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

Re: [Xen-devel] [V3 PATCH 7/9] x86/hvm: pkeys, add pkeys support for guest_walk_tables

On 18/12/15 08:21, Han, Huaitong wrote:
> On Wed, 2015-12-16 at 15:36 +0000, George Dunlap wrote:
>> [Adding Tim, the previous mm maintainer]
>> On 11/12/15 09:16, Wu, Feng wrote:
>>>>> +{
>>>>> +    void *xsave_addr;
>>>>> +    unsigned int pkru = 0;
>>>>> +    bool_t pkru_ad, pkru_wd;
>>>>> +
>>>>> +    bool_t uf = !!(pfec & PFEC_user_mode);
>>>>> +    bool_t wf = !!(pfec & PFEC_write_access);
>>>>> +    bool_t ff = !!(pfec & PFEC_insn_fetch);
>>>>> +    bool_t rsvdf = !!(pfec & PFEC_reserved_bit);
>>>>> +    bool_t pkuf  = !!(pfec & PFEC_prot_key);
>>>> So I'm just wondering out loud here -- is there actually any
>>>> situation
>>>> in which we would want guest_walk_tables to act differently than
>>>> the
>>>> real hardware?
>>>> That is, is there actually any situation where, pku is enabled,
>>>> the vcpu
>>>> is in long mode, PFEC_write_access and/or PFEC_page_present is
>>>> set, and
>>>> the pkey is non-zero, that we want guest_walk_tables() to only
>>>> check the
>>>> write-protect bit for the pte, and not also check the pkru?
>>>> Because if not, it seems like it would be much more robust to
>>>> simply
>>>> *always* check for pkru_ad if PFEC_page_present is set, and for
>>>> pkru_wd
>>>> if PFEC_write_access is set.
>>> guest_walk_tables() is also used by shadow code, though we don't
>>> plan to support pkeys for shadow now, however, in that case, the
>>> 'pfec'
>>> is generated by hardware, and the pkuf bit may be 0 or 1 depending
>>> on the real page fault. If we unconditionally check pkeys in
>>> guest_walk_tables(), it is not a good ideas for someone who may
>>> want to implement the pkeys for shadow in future, since we only
>>> need to check pkeys when the pkuf is set in pfec.
>> So you'll have to forgive me for being a bit slow here -- I stepped
>> away
>> from the mm code for a while, and a lot has changed since I agreed to
>> become mm maintainer earlier this year.
>> So here is what I see in the tree; please correct me if I missed
>> something important.
>> The function guest_walk_tables():
>>  1. Accepts a pfec argument which is meant to describe the *access
>> type*
>> that is happening
>>  2. Returns a value with the "wrong" flags (i.e., flags which needed
>> to
>> be set that were missing, or flags that needed to be clear that were
>> set).
>>  3. It checks reserved bits in the pagetable regardless of whether
>> PFEC_reserved_bit is set in the caller, and returns
>> if so.
>>  4. If PFEC_insn_fetch is set, it will only check for nx and smep if
>> those features are enabled for the guest.
>> The main callers are the various gva_to_gfn(), which accept a
>> *pointer*
>> to a pfec argument.  This pointer is actually bidirectional: its
>> value
>> passed to guest_walk_tables() to determine what access types are
>> checked; what is returned is meant to be a pfec value which can be
>> passed to a guest as part of a fault (and is by several callers). 
>>  But
>> it may also return the Xen-internal flags, PFEC_page_{paged,shared}.
>> And, importantly, it modifies the pfec value based on the bits that
>> are
>> returned from guest_walk_tables(): It will clear PFEC_present if
>> guest_walk_tables() returns _PAGE_PRESENT, and it will set
>> PFEC_reserved_bit if guest_walk_tables() returns _PAGE_INVALID_BIT.
>> The next logical level up are the
>> hvm_{copy,fetch}_{to,from}_guest_virt(), which also take a pfec
>> argument: but really the only purpose of the pfec argument is to
>> allow
>> the caller to add the PFEC_user_mode flag; the other access-type
>> flags
>> are set automatically by the functions themselves (e.g., "to" sets
>> PFEC_write_access, "fetch" sets PFEC_insn_fetch, &c).  Several
>> callers
>> set PFEC_present as well, but callers set any other bits.
>> (hvm_fetch_from_guest_virt() seems to only set PFEC_insn_fetch if nx
>> or
>> smep are enabled in the guest.  This seems inconsistent to me with
>> the
>> treatment of PFEC_reserved_bit: it seems like
>> hvm_fetch_from_guest_virt() should always pass in PFEC_insn_fetch,
>> particularly as guest_walk_tables() will already gate the checks
>> based
>> on whether nx or smep is enabled in the guest.  Tim, you know of any
>> reason for this?)
>> So there seems to me to be no reason to pass PFEC_prot_key into
>> guest_walk_tables() or gva_to_gfn().  The pfec value passed *into*
>> those
>> should simply indicate the type of memory access being done: present,
>> write, instruction fetch, user.
>> With your current series, guest_walk_tables() already checks for
>> pkeys
>> being enabled in the guest before checking for them in the
>> pagetables.
>> For shadow mode, these will be false, and so no checks will be done. 
>>  If
>> anyone ever implements pkeys for shadow mode, then these will be
>> enabled, and the checks will be done, without any intervention on the
>> part of the caller.
> I have understood it, but, the problem with shadow mode is that pfec
> may come from regs->error_codeïhardwareï, just like:
> rc = sh_walk_guest_tables(v, va, &gw, regs->error_code);
> so, when regs->error_code does not have PFEC_prot_key,
> guest_walk_tables may still check PKEY when codes is writen according
> to what you said, and it maybe return a different result.

Under what situation would the *hardware* not generate PFEC_prot_key,
but guest_walk_tables() would, under exactly the same conditions,
generate PFEC_prot_key?  Shouldn't guest_walk_tables() be made to behave
identically to the hardware?

The only situation where that might legitimately happen is when the
shadow pagetables and the guest pagetables diverge; for instance, when a
page had been write-protected because Xen thought it was a pagetable.
Then the guest pte might have write permission, but the pkey read-only
permission; but the shadow pte would not have write permission.  In that
case, the hardware might give PFEC_write_access but not PFEC_prot_key, I
suppose (haven't read the manual to be sure); but in that case, we
*want* guest_walk_tables() to detect and add PFEC_prot_key, don't we?


Xen-devel mailing list



Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.