[PATCH V2 2/2] mm/pgtable/debug: Add test validating architecture page table helpers

[Christophe Leroy]

Le 19/09/2019 à 06:56, Anshuman Khandual a écrit :
> 
> 
> On 09/18/2019 09:56 PM, Christophe Leroy wrote:
>>
>>
>> Le 18/09/2019 à 07:04, Anshuman Khandual a écrit :
>>>
>>>
>>> On 09/13/2019 03:31 PM, Christophe Leroy wrote:
>>>>
>>>>
>>>> Le 13/09/2019 à 11:02, Anshuman Khandual a écrit :
>>>>>
>>>>>>> +#if !defined(__PAGETABLE_PMD_FOLDED) && !defined(__ARCH_HAS_4LEVEL_HACK)
>>>>>>
>>>>>> #ifdefs have to be avoided as much as possible, see below
>>>>>
>>>>> Yeah but it has been bit difficult to avoid all these $ifdef because of the
>>>>> availability (or lack of it) for all these pgtable helpers in various config
>>>>> combinations on all platforms.
>>>>
>>>> As far as I can see these pgtable helpers should exist everywhere at least via asm-generic/ files.
>>>
>>> But they might not actually do the right thing.
>>>
>>>>
>>>> Can you spot a particular config which fails ?
>>>
>>> Lets consider the following example (after removing the $ifdefs around it)
>>> which though builds successfully but fails to pass the intended test. This
>>> is with arm64 config 4K pages sizes with 39 bits VA space which ends up
>>> with a 3 level page table arrangement.
>>>
>>> static void __init p4d_clear_tests(p4d_t *p4dp)
>>> {
>>>           p4d_t p4d = READ_ONCE(*p4dp);
>>
>> My suggestion was not to completely drop the #ifdef but to do like you did in pgd_clear_tests() for instance, ie to add the following test on top of the function:
>>
>>      if (mm_pud_folded(mm) || is_defined(__ARCH_HAS_5LEVEL_HACK))
>>          return;
>>
> 
> Sometimes this does not really work. On some platforms, combination of
> __PAGETABLE_PUD_FOLDED and __ARCH_HAS_5LEVEL_HACK decide whether the
> helpers such as __pud() or __pgd() is even available for that platform.
> Ideally it should have been through generic falls backs in include/*/
> but I guess there might be bugs on the platform or it has not been
> changed to adopt 5 level page table framework with required folding
> macros etc.

Yes. As I suggested below, most likely that's better to retain the 
#ifdef __ARCH_HAS_5LEVEL_HACK but change the #ifdef 
__PAGETABLE_PUD_FOLDED by a runtime test of mm_pud_folded(mm)

As pointed by Gerald, some arches don't have __PAGETABLE_PUD_FOLDED 
because they are deciding dynamically if they fold the level on not, but 
have mm_pud_folded(mm)

> 
>>>
>>>           p4d = __p4d(p4d_val(p4d) | RANDOM_ORVALUE);
>>>           WRITE_ONCE(*p4dp, p4d);
>>>           p4d_clear(p4dp);
>>>           p4d = READ_ONCE(*p4dp);
>>>           WARN_ON(!p4d_none(p4d));
>>> }
>>>
>>> The following test hits an error at WARN_ON(!p4d_none(p4d))
>>>
>>> [   16.757333] ------------[ cut here ]------------
>>> [   16.758019] WARNING: CPU: 11 PID: 1 at mm/arch_pgtable_test.c:187 arch_pgtable_tests_init+0x24c/0x474

[...]

>>> [   16.781282] ---[ end trace 042e6c40c0a3b038 ]---
>>>
>>> On arm64 (4K page size|39 bits VA|3 level page table)
>>>
>>> #elif CONFIG_PGTABLE_LEVELS == 3    /* Applicable here */
>>> #define __ARCH_USE_5LEVEL_HACK
>>> #include <asm-generic/pgtable-nopud.h>
>>>
>>> Which pulls in
>>>
>>> #include <asm-generic/pgtable-nop4d-hack.h>
>>>
>>> which pulls in
>>>
>>> #include <asm-generic/5level-fixup.h>
>>>
>>> which defines
>>>
>>> static inline int p4d_none(p4d_t p4d)
>>> {
>>>           return 0;
>>> }
>>>
>>> which will invariably trigger WARN_ON(!p4d_none(p4d)).
>>>
>>> Similarly for next test p4d_populate_tests() which will always be
>>> successful because p4d_bad() invariably returns negative.
>>>
>>> static inline int p4d_bad(p4d_t p4d)
>>> {
>>>           return 0;
>>> }
>>>
>>> static void __init p4d_populate_tests(struct mm_struct *mm, p4d_t *p4dp,
>>>                                         pud_t *pudp)
>>> {
>>>           p4d_t p4d;
>>>
>>>           /*
>>>            * This entry points to next level page table page.
>>>            * Hence this must not qualify as p4d_bad().
>>>            */
>>>           pud_clear(pudp);
>>>           p4d_clear(p4dp);
>>>           p4d_populate(mm, p4dp, pudp);
>>>           p4d = READ_ONCE(*p4dp);
>>>           WARN_ON(p4d_bad(p4d));
>>> }
>>>
>>> We should not run these tests for the above config because they are
>>> not applicable and will invariably produce same result.
>>>

[...]

>>>>
>>>> So it shouldn't be an issue. Maybe if a couple of arches miss them, the best would be to fix the arches, since that's the purpose of your testsuite isn't it ?
>>>
>>> The run time failures as explained previously is because of the folding which
>>> needs to be protected as they are not even applicable. The compile time
>>> failures are because pxx_populate() signatures are platform specific depending
>>> on how many page table levels they really support.
>>>
>>
>> So IIUC, the compiletime problem is around __ARCH_HAS_5LEVEL_HACK. For all #if !defined(__PAGETABLE_PXX_FOLDED), something equivalent to the following should make the trick.
>>
>>      if (mm_pxx_folded())
>>          return;
>>
>>
>> For the __ARCH_HAS_5LEVEL_HACK stuff, I think we should be able to regroup all impacted functions inside a single #ifdef __ARCH_HAS_5LEVEL_HACK
> 
> I was wondering if it will be better to
> 
> 1) Minimize all #ifdefs in the code which might fail on some platforms
> 2) Restrict proposed test module to platforms where it builds and runs
> 3) Enable other platforms afterwards after fixing their build problems or other requirements

I understand that __ARCH_HAS_5LEVEL_HACK is an HACK as its name 
suggests, so you can't expect all platforms to go for an HACK. I think 
you can keep a single #ifdef __ARCH_HAS_5LEVEL_HACK / #else / #endif and 
put all relevant tests inside it.

For things like __PAGETABLE_PXX_FOLDED dependancies, I still think that 
they can all be replaced by a runtime test of mm_pxx_folded().

Can you try that and see what problem remains ?

> 
> Would that be a better approach instead ?
> 

Based on the above, that might be the approach to take, yes.

Christophe