On Tue, Apr 24, 2018 at 01:19:15PM +1000, Sam Bobroff wrote:
> On Mon, Apr 23, 2018 at 11:06:35AM +0200, Cédric Le Goater wrote:
> > On 04/16/2018 06:09 AM, David Gibson wrote:
> > > On Thu, Apr 12, 2018 at 05:02:06PM +1000, Sam Bobroff wrote:
> > >> It is not currently possible to create the full number of possible
> > >> VCPUs (KVM_MAX_VCPUS) on Power9 with KVM-HV when the guest uses less
> > >> threads per core than it's core stride (or "VSMT mode"). This is
> > >> because the VCORE ID and XIVE offsets to grow beyond KVM_MAX_VCPUS
> > >> even though the VCPU ID is less than KVM_MAX_VCPU_ID.
> > >>
> > >> To address this, "pack" the VCORE ID and XIVE offsets by using
> > >> knowledge of the way the VCPU IDs will be used when there are less
> > >> guest threads per core than the core stride. The primary thread of
> > >> each core will always be used first. Then, if the guest uses more than
> > >> one thread per core, these secondary threads will sequentially follow
> > >> the primary in each core.
> > >>
> > >> So, the only way an ID above KVM_MAX_VCPUS can be seen, is if the
> > >> VCPUs are being spaced apart, so at least half of each core is empty
> > >> and IDs between KVM_MAX_VCPUS and (KVM_MAX_VCPUS * 2) can be mapped
> > >> into the second half of each core (4..7, in an 8-thread core).
> > >>
> > >> Similarly, if IDs above KVM_MAX_VCPUS * 2 are seen, at least 3/4 of
> > >> each core is being left empty, and we can map down into the second and
> > >> third quarters of each core (2, 3 and 5, 6 in an 8-thread core).
> > >>
> > >> Lastly, if IDs above KVM_MAX_VCPUS * 4 are seen, only the primary
> > >> threads are being used and 7/8 of the core is empty, allowing use of
> > >> the 1, 3, 5 and 7 thread slots.
> > >>
> > >> (Strides less than 8 are handled similarly.)
> > >>
> > >> This allows the VCORE ID or offset to be calculated quickly from the
> > >> VCPU ID or XIVE server numbers, without access to the VCPU structure.
> > >>
> > >> Signed-off-by: Sam Bobroff <sam.bobroff@xxxxxxxxxxx>
> > >> ---
> > >> Hello everyone,
> > >>
> > >> I've tested this on P8 and P9, in lots of combinations of host and guest
> > >> threading modes and it has been fine but it does feel like a "tricky"
> > >> approach, so I still feel somewhat wary about it.
> >
> > Have you done any migration ?
>
> No, but I will :-)
>
> > >> I've posted it as an RFC because I have not tested it with guest native-XIVE,
> > >> and I suspect that it will take some work to support it.
> >
> > The KVM XIVE device will be different for XIVE exploitation mode, same structures
> > though. I will send a patchset shortly.
>
> Great. This is probably where conflicts between the host and guest
> numbers will show up. (See dwg's question below.)
>
> > >> arch/powerpc/include/asm/kvm_book3s.h | 19 +++++++++++++++++++
> > >> arch/powerpc/kvm/book3s_hv.c | 14 ++++++++++----
> > >> arch/powerpc/kvm/book3s_xive.c | 9 +++++++--
> > >> 3 files changed, 36 insertions(+), 6 deletions(-)
> > >>
> > >> diff --git a/arch/powerpc/include/asm/kvm_book3s.h b/arch/powerpc/include/asm/kvm_book3s.h
> > >> index 376ae803b69c..1295056d564a 100644
> > >> --- a/arch/powerpc/include/asm/kvm_book3s.h
> > >> +++ b/arch/powerpc/include/asm/kvm_book3s.h
> > >> @@ -368,4 +368,23 @@ extern int kvmppc_h_logical_ci_store(struct kvm_vcpu *vcpu);
> > >> #define SPLIT_HACK_MASK 0xff000000
> > >> #define SPLIT_HACK_OFFS 0xfb000000
> > >>
> > >> +/* Pack a VCPU ID from the [0..KVM_MAX_VCPU_ID) space down to the
> > >> + * [0..KVM_MAX_VCPUS) space, while using knowledge of the guest's core stride
> > >> + * (but not it's actual threading mode, which is not available) to avoid
> > >> + * collisions.
> > >> + */
> > >> +static inline u32 kvmppc_pack_vcpu_id(struct kvm *kvm, u32 id)
> > >> +{
> > >> + const int block_offsets[MAX_SMT_THREADS] = {0, 4, 2, 6, 1, 5, 3, 7};
> > >
> > > I'd suggest 1,3,5,7 at the end rather than 1,5,3,7 - accomplishes
> > > roughly the same thing, but I think makes the pattern more obvious.
>
> OK.
>
> > >> + int stride = kvm->arch.emul_smt_mode > 1 ?
