lkml.org 
[lkml]   [2022]   [Mar]   [7]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
SubjectRe: [RFC PATCH 3/6] KVM: SVM: Implement demand page pinning
From
On 3/7/2022 1:18 AM, Mingwei Zhang wrote:
> On Tue, Jan 18, 2022, Nikunj A Dadhania wrote:
>> Use the memslot metadata to store the pinned data along with the pfns.
>> This improves the SEV guest startup time from O(n) to a constant by
>> deferring guest page pinning until the pages are used to satisfy nested
>> page faults. The page reference will be dropped in the memslot free
>> path.
>>
>> Remove the enc_region structure definition and the code which did
>> upfront pinning, as they are no longer needed in view of the demand
>> pinning support.
>>
>> Leave svm_register_enc_region() and svm_unregister_enc_region() as stubs
>> since qemu is dependent on this API.
>>
>> Signed-off-by: Nikunj A Dadhania <nikunj@amd.com>
>> ---

>> +
>> + /* Pin the page, KVM doesn't yet support page migration. */
>> + for (i = 0; i < npages; i++, rel_gfn++, pin_pfn++) {
>> + if (test_bit(rel_gfn, aslot->pinned_bitmap)) {
>> + old_pfn = aslot->pfns[rel_gfn];
>> + if (old_pfn == pin_pfn)
>> + continue;
>> +
>> + put_page(pfn_to_page(old_pfn));
>
> You need to flush the old pfn using VMPAGE_FLUSH before doing put_page.
> Normally, this should not happen. But if the user-level VMM is
> malicious, then it could just munmap() the region (not the memslot);
> mmap() it again; let the guest VM touches the page and you will see this
> path get executed.
>
> Clearly, this will slow down the faulting path if this happens. So,
> alternatively, you can register a hook in mmu_notifier and shoot a flush
> there according to the bitmap. Either way should work.
>

We can call sev_flush_guest_memory() before the put_page().

>> + }
>> +
>> + set_bit(rel_gfn, aslot->pinned_bitmap);
>> + aslot->pfns[rel_gfn] = pin_pfn;
>> + get_page(pfn_to_page(pin_pfn));
>> + }
>> +
>> + /*
>> + * Flush any cached lines of the page being added since "ownership" of
>> + * it will be transferred from the host to an encrypted guest.
>> + */
>> + clflush_cache_range(__va(pfn << PAGE_SHIFT), page_level_size(level));
>> +}
>> +
>> void sev_free_memslot(struct kvm *kvm, struct kvm_memory_slot *slot)
>> {
>> struct kvm_arch_memory_slot *aslot = &slot->arch;
>> + kvm_pfn_t *pfns;
>> + gfn_t gfn;
>> + int i;
>>
>> if (!sev_guest(kvm))
>> return;
>>
>> + if (!aslot->pinned_bitmap || !slot->arch.pfns)
>> + goto out;
>> +
>> + pfns = aslot->pfns;
>> +
>> + /*
>> + * Iterate the memslot to find the pinned pfn using the bitmap and drop
>> + * the pfn stored.
>> + */
>> + for (i = 0, gfn = slot->base_gfn; i < slot->npages; i++, gfn++) {
>> + if (test_and_clear_bit(i, aslot->pinned_bitmap)) {
>> + if (WARN_ON(!pfns[i]))
>> + continue;
>> +
>> + put_page(pfn_to_page(pfns[i]));
>
> Here, you get lucky that you don't have to flush the cache. However,
> this is because sev_free_memslots is called after the
> kvm_arch_destroy_vm, which flushes the cache system wise.

I have added wbinvd_on_all_cpus() just before the iteration in my new version.

Regards
Nikunj

\
 
 \ /
  Last update: 2022-03-07 08:09    [W:0.109 / U:2.732 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site