Sunday, August 6, 2017

Creating OpenBSD guest with libvirt/bhyve

I've been asked a couple of times how to setup OpenBSD guest with libvirt/bhyve. It's fairly straight-forward, but minimal libvirt version required for that is 3.5.0 because it's the first version that allows to specify vgaconf settings.

First, download the installation image: https://ftp.openbsd.org/pub/OpenBSD/6.1/amd64/install61.fs and create a disk image:

truncate -s 2G openbsd61.raw

Create an initial XML (pay attention to the vgaconf='off' bit):

Now with virsh do:

virsh # define /path/to/openbsd61.xml
Domain openbsd61 defined from /path/to/openbsd61.xml

virsh # start openbsd61
Domain openbsd61 started

virsh # vncdisplay openbsd61                                                                                                                                                                                       
127.0.0.1:0                                                                                                                                                                                                        

Using VNC client connect to the guest at the specified port:

vncviewer :0

This gives a boot loader prompt. For some reason, it doesn't set root device properly, so I have to do:

set device hd1a
boot

Then there'll be a number of installer questions (where I usually accept defaults) until it asks what disk should be used as a root device. It's important to check this carefully and choose the right one (you can find it by its size for example), otherwise it might just try to override the install image.

When the installation finishes, shutdown the guest, go back to virsh and edit the domain xml:

virsh # edit openbsd61

And remove the disk block corresponding to install61.fs. And then start domain again:

virsh # start openbsd61   
Domain openbsd61 started  

Now you should be able to connect to the guest via VNC or via SSH.

PS My sample xml uses the autoport feature that will be available only in libvirt 3.7.0. If you're using an earlier version, specify the VNC port explicitly:

<graphics type='vnc' port='5904'>

Sunday, May 21, 2017

Profiling OpenStack Horizon

OpenStack Horizon is a web based interface for managing OpenStack clouds. Just like many other OpenStack components, it's implemented in Python and uses the Django framework.

Sometimes you might notice that Horizon is sluggish and slow and it's not obvious what causes this. Natural thing to do in this case is to profile it to see problematic points.

There are at least two possible ways to profile that:

  • Use osprofiler, an "official" OpenStack Profiler,
  • Use standard Python profiling facilities like cProfile.

I recommend using the latter and I'll elaborate on that a little later.

Preparation

I think it's a good idea to profile Horizon in the same environment where you run your stage Horizon, i.e. pick one of the nodes where you have Horizon installed already (e.g. configured to be served using Apache) and bring your own instance of Horizon with the same configuration, but on a different port and using Django's built in http server.

For that purpose, you can copy over Horizon sources somewhere to your home directory (either an appropriate branch from git, or maybe grab sources from the package you use, that depends on your processes), copy over the existing config files (e.g. local_settings.py) and from top level source directory of horizon run:

~/horizon$ ./manage.py runserver 0.0.0.0:8000

This way your brand new Horizon instance will be available on port 8000. Also, you might want to check iptables configuration that it actually allows connections to this port.

Note: We're not using virtualenv because as we're running this on a Horizon node, we have all the dependencies installed via system packages.

Once this instance becomes accessible, we can start actual debugging.

Profiling

There's a wonderful Django Extensions package that makes Django profiling very easy. That could be installed via packages, e.g. on Ubuntu you can do:

#  apt-get install python-django-extensions

It could be installed using pip too in case your distro misses this package:

# pip install django-extensions

After installing the package, it's required to include it to the Horizon's INSTALLED_APPS list. While in the local copy of Horizon, open openstack_dashboard/settings.py, find INSTALLED_APPS and add 'django_extensions' to this list. The resulting setting should look something like this:

INSTALLED_APPS = [
    'openstack_dashboard',
    'sbr_reports',
    'sbr_log_processing',
    'django.contrib.contenttypes',
    'django.contrib.auth',
    'django.contrib.sessions',
    'django.contrib.messages',
    'django.contrib.staticfiles',
    'django.contrib.humanize',
    'django_pyscss',
    'openstack_dashboard.django_pyscss_fix',
    'compressor',
    'horizon',
    'openstack_auth',
    'django_extensions', #  for runprofileserver
]

Now it's possible to run the profile server. Terminate the old 'runserver' that was used to test basic configuration, create a directory for the profiling data and start the profiling server:

$ mkdir ~/horizon_profile_data_001
$ ./manage.py runprofileserver --use-cprofile --prof-path=/home/user/horizon_profile_data_001 0.0.0.0:8000

Note: looks like --prof-path doesn't decode "~", so have to pass the full path.

