Re: [PATCH Part2 RFC v4 25/40] KVM: SVM: Reclaim the guest pages when SEV-SNP VM terminates

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

 



On Fri, Jul 16, 2021, Brijesh Singh wrote:
> 
> On 7/16/21 3:09 PM, Sean Christopherson wrote:
> > On Wed, Jul 07, 2021, Brijesh Singh wrote:
> >> +			e = snp_lookup_page_in_rmptable(region->pages[i], &level);
> >> +			if (unlikely(!e))
> >> +				continue;
> >> +
> >> +			/* If its not a guest assigned page then skip it. */
> >> +			if (!rmpentry_assigned(e))
> >> +				continue;
> >> +
> >> +			/* Is the page part of a 2MB RMP entry? */
> >> +			if (level == PG_LEVEL_2M) {
> >> +				val.pagesize = RMP_PG_SIZE_2M;
> >> +				pfn &= ~(KVM_PAGES_PER_HPAGE(PG_LEVEL_2M) - 1);
> >> +			} else {
> >> +				val.pagesize = RMP_PG_SIZE_4K;
> > This raises yet more questions (for me) as to the interaction between Page-Size
> > and Hyperivsor-Owned flags in the RMP.  It also raises questions on the correctness
> > of zeroing the RMP entry if KVM_SEV_SNP_LAUNCH_START (in the previous patch).
> 
> I assume you mean the LAUNCH_UPDATE because that's when we need to
> perform the RMPUPDATE.

Doh, yes.

> The hypervisor owned means all zero in the RMP entry.

Figured out where I went wrong after reading the RMPUDPATE pseudocode.  RMPUPDATE
takes the page size as a parameter even though it unconditionally zeros the page
size flag in the RMP entry for unassigned pages.

A wrapper around rmpupdate() would definitely help, e.g. (though level might need
to be an "int" to avoid a bunch of casts).

  int rmp_make_shared(u64 pfn, enum pg_level level);

Wrappers for "private" and "firmware" would probably be helpful too.  And if you
do that, I think you can bury both "struct rmpupdate", rmpupdate(), and
X86_TO_RMP_PG_LEVEL() in arch/x86/kernel/sev.c.  snp_set_rmptable_state() might
need some refactoring to avoid three booleans, but I guess maybe that could be
an exception?  Not sure.  Anyways, was thinking something like:

  int rmp_make_private(u64 pfn, u64 gpa, enum pg_level level, int asid);
  int rmp_make_firmware(u64 pfn);

It would consolidate a bit of code, and more importantly it would give visual
cues to the reader, e.g. it's easy to overlook "val = {0}" meaning "make shared".

Side topic, what happens if a firmware entry is configured with page_size=1?

And one architectural question: what prevents a malicious VMM from punching a 4k
shared page into a 2mb private page?  E.g.

  rmpupdate(1 << 20, [private, 2mb]);
  rmpupdate(1 << 20 + 4096, [shared, 4kb]);

I don't see any checks in the pseudocode that will detect this, and presumably the
whole point of a 2mb private RMP entry is to not have to go walk the individual
4kb entries on a private access.

  NEW_RMP = READ_MEM.o [NEW_RMP_PTR]

  IF ((NEW_RMP.PAGE_SIZE == 2MB) && (SYSTEM_PA[20:12] != 0))  <-- not taken, 4kb entry
          EAX = FAIL_INPUT
          EXIT

  IF (!NEW_RMP.ASSIGNED && (NEW_RMP.IMMUTABLE || (NEW_RMP.ASID != 0))  <-- not taken, new entry valid
          EAX = FAIL_INPUT
          EXIT

  RMP_ENTRY_PA = RMP_BASE + 0x4000 + (SYSTEM_PA / 0x1000) * 16
  IF (RMP_ENTRY_PA > RMP_END)
          EAX = FAIL_INPUT
          EXIT

  // System address must have an RMP entry
  OLD_RMP = READ_MEM_PA.o [RMP_ENTRY_PA]
  IF (OLD_RMP.IMMUTABLE) <-- passes, private entry not immutable
          EAX = FAIL_PERMISSION
          EXIT

  IF (NEW_RMP.PAGE_SIZE == 4KB)
          IF ((SYSTEM_PA[20:12] == 0) && (OLD_RMP.PAGE_SIZE == 2MB)) <- not taken, PA[12] == 1
                  EAX = FAIL_OVERLAP
                  EXIT
  ELSE
          IF (Any 4KB RMP entry with (RMP.ASSIGNED == 1) exists in 2MB region)
                  EAX = FAIL_OVERLAP
                  EXIT
          ELSE
                  FOR (I = 1; I < 512, I++) {
                          temp_RMP = 0
                          temp_RMP.ASSIGNED = NEW_RMP.ASSIGNED
                          WRITE_MEM.o [RMP_ENTRY_PA + I * 16] = temp_RMP;
                  }



[Index of Archives]     [Linux Kernel Development]     [Linux USB Devel]     [Video for Linux]     [Linux Audio Users]     [Yosemite News]     [Linux Kernel]     [Linux SCSI]

  Powered by Linux