Re: [PATCH v10 1/9] mm: Introduce memfd_restricted system call to create restricted user memory

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On 1/24/23 00:38, Sean Christopherson wrote:
> On Mon, Jan 23, 2023, Huang, Kai wrote:
>> On Mon, 2023-01-23 at 15:03 +0100, Vlastimil Babka wrote:
>>> On 12/22/22 01:37, Huang, Kai wrote:
>>>>>> I argue that this page pinning (or page migration prevention) is not
>>>>>> tied to where the page comes from, instead related to how the page will
>>>>>> be used. Whether the page is restrictedmem backed or GUP() backed, once
>>>>>> it's used by current version of TDX then the page pinning is needed. So
>>>>>> such page migration prevention is really TDX thing, even not KVM generic
>>>>>> thing (that's why I think we don't need change the existing logic of
>>>>>> kvm_release_pfn_clean()). 
>>>>>>
>>>> This essentially boils down to who "owns" page migration handling, and sadly,
>>>> page migration is kinda "owned" by the core-kernel, i.e. KVM cannot handle page
>>>> migration by itself -- it's just a passive receiver.
>>>>
>>>> For normal pages, page migration is totally done by the core-kernel (i.e. it
>>>> unmaps page from VMA, allocates a new page, and uses migrate_pape() or a_ops-
>>>>> migrate_page() to actually migrate the page).
>>>> In the sense of TDX, conceptually it should be done in the same way. The more
>>>> important thing is: yes KVM can use get_page() to prevent page migration, but
>>>> when KVM wants to support it, KVM cannot just remove get_page(), as the core-
>>>> kernel will still just do migrate_page() which won't work for TDX (given
>>>> restricted_memfd doesn't have a_ops->migrate_page() implemented).
>>>>
>>>> So I think the restricted_memfd filesystem should own page migration handling,
>>>> (i.e. by implementing a_ops->migrate_page() to either just reject page migration
>>>> or somehow support it).
>>>
>>> While this thread seems to be settled on refcounts already, 
>>>
>>
>> I am not sure but will let Sean/Paolo to decide.
> 
> My preference is whatever is most performant without being hideous :-)
> 
>>> just wanted
>>> to point out that it wouldn't be ideal to prevent migrations by
>>> a_ops->migrate_page() rejecting them. It would mean cputime wasted (i.e.
>>> by memory compaction) by isolating the pages for migration and then
>>> releasing them after the callback rejects it (at least we wouldn't waste
>>> time creating and undoing migration entries in the userspace page tables
>>> as there's no mmap). Elevated refcount on the other hand is detected
>>> very early in compaction so no isolation is attempted, so from that
>>> aspect it's optimal.
>>
>> I am probably missing something,
> 
> Heh, me too, I could have sworn that using refcounts was the least efficient way
> to block migration.

Well I admit that due to my experience with it, I do mostly consider
migration through memory compaction POV, which is a significant user of
migration on random pages that's not requested by userspace actions on
specific ranges.

And compaction has in isolate_migratepages_block():

/*
 * Migration will fail if an anonymous page is pinned in memory,
 * so avoid taking lru_lock and isolating it unnecessarily in an
 * admittedly racy check.
 */
mapping = page_mapping(page);
if (!mapping && page_count(page) > page_mapcount(page))
        goto isolate_fail;

so that prevents migration of pages with elevated refcount very early,
before they are even isolated, so before migrate_pages() is called.

But it's true there are other sources of "random pages migration" - numa
balancing, demotion in lieu of reclaim... and I'm not sure if all have
such early check too.

Anyway, whatever is decided to be a better way than elevated refcounts,
would ideally be checked before isolation as well, as that's the most
efficient way.

>> but IIUC the checking of refcount happens at very last stage of page migration too 




[Index of Archives]     [Kernel Newbies]     [Security]     [Netfilter]     [Bugtraq]     [Linux FS]     [Yosemite Forum]     [MIPS Linux]     [ARM Linux]     [Linux Security]     [Linux RAID]     [Samba]     [Video 4 Linux]     [Device Mapper]     [Linux Resources]

  Powered by Linux