Now Horizon is running in profiling mode, so we can open it up in the browser and navigate to the problematic pages, for example "Admin -> System -> Networks".

Analyzing Results

If everything went file, in ~/horizon_profile_data_001 you'll see something like this:

~/horizon_profile_data_001$ ls |grep networks
admin.networks.001278ms.1495361937.prof
~/horizon_profile_data_001$

Now we can use the pstats module to analyze the results:

python -m pstats admin.networks.001278ms.1495361937.prof

This opens an interactive interface for exploring profiling data. If you're not familiar with that, a good start would be to do:

sort cumtime
stats 40

Here's how it looks for me:

Here you can get an idea how much time was spent in template rendering, Neutron client, etc.

Why not osprofiler

  • osprofiler only traces calls that were decorated with the profiler decorator (example). This means that if some function was not decorated, either intentionally or by mistake, you will not get details about it,
  • It uses JSON-formatted results, but I wasn't able to find a schema or documentation on it,
  • Its CLI is non-intuitive. Specifically, it provides osprofiler trace show command (as per docs), but there's no osprofiler trace list or other command to list traces,
  • It requires to have Mongo installation and therefore is harder to configure,

Also, the documentation stats that osprofiler is different from cProfile in a way that it skips random Python calls, however, it's not hard to strip those from the .prof file (but why would you need to do that though?), and it doesn't look like an advantage to me. Maybe it makes more sense for profiling components like Nova or Neutron though.

Further Reading

Saturday, April 1, 2017

Update on libvirt bhyve driver: UEFI support

New libvirt version 3.2.0 will be released really soon (maybe even before I publish this post), and I'm happy to announce that it brings UEFI support for the bhyve driver.

Specifically, starting with this version it's possible:

  • Point domains to the UEFI boot ROM to boot UEFI guests (instead of using grub-bhyve(1) for example)
  • Connect to guests through VNC (e.g. using vncviewer from net/tigervnc)
  • Use an USB tablet input device

I've updated the official documentation with description of these features and an example, scroll to Example config (Linux UEFI guest, VNC, tablet).

Gotchas: as you might now, bhyve's fbuf device accepts the vga option (e.g. vga=off etc), and different guest OSes might require different settings to make things work. It's not possible to change this using libvirt right now, so the default value (which is vga=io) is used, hence some guests will not work.

Kudos to Fabian Freyer for doing the majority of that work as a part of GSoC 2016.

Sunday, December 4, 2016

Building libvirt from git on macOS

Update 1 (Dec 20th, 2016): a patched version of Apple's rpcgen was committed to homebrew, please use this one instead. Proper command for it is:

./configure ac_cv_path_RPCGEN=$(brew --prefix rpcgen)/bin/rpcgen ...

Update 2 (Dec 21th, 2016): now it's possible to install libvirt's current git version using homebrew (thanks to Justin Clift!) with:

brew install --HEAD libvirt

Considering all that, this blog post doesn't make much sense anymore, just keeping that for sake of history.


Libvirt is available on macOS via homebrew, however, building a git version (rather than from a release tarball) might be a little tricky.

Recently, I've fixed some minor issues concerning building on macOS, so here is a little guide on building that on macOS.

The first thing needed is a relatively fresh rpcgen, as the one that comes with macOS is too old. I've created a homebrew formula for rpcgen from FreeBSD, you can grab it here: https://gist.github.com/novel/0d74cdbc7b71f60640a42b52c9cc1459 and install it. Please consult homebrew docs if you're not sure how to do that.

Now it's just a matter to point it to the proper rpcgen and build it as usual:

$ ./configure ac_cv_path_RPCGEN=/usr/local/bin/bsdrpcgen --without-sasl
$ make
$ make check

There are some tests failing, need to figure out why, but at least test suite is working.

PS The --without-sasl flag is not really necessary, it's just me being lazy to find out why it failed to find proper headers/libs for me at that time.

Sunday, November 13, 2016

Playing around with valgrind in libvirt

Valgrind is a well known memory debugging tool. It doesn't seem to officially support FreeBSD, but, luckily, it's available in FreeBSD ports.

Using it is not hard (at least on a basic level), but requires some patience and attention.

Fortunately, libvirt comes provides an easy interface for adding unit tests, and it could be used for running with valgrind as well. Obviously, tests should cover code paths we want to test, but it's still much more convenient than triggering all the stuff manually.

Individual tests could be executed like this:

VIR_TEST_RANGE=22 VIR_TEST_DEBUG=1 ./tests/bhyvexml2argvtest

bhyvexml2argvtest is a test that checks that the bhyve driver converts XML to a proper set of actual bhyve(8) commands. VIR_TEST_RANGE tests to run only test #22 from the suite, and VIR_TEST_DEBUG=1 makes it produce more verbose output.

