Re: [PATCH 1/6] mm: introduce kv[mz]alloc helpers

[John Hubbard <jhubbard@nvidia.com>]

On 01/19/2017 12:45 AM, Michal Hocko wrote:
> On Thu 19-01-17 00:37:08, John Hubbard wrote:
>>
>>
>> On 01/18/2017 12:21 AM, Michal Hocko wrote:
>>> On Tue 17-01-17 21:59:13, John Hubbard wrote:
> [...]
>>>>  * Reclaim modifiers - __GFP_NORETRY and __GFP_NOFAIL should not be passed in.
>>>>  * Passing in __GFP_REPEAT is supported, but note that it is ignored for small
>>>>  * (<=64KB) allocations, during the kmalloc attempt.
>>>
>>>> __GFP_REPEAT is fully
>>>>  * honored for  all allocation sizes during the second part: the vmalloc attempt.
>>>
>>> this is not true to be really precise because vmalloc doesn't respect
>>> the given gfp mask all the way down (look at the pte initialization).
>>>
>>
>> I'm having some difficulty in locating that pte initialization part, am I on
>> the wrong code path? Here's what I checked, before making the claim about
>> __GFP_REPEAT being honored:
>>
>> kvmalloc_node
>>   __vmalloc_node_flags
>>     __vmalloc_node
>>       __vmalloc_node_range
>>         __vmalloc_area_node
> 	    map_vm_area
> 	      vmap_page_range
> 	        vmap_page_range_noflush
> 		  vmap_pud_range
> 		    pud_alloc
> 		      __pud_alloc
> 		        pud_alloc_one
>
> pud will be allocated but the same pattern repeats on the pmd and pte
> levels. This is btw. one of the reasons why vmalloc with gfp flags is
> tricky!

Yes, I see that now, thank you for explaining, much appreciated. The flags are left 
way behind in the code path.

So that leaves us with maybe this for documentation?

  * Reclaim modifiers - __GFP_NORETRY and __GFP_NOFAIL should not be passed in.
  * Passing in __GFP_REPEAT is supported, and will cause the following behavior:
  * for larger (>64KB) allocations, the first part (kmalloc) will do some
  * retrying, before falling back to vmalloc.


>
> moreover
>>             alloc_pages_node
>
> this is order-0 request so...
>
>>               __alloc_pages_node
>>                 __alloc_pages
>>                   __alloc_pages_nodemask
>>                     __alloc_pages_slowpath
>>
>>
>> ...and __alloc_pages_slowpath does the __GFP_REPEAT handling:
>>
>>     /*
>>      * Do not retry costly high order allocations unless they are
>>      * __GFP_REPEAT
>>      */
>>     if (order > PAGE_ALLOC_COSTLY_ORDER && !(gfp_mask & __GFP_REPEAT))
>>         goto nopage;
>
> ... this doesn't apply
>

yes, true.

thanks
john h

>