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 Wed, Jan 18, 2023 at 04:16:41PM +0800,
Chao Peng <chao.p.peng@xxxxxxxxxxxxxxx> wrote:

> On Tue, Jan 17, 2023 at 04:34:15PM +0000, Sean Christopherson wrote:
> > On Tue, Jan 17, 2023, Chao Peng wrote:
> > > On Fri, Jan 13, 2023 at 09:54:41PM +0000, Sean Christopherson wrote:
> > > > > +	list_for_each_entry(notifier, &data->notifiers, list) {
> > > > > +		notifier->ops->invalidate_start(notifier, start, end);
> > > > 
> > > > Two major design issues that we overlooked long ago:
> > > > 
> > > >   1. Blindly invoking notifiers will not scale.  E.g. if userspace configures a
> > > >      VM with a large number of convertible memslots that are all backed by a
> > > >      single large restrictedmem instance, then converting a single page will
> > > >      result in a linear walk through all memslots.  I don't expect anyone to
> > > >      actually do something silly like that, but I also never expected there to be
> > > >      a legitimate usecase for thousands of memslots.
> > > > 
> > > >   2. This approach fails to provide the ability for KVM to ensure a guest has
> > > >      exclusive access to a page.  As discussed in the past, the kernel can rely
> > > >      on hardware (and maybe ARM's pKVM implementation?) for those guarantees, but
> > > >      only for SNP and TDX VMs.  For VMs where userspace is trusted to some extent,
> > > >      e.g. SEV, there is value in ensuring a 1:1 association.
> > > > 
> > > >      And probably more importantly, relying on hardware for SNP and TDX yields a
> > > >      poor ABI and complicates KVM's internals.  If the kernel doesn't guarantee a
> > > >      page is exclusive to a guest, i.e. if userspace can hand out the same page
> > > >      from a restrictedmem instance to multiple VMs, then failure will occur only
> > > >      when KVM tries to assign the page to the second VM.  That will happen deep
> > > >      in KVM, which means KVM needs to gracefully handle such errors, and it means
> > > >      that KVM's ABI effectively allows plumbing garbage into its memslots.
> > > 
> > > It may not be a valid usage, but in my TDX environment I do meet below
> > > issue.
> > > 
> > > kvm_set_user_memory AddrSpace#0 Slot#0 flags=0x4 gpa=0x0 size=0x80000000 ua=0x7fe1ebfff000 ret=0
> > > kvm_set_user_memory AddrSpace#0 Slot#1 flags=0x4 gpa=0xffc00000 size=0x400000 ua=0x7fe271579000 ret=0
> > > kvm_set_user_memory AddrSpace#0 Slot#2 flags=0x4 gpa=0xfeda0000 size=0x20000 ua=0x7fe1ec09f000 ret=-22
> > > 
> > > Slot#2('SMRAM') is actually an alias into system memory(Slot#0) in QEMU
> > > and slot#2 fails due to below exclusive check.
> > > 
> > > Currently I changed QEMU code to mark these alias slots as shared
> > > instead of private but I'm not 100% confident this is correct fix.
> > 
> > That's a QEMU bug of sorts.  SMM is mutually exclusive with TDX, QEMU shouldn't
> > be configuring SMRAM (or any SMM memslots for that matter) for TDX guests.
> 
> Thanks for the confirmation. As long as we only bind one notifier for
> each address, using xarray does make things simple.

In the past, I had patches for qemu to disable PAM and SMRAM, but they were
dropped for simplicity because SMRAM/PAM are disabled as reset state with unused
memslot registered. TDX guest bios(TDVF or EDK2) doesn't enable them.
Now we can revive them.
-- 
Isaku Yamahata <isaku.yamahata@xxxxxxxxx>



[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