by Claudio Imbrenda

[permalink] [raw]

Subject: Re: [PATCH v2 01/13] KVM: s390: pv: avoid stall notifications for some UVCs

On Thu, 29 Jul 2021 11:58:39 +0200
Janosch Frank <[email protected]> wrote:

> On 7/28/21 4:26 PM, Claudio Imbrenda wrote:
> > Improve make_secure_pte to avoid stalls when the system is heavily
> > overcommitted. This was especially problematic in
> > kvm_s390_pv_unpack, because of the loop over all pages that needed
> > unpacking.
> >
> > Also fix kvm_s390_pv_init_vm to avoid stalls when the system is
> > heavily overcommitted.
>
> Fixes tag?

will be in the next version

> >
> > Signed-off-by: Claudio Imbrenda <[email protected]>
> > ---
> > arch/s390/kernel/uv.c | 11 ++++++++---
> > arch/s390/kvm/pv.c | 2 +-
> > 2 files changed, 9 insertions(+), 4 deletions(-)
> >
> > diff --git a/arch/s390/kernel/uv.c b/arch/s390/kernel/uv.c
> > index aeb0a15bcbb7..fd0faa51c1bb 100644
> > --- a/arch/s390/kernel/uv.c
> > +++ b/arch/s390/kernel/uv.c
> > @@ -196,11 +196,16 @@ static int make_secure_pte(pte_t *ptep,
> > unsigned long addr, if (!page_ref_freeze(page, expected))
> > return -EBUSY;
> > set_bit(PG_arch_1, &page->flags);
> > - rc = uv_call(0, (u64)uvcb);
> > + rc = __uv_call(0, (u64)uvcb);
>
> We should exchange rc with cc since that's what we get back from
> __uv_call(). Technically we always get a cc but for the other
> functions it's only ever 0/1 which translates to success/error so rc
> is ok.

will be in the next version

> > page_ref_unfreeze(page, expected);
> > - /* Return -ENXIO if the page was not mapped, -EINVAL
> > otherwise */
> > - if (rc)
> > + /*
> > + * Return -ENXIO if the page was not mapped, -EINVAL for
> > other errors.
> > + * If busy or partially completed, return -EAGAIN.
> > + */
> > + if (rc == 1)
> > rc = uvcb->rc == 0x10a ? -ENXIO : -EINVAL;
> > + else if (rc > 1)
> > + rc = -EAGAIN;
> > return rc;
>
> Could you define the CCs in uv.h and check against the constants here
> so it's easier to understand that the rc > 1 checks against a "UV was
> busy please re-issue the call again" cc?
>
> Maybe also make it explicit for cc 2 and 3 instead of cc > 1

will be in the next version

> > }
> >
> > diff --git a/arch/s390/kvm/pv.c b/arch/s390/kvm/pv.c
> > index c8841f476e91..e007df11a2fe 100644
> > --- a/arch/s390/kvm/pv.c
> > +++ b/arch/s390/kvm/pv.c
> > @@ -196,7 +196,7 @@ int kvm_s390_pv_init_vm(struct kvm *kvm, u16
> > *rc, u16 *rrc) uvcb.conf_base_stor_origin =
> > (u64)kvm->arch.pv.stor_base; uvcb.conf_virt_stor_origin =
> > (u64)kvm->arch.pv.stor_var;
> > - cc = uv_call(0, (u64)&uvcb);
> > + cc = uv_call_sched(0, (u64)&uvcb);
> > *rc = uvcb.header.rc;
> > *rrc = uvcb.header.rrc;
> > KVM_UV_EVENT(kvm, 3, "PROTVIRT CREATE VM: handle %llx len
> > %llx rc %x rrc %x",
>

2021-07-29 13:47:15

by Janosch Frank

[permalink] [raw]

Subject: Re: [PATCH v2 02/13] KVM: s390: pv: leak the ASCE page when destroy fails

