Re: [Qemu-devel] [PATCH v10 2/5] i386: Populate AMD Processor Cache Information for cpuid 0x8000001D

[Moger, Babu]

Hi Eduardo, Please see my comments below.

> -----Original Message-----
> From: Eduardo Habkost [mailto:address@hidden
> Sent: Tuesday, May 22, 2018 8:54 AM
> To: Moger, Babu <address@hidden>
> Cc: address@hidden; address@hidden; address@hidden;
> address@hidden; address@hidden; address@hidden;
> address@hidden; address@hidden; address@hidden
> Subject: Re: [PATCH v10 2/5] i386: Populate AMD Processor Cache
> Information for cpuid 0x8000001D
> 
> On Mon, May 21, 2018 at 08:41:12PM -0400, Babu Moger wrote:
> > Add information for cpuid 0x8000001D leaf. Populate cache topology
> information
> > for different cache types(Data Cache, Instruction Cache, L2 and L3)
> supported
> > by 0x8000001D leaf. Please refer Processor Programming Reference (PPR)
> for AMD
> > Family 17h Model for more details.
> >
> > Signed-off-by: Babu Moger <address@hidden>
> > ---
> >  target/i386/cpu.c | 103
> ++++++++++++++++++++++++++++++++++++++++++++++++++++++
> >  target/i386/kvm.c |  29 +++++++++++++--
> >  2 files changed, 129 insertions(+), 3 deletions(-)
> >
> > diff --git a/target/i386/cpu.c b/target/i386/cpu.c
> > index d9773b6..1dd060a 100644
> > --- a/target/i386/cpu.c
> > +++ b/target/i386/cpu.c
> > @@ -336,6 +336,85 @@ static void
> encode_cache_cpuid80000006(CPUCacheInfo *l2,
> >      }
> >  }
> >
> 
> The number of variables here is large, so maybe we should
> document what each one mean so it's easier to review:
> 

Sure. Will add more comments.

> 
> > +/* Definitions used for building CPUID Leaf 0x8000001D and 0x8000001E */
> > +/* Please refer AMD64 Architecture Programmer’s Manual Volume 3 */
> > +#define MAX_CCX 2
> 
> CCX is "core complex", right?  A comment would be useful here.

Yes. It is core complex.  Will add comments.

> 
> > +#define MAX_CORES_IN_CCX 4
> > +#define MAX_NODES_EPYC 4
> 
> A comment explaining why it's OK to use a EPYC-specific constant
> here would be useful.

Sure.

> 
> 
> > +#define MAX_CORES_IN_NODE 8
> > +
> > +/* Number of logical processors sharing L3 cache */
> > +#define NUM_SHARING_CACHE(threads, num_sharing)   ((threads > 1) ?
> \
> > +                         (((num_sharing - 1) * threads) + 1)  : \
> > +                         (num_sharing - 1))
> 
> This formula is confusing to me.  If 4 cores are sharing the
> cache and threads==1, 4 logical processors share the cache, and
> we return 3.  Sounds OK.
> 
> But, if 4 cores are sharing the cache and threads==2, the number
> of logical processors sharing the cache is 8.  We should return
> 7.  The formula above returns (((4 - 1) * 2) + 1), which is
> correct.
> 
> But isn't it simpler to write this as:
> 
> #define NUM_SHARING_CACHE(threads, num_sharing) \
>         (((num_sharing) * (threads)) - 1)
> 
> 
> (Maybe the "- 1" part could be moved outside the macro for
> clarity.  See below.)

Yes, If we move -1 outside, then we could simplify it and we don’t need this 
macro. Will change it.

> 
> 
> > +/*
> > + * L3 Cache is shared between all the cores in a core complex.
> > + * Maximum cores that can share L3 is 4.
> > + */
> > +static int num_sharing_l3_cache(int nr_cores)
> 
> Can we document what exactly this function is going to return?
> This returns the number of cores sharing l3 cache, not the number
> of logical processors, correct?