Adding valgrind into the game is fairly simple:

VIR_TEST_RANGE=22 VIR_TEST_DEBUG=1 valgrind --trace-children=yes --tool=memcheck --leak-check=full   ./tests/bhyvexml2argvtest

Note: I also configured libvirt with '-O0 -g' in CFLAGS for debugging.

It produces a bunch of information about leaks, for example:

==28853== 128 bytes in 1 blocks are definitely lost in loss record 134 of 173
==28853==    at 0x4C246C0: malloc (in /usr/local/lib/valgrind/vgpreload_memcheck-amd64-freebsd.so)
==28853==    by 0x7D906D8: vasprintf_l (in /lib/libc.so.7)
==28853==    by 0x512876F: virVasprintfInternal (virstring.c:481)
==28853==    by 0x512886D: virAsprintfInternal (virstring.c:502)
==28853==    by 0x40D9A9: testCompareXMLToArgvHelper (bhyvexml2argvtest.c:109)
==28853==    by 0x40DFD8: virTestRun (testutils.c:180)
==28853==    by 0x40DD50: mymain (bhyvexml2argvtest.c:177)
==28853==    by 0x40FA77: virTestMain (testutils.c:992)
==28853==    by 0x40DE0E: main (bhyvexml2argvtest.c:193)
==28853== 
==28853== 128 bytes in 1 blocks are definitely lost in loss record 135 of 173
==28853==    at 0x4C246C0: malloc (in /usr/local/lib/valgrind/vgpreload_memcheck-amd64-freebsd.so)
==28853==    by 0x7D906D8: vasprintf_l (in /lib/libc.so.7)
==28853==    by 0x512876F: virVasprintfInternal (virstring.c:481)
==28853==    by 0x512886D: virAsprintfInternal (virstring.c:502)
==28853==    by 0x40D9F5: testCompareXMLToArgvHelper (bhyvexml2argvtest.c:111)
==28853==    by 0x40DFD8: virTestRun (testutils.c:180)
==28853==    by 0x40DD50: mymain (bhyvexml2argvtest.c:177)
==28853==    by 0x40FA77: virTestMain (testutils.c:992)
==28853==    by 0x40DE0E: main (bhyvexml2argvtest.c:193)

Moving from bottom to top of these traces, it's fair to assume that the issue is somewhere in testCompareXMLToArgvHelper. It looks like this:

 96 
 97 static int
 98 testCompareXMLToArgvHelper(const void *data)
 99 {
100     int ret = -1;
101     const struct testInfo *info = data;
102     char *xml = NULL;
103     char *args = NULL, *ldargs = NULL, *dmargs = NULL;
104 
105     if (virAsprintf(&xml, "%s/bhyvexml2argvdata/bhyvexml2argv-%s.xml",
106                     abs_srcdir, info->name) < 0 ||
107         virAsprintf(&args, "%s/bhyvexml2argvdata/bhyvexml2argv-%s.args",
108                     abs_srcdir, info->name) < 0 ||
109         virAsprintf(&ldargs, "%s/bhyvexml2argvdata/bhyvexml2argv-%s.ldargs",
110                     abs_srcdir, info->name) < 0 ||
111         virAsprintf(&dmargs, "%s/bhyvexml2argvdata/bhyvexml2argv-%s.devmap",
112                     abs_srcdir, info->name) < 0)
113         goto cleanup;
114 
115     ret = testCompareXMLToArgvFiles(xml, args, ldargs, dmargs, info->flags);
116 
117  cleanup:
118     VIR_FREE(xml);
119     VIR_FREE(args);
120     return ret;
121 }

We can see that we call virAsnprintf a few times to populate xml, args, ldargs and dmargs, but later, in the cleanup section, we only free xml and args, leaking ldargs and dmargs, just like valgrind reported.

Now trying to fix that by adding two VIR_FREE calls for leaked variables and try again.

