On Tue, Aug 14, 2012 at 10:58:07AM +0800, Xiao Guangrong wrote:
> On 08/14/2012 01:39 AM, Marcelo Tosatti wrote:
> > On Sat, Aug 11, 2012 at 11:36:20AM +0800, Xiao Guangrong wrote:
> >> On 08/11/2012 02:14 AM, Marcelo Tosatti wrote:
> >>> On Tue, Aug 07, 2012 at 05:47:15PM +0800, Xiao Guangrong wrote:
> >>>> Changelog:
> >>>> - introduce KVM_PFN_ERR_RO_FAULT instead of dummy page
> >>>> - introduce KVM_HVA_ERR_BAD and optimize error hva indicators
> >>>>
> >>>> The test case can be found at:
> >>>> http://lkml.indiana.edu/hypermail/linux/kernel/1207.2/00819/migrate-perf.tar.bz2
> >>>>
> >>>> In current code, if we map a readonly memory space from host to guest
> >>>> and the page is not currently mapped in the host, we will get a fault-pfn
> >>>> and async is not allowed, then the vm will crash.
> >>>>
> >>>> As Avi's suggestion, We introduce readonly memory region to map ROM/ROMD
> >>>> to the guest, read access is happy for readonly memslot, write access on
> >>>> readonly memslot will cause KVM_EXIT_MMIO exit.
> >>>
> >>> Memory slots whose QEMU mapping is write protected is supported
> >>> today, as long as there are no write faults.
> >>>
> >>> What prevents the use of mmap(!MAP_WRITE) to handle read-only memslots
> >>> again?
> >>>
> >>
> >> It is happy to map !write host memory space to the readonly memslot,
> >> and they can coexist as well.
> >>
> >> readonly memslot checks the write-permission by seeing slot->flags and
> >> !write memory checks the write-permission in hva_to_pfn() function
> >> which checks vma->flags. It is no conflict.
> > 
> > Yes, there is no conflict. The point is, if you can use the
> > mmap(PROT_READ) interface (supporting read faults on read-only slots)
> > for this behavior, what is the advantage of a new memslot flag?
> > 
> 
> You can get the discussion at:
> https://lkml.org/lkml/2012/5/22/228
> 
> > I'm not saying mmap(PROT_READ) is the best interface, i am just asking
> > why it is not.
> 
> My fault. :(
> 
> > 
> >>> The initial objective was to fix a vm crash, can you explain that
> >>> initial problem?
> >>>
> >>
> >> The issue was trigged by this code:
> >>
> >>                 } else {
> >>                         if (async && (vma->vm_flags & VM_WRITE))
> >>                                 *async = true;
> >>                         pfn = KVM_PFN_ERR_FAULT;
> >>                 }
> >>
> >> If the host memory region is readonly (!vma->vm_flags & VM_WRITE) and
> >> its physical page is swapped out (or the file data does not be read in),
> >> get_user_page_nowait will fail, above code reject to set async,
> >> then we will get a fault pfn and async=false.
> >>
> >> I guess this issue also exists in "QEMU write protected mapping" as
> >> you mentioned above.
> > 
> > Yes, it does. As far as i understand, what that check does from a high
> > level pov is:
> > 
> > - Did get_user_pages_nowait() fail due to a swapped out page (in which 
> > case we should try to swappin the page asynchronously), or due to 
> > another reason (for which case an error should be returned).
> > 
> > Using vma->vm_flags VM_WRITE for that is trying to guess why
> > get_user_pages_nowait() failed, because it (gup_nowait return values) 
> > does not provide sufficient information by itself.
> > 
> 
> That is exactly what i did in the first version. :)
> 
> You can see it and the reason why it switched to the new way (readonly memslot)
> in the above website (the first message in thread).

Userspace can create multiple mappings for the same memory region, for
example via shared memory (shm_open), and have different protections for
the two (or more) regions. I had old patch doing this, its attached.

> > Can't that be fixed separately? 
> > 
> > Another issue which is also present with the mmap(PROT_READ) scheme is
> > interaction with reexecute_instruction. That is, unless i am mistaken,
> > reexecute_instruction can succeed (return true) on a region that is
> > write protected. This breaks the "write faults on read-only slots exit
> > to userspace via EXIT_MMIO" behaviour.
> 
> Sorry, Why? After re-entry to the guest, it can not generate a correct MMIO?

reexecute_instruction validates presence of GPA by looking at registered
memslots. But if the access is a write, and userspace memory map is
read-only, reexecute_instruction should exit via MMIO.

That is, reexecute_instruction must validate GPA using registered
memslots AND additionaly userspace map permission, not only registered
memslot.