Re: [PATCH 1/5] x86/coco: Use slow_virt_to_phys() in page transition hypervisor callbacks

From: Tom Lendacky
Date: Mon Oct 02 2023 - 11:52:31 EST


On 9/29/23 13:19, Michael Kelley wrote:
In preparation for temporarily marking pages not present during a
transition between encrypted and decrypted, use slow_virt_to_phys()
in the hypervisor callbacks. As long as the PFN is correct,
slow_virt_to_phys() works even if the leaf PTE is not present.
The existing functions that depends on vmalloc_to_page() all
require that the leaf PTE be marked present, so they don't work.

Update the comments for slow_virt_to_phys() to note this broader usage
and the requirement to work even if the PTE is not marked present.

Signed-off-by: Michael Kelley <mikelley@xxxxxxxxxxxxx>
---
arch/x86/hyperv/ivm.c | 9 ++++++++-
arch/x86/kernel/sev.c | 8 +++++++-
arch/x86/mm/pat/set_memory.c | 13 +++++++++----
3 files changed, 24 insertions(+), 6 deletions(-)

diff --git a/arch/x86/hyperv/ivm.c b/arch/x86/hyperv/ivm.c
index c1088d3..084fab6 100644
--- a/arch/x86/hyperv/ivm.c
+++ b/arch/x86/hyperv/ivm.c
@@ -524,7 +524,14 @@ static bool hv_vtom_set_host_visibility(unsigned long kbuffer, int pagecount, bo
return false;
for (i = 0, pfn = 0; i < pagecount; i++) {
- pfn_array[pfn] = virt_to_hvpfn((void *)kbuffer + i * HV_HYP_PAGE_SIZE);
+ /*
+ * Use slow_virt_to_phys() because the PRESENT bit has been
+ * temporarily cleared in the PTEs. slow_virt_to_phys() works
+ * without the PRESENT bit while virt_to_hvpfn() or similar
+ * does not.
+ */
+ pfn_array[pfn] = slow_virt_to_phys((void *)kbuffer +
+ i * HV_HYP_PAGE_SIZE) >> HV_HYP_PAGE_SHIFT;
pfn++;
if (pfn == HV_MAX_MODIFY_GPA_REP_COUNT || i == pagecount - 1) {
diff --git a/arch/x86/kernel/sev.c b/arch/x86/kernel/sev.c
index 2787826..f5d6cec 100644
--- a/arch/x86/kernel/sev.c
+++ b/arch/x86/kernel/sev.c
@@ -784,7 +784,13 @@ static unsigned long __set_pages_state(struct snp_psc_desc *data, unsigned long
hdr->end_entry = i;
if (is_vmalloc_addr((void *)vaddr)) {
- pfn = vmalloc_to_pfn((void *)vaddr);
+ /*
+ * Use slow_virt_to_phys() because the PRESENT bit has been

Since __set_pages_state() is called by more than one path, please update this comment to add something like "because when called via the set_memory_encrypted() or set_memory_decrypted() path the PRESENT bit..."

Thanks,
Tom

+ * temporarily cleared in the PTEs. slow_virt_to_phys() works
+ * without the PRESENT bit while vmalloc_to_pfn() or similar
+ * does not.
+ */
+ pfn = slow_virt_to_phys((void *)vaddr) >> PAGE_SHIFT;
use_large_entry = false;
} else {
pfn = __pa(vaddr) >> PAGE_SHIFT;
diff --git a/arch/x86/mm/pat/set_memory.c b/arch/x86/mm/pat/set_memory.c
index bda9f12..8e19796 100644
--- a/arch/x86/mm/pat/set_memory.c
+++ b/arch/x86/mm/pat/set_memory.c
@@ -755,10 +755,15 @@ pmd_t *lookup_pmd_address(unsigned long address)
* areas on 32-bit NUMA systems. The percpu areas can
* end up in this kind of memory, for instance.
*
- * This could be optimized, but it is only intended to be
- * used at initialization time, and keeping it
- * unoptimized should increase the testing coverage for
- * the more obscure platforms.
+ * It is also used in callbacks for CoCo VM page transitions between private
+ * and shared because it works when the PRESENT bit is not set in the leaf
+ * PTE. In such cases, the state of the PTEs, including the PFN, is otherwise
+ * known to be valid, so the returned physical address is correct. The similar
+ * function vmalloc_to_pfn() can't be used because it requires the PRESENT bit.
+ *
+ * This could be optimized, but it is only used in paths that are not perf
+ * sensitive, and keeping it unoptimized should increase the testing coverage
+ * for the more obscure platforms.
*/
phys_addr_t slow_virt_to_phys(void *__virt_addr)
{