead indicate
its desire to do so to its caller.

Signed-off-by: Avi Kivity <avi@cloudius-systems.com>

Currently it fails with:

Traceback (most rTraceback (most recent call last):
  File "scripts/trace.py", line 433, in <module>
    args = parser.parse_args()
...
  File "/usr/lib/python2.7/argparse.py", line 601, in _expand_help
    return self._get_help_string(action) % params
ValueError: incomplete format

Signed-off-by: Tomasz Grabiec <tgrabiec@cloudius-systems.com>
Signed-off-by: Pekka Enberg <penberg@cloudius-systems.com>

Signed-off-by: Avi Kivity <avi@cloudius-systems.com>

"This series fixes problems that we see with Glaubers glommer/zfsbuffers
branch. The branch has code that maps read only pages directly from zfs
ARC cache. During heavy testing there are a lot of tlb flushes and page
faults and at some point test fails with stack corruption."

Signed-off-by: Avi Kivity <avi@cloudius-systems.com>

Page fault handler may call into zfs and in some cases this causes
exception stack overflow. Enlarge exception stack to 16k to prevent
overflow.

Signed-off-by: Gleb Natapov <gleb@cloudius-systems.com>

Interrupt may happen after exception stack is switched, but before
exception handler returns. If interrupt handler will causes exception
it will overwrite current stack.

Signed-off-by: Gleb Natapov <gleb@cloudius-systems.com>

switch_to() always sets exception pointer in tss to primary exception
stack, so when scheduling happens from an exception nested stack
configuration is lost. Fix it by saving and restoring current exception
stack from tss.

Signed-off-by: Gleb Natapov <gleb@cloudius-systems.com>

Signed-off-by: Avi Kivity <avi@cloudius-systems.com>

Signed-off-by: Avi Kivity <avi@cloudius-systems.com>

This patch fixes "perf stat memory_malloc". This used to crash: When the
tracepoint was first reached, we called per_cpu_counter::increment();
If this was the first time this function was called, the lazy symbol
resolution caused program::resolve() to be called, which among other things
allocates memory again - yielding endless recursion.

The fix is to call per_cpu_counter::increment() once, on some silly junk
counter, before setting the tracepoint probe. This will ensure that this
function will have already been resolved by the time it is called in an
actual tracepoint.

Fixes #239.

Signed-off-by: Nadav Har'El <nyh@cloudius-systems.com>
Signed-off-by: Avi Kivity <avi@cloudius-systems.com>

This line is useless.

Signed-off-by: Asias He <asias@cloudius-systems.com>

Writing to PORT_SACT is need for NCQ cmd only.

Signed-off-by: Asias He <asias@cloudius-systems.com>

Currently, only one cmd will be pending at most. The device is always
ready when we send the next cmd.

Signed-off-by: Asias He <asias@cloudius-systems.com>

This simplifies the cmd completion code a bit.

Signed-off-by: Asias He <asias@cloudius-systems.com>

This is a preparation for next patch. We will need the slot number in
wait_cmd_poll and wait_cmd_irq.

Signed-off-by: Asias He <asias@cloudius-systems.com>

Instead of wait for hardcoded slot 0, add slot parameter.

Signed-off-by: Asias He <asias@cloudius-systems.com>

It is cumbersome to have to type "file" on every trace command. Let's
use "tracefile" by default.

Before:

 trace extract file
 trace summary file
 trace prof file

After:

 trace extract
 trace summary
 trace prof

Signed-off-by: Tomasz Grabiec <tgrabiec@cloudius-systems.com>
Signed-off-by: Pekka Enberg <penberg@cloudius-systems.com>

Instead of:

  trace list | less

just do:

  trace list

Signed-off-by: Tomasz Grabiec <tgrabiec@cloudius-systems.com>
Signed-off-by: Pekka Enberg <penberg@cloudius-systems.com>

Often profile has only one tracepoint, let's save users some typing by
showing all by default.

Before:

  trace prof -t sched_wait

After:

  trace prof

Signed-off-by: Tomasz Grabiec <tgrabiec@cloudius-systems.com>
Signed-off-by: Pekka Enberg <penberg@cloudius-systems.com>