After running the test again, this warning is gone (yay!), so moving to the next one:

==29877== 296 (200 direct, 96 indirect) bytes in 1 blocks are definitely lost in loss record 146 of 171
==29877==    at 0x4C25C90: calloc (in /usr/local/lib/valgrind/vgpreload_memcheck-amd64-freebsd.so)
==29877==    by 0x50C47D8: virAllocVar (viralloc.c:560)
==29877==    by 0x510D3A1: virObjectNew (virobject.c:193)
==29877==    by 0x510D4FC: virObjectLockableNew (virobject.c:219)
==29877==    by 0x51EBFC1: virGetConnect (datatypes.c:127)
==29877==    by 0x40D680: testCompareXMLToArgvFiles (bhyvexml2argvtest.c:34)
==29877==    by 0x40DA1B: testCompareXMLToArgvHelper (bhyvexml2argvtest.c:115)
==29877==    by 0x40DFF0: virTestRun (testutils.c:180)
==29877==    by 0x40DD68: mymain (bhyvexml2argvtest.c:179)
==29877==    by 0x40FA8F: virTestMain (testutils.c:992)
==29877==    by 0x40DE26: main (bhyvexml2argvtest.c:195)
==29877== 

And indeed, it looks like connect object is leaking here:

 22 static int testCompareXMLToArgvFiles(const char *xml,
 23                                      const char *cmdline,
 24                                      const char *ldcmdline,
 25                                      const char *dmcmdline,
 26                                      unsigned int flags)
 27 {
 28     char *actualargv = NULL, *actualld = NULL, *actualdm = NULL;
 29     virDomainDefPtr vmdef = NULL;
 30     virCommandPtr cmd = NULL, ldcmd = NULL;
 31     virConnectPtr conn;
 32     int ret = -1;
 33 
 34     if (!(conn = virGetConnect()))
 35         goto out;
  --- skip ---
 82  out:
 83     VIR_FREE(actualargv);
 84     VIR_FREE(actualld);
 85     VIR_FREE(actualdm);
 86     virCommandFree(cmd);
 87     virCommandFree(ldcmd);
 88     virDomainDefFree(vmdef);
 89     return ret;
 90 }

So adding virObjectUnref call for the connection object and trying again. And after this there's only one warning left:

==30532== 7 bytes in 1 blocks are definitely lost in loss record 6 of 169
==30532==    at 0x4C246C0: malloc (in /usr/local/lib/valgrind/vgpreload_memcheck-amd64-freebsd.so)
==30532==    by 0x7D361DF: strdup (in /lib/libc.so.7)
==30532==    by 0x5128C71: virStrdup (virstring.c:714)
==30532==    by 0x410B8F: bhyveBuildNetArgStr (bhyve_command.c:94)
==30532==    by 0x411410: virBhyveProcessBuildBhyveCmd (bhyve_command.c:305)
==30532==    by 0x40D6FF: testCompareXMLToArgvFiles (bhyvexml2argvtest.c:46)
==30532==    by 0x40DA27: testCompareXMLToArgvHelper (bhyvexml2argvtest.c:116)
==30532==    by 0x40DFFC: virTestRun (testutils.c:180)
==30532==    by 0x40DD74: mymain (bhyvexml2argvtest.c:180)
==30532==    by 0x40FA9B: virTestMain (testutils.c:992)
==30532==    by 0x40DE32: main (bhyvexml2argvtest.c:196)

This one is a little harder. Reading the code of bhyveBuildNetArgStr, it's not exactly obvious where it leaks, because net->ifname was freed in every code path.

