Date:	Mon, 22 Jun 2009 14:03:59 -0400
From:	Gregory Haskins <ghaskins@...ell.com>
To:	"Michael S. Tsirkin" <mst@...hat.com>
CC:	Gregory Haskins <gregory.haskins@...il.com>, kvm@...r.kernel.org,
	linux-kernel@...r.kernel.org, avi@...hat.com,
	paulmck@...ux.vnet.ibm.com, davidel@...ilserver.org, mingo@...e.hu,
	rusty@...tcorp.com.au
Subject: Re: [KVM PATCH v3 3/3] KVM: Fix races in irqfd using new	eventfd_kref_get
 interface

Michael S. Tsirkin wrote:
> On Mon, Jun 22, 2009 at 01:31:29PM -0400, Gregory Haskins wrote:
>   
>> Michael S. Tsirkin wrote:
>>     
>>> On Mon, Jun 22, 2009 at 12:05:57PM -0400, Gregory Haskins wrote:
>>>   
>>>       
>>>> This patch fixes all known races in irqfd, and paves the way to restore
>>>> DEASSIGN support.  For details of the eventfd races, please see the patch
>>>> presumably commited immediately prior to this one.
>>>>
>>>> In a nutshell, we use eventfd_kref_get/put() to properly manage the
>>>> lifetime of the underlying eventfd.  We also use careful coordination
>>>> with our workqueue to ensure that all irqfd objects have terminated
>>>> before we allow kvm to shutdown.  The logic used for shutdown walks
>>>> all open irqfds and releases them.  This logic can be generalized in
>>>> the future to allow a subset of irqfds to be released, thus allowing
>>>> DEASSIGN support.
>>>>
>>>> Signed-off-by: Gregory Haskins <ghaskins@...ell.com>
>>>>     
>>>>         
>>> I think this patch is a shade too tricky. Some explanation why below.
>>>
>>> But I think irqfd_pop is a good idea.
>>>   
>>>       
>> Yeah, next we can add something like "irqfd_remove(gsi)" in a similar
>> way to do DEASSIGN.
>>
>>     
>>> Here's an alternative design sketch: add a list of irqfds to be shutdown
>>> in kvm, and create a single-threaded workqueue. To kill an irqfd, move
>>> it from list of live irqfds to list of dead irqfds, then schedule work
>>> on a workqueue that walks this list and kills irqfds.
>>>   
>>>       
>> Yeah, I actually thought of that too, and I think that will work.  But
>> then I realized flush_schedule_work does the same thing and its much
>> less code.  Perhaps it is also much less clear, too ;)  At the very
>> least, you have made me realize I need to comment better.
>>     
>
> Not really, it's impossible to document all races one have thought
> about and avoided.
>   

Heh, that is a very astute observation.

