> On 1 Aug 2020, at 00:03, Stefano Stabellini <sstabell...@kernel.org> wrote:
>
> On Fri, 31 Jul 2020, Bertrand Marquis wrote:
>> Sorry missed some points in my previous answer.
>>
>>> On 30 Jul 2020, at 22:50, Julien Grall <jul...@xen.org> wrote:
>>>
>>> Hi Bertrand,
>>>
>>> To avoid extra work on your side, I would recommend to wait a bit before
>>> sending a new version. It would be good to at least settle the conversation
>>> in v2 regarding the approach taken.
>>>
>>> On 30/07/2020 11:24, Bertrand Marquis wrote:
>>>> At the moment on Arm, a Linux guest running with KTPI enabled will
>>>> cause the following error when a context switch happens in user mode:
>>>> (XEN) p2m.c:1890: d1v0: Failed to walk page-table va 0xffffff837ebe0cd0
>>>> The error is caused by the virtual address for the runstate area
>>>> registered by the guest only being accessible when the guest is running
>>>> in kernel space when KPTI is enabled.
>>>> To solve this issue, this patch is doing the translation from virtual
>>>> address to physical address during the hypercall and mapping the
>>>> required pages using vmap. This is removing the conversion from virtual
>>>> to physical address during the context switch which is solving the
>>>> problem with KPTI.
>>>
>>> To echo what Jan said on the previous version, this is a change in a stable
>>> ABI and therefore may break existing guest. FAOD, I agree in principle with
>>> the idea. However, we want to explain why breaking the ABI is the *only*
>>> viable solution.
>>>
>>> From my understanding, it is not possible to fix without an ABI breakage
>>> because the hypervisor doesn't know when the guest will switch back from
>>> userspace to kernel space. The risk is the information provided by the
>>> runstate wouldn't contain accurate information and could affect how the
>>> guest handle stolen time.
>>>
>>> Additionally there are a few issues with the current interface:
>>> 1) It is assuming the virtual address cannot be re-used by the userspace.
>>> Thanksfully Linux have a split address space. But this may change with KPTI
>>> in place.
>>> 2) When update the page-tables, the guest has to go through an invalid
>>> mapping. So the translation may fail at any point.
>>>
>>> IOW, the existing interface can lead to random memory corruption and
>>> inacurracy of the stolen time.
>>>
>>>> This is done only on arm architecture, the behaviour on x86 is not
>>>> modified by this patch and the address conversion is done as before
>>>> during each context switch.
>>>> This is introducing several limitations in comparison to the previous
>>>> behaviour (on arm only):
>>>> - if the guest is remapping the area at a different physical address Xen
>>>> will continue to update the area at the previous physical address. As
>>>> the area is in kernel space and usually defined as a global variable this
>>>> is something which is believed not to happen. If this is required by a
>>>> guest, it will have to call the hypercall with the new area (even if it
>>>> is at the same virtual address).
>>>> - the area needs to be mapped during the hypercall. For the same reasons
>>>> as for the previous case, even if the area is registered for a different
>>>> vcpu. It is believed that registering an area using a virtual address
>>>> unmapped is not something done.
>>>
>>> This is not clear whether the virtual address refer to the current vCPU or
>>> the vCPU you register the runstate for. From the past discussion, I think
>>> you refer to the former. It would be good to clarify.
>>>
>>> Additionally, all the new restrictions should be documented in the public
>>> interface. So an OS developper can find the differences between the
>>> architectures.
>>>
>>> To answer Jan's concern, we certainly don't know all the guest OSes
>>> existing, however we also need to balance the benefit for a large majority
>>> of the users.
>>>
>>> From previous discussion, the current approach was deemed to be acceptable
>>> on Arm and, AFAICT, also x86 (see [1]).
>>>
>>> TBH, I would rather see the approach to be common. For that, we would an
>>> agreement from Andrew and Jan in the approach here. Meanwhile, I think this
>>> is the best approach to address the concern from Arm users.
>>>
>>>> inline functions in headers could not be used as the architecture
>>>> domain.h is included before the global domain.h making it impossible
>>>> to use the struct vcpu inside the architecture header.
>>>> This should not have any performance impact as the hypercall is only
>>>> called once per vcpu usually.
>>>> Signed-off-by: Bertrand Marquis <bertrand.marq...@arm.com>
>>>> ---
>>>> Changes in v2
>>>> - use vmap to map the pages during the hypercall.
>>>> - reintroduce initial copy during hypercall.