 46 static int
 47 bhyveBuildNetArgStr(virConnectPtr conn,
 48                     const virDomainDef *def,
 49                     virDomainNetDefPtr net,
 50                     virCommandPtr cmd,
 51                     bool dryRun)
 52 {
 53     char macaddr[VIR_MAC_STRING_BUFLEN];
 54     char *realifname = NULL;
 55     char *brname = NULL;
 56     char *nic_model = NULL;
 57     int ret = -1;
 58     virDomainNetType actualType = virDomainNetGetActualType(net);
 59 
 60     if (STREQ(net->model, "virtio")) {
 61         if (VIR_STRDUP(nic_model, "virtio-net") < 0)
 62             return -1;
 63     } else if (STREQ(net->model, "e1000")) {
 64         if ((bhyveDriverGetCaps(conn) & BHYVE_CAP_NET_E1000) != 0) {
 65             if (VIR_STRDUP(nic_model, "e1000") < 0)
 66                 return -1;
 67         } else {
 68             virReportError(VIR_ERR_CONFIG_UNSUPPORTED, "%s",
 69                            _("NIC model 'e1000' is not supported "
 70                              "by given bhyve binary"));
 71             return -1;
 72         }
 73     } else {
 74         virReportError(VIR_ERR_CONFIG_UNSUPPORTED,
 75                        _("NIC model '%s' is not supported"),
 76                        net->model);
 77         return -1;
 78     }
 79 
 80     if (actualType == VIR_DOMAIN_NET_TYPE_BRIDGE) {
 81         if (VIR_STRDUP(brname, virDomainNetGetActualBridgeName(net)) < 0)
 82             goto out;
 83     } else {
 84         virReportError(VIR_ERR_CONFIG_UNSUPPORTED,
 85                        _("Network type %d is not supported"),
 86                        virDomainNetGetActualType(net));
 87         goto out;
 88     }
 89 
 90     if (!net->ifname ||
 91         STRPREFIX(net->ifname, VIR_NET_GENERATED_PREFIX) ||
 92         strchr(net->ifname, '%')) {
 93         VIR_FREE(net->ifname);
 94         if (VIR_STRDUP(net->ifname, VIR_NET_GENERATED_PREFIX "%d") < 0)
 95             goto out;
 96     }
 97 
 98     if (!dryRun) {
 99         if (virNetDevTapCreateInBridgePort(brname, &net->ifname, &net->mac,
100                                            def->uuid, NULL, NULL, 0,
101                                            virDomainNetGetActualVirtPortProfile(net),
102                                            virDomainNetGetActualVlan(net),
103                                            VIR_NETDEV_TAP_CREATE_IFUP | VIR_NETDEV_TAP_CREATE_PERSIST) < 0)
104             goto out;
105 
106         realifname = virNetDevTapGetRealDeviceName(net->ifname);
107 
108         if (realifname == NULL)
109             goto out;
110 
111         VIR_DEBUG("%s -> %s", net->ifname, realifname);
112         /* hack on top of other hack: we need to set
113          * interface to 'UP' again after re-opening to find its
114          * name
115          */
116         if (virNetDevSetOnline(net->ifname, true) != 0)
117             goto out;
118     } else {
119         if (VIR_STRDUP(realifname, "tap0") < 0)
120             goto out;
121     }
122 
123 
124     virCommandAddArg(cmd, "-s");
125     virCommandAddArgFormat(cmd, "%d:0,%s,%s,mac=%s",
126                            net->info.addr.pci.slot, nic_model,
127                            realifname, virMacAddrFormat(&net->mac, macaddr));
128 
129     ret = 0;
130  out:
131     VIR_FREE(net->ifname);
132     VIR_FREE(realifname);
133     VIR_FREE(brname);
134     VIR_FREE(nic_model);
135 
136     return ret;
137 }

Also, I had an assumption (wrong!) that I was running with dryRun == true in the test. Anyway, when things go in unexpected ways, it's good to run lldb:

(lldb) breakpoint set --file bhyve_command.c --line 94
Breakpoint 1: where = bhyvexml2argvtest`bhyveBuildNetArgStr + 731 at bhyve_command.c:94, address = 0x0000000000410b3d

Then a couple of next's and we can set a watchpoint and see what happens:

(lldb) print net->ifname
(char *) $1 = 0x0000000807343678 "vnet%d"
(lldb) watchpoint set var net->ifname
Watchpoint created: Watchpoint 1: addr = 0x80727f108 size = 8 state = enabled type = w
    declare @ '/home/novel/code/libvirt/src/bhyve/bhyve_command.c:49'
    watchpoint spec = 'net->ifname'
    new value: 0x0000000807343678
(lldb) c
Process 53109 resuming

