Add _netserver_started to judge if need setup netserver during test.
Signed-off-by: NYiqiao Pu <ypu@redhat.com>

Now the guest statys collect functions are implemented and tested with
RHEL and Fedora guests. It may cause problem for other guests. So set
it default to no, only enable it for RHEL and Fedora guests.
Signed-off-by: NYiqiao Pu <ypu@redhat.com>

Signed-off-by: NLucas Meneghel Rodrigues <lmr@redhat.com>

Sometime the output from server include some unexpect letters include
some numbers. So the old method will failed to get the correct cpu
number.
Signed-off-by: NYiqiao Pu <ypu@redhat.com>

Signed-off-by: NAmos Kong <akong@redhat.com>

Currently we still wait client thread to stop after effective test
finishes, it wastes of time. This patch directly killes the threads,
then client_thread.join() will return immediately.
Suggested-by: NWenli Quan <wquan@redhat.com>
Signed-off-by: NAmos Kong <akong@redhat.com>

We do rounding in regression.py, so here we adjust the default
fbase to 5
Signed-off-by: NAmos Kong <akong@redhat.com>

User wants to collect sysinfo of host/guest, but different users
might have different requests, and many kinds of system might be
used. So I would like to use custom-built scripts to collect the
information that will make users happy.

User needs to add shell scripts to autotest/client/tests/virt/shared/scripts/,
and set the script path in config file.

eg:
 #log_hostinfo_script = scripts/rh_perf_log_hostinfo_script.sh
 #log_guestinfo_script = scripts/rh_perf_log_guestinfo_script.sh
 #log_guestinfo_exec = bash
 #log_guestinfo_path = /tmp/log_guestinfo.sh

Then the script output will be saved to $test.resultsdir/sysinfo,
analyzer can use 'diff' to compare the system information when
regression occurs.

I will send a pull request to add to example scripts to virt-tests repo.
Signed-off-by: NAmos Kong <akong@redhat.com>

  ethtool -L $ethname combined $N

  todo: enable multiqueue for Windows guest
Signed-off-by: NAmos Kong <akong@redhat.com>

When we enable multiple queue, there are multiple virtio
input/output interrupts in guest.

This patch counts interrupt for each queue, and output
a summary in result.

This patch also moves the volatile items to the end of
record list. Then the avg_update in perf.cfg should be
updated.
Signed-off-by: NAmos Kong <akong@redhat.com>

commit a84a267d fixed one typo, another typo also exists. It caused that
'rpkt_per_irq' and 'tpkt_per_exit' were lost in the result.
Signed-off-by: NAmos Kong <akong@redhat.com>

commit [1] introduced a regression, it wrongly update the server_ip
to ip address of control connection.

We have two nics, the first (virtio-net) is data connection, the second
nic (non virtio-net) is control connection.

[1] https://github.com/kongove/virt-test/commit/63ada005Signed-off-by: NAmos Kong <akong@redhat.com>

Signed-off-by: NAmos Kong <akong@redhat.com>

Some times when boot guest with two nics, the second nic can not
get ip address.
The wait_for_get_address will try to renew guest ip and try to get
guest ip via serial console. it is more safe than get_address
Signed-off-by: NYunping Zheng <yunzheng@redhat.com>

Signed-off-by: NLucas Meneghel Rodrigues <lmr@redhat.com>

The largest timeout_netperf_start should be int(l) * 0.5,
if netperf threads start all at once, it will still need
int(l) * 0.5 time waste. Therefore, the netperf_start_timeout
overhead is not needed.
Signed-off-by: NWenli Quan <wquan@redhat.com>

During the review of the repository split, it was pointed
out that 'agnostic' was a rather strange word, so we'll go
with 'generic'. I opted for not rebasing the PR, since I
had many problems when trying to do so.
Signed-off-by: NLucas Meneghel Rodrigues <lmr@redhat.com>

This will make it easier when we're splitting these
tests into their own repositories.
Signed-off-by: NLucas Meneghel Rodrigues <lmr@redhat.com>

There's no need to have the name of the "run_<name>" method (main
test entry point) depend on the test name. So, for simplicity sake,
let's just name the test entry point "run".

I have decided to not fallback gracefully to the previous naming
convention as I believe this can lead to confusion in the future.
Signed-off-by: NCleber Rosa <crosa@redhat.com>

There is a background tcpdump thread that records DHCP leases
in the env["address_cache"] dictionary. Sometimes, this thread
might write to this dictionary while the environment object is
being saved, triggering a RuntimeError in virt-test.

So the solution involves:

1) Creating a save lock, (a threading reentrant lock object),
that is acquired by the env.save() method when it wants to
save env.data to a pickle file, and for methods that alter
env.data.

2) Move the tcpdump thread to a env object attribute, rather
than one of the keys of env.data, given that the lock object
can't be pickled.

