lkml.org 
[lkml]   [2012]   [Dec]   [20]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
    Patch in this message
    /
    SubjectRe: [PATCH 0/7] KVM: Alleviate mmu_lock hold time when we start dirty logging
    From
    Date
    On Thu, 2012-12-20 at 23:35 +0900, Takuya Yoshikawa wrote:
    > On Thu, 20 Dec 2012 06:41:27 -0700
    > Alex Williamson <alex.williamson@redhat.com> wrote:
    >
    > > Hmm, isn't the fix as simple as:
    > >
    > > --- a/virt/kvm/kvm_main.c
    > > +++ b/virt/kvm/kvm_main.c
    > > @@ -847,7 +847,8 @@ int __kvm_set_memory_region(struct kvm *kvm,
    > > GFP_KERNEL);
    > > if (!slots)
    > > goto out_free;
    > > - }
    > > + } else
    > > + slots->generation = kvm->memslots->generation;
    > >
    > > /* map new memory slot into the iommu */
    > > if (npages) {
    > >
    > > Or even just slots->generation++ since we're holding the lock across all
    > > of this.
    >
    > Yes, the fix should work, but I do not want to update the
    > generation from outside of update_memslots().

    Ok, then:

    diff --git a/include/linux/kvm_host.h b/include/linux/kvm_host.h
    index 87089dd..c7b5061 100644
    --- a/include/linux/kvm_host.h
    +++ b/include/linux/kvm_host.h
    @@ -413,7 +413,8 @@ void kvm_exit(void);

    void kvm_get_kvm(struct kvm *kvm);
    void kvm_put_kvm(struct kvm *kvm);
    -void update_memslots(struct kvm_memslots *slots, struct kvm_memory_slot *new);
    +void update_memslots(struct kvm_memslots *slots, struct kvm_memory_slot *new,
    + u64 last_generation);

    static inline struct kvm_memslots *kvm_memslots(struct kvm *kvm)
    {
    diff --git a/virt/kvm/kvm_main.c b/virt/kvm/kvm_main.c
    index c4c8ec1..06961ea 100644
    --- a/virt/kvm/kvm_main.c
    +++ b/virt/kvm/kvm_main.c
    @@ -667,7 +667,8 @@ static void sort_memslots(struct kvm_memslots *slots)
    slots->id_to_index[slots->memslots[i].id] = i;
    }

    -void update_memslots(struct kvm_memslots *slots, struct kvm_memory_slot *new)
    +void update_memslots(struct kvm_memslots *slots, struct kvm_memory_slot *new,
    + u64 last_generation)
    {
    if (new) {
    int id = new->id;
    @@ -679,7 +680,7 @@ void update_memslots(struct kvm_memslots *slots, struct kvm_memory_slot *new)
    sort_memslots(slots);
    }

    - slots->generation++;
    + slots->generation = last_generation + 1;
    }

    static int check_memory_region_flags(struct kvm_userspace_memory_region *mem)
    @@ -814,7 +815,7 @@ int __kvm_set_memory_region(struct kvm *kvm,
    slot = id_to_memslot(slots, mem->slot);
    slot->flags |= KVM_MEMSLOT_INVALID;

    - update_memslots(slots, NULL);
    + update_memslots(slots, NULL, kvm->memslots->generation);

    old_memslots = kvm->memslots;
    rcu_assign_pointer(kvm->memslots, slots);
    @@ -862,7 +863,7 @@ int __kvm_set_memory_region(struct kvm *kvm,
    memset(&new.arch, 0, sizeof(new.arch));
    }

    - update_memslots(slots, &new);
    + update_memslots(slots, &new, kvm->memslots->generation);
    old_memslots = kvm->memslots;
    rcu_assign_pointer(kvm->memslots, slots);
    synchronize_srcu_expedited(&kvm->srcu);
    > > The original patch can be reverted, there are no following dependencies,
    > > but the idea was that we're making the memslot array larger, so there
    > > could be more pressure in allocating it, so let's not trivially do extra
    > > frees and allocs. Thanks,
    >
    > I agree that the current set_memory_region() is not good for frequent updates.
    > But the alloc/free is not a major factor at the moment: flushing shadows should
    > be more problematic.

    I don't understand why we'd throw away even a minor optimization that's
    so easy to fix. We're not only removing a free/alloc, but we're being
    more friendly to the cache by avoiding the extra memcpy. The above also
    makes the generation management a bit more explicit. Thanks,

    Alex



    \
     
     \ /
      Last update: 2012-12-20 16:41    [W:2.435 / U:0.816 seconds]
    ©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site