Watchpoint 1 hit:
old value: 0x0000000807343678
new value: 0x0000000000000000
Process 53109 stopped
* thread #1: tid = 100513, 0x0000000800ccadb6 libvirt.so.0`virStrdup(dest=0x000000080727f108, src="vnet0", report=true, domcode=57, filename="bhyvexml2argvmock.c", funcname="virNetDevTapCreateInBridgePort", linenr=32) + 43 at virstring.c:712, stop reason = watchpoint 1
    frame #0: 0x0000000800ccadb6 libvirt.so.0`virStrdup(dest=0x000000080727f108, src="vnet0", report=true, domcode=57, filename="bhyvexml2argvmock.c", funcname="virNetDevTapCreateInBridgePort", linenr=32) + 43 at virstring.c:712
   709            size_t linenr)
   710  {
   711      *dest = NULL;
-> 712      if (!src)
   713          return 0;
   714      if (!(*dest = strdup(src))) {
   715          if (report)
(lldb) bt
* thread #1: tid = 100513, 0x0000000800ccadb6 libvirt.so.0`virStrdup(dest=0x000000080727f108, src="vnet0", report=true, domcode=57, filename="bhyvexml2argvmock.c", funcname="virNetDevTapCreateInBridgePort", linenr=32) + 43 at virstring.c:712, stop reason = watchpoint 1
  * frame #0: 0x0000000800ccadb6 libvirt.so.0`virStrdup(dest=0x000000080727f108, src="vnet0", report=true, domcode=57, filename="bhyvexml2argvmock.c", funcname="virNetDevTapCreateInBridgePort", linenr=32) + 43 at virstring.c:712
    frame #1: 0x0000000800862168 bhyvexml2argvmock.so`virNetDevTapCreateInBridgePort(brname="virbr0", ifname=0x000000080727f108, macaddr=0x000000080727f004, vmuuid="\x04\x11㮰P+874\a\b", tunpath=0x0000000000000000, tapfd=0x0000000000000000, tapfdSize=0, virtPortProfile=0x0000000000000000, virtVlan=0x0000000000000000, fakeflags=9) + 83 at bhyvexml2argvmock.c:32
    frame #2: 0x0000000000410bf3 bhyvexml2argvtest`bhyveBuildNetArgStr(conn=0x000000080725e000, def=0x000000080727c000, net=0x000000080727f000, cmd=0x0000000807245280, dryRun=false) + 913 at bhyve_command.c:99
    frame #3: 0x00000000004113eb bhyvexml2argvtest`virBhyveProcessBuildBhyveCmd(conn=0x000000080725e000, def=0x000000080727c000, dryRun=false) + 610 at bhyve_command.c:304
    frame #4: 0x000000000040d70a bhyvexml2argvtest`testCompareXMLToArgvFiles(xml="/home/novel/code/libvirt/tests/bhyvexml2argvdata/bhyvexml2argv-net-e1000.xml", cmdline="/home/novel/code/libvirt/tests/bhyvexml2argvdata/bhyvexml2argv-net-e1000.args", ldcmdline="/home/novel/code/libvirt/tests/bhyvexml2argvdata/bhyvexml2argv-net-e1000.ldargs", dmcmdline="/home/novel/code/libvirt/tests/bhyvexml2argvdata/bhyvexml2argv-net-e1000.devmap", flags=0) + 225 at bhyvexml2argvtest.c:48
    frame #5: 0x000000000040da26 bhyvexml2argvtest`testCompareXMLToArgvHelper(data=0x000000000063e460) + 405 at bhyvexml2argvtest.c:117
    frame #6: 0x000000000040dfe3 bhyvexml2argvtest`virTestRun(title="BHYVE XML-2-ARGV net-e1000", body=(bhyvexml2argvtest`testCompareXMLToArgvHelper at bhyvexml2argvtest.c:101), data=0x000000000063e460) + 245 at testutils.c:180
    frame #7: 0x000000000040dd5b bhyvexml2argvtest`mymain + 789 at bhyvexml2argvtest.c:179
    frame #8: 0x000000000040fa82 bhyvexml2argvtest`virTestMain(argc=1, argv=0x00007fffffffe6e8, func=(bhyvexml2argvtest`mymain at bhyvexml2argvtest.c:127)) + 1675 at testutils.c:992
    frame #9: 0x000000000040de19 bhyvexml2argvtest`main(argc=1, argv=0x00007fffffffe6e8) + 50 at bhyvexml2argvtest.c:195
    frame #10: 0x000000000040d44f bhyvexml2argvtest`_start + 383
(lldb) 

But virStrdup is not really helpful, so going to the previous frame to see the logic behind that:

(lldb) fr s 1
frame #1: 0x0000000800862168 bhyvexml2argvmock.so`virNetDevTapCreateInBridgePort(brname="virbr0", ifname=0x000000080727f108, macaddr=0x000000080727f004, vmuuid="\x04\x11㮰P+874\a\b", tunpath=0x0000000000000000, tapfd=0x0000000000000000, tapfdSize=0, virtPortProfile=0x0000000000000000, virtVlan=0x0000000000000000, fakeflags=9) + 83 at bhyvexml2argvmock.c:32
   29                                      virNetDevVlanPtr virtVlan ATTRIBUTE_UNUSED,
   30                                      unsigned int fakeflags ATTRIBUTE_UNUSED)
   31   {
-> 32       if (VIR_STRDUP(*ifname, "vnet0") < 0)
   33           return -1;
   34       return 0;
   35   }
(lldb) 

Finally, things are getting clear: bhyveBuildNetArgStr was calling virNetDevTapCreateInBridgePort that is mocked for tests. And the mocked version does not free ifname before calling strdup on it, so it's never freed again. The rest of the fix is trivial, problem solved.

All in all, this appears like an interesting and useful exercise. This is probably too detailed, but hopefully it'll be useful for somebody who's just starting the journey of searching for memory leaks (like myself).

