Received-SPF: pass (google.com: domain of linux-kernel+bounces-37552-linux.lists.archive=gmail.com@vger.kernel.org designates 147.75.199.223 as permitted sender) client-ip=147.75.199.223;
Date: Wed, 24 Jan 2024 10:52:12 -0800
In-Reply-To: <Zavj4U2LYeOsnXOh@yilunxu-OptiPlex-7050>
Precedence: bulk
Mime-Version: 1.0
References: <20240110011533.503302-1-seanjc@google.com> <20240110011533.503302-4-seanjc@google.com>
 <Zavj4U2LYeOsnXOh@yilunxu-OptiPlex-7050>
Message-ID: <ZbFcXB5ctGOMEr22@google.com>
Subject: Re: [PATCH 3/4] KVM: Get reference to VM's address space in the async
 #PF worker
From: Sean Christopherson <seanjc@google.com>
To: Xu Yilun <yilun.xu@linux.intel.com>
Cc: Paolo Bonzini <pbonzini@redhat.com>, kvm@vger.kernel.org, linux-kernel@vger.kernel.org, 
	David Matlack <dmatlack@google.com>
Content-Type: text/plain; charset="us-ascii"

On Sat, Jan 20, 2024, Xu Yilun wrote:
> On Tue, Jan 09, 2024 at 05:15:32PM -0800, Sean Christopherson wrote:
> > Get a reference to the target VM's address space in async_pf_execute()
> > instead of gifting a reference from kvm_setup_async_pf().  Keeping the
> > address space alive just to service an async #PF is counter-productive,
> > i.e. if the process is exiting and all vCPUs are dead, then NOT doing
> > get_user_pages_remote() and freeing the address space asap is desirable.
> > 
> > Handling the mm reference entirely within async_pf_execute() also
> > simplifies the async #PF flows as a whole, e.g. it's not immediately
> > obvious when the worker task vs. the vCPU task is responsible for putting
> > the gifted mm reference.
> > 
> > Signed-off-by: Sean Christopherson <seanjc@google.com>
> > ---
> >  include/linux/kvm_host.h |  1 -
> >  virt/kvm/async_pf.c      | 32 ++++++++++++++++++--------------
> >  2 files changed, 18 insertions(+), 15 deletions(-)
> > 
> > diff --git a/include/linux/kvm_host.h b/include/linux/kvm_host.h
> > index 7e7fd25b09b3..bbfefd7e612f 100644
> > --- a/include/linux/kvm_host.h
> > +++ b/include/linux/kvm_host.h
> > @@ -238,7 +238,6 @@ struct kvm_async_pf {
> >  	struct list_head link;
> >  	struct list_head queue;
> >  	struct kvm_vcpu *vcpu;
> > -	struct mm_struct *mm;
> >  	gpa_t cr2_or_gpa;
> >  	unsigned long addr;
> >  	struct kvm_arch_async_pf arch;
> > diff --git a/virt/kvm/async_pf.c b/virt/kvm/async_pf.c
> > index d5dc50318aa6..c3f4f351a2ae 100644
> > --- a/virt/kvm/async_pf.c
> > +++ b/virt/kvm/async_pf.c
> > @@ -46,8 +46,8 @@ static void async_pf_execute(struct work_struct *work)
> >  {
> >  	struct kvm_async_pf *apf =
> >  		container_of(work, struct kvm_async_pf, work);
> > -	struct mm_struct *mm = apf->mm;
> >  	struct kvm_vcpu *vcpu = apf->vcpu;
> > +	struct mm_struct *mm = vcpu->kvm->mm;
> >  	unsigned long addr = apf->addr;
> >  	gpa_t cr2_or_gpa = apf->cr2_or_gpa;
> >  	int locked = 1;
> > @@ -56,16 +56,24 @@ static void async_pf_execute(struct work_struct *work)
> >  	might_sleep();
> >  
> >  	/*
> > -	 * This work is run asynchronously to the task which owns
> > -	 * mm and might be done in another context, so we must
> > -	 * access remotely.
> > +	 * Attempt to pin the VM's host address space, and simply skip gup() if
> > +	 * acquiring a pin fail, i.e. if the process is exiting.  Note, KVM
> > +	 * holds a reference to its associated mm_struct until the very end of
> > +	 * kvm_destroy_vm(), i.e. the struct itself won't be freed before this
> > +	 * work item is fully processed.
> >  	 */
> > -	mmap_read_lock(mm);
> > -	get_user_pages_remote(mm, addr, 1, FOLL_WRITE, NULL, &locked);
> > -	if (locked)
> > -		mmap_read_unlock(mm);
> > -	mmput(mm);
> > +	if (mmget_not_zero(mm)) {
> > +		mmap_read_lock(mm);
> > +		get_user_pages_remote(mm, addr, 1, FOLL_WRITE, NULL, &locked);
> > +		if (locked)
> > +			mmap_read_unlock(mm);
> > +		mmput(mm);
> > +	}
> >  
> > +	/*
> > +	 * Notify and kick the vCPU even if faulting in the page failed, e.g.
> 
> How about when the process is exiting? Could we just skip the following?

Maybe?  I'm not opposed to trimming this down even more, but I doubt it will make
much of a difference.  The vCPU can't be running so async_pf.lock shouldn't be
contended, no IPIs will be issued for kicks, etc.  So for this patch at least,
I want to take the most conservative approach while still cleaning up the mm_struct
usage.

> > +	 * so that the vCPU can retry the fault synchronously.
> > +	 */
> >  	if (IS_ENABLED(CONFIG_KVM_ASYNC_PF_SYNC))
> >  		kvm_arch_async_page_present(vcpu, apf);
> >  
> > @@ -129,10 +137,8 @@ void kvm_clear_async_pf_completion_queue(struct kvm_vcpu *vcpu)
> >  #ifdef CONFIG_KVM_ASYNC_PF_SYNC
> >  		flush_work(&work->work);
> >  #else
> > -		if (cancel_work_sync(&work->work)) {
> > -			mmput(work->mm);
> > +		if (cancel_work_sync(&work->work))
> >  			kmem_cache_free(async_pf_cache, work);
> > -		}
> >  #endif
> >  		spin_lock(&vcpu->async_pf.lock);
> >  	}
> > @@ -211,8 +217,6 @@ bool kvm_setup_async_pf(struct kvm_vcpu *vcpu, gpa_t cr2_or_gpa,
> >  	work->cr2_or_gpa = cr2_or_gpa;
> >  	work->addr = hva;
> >  	work->arch = *arch;
> > -	work->mm = current->mm;
> > -	mmget(work->mm);
> >  
> >  	INIT_WORK(&work->work, async_pf_execute);
> >  
> > -- 
> > 2.43.0.472.g3155946c3a-goog
> >