>   
>>>   
>>>       
>>>> ---
>>>>
>>>>  virt/kvm/eventfd.c |  144 ++++++++++++++++++++++++++++++++++++++++------------
>>>>  1 files changed, 110 insertions(+), 34 deletions(-)
>>>>
>>>> diff --git a/virt/kvm/eventfd.c b/virt/kvm/eventfd.c
>>>> index 9656027..67985cd 100644
>>>> --- a/virt/kvm/eventfd.c
>>>> +++ b/virt/kvm/eventfd.c
>>>> @@ -28,6 +28,7 @@
>>>>  #include <linux/file.h>
>>>>  #include <linux/list.h>
>>>>  #include <linux/eventfd.h>
>>>> +#include <linux/kref.h>
>>>>  
>>>>  /*
>>>>   * --------------------------------------------------------------------
>>>> @@ -36,26 +37,68 @@
>>>>   * Credit goes to Avi Kivity for the original idea.
>>>>   * --------------------------------------------------------------------
>>>>   */
>>>> +
>>>> +enum {
>>>> +	irqfd_flags_shutdown,
>>>> +};
>>>> +
>>>>  struct _irqfd {
>>>>  	struct kvm               *kvm;
>>>> +	struct kref              *eventfd;
>>>>     
>>>>         
>>> Yay, kref.
>>>
>>>   
>>>       
>>>>  	int                       gsi;
>>>>  	struct list_head          list;
>>>>  	poll_table                pt;
>>>>  	wait_queue_head_t        *wqh;
>>>>  	wait_queue_t              wait;
>>>> -	struct work_struct        inject;
>>>> +	struct work_struct        work;
>>>> +	unsigned long             flags;
>>>>     
>>>>         
>>> Just make it "int shutdown"?
>>>   
>>>       
>> Yep, that is probably fine but we will have to use an explicit wmb in
>> lieu of a set_bit operation.  NBD.
>>
>>     
>>>   
>>>       
>>>>  };
>>>>  
>>>>  static void
>>>> -irqfd_inject(struct work_struct *work)
>>>> +irqfd_release(struct _irqfd *irqfd)
>>>> +{
>>>> +	eventfd_kref_put(irqfd->eventfd);
>>>> +	kfree(irqfd);
>>>> +}
>>>> +
>>>> +static void
>>>> +irqfd_work(struct work_struct *work)
>>>>  {
>>>> -	struct _irqfd *irqfd = container_of(work, struct _irqfd, inject);
>>>> +	struct _irqfd *irqfd = container_of(work, struct _irqfd, work);
>>>>  	struct kvm *kvm = irqfd->kvm;
>>>>  
>>>> -	mutex_lock(&kvm->irq_lock);
>>>> -	kvm_set_irq(kvm, KVM_USERSPACE_IRQ_SOURCE_ID, irqfd->gsi, 1);
>>>> -	kvm_set_irq(kvm, KVM_USERSPACE_IRQ_SOURCE_ID, irqfd->gsi, 0);
>>>> -	mutex_unlock(&kvm->irq_lock);
>>>> +	if (!test_bit(irqfd_flags_shutdown, &irqfd->flags)) {
>>>>     
>>>>         
>>> Why is it safe to test this bit outside of any lock?
>>>   
>>>       
>> Because the ordering is guaranteed to set_bit(), schedule_work().  All
>> we need to do is make sure that the work-queue runs at least one more
>> time after the flag has been set.  (Of course, I could have screwed up
>> too, but that was my rationale).
>>
>>     
>>>   
>>>       
>>>> +		/* Inject an interrupt */
>>>> +		mutex_lock(&kvm->irq_lock);
>>>> +		kvm_set_irq(kvm, KVM_USERSPACE_IRQ_SOURCE_ID, irqfd->gsi, 1);
>>>> +		kvm_set_irq(kvm, KVM_USERSPACE_IRQ_SOURCE_ID, irqfd->gsi, 0);
>>>> +		mutex_unlock(&kvm->irq_lock);
>>>> +	} else {
>>>>     
>>>>         
>>> Not much shared code here - create a separate showdown work struct?
>>> They are cheap ...
>>>   
>>>       
>> We can't because we need to ensure that all inject-jobs complete before
>> release-jobs.  Reading the work-queue code, it would be a deadlock for
>> the release-job to do a flush_work(inject-job).  Therefore, both
>> workloads are encapsulated into a single job, and we ensure that the job
>> is launched at least one more time after the flag has been set.
>>     
>
> AFAIK schedule_work does not give you in-order guarantees - it's
> multithreaded. you will have to create a single-threaded workqueue
> if you want in order execution.
>   

Right, that was my understanding as well.  Thats why I do both tasks
from a single work-item ;)

>   
>> Of course, now that I wrote that,  I realize it was clear-as-mud in the
>> code and needs some commenting ;)
>>
>>     
>>>   
>>>       
>>>> +		/* shutdown the irqfd */
>>>> +		struct _irqfd *_irqfd = NULL;
>>>> +
>>>> +		mutex_lock(&kvm->lock);
>>>> +
>>>> +		if (!list_empty(&irqfd->list))
>>>> +			_irqfd = irqfd;
>>>> +
>>>> +		if (_irqfd)
>>>> +			list_del(&_irqfd->list);
>>>> +
>>>> +		mutex_unlock(&kvm->lock);
>>>> +
>>>> +		/*
>>>> +		 * If the item is not currently on the irqfds list, we know
>>>> +		 * we are running concurrently with the KVM side trying to
>>>> +		 * remove this item as well.
>>>>     
>>>>         
>>> We do? How? As far as I can see list is only empty after it has been
>>> created.  Generally, it would be better to either use a flag or use
>>> list_empty as an indication of going down, but not both.
>>>   
>>>       
>> I think you are mis-reading that.  list_empty(&irqfd->list) is the
>> individual irqfd list-item, not the kvm->irqfds list itself.  This
>> conditional is telling us whether the irqfd in question is on or off the
>> list (its effectively an irqfd-specific flag), not whether the global
>> list is empty.  Again, poor commenting on my part.
>>     
>
> Yes, but you do INIT_LIST_HEAD in a single place. Once you add
> irqfd->list to a list, it won't be empty until you init it again.
>   

Good point.  I need list_del_init() and then it would work, right?