Friday, November 4, 2016

bhyve vs VirtualBox benchmarking, part2

"There are in order of increasing severity: lies, damn lies, statistics, and computer benchmarks",
man 8 diskinfo

I got some feedback to my previous post about benchmarking of bhyve and VirtualBox:

So I decided to do some more tests and include e1000 for networking tests and try different drivers and image types for I/O tests.

Networking test

This is a little more extensive test than the one in my previous blog post, now it includes e1000 and virtio-net for both VirtualBox and bhyve. Setting for the test is still the same: iperf and bridged network mode. Commands remain the same, on VM side I run:

iperf -s

On the host, I run:

iperf -c $vmip

And calculate the average of 8 runs. Results:

And actual values (in Gbits/sec) are:

VirtualBoxbhyve
e10001.428751.57375
virtio-net0.432752.4

The most shocking part here is that in VirtualBox e1000 is more than 3x times faster than virtio-net. This seems a little strange, esp. considering that e1000 performance in bhyve is almost the same, but virtio-net in bhyve is approx. 1.5x times faster than e1000 (that's probably a very huge difference too, but at least it's expected virtio to be faster).

I/O testing

I decided to check things suggested in the tweet above and started with disk configuration. I've converted my image to the "fixed size" type image like this:

VBoxManage clonehd uefi_fbsd_20gigs.vdi uefi_fbsd_20gigs_fixed.vdi --variant Fixed

Then I conducted tests for the following configurations:

  • VirtualBox + fixed size image
  • VirtualBox + dynamic size image
  • bhyve + virtio-blk
  • bhyve + ahci-hd

The same image was used for all tests. Fixed size image was produced using the command specified above, raw image for bhyve was created from the VirtualBox image using the qemu-img(1) tool.

I started with bonnie++ test at results surprised me, to put it softly:

Here, we can see that bhyve with virtio-blk shows the best write speed (that is expected), but then shows the worst rewrite speed (which is a little surprising, but the gap is minimal) and shows the worst read speed (more than 2 times slower than VirtualBox; extremely surprising). After that I decided to take a few days break and then to try some different ways of benchmarking.

So, for read performance I used the diskinfo(8) tool this way:

diskinfo -tv ada0 | grep middle

and use the average of 16 runs. For writing, I used dd(1):

dd bs=1M count=2048 if=/dev/zero of=test conv=sync; sync; rm test

and also use the average of 16 runs. I got the following results:

And the numbers (all values are kbytes/sec):

vbox (fixed size img)vbox (dynamic img)bhyve (ahci-hd)bhyve (virtio-blk)
diskinfo1232397152877912960552647685
dd113737135088113889115924

Frankly, this didn't help to figure out state of things, because these results are kind of opposite to what bonnie++ showed: bhyve with virtio-blk shows very high read speed as demonstrated by diskinfo, 1.7x faster than VirtualBox. On the other hand, the numbers that diskinfo is showing are crazy: 2647 mbytes/sec. This feels more like RAM transfer rates, so it looks like some sort of caching is involved here.

As for the dd(1) test part, bhyve with virtio-blk is 16.5% slower than VirtualBox using dynmanic size image.

Conclusion

  • For VirtualBox, if choosing between e1000 and virtio-net, e1000 definitely provides better performance over virtio-net, at least on FreeBSD hosts with FreeBSD guests
  • virtio-net in bhyve is approx. 1.5x faster than VirtualBox with e1000
  • I'll refrain from comments on I/O tests.

Further Reading

Sunday, October 16, 2016

bhyve vs VirtualBox benchmark

I've always been curious how bhyve performance compares to other hypervisors, so I've decided to compare it with VirtualBox. Target audience of these projects is somewhat different, though. VirtualBox is targeted more to desktop users (as it's easy to use, has a nice GUI and guest additions for running GUI within a guest smoothly). And bhyve appears to be targeting more experienced users and operators as it's easier to create flexible configurations with it, as well as automate things. Anyway, it's still interesting to find out which one is faster.

Setup Overview

I used my development box for this benchmarking, and it has no additional load, so results should be more or less clean in that matter. It's running 12-CURRENT:

FreeBSD 12.0-CURRENT amd64

as of Oct, 4th.

