Re: [RFC v3 11/19] kunit: add Python libraries for handing KUnit config and kernel

Re: [RFC v3 11/19] kunit: add Python libraries for handing KUnit config and kernel

[Matthew Wilcox]

On Thu, Dec 06, 2018 at 12:32:47PM +0000, Kieran Bingham wrote:
> On 04/12/2018 20:47, Luis Chamberlain wrote:
> > On Mon, Dec 03, 2018 at 03:48:15PM -0800, Brendan Higgins wrote:
> >> On Thu, Nov 29, 2018 at 5:54 AM Kieran Bingham
> >> <kieran.bingham@ideasonboard.com> wrote:
> >>>
> >>> Hi Brendan,
> >>>
> >>> Thanks again for this series!
> >>>
> >>> On 28/11/2018 19:36, Brendan Higgins wrote:
> >>>> The ultimate goal is to create minimal isolated test binaries; in the
> >>>> meantime we are using UML to provide the infrastructure to run tests, so
> >>>> define an abstract way to configure and run tests that allow us to
> >>>> change the context in which tests are built without affecting the user.
> >>>> This also makes pretty and dynamic error reporting, and a lot of other
> >>>> nice features easier.
> >>>
> >>>
> >>> I wonder if we could somehow generate a shared library object
> >>> 'libkernel' or 'libumlinux' from a UM configured set of headers and
> >>> objects so that we could create binary targets directly ?
> >>
> >> That's an interesting idea. I think it would be difficult to figure
> >> out exactly where to draw the line of what goes in there and what
> >> needs to be built specific to a test a priori. Of course, that leads
> >> into the biggest problem in general, needed to know what I need to
> >> build to test the thing that I want to test.
> >>
> >> Nevertheless, I could definitely imagine that being useful in a lot of cases.
> > 
> > Whether or not we can abstract away the kernel into such a mechanism
> > with uml libraries is a good question worth exploring.
> > 
> > Developers working upstream do modify their kernels a lot, so we'd have
> > to update such libraries quite a bit, but I think that's fine too. The
> > *real* value I think from the above suggestion would be enterprise /
> > mobile distros or stable kernel maintainers which have a static kernel
> > they need to support for a relatively *long time*, consider a 10 year
> > time frame. Running unit tests without qemu with uml and libraries for
> > respective kernels seems real worthy.
> 
> I think any such library might be something generated by the kernel
> build system, so if someone makes substantial changes to a core
> component provided by the library - it can be up to them to build a
> corresponding userspace library as well.
> 
> We could also consider to only provide *static* libraries rather than
> dynamic. So any one building some userspace tool / test with this would
> be required to compile against (the version of) the kernel they expect
> perhaps... - much like we expect modules to be compiled currently.
> 
> And then the userspace binary would be sufficiently able to live it's
> life on it's own :)
> 
> > The overhead for testing a unit test for said targets, *ideally*, would
> > just be to to reboot into the system with such libraries available, a
> > unit test would just look for the respective uname -r library and mimic
> > that kernel, much the same way enterprise distributions today rely on
> > having debugging symbols available to run against crash / gdb. Having
> > debug modules / kernel for crash requires such effort already, so this
> > would just be an extra layer of other prospect tests.
> 
> Oh - although, yes - there are some good concepts there - but I'm a bit
> weary of how easy it would be to 'run' the said test against multiple
> kernel version libraries... there would be a lot of possible ABI
> conflicts perhaps.
> 
> My main initial idea for a libumlinux is to provide infrastructure such
> as our linked-lists and other kernel formatting so that we can take
> kernel code directly to userspace for test and debug (assuming that
> there are no hardware dependencies or things that we can't mock out)
> 
> I think all of this could complement kunit of course - this isn't
> suggesting an alternative implementation :-)

I suspect the reason Luis cc'd me on this is that we already have some
artisinally-crafted userspace kernel-mocking interfaces under tools/.
The tools/testing/radix-tree directory is the source of some of this,
but I've been moving pieces out into tools/ more generally where it
makes sense to.

We have liburcu already, which is good.  The main sticking points are:

 - No emulation of kernel thread interfaces
 - The kernel does not provide the ability to aggressively fail memory
   allocations (which is useful when trying to exercise the memory failure
   paths).
 - printk has started adding a lot of %pX enhancements which printf
   obviously doesn't know about.
 - No global pseudo-random number generator in the kernel.  Probably
   we should steal the i915 one.

I know Dan Williams has also done a lot of working mocking kernel
interfaces for libnvdimm.

