Date:	Sat, 23 May 2009 00:57:54 -0400
From:	Gregory Haskins <ghaskins@...ell.com>
To:	Marcelo Tosatti <mtosatti@...hat.com>
CC:	kvm@...r.kernel.org, linux-kernel@...r.kernel.org, avi@...hat.com
Subject: Re: [KVM PATCH v3 4/4] kvm: add iosignalfd support

Hi Marcelo,

Marcelo Tosatti wrote:
> On Thu, May 21, 2009 at 12:51:29PM -0400, Gregory Haskins wrote:
>   
>> iosignalfd is a mechanism to register PIO/MMIO regions to trigger an eventfd
>> signal when written to by a guest.  Host userspace can register any arbitrary
>> IO address with a corresponding eventfd and then pass the eventfd to a
>> specific end-point of interest for handling.
>>
>> Normal IO requires a blocking round-trip since the operation may cause
>> side-effects in the emulated model or may return data to the caller.
>> Therefore, an IO in KVM traps from the guest to the host, causes a VMX/SVM
>> "heavy-weight" exit back to userspace, and is ultimately serviced by qemu's
>> device model synchronously before returning control back to the vcpu.
>>
>> However, there is a subclass of IO which acts purely as a trigger for
>> other IO (such as to kick off an out-of-band DMA request, etc).  For these
>> patterns, the synchronous call is particularly expensive since we really
>> only want to simply get our notification transmitted asychronously and
>> return as quickly as possible.  All the sychronous infrastructure to ensure
>> proper data-dependencies are met in the normal IO case are just unecessary
>> overhead for signalling.  This adds additional computational load on the
>> system, as well as latency to the signalling path.
>>
>> Therefore, we provide a mechanism for registration of an in-kernel trigger
>> point that allows the VCPU to only require a very brief, lightweight
>> exit just long enough to signal an eventfd.  This also means that any
>> clients compatible with the eventfd interface (which includes userspace
>> and kernelspace equally well) can now register to be notified. The end
>> result should be a more flexible and higher performance notification API
>> for the backend KVM hypervisor and perhipheral components.
>>
>> To test this theory, we built a test-harness called "doorbell".  This
>> module has a function called "doorbell_ring()" which simply increments a
>> counter for each time the doorbell is signaled.  It supports signalling
>> from either an eventfd, or an ioctl().
>>
>> We then wired up two paths to the doorbell: One via QEMU via a registered
>> io region and through the doorbell ioctl().  The other is direct via iosignalfd.
>>
>> You can download this test harness here:
>>
>> ftp://ftp.novell.com/dev/ghaskins/doorbell.tar.bz2
>>
>> The measured results are as follows:
>>
>> qemu-mmio:       110000 iops, 9.09us rtt
>> iosignalfd-mmio: 200100 iops, 5.00us rtt
>> iosignalfd-pio:  367300 iops, 2.72us rtt
>>
>> I didn't measure qemu-pio, because I have to figure out how to register a
>> PIO region with qemu's device model, and I got lazy.  However, for now we
>> can extrapolate based on the data from the NULLIO runs of +2.56us for MMIO,
>> and -350ns for HC, we get:
>>
>> qemu-pio:      153139 iops, 6.53us rtt
>> iosignalfd-hc: 412585 iops, 2.37us rtt
>>
>> these are just for fun, for now, until I can gather more data.
>>
>> Here is a graph for your convenience:
>>
>> http://developer.novell.com/wiki/images/7/76/Iofd-chart.png
>>
>> The conclusion to draw is that we save about 4us by skipping the userspace
>> hop.
>>
>> --------------------
>>
>> Signed-off-by: Gregory Haskins <ghaskins@...ell.com>
>> ---
>>
>>  arch/x86/kvm/x86.c       |    1 
>>  include/linux/kvm.h      |   15 ++++
>>  include/linux/kvm_host.h |   10 ++-
>>  virt/kvm/eventfd.c       |  165 ++++++++++++++++++++++++++++++++++++++++++++++
>>  virt/kvm/kvm_main.c      |   11 +++
>>  5 files changed, 198 insertions(+), 4 deletions(-)
>>
>> diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
>> index 98c2434..cee63ff 100644
>> --- a/arch/x86/kvm/x86.c
>> +++ b/arch/x86/kvm/x86.c
>> @@ -1085,6 +1085,7 @@ int kvm_dev_ioctl_check_extension(long ext)
>>  	case KVM_CAP_IRQ_INJECT_STATUS:
>>  	case KVM_CAP_ASSIGN_DEV_IRQ:
>>  	case KVM_CAP_IRQFD:
>> +	case KVM_CAP_IOSIGNALFD:
>>  		r = 1;
>>  		break;
>>  	case KVM_CAP_COALESCED_MMIO:
>> diff --git a/include/linux/kvm.h b/include/linux/kvm.h
>> index 8f53f24..8162466 100644
>> --- a/include/linux/kvm.h
>> +++ b/include/linux/kvm.h
>> @@ -292,6 +292,19 @@ struct kvm_guest_debug {
>>  	struct kvm_guest_debug_arch arch;
>>  };
>>  
>> +#define KVM_IOSIGNALFD_FLAG_DEASSIGN  (1 << 0)
