On Wed, 2024-07-10 at 11:20 +0100, David Woodhouse wrote:
> On Wed, 2024-07-10 at 10:49 +0100, Patrick Roy wrote:
>> On 7/9/24 15:36, David Woodhouse wrote:
>
> I did? It isn't September yet, surely?
Argh, thanks for letting me know, I think I've whacked some sense into
my mail client now :)
>>> On Tue, 2024-07-09 at 14:20 +0100, Patrick Roy wrote:
>>>> KVM uses gfn_to_pfn_caches to cache translations from gfn all the way to
>>>> the pfn (for example, kvm-clock caches the page storing the page used
>>>> for guest/host communication this way). Unlike the gfn_to_hva_cache,
>>>> where no equivalent caching semantics were possible to gmem-backed gfns
>>>> (see also 858e8068a750 ("kvm: pfncache: enlighten about gmem")), here it
>>>> is possible to simply cache the pfn returned by `kvm_gmem_get_pfn`.
>>>>
>>>> Additionally, gfn_to_pfn_caches now invalidate whenever a cached gfn's
>>>> attributes are flipped from shared to private (or vice-versa).
>>>>
>>>> Signed-off-by: Patrick Roy <roypat@xxxxxxxxxxxx>
>>>
>>> I can't see how this is safe from race conditions.
>>>
>>> When the GPC is invalidated from gfn_to_pfn_cache_invalidate_start()
>>> its *write* lock is taken and gpc->valid is set to false.
>>>
>>> In parallel, any code using the GPC to access guest memory will take
>>> the *read* lock, call kvm_gpc_check(), and then go ahead and use the
>>> pointer to its heart's content until eventually dropping the read lock.
>>>
>>> Since invalidation takes the write lock, it has to wait until the GPC
>>> is no longer in active use, and the pointer cannot be being
>>> dereferenced.
>>>
>>> How does this work for the kvm_mem_is_private() check. You've added a
>>> check in kvm_gpc_check(), but what if the pfn is made private
>>> immediately *after* that check? Unless the code path which makes the
>>> pfn private also takes the write lock, how is it safe?
>>
>> Ah, you're right - I did in fact overlook this. I do think that it works
>> out though: kvm_vm_set_mem_attributes, which is used for flipping
>> between shared/private, registers the range which had its attributes
>> changed for invalidation, and thus gfn_to_pfn_cache_invalidate_start
>> should get called for it (although I have to admit I do not immediately
>> see what the exact callstack for this looks like, so maybe I am
>> misunderstanding something about invalidation here?).
>
> In that case, wouldn't that mean the explicit checks on gpc->is_private
> matching kvm_mem_is_private() would be redundant and you can remove
> them because you can trust that gpc->valid would be cleared?
>
Right, yes, it would indeed mean that. I'll double-check my assumption
about the whole invalidation thing and adjust the code for the next
iteration!
