Re: [libvirt] [RFC PATCH 0/3] Implement mockup capabilities cache in QEMU tests
Hello! > It's just an assert. We have those for various reasons and extra > carefulness cannot hurt. I don't think it needs differentiating > between whether the code is ran in tests or not. It would need differentiating because in tests we should not talk to the real filesystem, while in real environment it's OK. So we would need some "extern bool qemuTestMode". Well, anyway, take a look at my new implementation, i sent it yesterday. It uses a similar global variable, with only a 2-line change in the library code. The basic idea is to override 'binary' name in virQEMUCapsCacheLookup(), so that we can force it to return particular capability set no matter what def->emulator is. This very simple change actually allows us to do lots of nice things, and improve our test suite even further by removing hardcoded capabilities at all. I cc'ed to you, but just in case: http://www.redhat.com/archives/libvir-list/2015-September/msg00084.html Kind regards, Pavel Fedin Expert Engineer Samsung Electronics Research center Russia -- libvir-list mailing list libvir-list@redhat.com https://www.redhat.com/mailman/listinfo/libvir-list
Re: [libvirt] [RFC PATCH 0/3] Implement mockup capabilities cache in QEMU tests
Hello! > The current xmlopt structure has two callbacks in parser > configuration, one for post-parsing of the whole domain XML and one > for device XML. We could add new (optional) callback in this > structure that would be called before other post-parse callbacks, but > still after the xml is parsed. This would not be utilized anywhere > else than in tests We don't need to add this because we already have domain post-parse callback. And this is what i actually disliked. There is a very simple way to patch this callback. Currently in our tests we call virQEMUDriverCreateXMLConf(), which returns a structure, containing our callbacks. Actually we could make a wrapper around this function which would replace domainPostParseCallback with own pointer, remembering the original one. Then our function would take emulator name from the parsed XML, insert capabilities cache, and proceed to original function. > We should be really careful about not depending on the host in tests, > so any assert like this would be much appreciated. To tell the truth i'm not a big fan of test-only code in library body. But, i agree with this, and i see no other way round. I think we could add some global Boolean variable like bool testMode, which would be set to true by test suite. Actually, this variable could even modify capsCache behavior to ignore binary name at all, and return, let's say, the first entry. But we should be really careful about it because i believe we want to exercise the actual library code, and not some test-only code. Kind regards, Pavel Fedin Expert Engineer Samsung Electronics Research center Russia -- libvir-list mailing list libvir-list@redhat.com https://www.redhat.com/mailman/listinfo/libvir-list
Re: [libvirt] [RFC PATCH 0/3] Implement mockup capabilities cache in QEMU tests
On Thu, Sep 03, 2015 at 09:51:11AM +0300, Pavel Fedin wrote: Hello! The current xmlopt structure has two callbacks in parser configuration, one for post-parsing of the whole domain XML and one for device XML. We could add new (optional) callback in this structure that would be called before other post-parse callbacks, but still after the xml is parsed. This would not be utilized anywhere else than in tests We don't need to add this because we already have domain post-parse callback. And this is what i actually disliked. There is a very simple way to patch this callback. Currently in our tests we call virQEMUDriverCreateXMLConf(), which returns a structure, containing our callbacks. Actually we could make a wrapper around this function which would replace domainPostParseCallback with own pointer, remembering the original one. Then our function would take emulator name from the parsed XML, insert capabilities cache, and proceed to original function. It's just that we would have to replace and patch both domain callback and domaindevice callback (domaindevice callback can be called without domain callback) and I thought it would be nicer to have it in one place. But I think both works for me. We should be really careful about not depending on the host in tests, so any assert like this would be much appreciated. To tell the truth i'm not a big fan of test-only code in library body. But, i agree with this, and i see no other way round. I think we could add some global Boolean variable like bool testMode, which would be set to true by test suite. Actually, this variable could even modify capsCache behavior to ignore binary name at all, and return, let's say, the first entry. But we should be really careful about it because i believe we want to exercise the actual library code, and not some test-only code. It's just an assert. We have those for various reasons and extra carefulness cannot hurt. I don't think it needs differentiating between whether the code is ran in tests or not. That defeats the purpose of our tests. It'd be better to resort to mocking some simple function if needed. Kind regards, Pavel Fedin Expert Engineer Samsung Electronics Research center Russia signature.asc Description: PGP signature -- libvir-list mailing list libvir-list@redhat.com https://www.redhat.com/mailman/listinfo/libvir-list
Re: [libvirt] [RFC PATCH 0/3] Implement mockup capabilities cache in QEMU tests
