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.
>
> 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.
>
> 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.
> + 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.
> + 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.
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.
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?
> +
> #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.
If those are host side references, I guess they don't need updates for
this.
But if that's the case, then how does indexing into the same array
with both host and guest server numbers make sense?
> 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