Re: [RFC v3 11/19] kunit: add Python libraries for handing KUnit config and kernel

[Kieran Bingham]

Hi Matthew,

On 06/12/2018 15:37, Matthew Wilcox wrote:
> On Thu, Dec 06, 2018 at 12:32:47PM +0000, Kieran Bingham wrote:
>> On 04/12/2018 20:47, Luis Chamberlain wrote:
>>> On Mon, Dec 03, 2018 at 03:48:15PM -0800, Brendan Higgins wrote:
>>>> On Thu, Nov 29, 2018 at 5:54 AM Kieran Bingham
>>>> <kieran.bingham@ideasonboard.com> wrote:
>>>>>
>>>>> Hi Brendan,
>>>>>
>>>>> Thanks again for this series!
>>>>>
>>>>> On 28/11/2018 19:36, Brendan Higgins wrote:
>>>>>> The ultimate goal is to create minimal isolated test binaries; in the
>>>>>> meantime we are using UML to provide the infrastructure to run tests, so
>>>>>> define an abstract way to configure and run tests that allow us to
>>>>>> change the context in which tests are built without affecting the user.
>>>>>> This also makes pretty and dynamic error reporting, and a lot of other
>>>>>> nice features easier.
>>>>>
>>>>>
>>>>> I wonder if we could somehow generate a shared library object
>>>>> 'libkernel' or 'libumlinux' from a UM configured set of headers and
>>>>> objects so that we could create binary targets directly ?
>>>>
>>>> That's an interesting idea. I think it would be difficult to figure
>>>> out exactly where to draw the line of what goes in there and what
>>>> needs to be built specific to a test a priori. Of course, that leads
>>>> into the biggest problem in general, needed to know what I need to
>>>> build to test the thing that I want to test.
>>>>
>>>> Nevertheless, I could definitely imagine that being useful in a lot of cases.
>>>
>>> Whether or not we can abstract away the kernel into such a mechanism
>>> with uml libraries is a good question worth exploring.
>>>
>>> Developers working upstream do modify their kernels a lot, so we'd have
>>> to update such libraries quite a bit, but I think that's fine too. The
>>> *real* value I think from the above suggestion would be enterprise /
>>> mobile distros or stable kernel maintainers which have a static kernel
>>> they need to support for a relatively *long time*, consider a 10 year
>>> time frame. Running unit tests without qemu with uml and libraries for
>>> respective kernels seems real worthy.
>>
>> I think any such library might be something generated by the kernel
>> build system, so if someone makes substantial changes to a core
>> component provided by the library - it can be up to them to build a
>> corresponding userspace library as well.
>>
>> We could also consider to only provide *static* libraries rather than
>> dynamic. So any one building some userspace tool / test with this would
>> be required to compile against (the version of) the kernel they expect
>> perhaps... - much like we expect modules to be compiled currently.
>>
>> And then the userspace binary would be sufficiently able to live it's
>> life on it's own :)
>>
>>> The overhead for testing a unit test for said targets, *ideally*, would
>>> just be to to reboot into the system with such libraries available, a
>>> unit test would just look for the respective uname -r library and mimic
>>> that kernel, much the same way enterprise distributions today rely on
>>> having debugging symbols available to run against crash / gdb. Having
>>> debug modules / kernel for crash requires such effort already, so this
>>> would just be an extra layer of other prospect tests.
>>
>> Oh - although, yes - there are some good concepts there - but I'm a bit
>> weary of how easy it would be to 'run' the said test against multiple
>> kernel version libraries... there would be a lot of possible ABI
>> conflicts perhaps.
>>
>> My main initial idea for a libumlinux is to provide infrastructure such
>> as our linked-lists and other kernel formatting so that we can take
>> kernel code directly to userspace for test and debug (assuming that
>> there are no hardware dependencies or things that we can't mock out)
>>
>> I think all of this could complement kunit of course - this isn't
>> suggesting an alternative implementation :-)
> 
> I suspect the reason Luis cc'd me on this is that we already have some
> artisinally-crafted userspace kernel-mocking interfaces under tools/.

Aha - excellent - I had hoped to grab you at Plumbers to talk about
this, after hearing you mention something at your Xarray talk - but
didn't seem to find a suitable time.

> The tools/testing/radix-tree directory is the source of some of this,
> but I've been moving pieces out into tools/ more generally where it
> makes sense to.

Sounds like we already have a starting point then.


> We have liburcu already, which is good.  The main sticking points are:
> 
>  - No emulation of kernel thread interfaces

Scheduling finesse aside, This shouldn't be too hard to emulate/wrap
with pthreads?