> > >> + kvm->arch.emul_smt_mode : kvm->arch.smt_mode;
> > >
> > > AFAICT from BUG_ON()s etc. at the callsites, kvm->arch.smt_mode must
> > > always be 1 when this is called, so the conditional here doesn't seem
> > > useful.
>
> Ah yes, right. (That was an older version when I was thinking of using
> it for P8 as well but that didn't seem to be a good idea.)
>
> > >> + int block = (id / KVM_MAX_VCPUS) * (MAX_SMT_THREADS / stride);
> > >> + u32 packed_id;
> > >> +
> > >> + BUG_ON(block >= MAX_SMT_THREADS);
> > >> + packed_id = (id % KVM_MAX_VCPUS) + block_offsets[block];
> > >> + BUG_ON(packed_id >= KVM_MAX_VCPUS);
> > >> + return packed_id;
> > >> +}
> > >
> > > It took me a while to wrap my head around the packing function, but I
> > > think I got there in the end. It's pretty clever.
>
> Thanks, I'll try to add a better description as well :-)
>
> > > One thing bothers me, though. This certainly packs things under
> > > KVM_MAX_VCPUS, but not necessarily under the actual number of vcpus.
> > > e.g. KVM_MAC_VCPUS==16, 8 vcpus total, stride 8, 2 vthreads/vcore (as
> > > qemu sees it), gives both unpacked IDs (0, 1, 8, 9, 16, 17, 24, 25)
> > > and packed ids of (0, 1, 8, 9, 4, 5, 12, 13) - leaving 2, 3, 6, 7
> > > etc. unused.
>
> That's right. The property it provides is that all the numbers are under
> KVM_MAX_VCPUS (which, see below, is the size of the fixed areas) not
> that they are sequential.
>
> > > So again, the question is what exactly are these remapped IDs useful
> > > for. If we're indexing into a bare array of structures of size
> > > KVM_MAX_VCPUS then we're *already* wasting a bunch of space by having
> > > more entries than vcpus. If we're indexing into something sparser,
> > > then why is the remapping worthwhile?
>
> Well, here's my thinking:
>
> At the moment, kvm->vcores[] and xive->vp_base are both sized by NR_CPUS
> (via KVM_MAX_VCPUS and KVM_MAX_VCORES which are both NR_CPUS). This is
> enough space for the maximum number of VCPUs, and some space is wasted
> when the guest uses less than this (but KVM doesn't know how many will
> be created, so we can't do better easily). The problem is that the
> indicies overflow before all of those VCPUs can be created, not that
> more space is needed.
>
> We could fix the overflow by expanding these areas to KVM_MAX_VCPU_ID
> but that will use 8x the space we use now, and we know that no more than
> KVM_MAX_VCPUS will be used so all this new space is basically wasted.
>
> So remapping seems better if it will work. (Ben H. was strongly against
> wasting more XIVE space if possible.)
Hm, ok. Are the relevant arrays here per-VM, or global? Or some of both?
> In short, remapping provides a way to allow the guest to create it's full set
> of VCPUs without wasting any more space than we do currently, without
> having to do something more complicated like tracking used IDs or adding
> additional KVM CAPs.
>
> > >> +
> > >> #endif /* __ASM_KVM_BOOK3S_H__ */
> > >> diff --git a/arch/powerpc/kvm/book3s_hv.c b/arch/powerpc/kvm/book3s_hv.c
> > >> index 9cb9448163c4..49165cc90051 100644
> > >> --- a/arch/powerpc/kvm/book3s_hv.c
> > >> +++ b/arch/powerpc/kvm/book3s_hv.c
> > >> @@ -1762,7 +1762,7 @@ static int threads_per_vcore(struct kvm *kvm)
> > >> return threads_per_subcore;
> > >> }
> > >>
> > >> -static struct kvmppc_vcore *kvmppc_vcore_create(struct kvm *kvm, int core)
> > >> +static struct kvmppc_vcore *kvmppc_vcore_create(struct kvm *kvm, int id)
> > >> {
> > >> struct kvmppc_vcore *vcore;
> > >>
> > >> @@ -1776,7 +1776,7 @@ static struct kvmppc_vcore *kvmppc_vcore_create(struct kvm *kvm, int core)
> > >> init_swait_queue_head(&vcore->wq);
> > >> vcore->preempt_tb = TB_NIL;
> > >> vcore->lpcr = kvm->arch.lpcr;
> > >> - vcore->first_vcpuid = core * kvm->arch.smt_mode;
> > >> + vcore->first_vcpuid = id;
> > >> vcore->kvm = kvm;
> > >> INIT_LIST_HEAD(&vcore->preempt_list);
> > >>
> > >> @@ -1992,12 +1992,18 @@ static struct kvm_vcpu *kvmppc_core_vcpu_create_hv(struct kvm *kvm,
