Re: [PATCH] perf: map pages in advance

[David Hildenbrand <david@redhat.com>]

On 29.11.24 15:35, Lorenzo Stoakes wrote:
> On Fri, Nov 29, 2024 at 03:09:49PM +0100, David Hildenbrand wrote:
>>>>> I just tested it and write() works fine, it uses uaccess afaict as part of the
>>>>> lib/iov_iter.c code:
>>>>>
>>>>> generic_perform_write()
>>>>> -> copy_folio_from_iter_atomic()
>>>>> -> copy_page_from_iter_atomic()
>>>>> -> __copy_from_iter()
>>>>> -> copy_from_user_iter()
>>>>> -> raw_copy_from_user()
>>>>> -> copy_user_generic()
>>>>> -> [uaccess asm]
>>>>>
>>>>
>>>> Ah yes. O_DIRECT is the problematic bit I suspect, which will use GUP.
>>>>
>>>> Ptrace access and friends should also no longer work on these pages, likely
>>>> that's tolerable.
>>>
>>> Yeah Peter can interject if not, but I'd be _very_ surprised if anybody expects
>>> to be able to ptrace perf counter mappings in another process (it'd be kind of
>>> totally insane to do that anyway since it's a ring buffer that needs special
>>> handing anyway).
>>
>> I think so as well. Disallowing GUP has some side effects, like not getting
>> these pages included in a coredump etc ... at least I hope nobody uses
>> O_DIRECT on them.
> 
> We set VM_DONTDUMP anyway (set by remap_pfn_range() also) so this part won't be
> a problem, and I can't see anybody using O_DIRECT on them, sensibly.

Ah, even better.

>>
>> Actually, the whole thing is on my todo list, because messing with the RMAP
>> with vm_insert_pages() doesn't make any sense (whereby the refcount might!).
>> See the TODO I added in __folio_rmap_sanity_checks().
> 
> How long is your TODO list I wonder? :)) I imagine it requires a huge page to
> map at this point..

:D

Too long, but as some of these TODOs stand in the memdesc way, 
fortunately other people might be able to give a helping hand at some 
point ;)

I'll play with using a page type for some of these "simple" cases and 
see how hard it will get.

> 
>>
>> --
>> Cheers,
>>
>> David / dhildenb
>>
> 
> Cool will respin and send out shortly with an appropriately 'forgive us father
> for we have sinned' comment.
> 
> Thanks for taking the time to discuss this! Much appreciated.

Thanks for bearing with me. Whenever PFNMAP/MIXEDMAP is involved it gets 
tricky, and as soon as PFNMAP/MIXEDMAP is inevitable, things get ugly. :)

-- 
Cheers,

David / dhildenb