>  - The kernel does not provide the ability to aggressively fail memory
>    allocations (which is useful when trying to exercise the memory failure
>    paths).

Fault injection throughout would certainly be a valuable addition to any
unit-testing.

Wrapping tests into a single userspace binary could facilitate further
memory leak checking or other valgrind facilities too.



>  - printk has started adding a lot of %pX enhancements which printf
>    obviously doesn't know about.

Wrapping through User-mode linux essentially provides this already
though. In fact I guess that goes for the thread interfaces topic above too.


>  - No global pseudo-random number generator in the kernel.  Probably
>    we should steal the i915 one.
> 
> I know Dan Williams has also done a lot of working mocking kernel
> interfaces for libnvdimm.


Thanks for the references - more to investigate.

-- 
Regards
--
Kieran

Re: [RFC v3 11/19] kunit: add Python libraries for handing KUnit config and kernel

[Petr Mladek]

On Thu 2018-12-06 07:37:18, Matthew Wilcox wrote:
> On Thu, Dec 06, 2018 at 12:32:47PM +0000, Kieran Bingham wrote:
> > On 04/12/2018 20:47, Luis Chamberlain wrote:
> > > On Mon, Dec 03, 2018 at 03:48:15PM -0800, Brendan Higgins wrote:
> > >> On Thu, Nov 29, 2018 at 5:54 AM Kieran Bingham
> > >> <kieran.bingham@ideasonboard.com> wrote:
> > > Developers working upstream do modify their kernels a lot, so we'd have
> > > to update such libraries quite a bit, but I think that's fine too. The
> > > *real* value I think from the above suggestion would be enterprise /
> > > mobile distros or stable kernel maintainers which have a static kernel
> > > they need to support for a relatively *long time*, consider a 10 year
> > > time frame. Running unit tests without qemu with uml and libraries for
> > > respective kernels seems real worthy.
> > 
> > I think any such library might be something generated by the kernel
> > build system, so if someone makes substantial changes to a core
> > component provided by the library - it can be up to them to build a
> > corresponding userspace library as well.
> > 
> > My main initial idea for a libumlinux is to provide infrastructure such
> > as our linked-lists and other kernel formatting so that we can take
> > kernel code directly to userspace for test and debug (assuming that
> > there are no hardware dependencies or things that we can't mock out)
> 
> We have liburcu already, which is good.  The main sticking points are:
> 
>  - printk has started adding a lot of %pX enhancements which printf
>    obviously doesn't know about.

I wonder how big problem it is and if it is worth using another
approach.

An alternative would be to replace them with helper functions
the would produce the same string. The meaning would be easier
to understand. But concatenating with the surrounding text
would be less elegant. People might start using pr_cont()
that is problematic (mixed lines).

Also the %pX formats are mostly used to print context of some
structures. Even the helper functions would need some maintenance
to keep them compatible.

BTW: The printk() feature has been introduced 10 years ago by
the commit 4d8a743cdd2690c0bc8 ("vsprintf: add infrastructure
support for extended '%p' specifiers").

Best Regards,
Petr

Re: [RFC v3 11/19] kunit: add Python libraries for handing KUnit config and kernel

[Steven Rostedt]

On Tue, 11 Dec 2018 15:09:26 +0100
Petr Mladek <pmladek@suse.com> wrote:

> > We have liburcu already, which is good.  The main sticking points are:
> > 
> >  - printk has started adding a lot of %pX enhancements which printf
> >    obviously doesn't know about.  
> 
> I wonder how big problem it is and if it is worth using another
> approach.

No, please do not change the %pX approach.

> 
> An alternative would be to replace them with helper functions
> the would produce the same string. The meaning would be easier
> to understand. But concatenating with the surrounding text
> would be less elegant. People might start using pr_cont()
> that is problematic (mixed lines).
> 
> Also the %pX formats are mostly used to print context of some
> structures. Even the helper functions would need some maintenance
> to keep them compatible.
> 
> BTW: The printk() feature has been introduced 10 years ago by
> the commit 4d8a743cdd2690c0bc8 ("vsprintf: add infrastructure
> support for extended '%p' specifiers").

trace-cmd and perf know about most of the %pX data and how to read it.
Perhaps we can extend the libtraceevent library to export a generic way
to read data from printk() output for other tools to use.