> > >> mutex_lock(&kvm->lock);
> > >> vcore = NULL;
> > >> err = -EINVAL;
> > >> - core = id / kvm->arch.smt_mode;
> > >> + if (cpu_has_feature(CPU_FTR_ARCH_300)) {
> > >> + BUG_ON(kvm->arch.smt_mode != 1);
> > >> + core = kvmppc_pack_vcpu_id(kvm, id);
> > >> + } else {
> > >> + core = id / kvm->arch.smt_mode;
> > >> + }
> > >> if (core < KVM_MAX_VCORES) {
> > >> vcore = kvm->arch.vcores[core];
> > >> + BUG_ON(cpu_has_feature(CPU_FTR_ARCH_300) && vcore);
> > >> if (!vcore) {
> > >> err = -ENOMEM;
> > >> - vcore = kvmppc_vcore_create(kvm, core);
> > >> + vcore = kvmppc_vcore_create(kvm, id & ~(kvm->arch.smt_mode - 1));
> > >> kvm->arch.vcores[core] = vcore;
> > >> kvm->arch.online_vcores++;
> > >> }
> > >> diff --git a/arch/powerpc/kvm/book3s_xive.c b/arch/powerpc/kvm/book3s_xive.c
> > >> index f9818d7d3381..681dfe12a5f3 100644
> > >> --- a/arch/powerpc/kvm/book3s_xive.c
> > >> +++ b/arch/powerpc/kvm/book3s_xive.c
> > >> @@ -317,6 +317,11 @@ static int xive_select_target(struct kvm *kvm, u32 *server, u8 prio)
> > >> return -EBUSY;
> > >> }
> > >>
> > >> +static u32 xive_vp(struct kvmppc_xive *xive, u32 server)
> > >> +{
> > >> + return xive->vp_base + kvmppc_pack_vcpu_id(xive->kvm, server);
> > >> +}
> > >> +
> > >
> > > I'm finding the XIVE indexing really baffling. There are a bunch of
> > > other places where the code uses (xive->vp_base + NUMBER) directly.
>
> Ugh, yes. It looks like I botched part of my final cleanup and all the
> cases you saw in kvm/book3s_xive.c should have been replaced with a call to
> xive_vp(). I'll fix it and sorry for the confusion.
Ok.
> > This links the QEMU vCPU server NUMBER to a XIVE virtual processor number
> > in OPAL. So we need to check that all used NUMBERs are, first, consistent
> > and then, in the correct range.
>
> Right. My approach was to allow XIVE to keep using server numbers that
> are equal to VCPU IDs, and just pack down the ID before indexing into
> the vp_base area.
>
> > > If those are host side references, I guess they don't need updates for
> > > this.
>
> These are all guest side references.
>
> > > But if that's the case, then how does indexing into the same array
> > > with both host and guest server numbers make sense?
>
> Right, it doesn't make sense to mix host and guest server numbers when
> we're remapping only the guest ones, but in this case (without native
> guest XIVE support) it's just guest ones.
Right. Will this remapping be broken by guest-visible XIVE? That is
for the guest visible XIVE are we going to need to expose un-remapped
XIVE server IDs to the guest?
> > yes. VPs are allocated with KVM_MAX_VCPUS :
> >
> > xive->vp_base = xive_native_alloc_vp_block(KVM_MAX_VCPUS);
> >
> > but
> >
> > #define KVM_MAX_VCPU_ID (threads_per_subcore * KVM_MAX_VCORES)
> >
> > WE would need to change the allocation of the VPs I guess.
>
> Yes, this is one of the structures that overflow if we don't pack the IDs.
>
> > >> static u8 xive_lock_and_mask(struct kvmppc_xive *xive,
> > >> struct kvmppc_xive_src_block *sb,
> > >> struct kvmppc_xive_irq_state *state)
> > >> @@ -1084,7 +1089,7 @@ int kvmppc_xive_connect_vcpu(struct kvm_device *dev,
> > >> pr_devel("Duplicate !\n");
> > >> return -EEXIST;
> > >> }
> > >> - if (cpu >= KVM_MAX_VCPUS) {
> > >> + if (cpu >= KVM_MAX_VCPU_ID) {>>
> > >> pr_devel("Out of bounds !\n");
> > >> return -EINVAL;
> > >> }
> > >> @@ -1098,7 +1103,7 @@ int kvmppc_xive_connect_vcpu(struct kvm_device *dev,
> > >> xc->xive = xive;
> > >> xc->vcpu = vcpu;
> > >> xc->server_num = cpu;
> > >> - xc->vp_id = xive->vp_base + cpu;
> > >> + xc->vp_id = xive_vp(xive, cpu);
> > >> xc->mfrr = 0xff;
> > >> xc->valid = true;
> > >>
> > >
> >
--
David Gibson | I'll have my music baroque, and my code
david AT gibson.dropbear.id.au | minimalist, thank you. NOT _the_ _other_
| _way_ _around_!
http://www.ozlabs.org/~dgibson
Attachment:
signature.asc
Description: PGP signature