>   
>>>   
>>>       
>>>>  Since the KVM side should be
>>>> +		 * holding the reference now, and will block behind a
>>>> +		 * flush_work(), lets just let them do the release() for us
>>>> +		 */
>>>> +		if (!_irqfd)
>>>> +			return;
>>>> +
>>>> +		irqfd_release(_irqfd);
>>>> +	}
>>>>  }
>>>>  
>>>>  static int
>>>> @@ -65,25 +108,20 @@ irqfd_wakeup(wait_queue_t *wait, unsigned mode, int sync, void *key)
>>>>  	unsigned long flags = (unsigned long)key;
>>>>  
>>>>  	/*
>>>> -	 * Assume we will be called with interrupts disabled
>>>> +	 * called with interrupts disabled
>>>>  	 */
>>>> -	if (flags & POLLIN)
>>>> -		/*
>>>> -		 * Defer the IRQ injection until later since we need to
>>>> -		 * acquire the kvm->lock to do so.
>>>> -		 */
>>>> -		schedule_work(&irqfd->inject);
>>>> -
>>>>  	if (flags & POLLHUP) {
>>>>  		/*
>>>> -		 * for now, just remove ourselves from the list and let
>>>> -		 * the rest dangle.  We will fix this up later once
>>>> -		 * the races in eventfd are fixed
>>>> +		 * ordering is important: shutdown flag must be visible
>>>> +		 * before we schedule
>>>>  		 */
>>>>  		__remove_wait_queue(irqfd->wqh, &irqfd->wait);
>>>> -		irqfd->wqh = NULL;
>>>> +		set_bit(irqfd_flags_shutdown, &irqfd->flags);
>>>>     
>>>>         
>>> So what happens if a previously scheduled work runs on irqfd
>>> and sees this flag?
>>>       
>> My original thought was "thats ok", but now that you mention it I am not
>> so sure.  Ill give it some more thought because maybe you are on to
>> something.
>>
>>     
>>>  And note that multiple works can run on irqfd
>>> in parallel.
>>>   
>>>       
>> They can?  I thought work-queue items were guaranteed to only schedule
>> once?  If what you say is true, its broken, I agree, and Ill need to
>> revisit.  Let me get back to you.
>>     
>>>   
>>>       
>>>>  	}
>>>>  
>>>> +	if (flags & (POLLHUP | POLLIN))
>>>> +		schedule_work(&irqfd->work);
>>>> +
>>>>  	return 0;
>>>>  }
>>>>  
>>>> @@ -102,6 +140,7 @@ kvm_irqfd(struct kvm *kvm, int fd, int gsi, int flags)
>>>>  {
>>>>  	struct _irqfd *irqfd;
>>>>  	struct file *file = NULL;
>>>> +	struct kref *kref = NULL;
>>>>  	int ret;
>>>>  	unsigned int events;
>>>>  
>>>> @@ -112,7 +151,7 @@ kvm_irqfd(struct kvm *kvm, int fd, int gsi, int flags)
>>>>  	irqfd->kvm = kvm;
>>>>  	irqfd->gsi = gsi;
>>>>  	INIT_LIST_HEAD(&irqfd->list);
>>>> -	INIT_WORK(&irqfd->inject, irqfd_inject);
>>>> +	INIT_WORK(&irqfd->work, irqfd_work);
>>>>  
>>>>  	file = eventfd_fget(fd);
>>>>  	if (IS_ERR(file)) {
>>>> @@ -133,11 +172,13 @@ kvm_irqfd(struct kvm *kvm, int fd, int gsi, int flags)
>>>>  	list_add_tail(&irqfd->list, &kvm->irqfds);
>>>>  	mutex_unlock(&kvm->lock);
>>>>  
>>>> -	/*
>>>> -	 * Check if there was an event already queued
>>>> -	 */
>>>> -	if (events & POLLIN)
>>>> -		schedule_work(&irqfd->inject);
>>>> +	kref = eventfd_kref_get(file);
>>>> +	if (IS_ERR(file)) {
>>>> +		ret = PTR_ERR(file);
>>>> +		goto fail;
>>>> +	}
>>>> +
>>>> +	irqfd->eventfd = kref;
>>>>  
>>>>  	/*
>>>>  	 * do not drop the file until the irqfd is fully initialized, otherwise
>>>> @@ -145,9 +186,18 @@ kvm_irqfd(struct kvm *kvm, int fd, int gsi, int flags)
>>>>  	 */
>>>>  	fput(file);
>>>>  
>>>> +	/*
>>>> +	 * Check if there was an event already queued
>>>> +	 */
>>>>     
>>>>         
>>> This comment seems to confuse more that it clarifies:
>>> queued where? eventfd only counts... Just kill the comment?
>>>
>>>   
>>>       
>> non-zero values in eventfd are "queued" as a signal.  This test just
>> checks if an interrupt was already injected before we registered.
>>     
>
> After have understood the code I see what you mean, but the comment
> wasn't helpful and is better left out.
>   

