Re: [PATCH V12] mm/debug: Add tests validating architecture page table helpers

On 01/28/2020 09:03 AM, Qian Cai wrote:

quoted

On Jan 27, 2020, at 10:06 PM, Anshuman Khandual [off-list ref] wrote:



On 01/28/2020 07:41 AM, Qian Cai wrote:

quoted

quoted

On Jan 27, 2020, at 8:28 PM, Anshuman Khandual [off-list ref] wrote:

This adds tests which will validate architecture page table helpers and
other accessors in their compliance with expected generic MM semantics.
This will help various architectures in validating changes to existing
page table helpers or addition of new ones.

This test covers basic page table entry transformations including but not
limited to old, young, dirty, clean, write, write protect etc at various
level along with populating intermediate entries with next page table page
and validating them.

Test page table pages are allocated from system memory with required size
and alignments. The mapped pfns at page table levels are derived from a
real pfn representing a valid kernel text symbol. This test gets called
right after page_alloc_init_late().

This gets build and run when CONFIG_DEBUG_VM_PGTABLE is selected along with
CONFIG_VM_DEBUG. Architectures willing to subscribe this test also need to
select CONFIG_ARCH_HAS_DEBUG_VM_PGTABLE which for now is limited to x86 and
arm64. Going forward, other architectures too can enable this after fixing
build or runtime problems (if any) with their page table helpers.

Hello Qian,

quoted

What’s the value of this block of new code? It only supports x86 and arm64
which are supposed to be good now.

We have been over the usefulness of this code many times before as the patch is
already in it's V12. Currently it is enabled on arm64, x86 (except PAE), arc and
ppc32. There are build time or runtime problems with other archs which prevent

I am not sure if I care too much about arc and ppc32 which are pretty much legacy
platforms.

Okay but FWIW the maintainers for all these enabled platforms cared for this test
at the least and really helped in shaping the test to it's current state. Besides
I am still failing to understand your point here about evaluating particular feature's
usefulness based on it's support on relative and perceived importance of some platforms
compared to others. Again the idea is to integrate all platforms eventually but we had
discovered build and runtime issues which needs to be resolved at platform level first.
Unless I am mistaken, debug feature like this which is putting down a framework while
also benefiting some initial platforms to start with, will be a potential candidate for
eventual inclusion in the mainline. Otherwise, please point to any other agreed upon
community criteria for debug feature's mainline inclusion which I will try to adhere.
I wonder if all other similar debug features from the past ever met 'the all inclusive
at the beginning' criteria which you are trying to propose here. This test also adds a
feature file, enlisting all supported archs as suggested by Ingo for the exact same
reason. This is not the first time, a feature is listing out archs which are supported
and archs which are not.

quoted

enablement of this test (for the moment) but then the goal is to integrate all
of them going forward. The test not only validates platform's adherence to the
expected semantics from generic MM but also helps in keeping it that way during
code changes in future as well.

Another option maybe to get some decent arches on board first before merging this
thing, so it have more changes to catch regressions for developers who might run this. 

quoted

quoted

Did those tests ever find any regression or this is almost only useful for new

The test has already found problems with s390 page table helpers.

Hmm, that is pretty weak where s390 is not even official supported with this version.

And there were valid reasons why s390 could not be enabled just yet as explained by s390
folks during our previous discussions. I just pointed out an example where this test was
useful as you had asked previously. Not being official supported in this version does
not take away the fact the it was indeed useful for that platform in discovering a bug.

quoted

quoted

architectures which only happened once in a few years?

Again, not only it validates what exist today but its also a tool to make
sure that all platforms continue adhere to a common agreed upon semantics
as reflected through the tests here.

quoted

The worry if not many people will use this config and code those that much in

Debug features or tests in the kernel are used when required. These are never or
should not be enabled by default. AFAICT this is true even for entire DEBUG_VM
packaged tests. Do you have any particular data or precedence to substantiate
the fact that this test will be used any less often than the other similar ones
in the tree ? I can only speak for arm64 platform but the very idea for this
test came from Catalin when we were trying to understand the semantics for THP
helpers while enabling THP migration without split. Apart from going over the
commit messages from the past, there were no other way to figure out how any
particular page table helper is suppose to change given page table entry. This
test tries to formalize those semantics.

I am thinking about how we made so many mistakes before by merging too many of
those debugging options that many of them have been broken for many releases
proving that nobody actually used them regularly. We don’t need to repeat the same

Again will ask for some data to substantiate these claims. Though I am not really
sure but believe that there are integration test frameworks out there which regularly
validates each of these code path on multiple platforms. One such automation found
that V11 of the test was broken on X86 PAE platform which I fixed. Nonetheless, I can
speak only for arm64 platform and we intend to use this test to validate arm64 exported
page table helpers. Citing unsubstantiated past examples should not really block these
enabled platforms (arm64 at the very least) from getting this debug feature which has
already demonstrated it's usefulness during arm64 THP migration development and on s390
platforms as well.

mistake again. I am actually thinking about to remove things like  page_poisoning often
which is almost are never found any bug recently and only cause pains when interacting
with other new features that almost nobody will test them together to begin with.
We even have some SLUB debugging code sit there for almost 15 years that almost
nobody used it and maintainers refused to remove it.

Unlike those, the proposed test here is isolated as a stand alone test and stays clear
off from any other code path. I have not been involved in or aware of the usefulness of
existing MM debug features and hence will just leave them upto the judgment of the
maintainers whether to keep or discard them.

quoted

quoted

the future because it is inefficient to find bugs, it will simply be rotten

Could you be more specific here ? What parts of the test are inefficient ? I
am happy to improve upon the test. Do let me know you if you have suggestions.

quoted

like a few other debugging options out there we have in the mainline that

will be a pain to remove later on.

quoted

Even though I am not agreeing to your assessment about the usefulness of the
test without any substantial data backing up the claims, the test case in
itself is very much compartmentalized, staying clear from generic MM and
debug_vm_pgtable() is only function executing the test which is getting
called from kernel_init_freeable() path.

I am thinking exactly the other way around. You are proposing to merge this tests
without proving how useful it will be able to find regressions for future developers
to make sure it will actually get used.

As I had mentioned before, the test attempts to formalize page table helper semantics
as expected from generic MM code paths and intend to catch deviations when enabled on
a given platform. How else should we test semantics errors otherwise ? There are past
examples of usefulness for this procedure on arm64 and on s390. I am wondering how
else to prove the usefulness of a debug feature if these references are not enough.