Yes. It is the number of cores.  Will fix it.

> 
> 
> > +{
> > +    int i, nodes = 1;
> > +
> > +    /* Check if we can fit all the cores in one CCX */
> > +    if (nr_cores <= MAX_CORES_IN_CCX) {
> > +        return nr_cores;
> > +    }
> > +    /*
> > +     * Figure out the number of nodes(or dies) required to build
> > +     * this config. Max cores in a node is 8
> > +     */
> > +    for (i = nodes; i <= MAX_NODES_EPYC; i++) {
> > +        if (nr_cores <= (i * MAX_CORES_IN_NODE)) {
> > +            nodes = i;
> > +            break;
> > +        }
> > +        /* We support nodes 1, 2, 4 */
> > +        if (i == 3) {
> > +            continue;
> > +        }
> > +    }
> 
> "continue" as the very last statement of a for loop does nothing,
> so it looks like this could be written as:

In real hardware number of nodes 3 is not a valid configuration. I was trying 
to avoid 3 there. Yes, we can achieve this with DIV_ROUND_UP like below.

> 
>     for (i = nodes; i <= MAX_NODES_EPYC; i++) {
>         if (nr_cores <= (i * MAX_CORES_IN_NODE)) {
>             nodes = i;
>             break;
>         }
>     }
> 
> which in turn seems to be the same as:
> 
>     nodes = DIV_ROUND_UP(nr_cores, MAX_CORES_IN_NODE);
>     nodes = MIN(nodes, MAX_NODES_EPYC)
> 
> But, is this really what we want here?

Number of nodes supported is 1, 2 or 4.  Hardware does not support 3. That is 
what I was trying to achieve there.  

DIV_ROUND_UP will work with check for 3.  If it is 3 then make nodes = 4.  Will 
change it.

MIN(nodes, MAX_NODES_EPYC) is not required as I have added a check in patch 4/5 
to check topology(function verify_topology).
If we go beyond 4 nodes then I am disabling topoext feature. 


> 
> 
> > +    /* Spread the cores accros all the CCXs and return max cores in a ccx 
> > */
> > +    return (nr_cores / (nodes * MAX_CCX)) +
> > +            ((nr_cores % (nodes * MAX_CCX)) ? 1 : 0);
> 
> This also seems to be the same as DIV_ROUND_UP?
> 
>     return DIV_ROUND_UP(nr_cores, nodes * MAX_CCX);
> 
Yes.  DIV_ROUND_UP  will work.

> I didn't confirm the logic is valid, though, because I don't know
> what we should expect.  What is the expected return value of this
> function in the following cases?
> 
>  -smp 24,sockets=2,cores=12,threads=1

This should return 3(DIV_ROUND_UP(12, 2 * 2). We can fit in 2 nodes, with 4 
core complexes. There will be 3 cores in each core complex.

>  -smp 64,sockets=2,cores=32,threads=1

This should return 4(DIV_ROUND_UP(32, 4 * 2).. We can fit it in 4 nodes with 
total 8 core complexes. There will be 4 cores in each core complex.

> 
> 
> > +}
> > +
> > +/* Encode cache info for CPUID[8000001D] */
> > +static void encode_cache_cpuid8000001d(CPUCacheInfo *cache,
> CPUState *cs,
> > +                                uint32_t *eax, uint32_t *ebx,
> > +                                uint32_t *ecx, uint32_t *edx)
> > +{
> > +    uint32_t num_share_l3;
> > +    assert(cache->size == cache->line_size * cache->associativity *
> > +                          cache->partitions * cache->sets);
> > +
> > +    *eax = CACHE_TYPE(cache->type) | CACHE_LEVEL(cache->level) |
> > +               (cache->self_init ? CACHE_SELF_INIT_LEVEL : 0);
> > +
> > +    /* L3 is shared among multiple cores */
> > +    if (cache->level == 3) {
> > +        num_share_l3 = num_sharing_l3_cache(cs->nr_cores);
> > +        *eax |= (NUM_SHARING_CACHE(cs->nr_threads, num_share_l3) <<
> 14);
> 
> Considering that the line below has an explicit "- 1", I think
> the "- 1" part could be moved outside the NUM_SHARING_CACHE
> macro, and used explicitly here.
> 
> But then the NUM_SHARING_CACHE would be just a simple
> multiplication, so this could be simply written as:
> 
>     /* num_sharing_l3_cache() renamed to cores_sharing_l3_cache() */
>     uint32_t l3_cores = cores_sharing_l3_cache(cs->nr_cores);
>     uint32_t l3_logical_processors = l3_cores * cs->nr_threads;
>     *eax |= (l3_logical_processors - 1) << 14;