>>>> Changes in v3
>>>> - Fix Coding style
>>>> - Fix vaddr printing on arm32
>>>> - use write_atomic to modify state_entry_time update bit (only
>>>> in guest structure as the bit is not used inside Xen copy)
>>>> ---
>>>> xen/arch/arm/domain.c | 161 ++++++++++++++++++++++++++++++-----
>>>> xen/arch/x86/domain.c | 29 ++++++-
>>>> xen/arch/x86/x86_64/domain.c | 4 +-
>>>> xen/common/domain.c | 19 ++---
>>>> xen/include/asm-arm/domain.h | 9 ++
>>>> xen/include/asm-x86/domain.h | 16 ++++
>>>> xen/include/xen/domain.h | 5 ++
>>>> xen/include/xen/sched.h | 16 +---
>>>> 8 files changed, 206 insertions(+), 53 deletions(-)
>>>> diff --git a/xen/arch/arm/domain.c b/xen/arch/arm/domain.c
>>>> index 31169326b2..8b36946017 100644
>>>> --- a/xen/arch/arm/domain.c
>>>> +++ b/xen/arch/arm/domain.c
>>>> @@ -19,6 +19,7 @@
>>>> #include <xen/sched.h>
>>>> #include <xen/softirq.h>
>>>> #include <xen/wait.h>
>>>> +#include <xen/vmap.h>
>>>> #include <asm/alternative.h>
>>>> #include <asm/cpuerrata.h>
>>>> @@ -275,36 +276,156 @@ static void ctxt_switch_to(struct vcpu *n)
>>>> virt_timer_restore(n);
>>>> }
>>>> -/* Update per-VCPU guest runstate shared memory area (if registered). */
>>>> -static void update_runstate_area(struct vcpu *v)
>>>> +static void cleanup_runstate_vcpu_locked(struct vcpu *v)
>>>> {
>>>> - void __user *guest_handle = NULL;
>>>> + if ( v->arch.runstate_guest )
>>>> + {
>>>> + vunmap((void *)((unsigned long)v->arch.runstate_guest &
>>>> PAGE_MASK));
>>>> +
>>>> + put_page(v->arch.runstate_guest_page[0]);
>>>> +
>>>> + if ( v->arch.runstate_guest_page[1] )
>>>> + put_page(v->arch.runstate_guest_page[1]);
>>>> +
>>>> + v->arch.runstate_guest = NULL;
>>>> + }
>>>> +}
>>>> +
>>>> +void arch_vcpu_cleanup_runstate(struct vcpu *v)
>>>> +{
>>>> + spin_lock(&v->arch.runstate_guest_lock);
>>>> +
>>>> + cleanup_runstate_vcpu_locked(v);
>>>> +
>>>> + spin_unlock(&v->arch.runstate_guest_lock);
>>>> +}
>>>> +
>>>> +static int setup_runstate_vcpu_locked(struct vcpu *v, vaddr_t vaddr)
>>>> +{
>>>> + unsigned int offset;
>>>> + mfn_t mfn[2];
>>>> + struct page_info *page;
>>>> + unsigned int numpages;
>>>> struct vcpu_runstate_info runstate;
>>>> + void *p;
>>>> - if ( guest_handle_is_null(runstate_guest(v)) )
>>>> - return;
>>>> + /* user can pass a NULL address to unregister a previous area */
>>>> + if ( vaddr == 0 )
>>>> + return 0;
>>>> +
>>>> + offset = vaddr & ~PAGE_MASK;
>>>> +
>>>> + /* provided address must be aligned to a 64bit */
>>>> + if ( offset % alignof(struct vcpu_runstate_info) )
>>>
>>> This new restriction wants to be explained in the commit message and public
>>> header.
>>
>> ok
>>
>>>
>>>> + {
>>>> + gprintk(XENLOG_WARNING, "Cannot map runstate pointer at
>>>> 0x%"PRIvaddr
>>>> + ": Invalid offset\n", vaddr);
>>>
>>> We usually enforce 80 character per lines except for format string. So it
>>> is easier to grep them in the code.
>>
>> Ok i will fix this one and the following ones.
>> But here PRIvaddr would break any attempt to grep something in fact.