On 7/29/21 2:54 PM, Claudio Imbrenda wrote:
> On Thu, 29 Jul 2021 12:41:05 +0200
> Janosch Frank <[email protected]> wrote:
>
>> On 7/28/21 4:26 PM, Claudio Imbrenda wrote:
>>> When a protected VM is created, the topmost level of page tables of
>>> its ASCE is marked by the Ultravisor; any attempt to use that
>>> memory for protected virtualization will result in failure.
>>>
>>> Only a successful Destroy Configuration UVC will remove the marking.
>>>
>>> When the Destroy Configuration UVC fails, the topmost level of page
>>> tables of the VM does not get its marking cleared; to avoid issues
>>> it must not be used again.
>>>
>>> Since the page becomes in practice unusable, we set it aside and
>>> leak it.
>>>
>>> Signed-off-by: Claudio Imbrenda <[email protected]>
>>> ---
>>> arch/s390/kvm/pv.c | 53
>>> +++++++++++++++++++++++++++++++++++++++++++++- 1 file changed, 52
>>> insertions(+), 1 deletion(-)
>>>
>>> diff --git a/arch/s390/kvm/pv.c b/arch/s390/kvm/pv.c
>>> index e007df11a2fe..1ecdc1769ed9 100644
>>> --- a/arch/s390/kvm/pv.c
>>> +++ b/arch/s390/kvm/pv.c
>>> @@ -155,6 +155,55 @@ static int kvm_s390_pv_alloc_vm(struct kvm
>>> *kvm) return -ENOMEM;
>>> }
>>>
>>> +/*
>>> + * Remove the topmost level of page tables from the list of page
>>> tables of
>>> + * the gmap.
>>> + * This means that it will not be freed when the VM is torn down,
>>> and needs
>>> + * to be handled separately by the caller, unless an intentional
>>> leak is
>>> + * intended.
>>> + */
>>> +static void kvm_s390_pv_remove_old_asce(struct kvm *kvm)
>>> +{
>>> + struct page *old;
>>> +
>>> + old = virt_to_page(kvm->arch.gmap->table);
>>> + list_del(&old->lru);
>>> + /* in case the ASCE needs to be "removed" multiple times */
>>> + INIT_LIST_HEAD(&old->lru);
>>> +}
>>> +
>>> +/*
>>> + * Try to replace the current ASCE with another equivalent one.
>>> + * If the allocation of the new top level page table fails, the
>>> ASCE is not
>>> + * replaced.
>>> + * In any case, the old ASCE is removed from the list, therefore
>>> the caller
>>> + * has to make sure to save a pointer to it beforehands, unless an
>>> + * intentional leak is intended.
>>> + */
>>> +static int kvm_s390_pv_replace_asce(struct kvm *kvm)
>>> +{
>>> + unsigned long asce;
>>> + struct page *page;
>>> + void *table;
>>> +
>>> + kvm_s390_pv_remove_old_asce(kvm);
>>> +
>>> + page = alloc_pages(GFP_KERNEL_ACCOUNT, CRST_ALLOC_ORDER);
>>> + if (!page)
>>> + return -ENOMEM;
>>> + list_add(&page->lru, &kvm->arch.gmap->crst_list);
>>> +
>>> + table = page_to_virt(page);
>>> + memcpy(table, kvm->arch.gmap->table, 1UL <<
>>> (CRST_ALLOC_ORDER + PAGE_SHIFT));
>>
>> Don't we want to memcpy first and then add it to the list?
>> The gmap is still active per-se so I think we want to take the
>> guest_table_lock for the list_add here.
>
> doesn't really make a difference, it is not actually used until a few
> lines later

Still when you touch the patch, then just move that around please :)

>
> also, the list is only ever touched here, during guest creation and
> destruction; IIRC in all those cases we hold kvm->lock

The crst_list is also used in gmap_alloc_table() coming from
__gmap_link() when touching the top level entries. To be fair there
shouldn't be asynchronous things that trigger a fault/link but we have
valid references to the gmap so at the very least I want a comment why
this is totally fine like we have for gmap_free().

Hrm, I'll need to brood over this a bit more, it has been too long.

>
>>> +
>>> + asce = (kvm->arch.gmap->asce & ~PAGE_MASK) | __pa(table);
>>> + WRITE_ONCE(kvm->arch.gmap->asce, asce);
>>> + WRITE_ONCE(kvm->mm->context.gmap_asce, asce);
>>> + WRITE_ONCE(kvm->arch.gmap->table, table);
>>
>> If I remember correctly those won't need locks but I'm not 100% sure
>> so please have a look at that.
>
> it should not need locks, the VM is in use, so it can't disappear under
> our feet.
>
>>> +
>>> + return 0;
>>> +}
>>
>> That should both be in gmap.c
>
> why?

Because we do gmap stuff and I don't want to pollute pv.c with memory
things if we don't have to.

Btw. s390_reset_acc() is also in gmap.c

>
>>> +
>>> /* this should not fail, but if it does, we must not free the
>>> donated memory */ int kvm_s390_pv_deinit_vm(struct kvm *kvm, u16
>>> *rc, u16 *rrc) {
>>> @@ -169,9 +218,11 @@ int kvm_s390_pv_deinit_vm(struct kvm *kvm, u16
>>> *rc, u16 *rrc) atomic_set(&kvm->mm->context.is_protected, 0);
>>> KVM_UV_EVENT(kvm, 3, "PROTVIRT DESTROY VM: rc %x rrc %x",
>>> *rc, *rrc); WARN_ONCE(cc, "protvirt destroy vm failed rc %x rrc
>>> %x", *rc, *rrc);
>>> - /* Inteded memory leak on "impossible" error */
>>> + /* Intended memory leak on "impossible" error */
>>> if (!cc)
>>> kvm_s390_pv_dealloc_vm(kvm);
>>> + else
>>> + kvm_s390_pv_replace_asce(kvm);
>>> return cc ? -EIO : 0;
>>> }
>>>
>>>
>>
>