>> +#define KVM_IOSIGNALFD_FLAG_PIO       (1 << 1)
>> +#define KVM_IOSIGNALFD_FLAG_COOKIE    (1 << 2)
>> +
>> +struct kvm_iosignalfd {
>> +	__u64 cookie;
>> +	__u64 addr;
>> +	__u32 len;
>> +	__u32 fd;
>> +	__u32 flags;
>> +	__u8  pad[12];
>> +};
>> +
>>  #define KVM_TRC_SHIFT           16
>>  /*
>>   * kvm trace categories
>> @@ -419,6 +432,7 @@ struct kvm_trace_rec {
>>  #define KVM_CAP_MCE 31
>>  #endif
>>  #define KVM_CAP_IRQFD 32
>> +#define KVM_CAP_IOSIGNALFD 33
>>  
>>  #ifdef KVM_CAP_IRQ_ROUTING
>>  
>> @@ -525,6 +539,7 @@ struct kvm_irqfd {
>>  			_IOW(KVMIO, 0x74, struct kvm_assigned_msix_entry)
>>  #define KVM_DEASSIGN_DEV_IRQ       _IOW(KVMIO, 0x75, struct kvm_assigned_irq)
>>  #define KVM_IRQFD                  _IOW(KVMIO, 0x76, struct kvm_irqfd)
>> +#define KVM_IOSIGNALFD             _IOW(KVMIO, 0x77, struct kvm_iosignalfd)
>>  
>>  /*
>>   * ioctls for vcpu fds
>> diff --git a/include/linux/kvm_host.h b/include/linux/kvm_host.h
>> index 7dcae4b..5b2be86 100644
>> --- a/include/linux/kvm_host.h
>> +++ b/include/linux/kvm_host.h
>> @@ -138,6 +138,7 @@ struct kvm {
>>  	struct kvm_io_bus pio_bus;
>>  #ifdef CONFIG_HAVE_KVM_EVENTFD
>>  	struct list_head irqfds;
>> +	struct list_head iosignalfds;
>>  #endif
>>  	struct kvm_vm_stat stat;
>>  	struct kvm_arch arch;
>> @@ -535,19 +536,24 @@ static inline void kvm_free_irq_routing(struct kvm *kvm) {}
>>  
>>  #ifdef CONFIG_HAVE_KVM_EVENTFD
>>  
>> -void kvm_irqfd_init(struct kvm *kvm);
>> +void kvm_eventfd_init(struct kvm *kvm);
>>  int kvm_irqfd(struct kvm *kvm, int fd, int gsi, int flags);
>>  void kvm_irqfd_release(struct kvm *kvm);
>> +int kvm_iosignalfd(struct kvm *kvm, struct kvm_iosignalfd *args);
>>  
>>  #else
>>  
>> -static inline void kvm_irqfd_init(struct kvm *kvm) {}
>> +static inline void kvm_eventfd_init(struct kvm *kvm) {}
>>  static inline int kvm_irqfd(struct kvm *kvm, int fd, int gsi, int flags)
>>  {
>>  	return -EINVAL;
>>  }
>>  
>>  static inline void kvm_irqfd_release(struct kvm *kvm) {}
>> +static inline int kvm_iosignalfd(struct kvm *kvm, struct kvm_iosignalfd *args)
>> +{
>> +	return -EINVAL;
>> +}
>>  
>>  #endif /* CONFIG_HAVE_KVM_EVENTFD */
>>  
>> diff --git a/virt/kvm/eventfd.c b/virt/kvm/eventfd.c
>> index c63ff6a..7700e39 100644
>> --- a/virt/kvm/eventfd.c
>> +++ b/virt/kvm/eventfd.c
>> @@ -21,12 +21,16 @@
>>   */
>>  
>>  #include <linux/kvm_host.h>
>> +#include <linux/kvm.h>
>>  #include <linux/workqueue.h>
>>  #include <linux/syscalls.h>
>>  #include <linux/wait.h>
>>  #include <linux/poll.h>
>>  #include <linux/file.h>
>>  #include <linux/list.h>
>> +#include <linux/eventfd.h>
>> +
>> +#include "iodev.h"
>>  
>>  /*
>>   * --------------------------------------------------------------------
>> @@ -207,9 +211,10 @@ kvm_deassign_irqfd(struct kvm *kvm, int fd, int gsi)
>>  }
>>  
>>  void
>> -kvm_irqfd_init(struct kvm *kvm)
>> +kvm_eventfd_init(struct kvm *kvm)
>>  {
>>  	INIT_LIST_HEAD(&kvm->irqfds);
>> +	INIT_LIST_HEAD(&kvm->iosignalfds);
>>  }
>>  
>>  int
>> @@ -232,3 +237,161 @@ kvm_irqfd_release(struct kvm *kvm)
>>  		irqfd_release(irqfd);
>>  	}
>>  }
>> +
>> +/*
>> + * --------------------------------------------------------------------
>> + * iosignalfd: translate a PIO/MMIO memory write to an eventfd signal.
>> + *
>> + * userspace can register a PIO/MMIO address with an eventfd for recieving
>> + * notification when the memory has been touched.
>> + * --------------------------------------------------------------------
>> + */
>> +
>> +struct _iosignalfd {
>> +	u64                  cookie;
>> +	u64                  addr;
>> +	size_t               length;
>> +	struct file         *file;
>> +	struct list_head     list;
>> +	struct kvm_io_device dev;
>> +};
>> +
>> +static int
>> +iosignalfd_in_range(struct kvm_io_device *this, gpa_t addr, int len,
>> +		    int is_write)
>> +{
>> +	struct _iosignalfd *p = (struct _iosignalfd *)this->private;
>> +
>> +	return ((addr >= p->addr && (addr < p->addr + p->length)));
>> +}
>> +
>> +/* writes trigger an event */
>> +static void
>> +iosignalfd_write(struct kvm_io_device *this, gpa_t addr, int len,
>> +		 const void *val)
>> +{
>> +	struct _iosignalfd *iosignalfd = (struct _iosignalfd *)this->private;
>> +
>> +	eventfd_signal(iosignalfd->file, 1);
>> +}
>> +
>> +/* reads return all zeros */
>> +static void
>> +iosignalfd_read(struct kvm_io_device *this, gpa_t addr, int len, void *val)
>> +{
>> +	memset(val, 0, len);
>> +}
>>     
>
> Gregory,
>
> Can you explain the reasoning behind limiting the interface to 
> write-only ranges, with reads returning zero.
>   