Ok.  What if I say "Check if an interrupt is already pending before we
registered the callback" ;)

>   
>>>> +	if (events & POLLIN)
>>>> +		schedule_work(&irqfd->work);
>>>> +
>>>>  	return 0;
>>>>  
>>>>  fail:
>>>> +	if (kref && !IS_ERR(kref))
>>>> +		eventfd_kref_put(kref);
>>>> +
>>>>  	if (file && !IS_ERR(file))
>>>>  		fput(file);
>>>>     
>>>>         
>>> let's add a couple more labels and avoid the kref/file check
>>> and the initialization above?
>>>   
>>>       
>> I think that just makes it more confusing, personally.  But I will give
>> it some thought.
>>
>>     
>>>   
>>>       
>>>>  
>>>> @@ -161,21 +211,47 @@ kvm_irqfd_init(struct kvm *kvm)
>>>>  	INIT_LIST_HEAD(&kvm->irqfds);
>>>>  }
>>>>  
>>>> +static struct _irqfd *
>>>> +irqfd_pop(struct kvm *kvm)
>>>> +{
>>>> +	struct _irqfd *irqfd = NULL;
>>>> +
>>>> +	mutex_lock(&kvm->lock);
>>>> +
>>>> +	if (!list_empty(&kvm->irqfds)) {
>>>> +		irqfd = list_first_entry(&kvm->irqfds, struct _irqfd, list);
>>>> +		list_del(&irqfd->list);
>>>> +	}
>>>> +
>>>> +	mutex_unlock(&kvm->lock);
>>>> +
>>>> +	return irqfd;
>>>> +}
>>>> +
>>>>  void
>>>>  kvm_irqfd_release(struct kvm *kvm)
>>>>  {
>>>> -	struct _irqfd *irqfd, *tmp;
>>>> +	struct _irqfd *irqfd;
>>>>  
>>>> -	list_for_each_entry_safe(irqfd, tmp, &kvm->irqfds, list) {
>>>> -		if (irqfd->wqh)
>>>> -			remove_wait_queue(irqfd->wqh, &irqfd->wait);
>>>> +	while ((irqfd = irqfd_pop(kvm))) {
>>>>  
>>>> -		flush_work(&irqfd->inject);
>>>> +		remove_wait_queue(irqfd->wqh, &irqfd->wait);
>>>>  
>>>> -		mutex_lock(&kvm->lock);
>>>> -		list_del(&irqfd->list);
>>>> -		mutex_unlock(&kvm->lock);
>>>> +		/*
>>>> +		 * We guarantee there will be no more notifications after
>>>> +		 * the remove_wait_queue returns.  Now lets make sure we
>>>> +		 * synchronize behind any outstanding work items before
>>>> +		 * releasing the resources
>>>> +		 */
>>>> +		flush_work(&irqfd->work);
>>>>  
>>>> -		kfree(irqfd);
>>>> +		irqfd_release(irqfd);
>>>>  	}
>>>> +
>>>> +	/*
>>>> +	 * We need to wait in case there are any outstanding work-items
>>>> +	 * in flight that had already removed themselves from the list
>>>> +	 * prior to entry to this function
>>>> +	 */
>>>>     
>>>>         
>>> Looks scary. Why doesn't the flush above cover all cases?
>>>   
>>>       
>> The path inside the while() is for when KVM wins the race and finds the
>> item in the list.  It atomically removes it, and is responsible for
>> freeing it in a coordinated way.  In this case, we must block with the
>> flush_work() before we can irqfd_release() so that we do not yank the
>> memory out from under a running work-item.
>>
>> The flush_scheduled_work() is for when eventfd wins the race and has
>> already removed itself from the list in the "shutdown" path in the
>> work-item.  We want to make sure that kvm_irqfd_release() cannot return
>> until all work-items have exited to prevent something like the kvm.ko
>> module unloading while the work-item is still in flight.
>> Thanks Michael,
>> -Greg
>>     
>>>   
>>>       
>>>> +	flush_scheduled_work();
>>>>  }
>>>>     
>>>>         
>>> --
>>> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
>>> the body of a message to majordomo@...r.kernel.org
>>> More majordomo info at  http://vger.kernel.org/majordomo-info.html
>>> Please read the FAQ at  http://www.tux.org/lkml/
>>>   
>>>       
>>     
>
>
>   



Download attachment "signature.asc" of type "application/pgp-signature" (267 bytes)

Powered by blists - more mailing lists