With the implementation of env.save_lock, we can also fix
the problem identified by the locking unittest. The base of
the solution was Pavel's idea, including a good deal of the
implementation, so kudos to him.
Signed-off-by: NPavel Hrdina <phrdina@redhat.com>
Signed-off-by: NLucas Meneghel Rodrigues <lmr@redhat.com>

We compute rpkt_per_exit by tx_pkts, and compute tpkt_per_exit
by rx_pkts. It's wrong.
Signed-off-by: NAmos Kong <akong@redhat.com>

I have noticed this function was QEMU specific only
afterwards.
Signed-off-by: NLucas Meneghel Rodrigues <lmr@redhat.com>

This was probably overlooked when making the transition
to the new environment class. With this, we can simply
remove the old API.
Signed-off-by: NLucas Meneghel Rodrigues <lmr@redhat.com>

Get netperf src package from web;
Make the script compatible with netperf 2.6.0;
Modify the method of get winutils vol;
Modify the method of start netserver in windows guest
Using wait_for  to wait until all net perf clients start to work;
Reorganize the configure file.
Signed-off-by: NYunping Zheng <yunzheng@redhat.com>

Make virt-test to have a docstring style mostly compliant with
the sphinx's [1] format, reStructured text. This means that
we have to trade all '@' tags to be ':', mostly.
Signed-off-by: NLucas Meneghel Rodrigues <lmr@redhat.com>

With the help of the PEP8 script, a semi-automated clean
of the code base was made, and effectively changing some
of the coding style rules of our code base:

 * No more 2 lines spacing among functions
 * No more combined module imports

In the end, it was going to be too difficult to take
into consideration E501 (line length > 80) errors and
W601 (.has_key() is deprecated), so these were ignored,
at least for now.
Signed-off-by: NLucas Meneghel Rodrigues <lmr@redhat.com>

In some host 'service iptables stop' command exit with error code 5.
It will fail our case.
Signed-off-by: NFeng Yang <fyang@redhat.com>

this patch:
1.fix the issue netperf cannot login with two nics,
2.add  netperf test support tcp_crr protocol
3.add an automated way to get the cdrom drive number.
4.modify the file path of netperf.tar.gz. now should
put the src file under "virt/shared/deps"
5.add more step info

Process some output value before use them to igonre the problem caused
by different guest os version and session setup.
Signed-off-by: NYiqiao Pu <ypu@redhat.com>

This will help to remove the blank lines in the results file. And
make the format suitable for further result analyze.
Signed-off-by: NYiqiao Pu <ypu@redhat.com>
Acked-by: NFeng Yang <fyang@redhat.com>

Currently we use many threads/heavy control sessions to launch many
clients, Autotest takes too much of the host resources.

This patch introduces an agent script, it'll be executed inside the
client system to launch multiple sessions, and Autotest only needs
to use one background thread to control clients.

We can consider to process the raw results in this agent script in
future, if it can help to simplify the case.
Signed-off-by: NAmos Kong <akong@redhat.com>

This patch was wrongly dropped in original refactor.

Tcpdump brings extra latency, it causes the guest to batch
more bytes per packet. We should stop tcpdump to drop the
effect, it will be restarted at the beginning of next case.
Reported-by: NWenli Quan <wquan@redhat.com>
Signed-off-by: NAmos Kong <akong@redhat.com>
Acked-by: NWenli Quan <wquan@redhat.com>

Result file was not copied from remote machine to localhost
when checking if all threads work, this patch fixed it.
And add one more client session for controlling.
Signed-off-by: NAmos Kong <akong@redhat.com>

In order to be consistent with previous changes.
Signed-off-by: NLucas Meneghel Rodrigues <lmr@redhat.com>

Enable demo mode of netperf, process mixed demo results to drop
unstable factors, only compute the effective demo results when
all the threads work. This would make performance result more
stable.

We start many clients and watch the demo output, if all the threads
really work, then start to record the interrupt/packet number, and
count the cpu usage. After a fixed test time, record the final
interrupt/packet number.
Signed-off-by: NAmos Kong <akong@redhat.com>

Currently we use '-b' parameter of netperf to test parallel traffic,
client would send many requests at start of _RR tests, but it's not
real parallelism, we want to use demo mode to improve the
stabilization. This patch removes '-b' parameter and uses multiple
threads for TCP_RR testing.
Signed-off-by: NAmos Kong <akong@redhat.com>

Rename 'client' to 'clients'
Signed-off-by: NAmos Kong <akong@redhat.com>

In netperf-2.6.0, we can assign two arguments (local send size &
remote send size) for TCP_MAERTS by '-m $local_send_size,remote_send_size'

Currently we only assigned useless local send size, this patch corrects
the parameters for TCP_MAERTS.
Signed-off-by: NAmos Kong <akong@redhat.com>

I already updated the netperf package to 2.6.0,
this patch updated the config file and file path.

wait_before_data.patch is no longer requested.
Signed-off-by: NAmos Kong <akong@redhat.com>

Signed-off-by: NLucas Meneghel Rodrigues <lmr@redhat.com>