On Tue, Sep 01, 2015 at 05:19:39PM -0400, Cole Robinson wrote: On 08/18/2015 05:40 AM, Pavel Fedin wrote: Since commit e8d55172544c1fafe31a9e09346bdebca4f0d6f9 qemu driver checks emulator capabilities during domain XML post-parse. However, test suite does not initialize it, therefore a condition to skip all checks if there is no cache supplied was added. This is actually a hack, whose sole purpose is to make existing test suite working. Additionally, it prevents from writing new tests for this particular functionality. This series attempts to solve this problem by implementing proper cache mockup in test suite. The main idea is to create a cache in standard way and put there a pre-defined capabilities set (which tests already have). The main problem here is to know emulator binary name, which is contained in the source XML. However, we have to create our cache before reading the XML. The simplest way to resolve this is to assume particular binary name from test name. Currently tests which assume cross-architecture binary are all prefixed with the architecture name (with one exception of "keywrap" tests which all assume /usr/bin/qemu-system-s390x and do not have "s390-" prefix in their name). This scheme works fine, unless we use "native" emulator binary. Here we have a mess. Most newer tests use /usr/bin/qemu, however there is a large number of tests which use /usr/libexec/qemu-kvm or /usr/bin/kvm (i guess these are leftovers from the epoch when qemu-kvm was a separate fork of qemu). This is currently not handled in any way, and these tests may report errors due to missing binaries (because virQEMUCapsCacheLookup() attempts to populate the cache automatically by querying the binary if not already known). There are several possible ways to resolve this: a) Add all possible names as aliases for /usr/bin/qemu b) Forbid to use oldstyle names at all in these tests c) Declare some prefix like "kvm-" for those tests who want to use /usr/libexec/qemu-kvm. Again, this would ban /usr/bin/kvm and /usr/bin/qemu-kvm (if not using aliases like in (b) d) Hardcode (optional) emulator name per test. IMHO a bad idea because number of tests is huge. e) Do some preparsing of the XML and extract binary name from it. Again, i disliked it for not being simple enough. With the current approach I think e) is the way to go... yeah it's redundant parsing but better than trying to maintain a whitelist or reformat a bunch of test cases to match this. And for example these patches as is cause breakage in other tests, like qemuargv2xml, and I think it's related to the incomplete whitelist. Using libvirt helpers it shouldn't be too hard to parse out the emulator path, basically just virXMLParseStringCtxt and then emulator = virXPathString("string(./devices/emulator[1])", ctxt); Maybe other people have better ideas about how to mock this out though. I also thought about an alternate implementation which would patch postParseCallback and insert own function there which builds a cache. At this point binary name is already known from the XML. However, such a design looks like an ugly hack by itself, so i stopped going in this direction. Yeah that sounds quite intense. Maybe there's some way to override a function with our own definition from the test suite, but that's outside my realm of knowledge. The current xmlopt structure has two callbacks in parser configuration, one for post-parsing of the whole domain XML and one for device XML. We could add new (optional) callback in this structure that would be called before other post-parse callbacks, but still after the xml is parsed. This would not be utilized anywhere else than in tests, currently, but I don't see it as ugly and it might be beneficial e.g. for some drivers or in the future. Another benefit is that it does not add that much of a code as it would be called only in virDomainDefPostParse() and virDomainDeviceDefParse(), plus the time complexity can be kept at minimum if the function called is no-op in case the cache is already built. Actually populating the cache does exercise the most qemu code though, so that's one benefit. Another minor bit to consider is possibly adding an assert in qemu_capabilities.c to ensure that we never try to actually probe the host FS from the test suite, since we don't want the test suite to accidentally become dependent on host state. Maybe there's a test suite environment variable we can check for. We should be really careful about not depending on the host in tests, so any assert like this would be much appreciated. - Cole signature.asc Description: PGP signature -- libvir-list mailing list libvir-list@redhat.com https://www.redhat.com/mailman/listinfo/libvir-list
Re: [libvirt] [RFC PATCH 0/3] Implement mockup capabilities cache in QEMU tests
On 08/18/2015 05:40 AM, Pavel Fedin wrote: > Since commit e8d55172544c1fafe31a9e09346bdebca4f0d6f9 qemu driver checks > emulator capabilities during domain XML post-parse. However, test suite > does not initialize it, therefore a condition to skip all checks if there > is no cache supplied was added. This is actually a hack, whose sole > purpose is to make existing test suite working. Additionally, it prevents > from writing new tests for this particular functionality. > > This series attempts to solve this problem by implementing proper cache > mockup in test suite. The main idea is to create a cache in standard way > and put there a pre-defined capabilities set (which tests already have). > > The main problem here is to know emulator binary name, which is contained > in the source XML. However, we have to create our cache before reading the > XML. The simplest way to resolve this is to assume particular binary name > from test name. Currently tests which assume cross-architecture binary are > all prefixed with the architecture name (with one exception of "keywrap" > tests which all assume /usr/bin/qemu-system-s390x and do not have "s390-" > prefix in their name). > > This scheme works fine, unless we use "native" emulator binary. Here we > have a mess. Most newer tests use /usr/bin/qemu, however there is a large > number of tests which use /usr/libexec/qemu-kvm or /usr/bin/kvm (i guess > these are leftovers from the epoch when qemu-kvm was a separate fork of > qemu). This is currently not handled in any way, and these tests may > report errors due to missing binaries (because virQEMUCapsCacheLookup() > attempts to populate the cache automatically by querying the binary if > not already known). > > There are several possible ways to resolve this: > a) Add all possible names as aliases for /usr/bin/qemu > b) Forbid to use oldstyle names at all in these tests > c) Declare some prefix like "kvm-" for those tests who want to use >/usr/libexec/qemu-kvm. Again, this would ban /usr/bin/kvm and >/usr/bin/qemu-kvm (if not using aliases like in (b) > d) Hardcode (optional) emulator name per test. IMHO a bad idea because >number of tests is huge. > e) Do some preparsing of the XML and extract binary name from it. Again, >i disliked it for not being simple enough. > With the current approach I think e) is the way to go... yeah it's redundant parsing but better than trying to maintain a whitelist or reformat a bunch of test cases to match this. And for example these patches as is cause breakage in other tests, like qemuargv2xml, and I think it's related to the incomplete whitelist. Using libvirt helpers it shouldn't be too hard to parse out the emulator path, basically just virXMLParseStringCtxt and then emulator = virXPathString("string(./devices/emulator[1])", ctxt); Maybe other people have better ideas about how to mock this out though. > I also thought about an alternate implementation which would patch > postParseCallback and insert own function there which builds a cache. At > this point binary name is already known from the XML. However, such a > design looks like an ugly hack by itself, so i stopped going in this > direction. > Yeah that sounds quite intense. Maybe there's some way to override a function with our own definition from the test suite, but that's outside my realm of knowledge. Actually populating the cache does exercise the most qemu code though, so that's one benefit. Another minor bit to consider is possibly adding an assert in qemu_capabilities.c to ensure that we never try to actually probe the host FS from the test suite, since we don't want the test suite to accidentally become dependent on host state. Maybe there's a test suite environment variable we can check for. - Cole -- libvir-list mailing list libvir-list@redhat.com https://www.redhat.com/mailman/listinfo/libvir-list
[libvirt] [RFC PATCH 0/3] Implement mockup capabilities cache in QEMU tests
Since commit e8d55172544c1fafe31a9e09346bdebca4f0d6f9 qemu driver checks emulator capabilities during domain XML post-parse. However, test suite does not initialize it, therefore a condition to skip all checks if there is no cache supplied was added. This is actually a hack, whose sole purpose is to make existing test suite working. Additionally, it prevents from writing new tests for this particular functionality. This series attempts to solve this problem by implementing proper cache mockup in test suite. The main idea is to create a cache in standard way and put there a pre-defined capabilities set (which tests already have). The main problem here is to know emulator binary name, which is contained in the source XML. However, we have to create our cache before reading the XML. The simplest way to resolve this is to assume particular binary name from test name. Currently tests which assume cross-architecture binary are all prefixed with the architecture name (with one exception of keywrap tests which all assume /usr/bin/qemu-system-s390x and do not have s390- prefix in their name). This scheme works fine, unless we use native emulator binary. Here we have a mess. Most newer tests use /usr/bin/qemu, however there is a large number of tests which use /usr/libexec/qemu-kvm or /usr/bin/kvm (i guess these are leftovers from the epoch when qemu-kvm was a separate fork of qemu). This is currently not handled in any way, and these tests may report errors due to missing binaries (because virQEMUCapsCacheLookup() attempts to populate the cache automatically by querying the binary if not already known). There are several possible ways to resolve this: a) Add all possible names as aliases for /usr/bin/qemu b) Forbid to use oldstyle names at all in these tests c) Declare some prefix like kvm- for those tests who want to use /usr/libexec/qemu-kvm. Again, this would ban /usr/bin/kvm and /usr/bin/qemu-kvm (if not using aliases like in (b) d) Hardcode (optional) emulator name per test. IMHO a bad idea because number of tests is huge. e) Do some preparsing of the XML and extract binary name from it. Again, i disliked it for not being simple enough. I also thought about an alternate implementation which would patch postParseCallback and insert own function there which builds a cache. At this point binary name is already known from the XML. However, such a design looks like an ugly hack by itself, so i stopped going in this direction. Comments and opinions are welcome. Pavel Fedin (3): Implement virQEMUCapsCache mockup Use mockup cache Removed unneeded check src/qemu/qemu_capabilities.c | 10 +- src/qemu/qemu_capspriv.h | 36 +++ src/qemu/qemu_domain.c | 5 + tests/qemuagenttest.c| 9 - tests/qemuargv2xmltest.c | 5 + tests/qemuhotplugtest.c | 23 ++ tests/qemuxml2argvtest.c | 5 + tests/qemuxml2xmltest.c | 6 ++ tests/qemuxmlnstest.c| 5 + tests/testutilsqemu.c| 45 tests/testutilsqemu.h| 3 +++ 11 files changed, 130 insertions(+), 22 deletions(-) create mode 100644 src/qemu/qemu_capspriv.h mode change 100644 = 100755 tests/qemuagenttest.c mode change 100644 = 100755 tests/qemuhotplugtest.c mode change 100644 = 100755 tests/qemuxml2xmltest.c mode change 100644 = 100755 tests/testutilsqemu.c -- 2.1.4 -- libvir-list mailing list libvir-list@redhat.com https://www.redhat.com/mailman/listinfo/libvir-list