Well, a write seems more closely modeling the action of emitting
something out from the guest, as opposed to a read which is technically
pulling something in.  Many real hardware devices work this way (e.g.
writes to a specific register trigger an action (dma, etc), whereas
reads return all zeros, all ones, or an undefined pattern, but either
way are a no-op) so I was mimicking their design here.  I guess one
benefit of this approach is that errant reads (which are more probable
than errant writes) do not cause unintended side-effects.

That said, I do not have a strong opinion regarding this distinction, so
if you prefer to see something else  (like reads also triggering an
event too) I am fine with that.

> Is that because it fits the use cases in mind for iosignalfd?
>   

No, not specifically.  All I care about is being able to trigger the
eventfd, and afaik an ioread() or iowrite() could do that equally well.

I suppose the possibility exists that a read vs write may offer
different latency characteristics, too.  Before I could figure that out
I would need to alter some of my test harnesses to quantify differences
of that nature.  I am guessing they are about equivalent, but that is
just a guess.
> Not that I have a better suggestion at the moment, just trying
> to understand.
>
> Also, the heavy-weight exit avoidance assumes that the action signalled
> will be serviced by the qemu device model in a separate CPU, otherwise
> there is no gain, is that correct?
>   

Not necessarily.  For one, it doesn't have to be a qemu device model per
se (for instance, it could be a vbus device model, or it could be
another guests irqfd to an in-kernel LAPIC, etc).  Second, there are
things we can do right in the same vcpu context that do not involve a
context switch (per se) or a return to userspace.  For instance, we
could potentially directly egress a packet (in a vbus device model), or
inject an interrupt to a different guest (via irqfd).  None of these
things require a qemu device model or a separate CPU, yet they all avoid
various overheads such as the heavy-weight exit.

That said, one intended use case for this construct *is* to wake up some
other thread to do something for us concurrently.  In this particular
application your observation is absolutely correct.  E.g. if the task
cannot be woken up to run in parallel to the vcpu thread, we may
unavoidably still take a heavy-weight exit anyway when/if the vcpu gets
preempted (and thus any gains are diminished).

However, it should be noted that even when this is the case, there still
may be gains in other places besides any (missed) gains from avoiding
the heavy-weight exit.  For instance, if the qemu MMIO/PIO handler wakes
a thread on a uP box, it may still be faster to wake it from iosignalfd
because we still get to the wakeup code via a shorter path, even though
both paths ultimately may need a heavy-weight to actually run the thread.

HTH, and thank you for the review,
-Greg

 



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

Powered by blists - more mailing lists