kvm mmu notifier update
I updated the mmu notifier patch. There was a problem in shutting down the vm in mmu_notifier_release instead of waiting the last filp->release, because all vcpus will be freed and the filp->private_data will point to already freed memory when the vcpu fd is closed.
It seems we may really not need to do anything in the mmu notifier release method, because for us the shadow pagetables are meaningless if no guest can run, and the ->release method is only invoked when all tasks with current->mm == kvm->mm already quit. After that the guest can't possibly run anymore, the ioctl becomes useless too. So I changed the code to only invalidate the root of the spte radix tree in every vcpu, just for debugging. If any guest attempts to run after mmu notifier release runs, we'll then notice. No spte can be established after ->release returns.
Probably ->release shouldn't be mandatory to implement, but from a different point of view it may also payoff to make all methods mandatory to be implemented as a microoptimization to avoid the null pointer check before invoking the notifier (and in the future to also fail registration if the api is extended and a module isn't updated, to decrease the risk of runtime failure).
arch/x86/kvm/Kconfig +1 -0
@@ -21,6 +21,7 @@ config KVM tristate "Kernel-based Virtual Machine (KVM) support" depends on HAVE_KVM select PREEMPT_NOTIFIERS + select MMU_NOTIFIER select ANON_INODES ---help--- Support hosting fully virtualized guest machines using hardware
arch/x86/kvm/mmu.c +165 -20
313 lines not on this page · raw message
arch/x86/kvm/paging_tmpl.h +19 -0
@@ -263,6 +263,12 @@ static void FNAME(update_pte)(struct kvm_vcpu *vcpu, struct kvm_mmu_page *page, pfn = vcpu->arch.update_pte.pfn;if (is_error_pfn (pfn))return ; +if (unlikely (atomic_read (&vcpu->kvm->arch.mmu_notifier_count))) +return ; +smp_rmb (); +if (unlikely (atomic_read (&vcpu->kvm->arch.mmu_notifier_seq) != + vcpu->arch.update_pte.mmu_seq)) +return ;kvm_get_pfn (pfn); mmu_set_spte(vcpu, spte, page->role.access, pte_access, 0, 0, gpte & PT_DIRTY_MASK, NULL, largepage, gpte_to_gfn(gpte),
@@ -380,6 +386,7 @@ static int FNAME(page_fault)(struct kvm_vcpu *vcpu, gva_t addr,int r;pfn_t pfn;int largepage =0 ; +int mmu_seq;pgprintk ("%s: addr %lx err %x\n" , __func__, addr, error_code);kvm_mmu_audit (vcpu,"pre page fault" );
@@ -413,6 +420,8 @@ static int FNAME(page_fault)(struct kvm_vcpu *vcpu, gva_t addr, largepage =1 ; } } + mmu_seq =atomic_read (&vcpu->kvm->arch.mmu_notifier_seq); +/* implicit mb(), we'll read before PT lock is unlocked */ pfn =gfn_to_pfn (vcpu->kvm, walker.gfn);up_read (¤t->mm->mmap_sem);
@@ -424,6 +433,11 @@ static int FNAME(page_fault)(struct kvm_vcpu *vcpu, gva_t addr, }spin_lock (&vcpu->kvm->mmu_lock); +if (unlikely (atomic_read (&vcpu->kvm->arch.mmu_notifier_count))) + goto out_unlock; +smp_rmb (); +if (unlikely (atomic_read (&vcpu->kvm->arch.mmu_notifier_seq) != mmu_seq)) + goto out_unlock;kvm_mmu_free_some_pages (vcpu); shadow_pte =FNAME (fetch)(vcpu, addr, &walker, user_fault, write_fault, largepage, &write_pt, pfn);
@@ -439,6 +453,11 @@ static int FNAME(page_fault)(struct kvm_vcpu *vcpu, gva_t addr,spin_unlock (&vcpu->kvm->mmu_lock);return write_pt; + +out_unlock: +spin_unlock (&vcpu->kvm->mmu_lock); +kvm_release_pfn_clean (pfn); +return 0 ; }static gpa_t FNAME (gva_to_gpa )(struct kvm_vcpu *vcpu,gva_t vaddr)
arch/x86/kvm/x86.c +172 -8
216 lines not on this page · raw message
include/asm-x86/kvm_host.h +8 -0
@@ -13,6 +13,7 @@#include <linux/types.h> #include <linux/mm.h> +#include <linux/mmu_notifier.h> #include <linux/kvm.h> #include <linux/kvm_para.h>
@@ -247,6 +248,7 @@ struct kvm_vcpu_arch {gfn_t gfn;/* presumed gfn during guest pte update */ pfn_t pfn;/* pfn corresponding to that gfn */ int largepage; +int mmu_seq; } update_pte;struct i387_fxsave_struct host_fx_image;
@@ -317,6 +319,10 @@ struct kvm_arch{struct page *ept_identity_pagetable;bool ept_identity_pagetable_done; + +struct mmu_notifier mmu_notifier; +atomic_t mmu_notifier_seq; +atomic_t mmu_notifier_count; };struct kvm_vm_stat {
@@ -441,6 +447,8 @@ void kvm_mmu_set_base_ptes(u64 base_pte);void kvm_mmu_set_mask_ptes (u64 user_mask,u64 accessed_mask,u64 dirty_mask,u64 nx_mask,u64 x_mask); +int kvm_unmap_hva (struct kvm *kvm,unsigned long hva); +int kvm_age_hva (struct kvm *kvm,unsigned long hva);int kvm_mmu_reset_context (struct kvm_vcpu *vcpu);void kvm_mmu_slot_remove_write_access (struct kvm *kvm,int slot);void kvm_mmu_zap_all (struct kvm *kvm);
-------------------------------------------------------------------------
This SF.net email is sponsored by the 2008 JavaOne(SM) Conference Don't miss this year's exciting event. There's still time to save $100. Use priority code J8TL2D2.
http://ad.doubleclick.net/clk;198757673;13503038;p?http://java.sun.com/javaone