On 28/02/2019 18:55, Paul Burton wrote:
> Hi Steven,
>
> On Thu, Feb 28, 2019 at 12:11:24PM +0000, Steven Price wrote:
>> On 28/02/2019 02:15, Paul Burton wrote:
>>> On Wed, Feb 27, 2019 at 05:05:45PM +0000, Steven Price wrote:
>>>> For mips, we don't support large pages on 32 bit so add stubs returning 0.
>>>
>>> So far so good :)
>>>
>>>> For 64 bit look for _PAGE_HUGE flag being set. This means exposing the
>>>> flag when !CONFIG_MIPS_HUGE_TLB_SUPPORT.
>>>
>>> Here I have to ask why? We could just return 0 like the mips32 case when
>>> CONFIG_MIPS_HUGE_TLB_SUPPORT=n, let the compiler optimize the whole
>>> thing out and avoid redundant work at runtime.
>>>
>>> This could be unified too in asm/pgtable.h - checking for
>>> CONFIG_MIPS_HUGE_TLB_SUPPORT should be sufficient to cover the mips32
>>> case along with the subset of mips64 configurations without huge pages.
>>
>> The intention here is to define a new set of macros/functions which will
>> always tell us whether we're at the leaf of a page table walk, whether
>> or not huge pages are compiled into the kernel. Basically this allows
>> the page walking code to be used on page tables other than user space,
>> for instance the kernel page tables (which e.g. might use a large
>> mapping for linear memory even if huge pages are not compiled in) or
>> page tables from firmware (e.g. EFI on arm64).
>>
>> I'm not familiar enough with mips to know how it handles things like the
>> linear map so I don't know how relevant that is, but I'm trying to
>> introduce a new set of functions which differ from the existing
>> p?d_huge() macros by not depending on whether these mappings could exist
>> for a user space VMA (i.e. not depending on HUGETLB support and existing
>> for all levels that architecturally they can occur at).
>
> Thanks for the explanation - the background helps.
>
> Right now for MIPS, with one exception, there'll be no difference
> between a page being huge or large. So for the vast majority of kernels
> with CONFIG_MIPS_HUGE_TLB_SUPPORT=n we should just return 0.
>
> The one exception I mentioned is old SGI IP27 support, which allows the
> kernel to be mapped through the TLB & does that using 2x 16MB pages when
> CONFIG_MAPPED_KERNEL=y. However even there your patch as-is won't pick
> up on that for 2 reasons:
>
> 1) The pages in question don't appear to actually be recorded in the
> page tables - they're just written straight into the TLB as wired
> entries (ie. entries that will never be evicted).
>
> 2) Even if they were in the page tables the _PAGE_HUGE bit isn't set.
>
> Since those pages aren't recorded in the page tables anyway we'd either
> need to:
>
> a) Add them to the page tables, and set the _PAGE_HUGE bit.
>
> b) Ignore them if the code you're working on won't be operating on the
> memory mapping the kernel.
>
> For other platforms the kernel is run from unmapped memory, and for all
> cases including IP27 the kernel will use unmapped memory to access
> lowmem or peripherals when possible. That is, MIPS has virtual address
> regions ((c)kseg[01] or xkphys) which are architecturally defined as
> linear maps to physical memory & so VA->PA translation doesn't use the
> TLB at all.
>
> So my thought would be that for almost everything we could just do:
>
> #define pmd_large(pmd) pmd_huge(pmd)
> #define pud_large(pmd) pud_huge(pmd)
>
> And whether we need to do anything about IP27 depends on whether a) or
> b) is chosen above.
>
> Or alternatively you could do something like:
>
> #ifdef _PAGE_HUGE
>
> static inline int pmd_large(pmd_t pmd)
> {
> return (pmd_val(pmd) & _PAGE_HUGE) != 0;
> }
>
> static inline int pud_large(pud_t pud)
> {
> return (pud_val(pud) & _PAGE_HUGE) != 0;
> }
>
> #else
> # define pmd_large(pmd) 0
> # define pud_large(pud) 0
> #endif
>
> That would cover everything except for the IP27, but would make it pick
> up the IP27 kernel pages automatically if someone later defines
> _PAGE_HUGE for IP27 CONFIG_MAPPED_KERNEL=y & makes use of it for those
> pages.
Thanks for the detailed explanation. I think my preference is your
change above (#ifdef _PAGE_HUGE) because I'm trying to stop people
thinking p?d_large==p?d_huge. MIPS is a little different from other
architectures in that the hardware doesn't walk the page tables, so
there isn't a definitive answer as to whether there is a 'huge' bit in
the tables or not - it actually does depend on the kernel configuration.
For the IP27 case I think the current situation is probably fine - the
intention is to walk the page tables, so even though the TLBs (and
therefore the actual translations) might not match, at least p?d_large
will accurately tell when the leaf of the page table tree has been
reached. And as you say, using _PAGE_HUGE as the #ifdef means that
should support be added the code should automatically make use of it.
Thanks for your help,
Steve
