From mboxrd@z Thu Jan 1 00:00:00 1970 From: Alex Williamson Subject: Re: VFIO based vGPU(was Re: [Announcement] 2015-Q3 release of XenGT - a Mediated ...) Date: Tue, 26 Jan 2016 20:07:48 -0700 Message-ID: <1453864068.3107.3.camel@redhat.com> References: <569C5071.6080004@intel.com> <1453092476.32741.67.camel@redhat.com> <569CA8AD.6070200@intel.com> <1453143919.32741.169.camel@redhat.com> <569F4C86.2070501@intel.com> <56A6083E.10703@intel.com> <1453757426.32741.614.camel@redhat.com> <56A72313.9030009@intel.com> <56A77D2D.40109@gmail.com> <1453826249.26652.54.camel@redhat.com> <1453844613.18049.1.camel@redhat.com> <1453846073.18049.3.camel@redhat.com> <1453847250.18049.5.camel@redhat.com> <1453848975.18049.7.camel@redhat.com> <56A821AD.5090606@intel.com> Mime-Version: 1.0 Content-Type: text/plain; charset="UTF-8" Content-Transfer-Encoding: quoted-printable Cc: Yang Zhang , "Ruan, Shuai" , "Tian, Kevin" , Neo Jia , "kvm@vger.kernel.org" , "igvt-g@lists.01.org" , qemu-devel , Gerd Hoffmann , Paolo Bonzini , "Lv, Zhiyuan" To: Jike Song Return-path: In-Reply-To: <56A821AD.5090606@intel.com> List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Errors-To: qemu-devel-bounces+gceq-qemu-devel=gmane.org@nongnu.org Sender: qemu-devel-bounces+gceq-qemu-devel=gmane.org@nongnu.org List-Id: kvm.vger.kernel.org On Wed, 2016-01-27 at 09:47 +0800, Jike Song wrote: > On 01/27/2016 06:56 AM, Alex Williamson wrote: > > On Tue, 2016-01-26 at 22:39 +0000, Tian, Kevin wrote: > > > > From: Alex Williamson [mailto:alex.williamson@redhat.com] > > > > Sent: Wednesday, January 27, 2016 6:27 AM > > > > =C2=A0 > > > > On Tue, 2016-01-26 at 22:15 +0000, Tian, Kevin wrote: > > > > > > From: Alex Williamson [mailto:alex.williamson@redhat.com] > > > > > > Sent: Wednesday, January 27, 2016 6:08 AM > > > > > > =C2=A0 > > > > > > > > > > =C2=A0 > > > > > > > > > =C2=A0 > > > > > > > > > Today KVMGT (not using VFIO yet) registers I/O emulatio= n callbacks to > > > > > > > > > KVM, so VM MMIO access will be forwarded to KVMGT direc= tly for > > > > > > > > > emulation in kernel. If we reuse above R/W flags, the w= hole emulation > > > > > > > > > path would be unnecessarily long with obvious performan= ce impact. We > > > > > > > > > either need a new flag here to indicate in-kernel emula= tion (bias from > > > > > > > > > passthrough support), or just hide the region alternati= vely (let KVMGT > > > > > > > > > to handle I/O emulation itself like today). > > > > > > > > =C2=A0 > > > > > > > > That sounds like a future optimization TBH.=C2=A0=C2=A0Th= ere's very strict > > > > > > > > layering between vfio and kvm.=C2=A0=C2=A0Physical device= assignment could make > > > > > > > > use of it as well, avoiding a round trip through userspac= e when an > > > > > > > > ioread/write would do.=C2=A0=C2=A0Userspace also needs to= orchestrate those kinds > > > > > > > > of accelerators, there might be cases where userspace wan= ts to see those > > > > > > > > transactions for debugging or manipulating the device.=C2= =A0=C2=A0We can't simply > > > > > > > > take shortcuts to provide such direct access.=C2=A0=C2=A0= Thanks, > > > > > > > > =C2=A0 > > > > > > > =C2=A0 > > > > > > > But we have to balance such debugging flexibility and accep= table performance. > > > > > > > To me the latter one is more important otherwise there'd be= no real usage > > > > > > > around this technique, while for debugging there are other = alternative (e.g. > > > > > > > ftrace) Consider some extreme case with 100k traps/second a= nd then see > > > > > > > how much impact a 2-3x longer emulation path can bring... > > > > > > =C2=A0 > > > > > > Are you jumping to the conclusion that it cannot be done with= proper > > > > > > layering in place?=C2=A0=C2=A0Performance is important, but i= t's not an excuse to > > > > > > abandon designing interfaces between independent components.=C2= =A0=C2=A0Thanks, > > > > > > =C2=A0 > > > > > =C2=A0 > > > > > Two are not controversial. My point is to remove unnecessary lo= ng trip > > > > > as possible. After another thought, yes we can reuse existing r= ead/write > > > > > flags: > > > > > =C2=A0 - KVMGT will expose a private control variable whether i= n-kernel > > > > > delivery is required; > > > > =C2=A0 > > > > But in-kernel delivery is never *required*.