It has Intel i5 CPU, 16 gigs of RAM and some old 7200 IDE HDD:

CPU: Intel(R) Core(TM) i5-4690 CPU @ 3.50GHz (3491.99-MHz K8-class CPU)
  Origin="GenuineIntel"  Id=0x306c3  Family=0x6  Model=0x3c  Stepping=3
  Features=0xbfebfbff<FPU,VME,DE,PSE,TSC,MSR,PAE,MCE,CX8,APIC,SEP,MTRR,PGE,MCA,CMOV,PAT,PSE36,CLFLUSH,DTS,ACPI,MMX,FXSR,SSE,SSE2,SS,HTT,TM,PBE>
  Features2=0x7ffafbff<SSE3,PCLMULQDQ,DTES64,MON,DS_CPL,VMX,SMX,EST,TM2,SSSE3,SDBG,FMA,CX16,xTPR,PDCM,PCID,SSE4.1,SSE4.2,x2APIC,MOVBE,POPCNT,TSCDLT,AESNI,XSAVE,OSXSAVE,AVX,F16C,RDRAND>
  AMD Features=0x2c100800<SYSCALL,NX,Page1GB,RDTSCP,LM>
  AMD Features2=0x21<LAHF,ABM>
  Structured Extended Features=0x2fbb<FSGSBASE,TSCADJ,BMI1,HLE,AVX2,SMEP,BMI2,ERMS,INVPCID,RTM,NFPUSG>
  XSAVE Features=0x1<XSAVEOPT>
  VT-x: PAT,HLT,MTF,PAUSE,EPT,UG,VPID
  TSC: P-state invariant, performance statistics
real memory  = 17179869184 (16384 MB)
avail memory = 16448757760 (15686 MB)
ada0: <ST3200822A 3.01> ATA-6 device

For tests I used two VMs, one in bhyve and one in VirtualBox. A bhyve VM was started like this:

bhyve -c 2 -m 4G -w -H -S \
        -s 0,hostbridge \
        -s 4,ahci-hd,/home/novel/img/uefi_fbsd_20gigs.raw \
        -s 5,virtio-net,tap1 \
        -s 29,fbuf,tcp=0.0.0.0:5900,w=800,h=600,wait \
        -s 30,xhci,tablet \
        -s 31,lpc -l com1,stdio \
        -l bootrom,/usr/local/share/uefi-firmware/BHYVE_UEFI.fd \
        vm0

Main pieces of the VirtualBox VM configuration are displayed on these screenshots:

Guest is:

FreeBSD 11.0-RELEASE-p1 amd64

on UFS. VirtualBox version 5.1.6 r110634 installed via pkg(8).

World Compilation

The first test is running:

make buildworld buildkernel -j4

on releng/11.0 source tree.

The result is that bhyve is approx. 15% slower (106 minutes vs 92 minutes for VirtualBox):

iperf test

The next test is network performance check using the iperf(8) tool. In order to avoid interaction with hardware NICs and switches (with unpredictable load), all the traffic for testing stays withing the host system:

vm# iperf -s
host# iperf -c $vmip

As a reminder, both bhyve and VirtualBox VMs are configured to use bridged networking. Also, both are using virtio-net NICs.

Result is a little strange because bhyve appears to be more than 4 times faster here:

That seemed to be strange to me, so I ran the test multiple times for both VirtualBox and bhyve, however, all the time the results were pretty close. I'm yet to find out if that's really correct result or something wrong with my testing or setup.

bonnie++ test

bonnie++ is a benchmarking tool for hard drivers and filesystems. Let's jump to the results right away:

This makes bhyve write and read speeds approx. 15% higher than VirtualBox. There's a note though: for VirtualBox VM I used VDI disk format (as it's native for VirtualBox) and SATA emulation because there's no virtio-blk support in VirtualBox. In case of bhyve I also used SATA emulation (I actually intended to use virtio-blk here, but forgot to change commandline), but on a raw image instead.

xz(1) on memory disk test

The last test is to unpack and pack an XZ archive on memory disk. I used memory disk specifically to exclude disk I/O from the equation. For a sample archive I choose:

ftp://ftp.freebsd.org/pub/FreeBSD/releases/ISO-IMAGES/11.0/FreeBSD-11.0-RELEASE-amd64-memstick.img.xz

And the commands to unpack and pack were:

unxz FreeBSD-11.0-RELEASE-amd64-memstick.img.xz 
xz FreeBSD-11.0-RELEASE-amd64-memstick.img

Results are almost the same here:

Summary

  • CPU and RAM performance seems to be identical (though it's not clear why buildworld takes longer on bhyve)
  • I/O performance is 15% better with bhyve
  • Networking performance is 4x better with bhyve (this looks suspicious and requires additional research; also, I'm not familiar with bridged networking implementation in VirtualBox, maybe it could explain the difference)

Further Reading

PS: Initially I was going to use phoronix-test-suite. However, it appears that a lot of important tests fail to run on FreeBSD. The ones that did run, such as apache or sqlite tests, do not seem very representative to me. I'll probably try to run similar tests but with Linux guest.