>>
>>>
>>>> + return -EINVAL;
>>>> + }
>>>> +
>>>> + page = get_page_from_gva(v, vaddr, GV2M_WRITE);
>>>> + if ( !page )
>>>> + {
>>>> + gprintk(XENLOG_WARNING, "Cannot map runstate pointer at
>>>> 0x%"PRIvaddr
>>>> + ": Page is not mapped\n", vaddr);
>>>> + return -EINVAL;
>>>> + }
>>>> +
>>>> + mfn[0] = page_to_mfn(page);
>>>> + v->arch.runstate_guest_page[0] = page;
>>>> +
>>>> + if ( offset > (PAGE_SIZE - sizeof(struct vcpu_runstate_info)) )
>>>> + {
>>>> + /* guest area is crossing pages */
>>>> + page = get_page_from_gva(v, vaddr + PAGE_SIZE, GV2M_WRITE);
>>>> + if ( !page )
>>>> + {
>>>> + put_page(v->arch.runstate_guest_page[0]);
>>>> + gprintk(XENLOG_WARNING,
>>>> + "Cannot map runstate pointer at 0x%"PRIvaddr
>>>> + ": 2nd Page is not mapped\n", vaddr);
>>>> + return -EINVAL;
>>>> + }
>>>> + mfn[1] = page_to_mfn(page);
>>>> + v->arch.runstate_guest_page[1] = page;
>>>> + numpages = 2;
>>>> + }
>>>> + else
>>>> + {
>>>> + v->arch.runstate_guest_page[1] = NULL;
>>>> + numpages = 1;
>>>> + }
>>>> - memcpy(&runstate, &v->runstate, sizeof(runstate));
>>>> + p = vmap(mfn, numpages);
>>>> + if ( !p )
>>>> + {
>>>> + put_page(v->arch.runstate_guest_page[0]);
>>>> + if ( numpages == 2 )
>>>> + put_page(v->arch.runstate_guest_page[1]);
>>>> - if ( VM_ASSIST(v->domain, runstate_update_flag) )
>>>> + gprintk(XENLOG_WARNING, "Cannot map runstate pointer at
>>>> 0x%"PRIvaddr
>>>> + ": vmap error\n", vaddr);
>>>> + return -EINVAL;
>>>> + }
>>>> +
>>>> + v->arch.runstate_guest = p + offset;
>>>> +
>>>> + if (v == current)
>>>> + memcpy(v->arch.runstate_guest, &v->runstate, sizeof(v->runstate));
>>>> + else
>>>> {
>>>> - guest_handle = &v->runstate_guest.p->state_entry_time + 1;
>>>> - guest_handle--;
>>>> - runstate.state_entry_time |= XEN_RUNSTATE_UPDATE;
>>>> - __raw_copy_to_guest(guest_handle,
>>>> - (void *)(&runstate.state_entry_time + 1) - 1,
>>>> 1);
>>>> - smp_wmb();
>>>> + vcpu_runstate_get(v, &runstate);
>>>> + memcpy(v->arch.runstate_guest, &runstate, sizeof(v->runstate));
>>>> }
>>>> - __copy_to_guest(runstate_guest(v), &runstate, 1);
>>>> + return 0;
>>>> +}
>>>> +
>>>> +int arch_vcpu_setup_runstate(struct vcpu *v,
>>>> + struct vcpu_register_runstate_memory_area
>>>> area)
>>>> +{
>>>> + int rc;
>>>> +
>>>> + spin_lock(&v->arch.runstate_guest_lock);
>>>> +
>>>> + /* cleanup if we are recalled */
>>>> + cleanup_runstate_vcpu_locked(v);
>>>> +
>>>> + rc = setup_runstate_vcpu_locked(v, (vaddr_t)area.addr.v);
>>>> +
>>>> + spin_unlock(&v->arch.runstate_guest_lock);
>>>> - if ( guest_handle )
>>>> + return rc;
>>>> +}
>>>> +
>>>> +
>>>> +/* Update per-VCPU guest runstate shared memory area (if registered). */
>>>> +static void update_runstate_area(struct vcpu *v)
>>>> +{
>>>> + spin_lock(&v->arch.runstate_guest_lock);
>>>> +
>>>> + if ( v->arch.runstate_guest )
>>>> {
>>>> - runstate.state_entry_time &= ~XEN_RUNSTATE_UPDATE;
>>>> - smp_wmb();
>>>> - __raw_copy_to_guest(guest_handle,
>>>> - (void *)(&runstate.state_entry_time + 1) - 1,
>>>> 1);
>>>> + if ( VM_ASSIST(v->domain, runstate_update_flag) )
>>>> + {
>>>> + v->runstate.state_entry_time |= XEN_RUNSTATE_UPDATE;
>>>> + write_atomic(&(v->arch.runstate_guest->state_entry_time),
>>>> + v->runstate.state_entry_time);
>>>
>>> NIT: You want to indent v-> at the same level as the argument from the
>>> first line.
>>
>> Ok
>>
>>>
>>> Also, I think you are missing a smp_wmb() here.
>>
>> The atomic operation itself would not need a barrier.
>> I do not see why you think a barrier is needed here.
>> For the internal structure ?
>
> We need to make sure the other-end sees the XEN_RUNSTATE_UPDATE change
> before other changes. Otherwise, due to cpu reordering, the writes could
> be seen in reverse order. (Technically the reader would have to use a
> read-barrier but that's a separate topic.)
I will add a barrier before the atomic.
Cheers
Bertrand