-- Steve

Re: [RFC v3 11/19] kunit: add Python libraries for handing KUnit config and kernel

[Anton Ivanov]

On 12/11/18 2:41 PM, Steven Rostedt wrote:
> On Tue, 11 Dec 2018 15:09:26 +0100
> Petr Mladek <pmladek@suse.com> wrote:
>
>>> We have liburcu already, which is good.  The main sticking points are:
>>>
>>>   - printk has started adding a lot of %pX enhancements which printf
>>>     obviously doesn't know about.
>> I wonder how big problem it is and if it is worth using another
>> approach.
> No, please do not change the %pX approach.
>
>> An alternative would be to replace them with helper functions
>> the would produce the same string. The meaning would be easier
>> to understand. But concatenating with the surrounding text
>> would be less elegant. People might start using pr_cont()
>> that is problematic (mixed lines).
>>
>> Also the %pX formats are mostly used to print context of some
>> structures. Even the helper functions would need some maintenance
>> to keep them compatible.
>>
>> BTW: The printk() feature has been introduced 10 years ago by
>> the commit 4d8a743cdd2690c0bc8 ("vsprintf: add infrastructure
>> support for extended '%p' specifiers").
> trace-cmd and perf know about most of the %pX data and how to read it.
> Perhaps we can extend the libtraceevent library to export a generic way
> to read data from printk() output for other tools to use.

Going back for a second to using UML for this. UML console at present is 
interrupt driven - it emulates serial IO using several different 
back-ends (file descriptors, xterm or actual tty/ptys). Epoll events on 
the host side are used to trigger the UML interrupts - both read and write.

This works OK for normal use, but may result in all kinds of interesting 
false positives/false negatives when UML is used to run unit tests 
against a change which changes interrupt behavior.

IMO it may be useful to consider some alternatives specifically for unit 
test coverage purposes where printk and/or the whole console output 
altogether bypass some of the IRQ driven semantics.

-- 

Anton R. Ivanov

Cambridge Greys Limited, England and Wales company No 10273661
http://www.cambridgegreys.com/

Re: [RFC v3 11/19] kunit: add Python libraries for handing KUnit config and kernel

[Brendan Higgins]

On Tue, Dec 11, 2018 at 9:02 AM Anton Ivanov
<anton.ivanov@cambridgegreys.com> wrote:
>
>
> On 12/11/18 2:41 PM, Steven Rostedt wrote:
> > On Tue, 11 Dec 2018 15:09:26 +0100
> > Petr Mladek <pmladek@suse.com> wrote:
> >
> >>> We have liburcu already, which is good.  The main sticking points are:
> >>>
> >>>   - printk has started adding a lot of %pX enhancements which printf
> >>>     obviously doesn't know about.
> >> I wonder how big problem it is and if it is worth using another
> >> approach.
> > No, please do not change the %pX approach.
> >
> >> An alternative would be to replace them with helper functions
> >> the would produce the same string. The meaning would be easier
> >> to understand. But concatenating with the surrounding text
> >> would be less elegant. People might start using pr_cont()
> >> that is problematic (mixed lines).
> >>
> >> Also the %pX formats are mostly used to print context of some
> >> structures. Even the helper functions would need some maintenance
> >> to keep them compatible.
> >>
> >> BTW: The printk() feature has been introduced 10 years ago by
> >> the commit 4d8a743cdd2690c0bc8 ("vsprintf: add infrastructure
> >> support for extended '%p' specifiers").
> > trace-cmd and perf know about most of the %pX data and how to read it.
> > Perhaps we can extend the libtraceevent library to export a generic way
> > to read data from printk() output for other tools to use.
>
> Going back for a second to using UML for this. UML console at present is
> interrupt driven - it emulates serial IO using several different
> back-ends (file descriptors, xterm or actual tty/ptys). Epoll events on
> the host side are used to trigger the UML interrupts - both read and write.
>
> This works OK for normal use, but may result in all kinds of interesting
> false positives/false negatives when UML is used to run unit tests
> against a change which changes interrupt behavior.
>
> IMO it may be useful to consider some alternatives specifically for unit
> test coverage purposes where printk and/or the whole console output
> altogether bypass some of the IRQ driven semantics.

Whoops, sorry, didn't see your comment before I went on vacation.

I completely agree. It is also annoying when trying to test other
really low level parts of the kernel. I would really like to get KUnit
to the point where it does not have any dependencies on anything in
the kernel, but that is very challenging for many reasons. This
loosely relates to what Luis, myself, and others have talked about in
other threads about having a stricter notion of code dependencies in
the kernel. Thinking about it now, I suspect it might be easier to
limit KUnit's dependency on kernel infrastructure first; that could
kind of motivate the later work.