[PATCH v2 3/3] x86: Support huge vmalloc mappings

[Christophe Leroy]

Le 28/12/2021 à 11:26, Kefeng Wang a écrit :
> 
> On 2021/12/27 23:56, Dave Hansen wrote:
>> On 12/27/21 6:59 AM, Kefeng Wang wrote:
>>> This patch select HAVE_ARCH_HUGE_VMALLOC to let X86_64 and X86_PAE
>>> support huge vmalloc mappings.
>> In general, this seems interesting and the diff is simple.  But, I don't
>> see _any_ x86-specific data.  I think the bare minimum here would be a
>> few kernel compiles and some 'perf stat' data for some TLB events.
> 
> When the feature supported on ppc,
> 
> commit 8abddd968a303db75e4debe77a3df484164f1f33
> Author: Nicholas Piggin <npiggin at gmail.com>
> Date:   Mon May 3 19:17:55 2021 +1000
> 
>      powerpc/64s/radix: Enable huge vmalloc mappings
> 
>      This reduces TLB misses by nearly 30x on a `git diff` workload on a
>      2-node POWER9 (59,800 -> 2,100) and reduces CPU cycles by 0.54%, due
>      to vfs hashes being allocated with 2MB pages.
> 
> But the data could be different on different machine/arch.
> 
>>> diff --git a/arch/x86/kernel/module.c b/arch/x86/kernel/module.c
>>> index 95fa745e310a..6bf5cb7d876a 100644
>>> --- a/arch/x86/kernel/module.c
>>> +++ b/arch/x86/kernel/module.c
>>> @@ -75,8 +75,8 @@ void *module_alloc(unsigned long size)
>>>       p = __vmalloc_node_range(size, MODULE_ALIGN,
>>>                       MODULES_VADDR + get_module_load_offset(),
>>> -                    MODULES_END, gfp_mask,
>>> -                    PAGE_KERNEL, VM_DEFER_KMEMLEAK, NUMA_NO_NODE,
>>> +                    MODULES_END, gfp_mask, PAGE_KERNEL,
>>> +                    VM_DEFER_KMEMLEAK | VM_NO_HUGE_VMAP, NUMA_NO_NODE,
>>>                       __builtin_return_address(0));
>>>       if (p && (kasan_module_alloc(p, size, gfp_mask) < 0)) {
>>>           vfree(p);
>> To figure out what's going on in this hunk, I had to look at the cover
>> letter (which I wasn't cc'd on).  That's not great and it means that
>> somebody who stumbles upon this in the code is going to have a really
>> hard time figuring out what is going on.  Cover letters don't make it
>> into git history.
> Sorry for that, will add more into arch's patch changelog.
>> This desperately needs a comment and some changelog material in *this*
>> patch.
>>
>> But, even the description from the cover letter is sparse:
>>
>>> There are some disadvantages about this feature[2], one of the main
>>> concerns is the possible memory fragmentation/waste in some scenarios,
>>> also archs must ensure that any arch specific vmalloc allocations that
>>> require PAGE_SIZE mappings(eg, module alloc with STRICT_MODULE_RWX)
>>> use the VM_NO_HUGE_VMAP flag to inhibit larger mappings.
>> That just says that x86 *needs* PAGE_SIZE allocations.  But, what
>> happens if VM_NO_HUGE_VMAP is not passed (like it was in v1)?  Will the
>> subsequent permission changes just fragment the 2M mapping?
>> .
> 
> Yes, without VM_NO_HUGE_VMAP, it could fragment the 2M mapping.
> 
> When module alloc with STRICT_MODULE_RWX on x86, it calls 
> __change_page_attr()
> 
> from set_memory_ro/rw/nx which will split large page, so there is no 
> need to make
> 
> module alloc with HUGE_VMALLOC.
> 

Maybe there is no need to perform the module alloc with HUGE_VMALLOC, 
but it least it would still work if you do so.

Powerpc did add VM_NO_HUGE_VMAP temporarily and for some reason which is 
  explained in a comment.

If x86 already has the necessary logic to handle it, why add 
VM_NO_HUGE_VMAP ?

Christophe