=C2=A0=C2=A0Wouldn't u= serspace want to > > > > deliver in-kernel any time it possibly could? > > > > =C2=A0 > > > > > =C2=A0 - when the variable is true, KVMGT will register in-kern= el MMIO > > > > > emulation callbacks then VM MMIO request will be delivered to K= VMGT > > > > > directly; > > > > > =C2=A0 - when the variable is false, KVMGT will not register an= ything. > > > > > VM MMIO request will then be delivered to Qemu and then ioread/= write > > > > > will be used to finally reach KVMGT emulation logic; > > > > =C2=A0 > > > > No, that means the interface is entirely dependent on a backdoor = through > > > > KVM.=C2=A0=C2=A0Why can't userspace (QEMU) do something like regi= ster an MMIO > > > > region with KVM handled via a provided file descriptor and offset= , > > > > couldn't KVM then call the file ops without a kernel exit?=C2=A0=C2= =A0Thanks, > > > > =C2=A0 > > > =C2=A0 > > > Could you elaborate this thought? If it can achieve the purpose w/o > > > a kernel exit definitely we can adapt to it. :-) > >=C2=A0 > > I only thought of it when replying to the last email and have been do= ing > > some research, but we already do quite a bit of synchronization throu= gh > > file descriptors.=C2=A0=C2=A0The kvm-vfio pseudo device uses a group = file > > descriptor to ensure a user has access to a group, allowing some degr= ee > > of interaction between modules.=C2=A0=C2=A0Eventfds and irqfds alread= y make use of > > f_ops on file descriptors to poke data.=C2=A0=C2=A0So, if KVM had inf= ormation that > > an MMIO region was backed by a file descriptor for which it already h= as > > a reference via fdget() (and verified access rights and whatnot), the= n > > it ought to be a simple matter to get to f_ops->read/write knowing th= e > > base offset of that MMIO region.=C2=A0=C2=A0Perhaps it could even sim= ply use > > __vfs_read/write().=C2=A0=C2=A0Then we've got a proper reference to t= he file > > descriptor for ownership purposes and we've transparently jumped acro= ss > > modules without any implicit knowledge of the other end.=C2=A0=C2=A0C= ould it work? >=C2=A0 > This is OK for KVMGT, from fops to vgpu device-model would always be si= mple. > The only question is, how is KVM hypervisor supposed to get the fd on V= M-exitings? Hi Jike, Sorry, I don't understand "on VM-exiting".=C2=A0=C2=A0KVM would hold a re= ference to the fd via fdget(), so the vfio device wouldn't be closed until the VM exits and KVM releases that reference. > copy-and-paste the current implementation of vcpu_mmio_write(), seems > nothing but GPA and len are provided: I presume that an MMIO region is already registered with a GPA and length, the additional information necessary would be a file descriptor and offset into the file descriptor for the base of the MMIO space. >=C2=A0 static int vcpu_mmio_write(struct kvm_vcpu *vcpu, gpa_t addr, int= len, >=C2=A0 =C2=A0=C2=A0=C2=A0const void *v) >=C2=A0 { >=C2=A0 int handled =3D 0; >=C2=A0 int n; >=C2=A0 >=C2=A0 do { >=C2=A0 n =3D min(len, 8); >=C2=A0 if (!(vcpu->arch.apic && >=C2=A0 =C2=A0=C2=A0=C2=A0=C2=A0=C2=A0=C2=A0!kvm_iodevice_write(vcpu, &= vcpu->arch.apic->dev, addr, n, v)) >=C2=A0 =C2=A0=C2=A0=C2=A0=C2=A0&& kvm_io_bus_write(vcpu, KVM_MMIO_BUS,= addr, n, v)) >=C2=A0 break; >=C2=A0 handled +=3D n; >=C2=A0 addr +=3D n; >=C2=A0 len -=3D n; >=C2=A0 v +=3D n; >=C2=A0 } while (len); >=C2=A0 >=C2=A0 return handled; >=C2=A0 } >=C2=A0 > If we back a GPA range with a fd, this will also be a 'backdoor'? KVM would simply be able to service the MMIO access using the provided fd and offset.=C2=A0=C2=A0It's not a back door because we will have creat= ed an API for KVM to have a file descriptor and offset registered (by userspace) to handle the access.=C2=A0=C2=A0Also, KVM does not know the file descrip= tor is handled by a VFIO device and VFIO doesn't know the read/write accesses is initiated by KVM.=C2=A0=C2=A0Seems like the question is whether we can= fit something like that into the existing KVM MMIO bus/device handlers in-kernel.=C2=A0=C2=A0Thanks, Alex