Yes. Will make these changes.

> 
> > +    } else {
> > +        *eax |= ((cs->nr_threads - 1) << 14);
> > +    }
> > +
> > +    assert(cache->line_size > 0);
> > +    assert(cache->partitions > 0);
> > +    assert(cache->associativity > 0);
> > +    /* We don't implement fully-associative caches */
> > +    assert(cache->associativity < cache->sets);
> > +    *ebx = (cache->line_size - 1) |
> > +           ((cache->partitions - 1) << 12) |
> > +           ((cache->associativity - 1) << 22);
> > +
> > +    assert(cache->sets > 0);
> > +    *ecx = cache->sets - 1;
> > +
> > +    *edx = (cache->no_invd_sharing ? CACHE_NO_INVD_SHARING : 0) |
> > +           (cache->inclusive ? CACHE_INCLUSIVE : 0) |
> > +           (cache->complex_indexing ? CACHE_COMPLEX_IDX : 0);
> > +}
> > +
> >  /*
> >   * Definitions of the hardcoded cache entries we expose:
> >   * These are legacy cache values. If there is a need to change any
> > @@ -4005,6 +4084,30 @@ void cpu_x86_cpuid(CPUX86State *env,
> uint32_t index, uint32_t count,
> >              *edx = 0;
> >          }
> >          break;
> > +    case 0x8000001D:
> > +        *eax = 0;
> > +        switch (count) {
> > +        case 0: /* L1 dcache info */
> > +            encode_cache_cpuid8000001d(env->cache_info_amd.l1d_cache,
> cs,
> > +                                       eax, ebx, ecx, edx);
> > +            break;
> > +        case 1: /* L1 icache info */
> > +            encode_cache_cpuid8000001d(env->cache_info_amd.l1i_cache, cs,
> > +                                       eax, ebx, ecx, edx);
> > +            break;
> > +        case 2: /* L2 cache info */
> > +            encode_cache_cpuid8000001d(env->cache_info_amd.l2_cache, cs,
> > +                                       eax, ebx, ecx, edx);
> > +            break;
> > +        case 3: /* L3 cache info */
> > +            encode_cache_cpuid8000001d(env->cache_info_amd.l3_cache, cs,
> > +                                       eax, ebx, ecx, edx);
> > +            break;
> > +        default: /* end of info */
> > +            *eax = *ebx = *ecx = *edx = 0;
> > +            break;
> > +        }
> > +        break;
> >      case 0xC0000000:
> >          *eax = env->cpuid_xlevel2;
> >          *ebx = 0;
> > diff --git a/target/i386/kvm.c b/target/i386/kvm.c
> > index d6666a4..a8bf7eb 100644
> > --- a/target/i386/kvm.c
> > +++ b/target/i386/kvm.c
> > @@ -979,9 +979,32 @@ int kvm_arch_init_vcpu(CPUState *cs)
> >          }
> >          c = &cpuid_data.entries[cpuid_i++];
> >
> > -        c->function = i;
> > -        c->flags = 0;
> > -        cpu_x86_cpuid(env, i, 0, &c->eax, &c->ebx, &c->ecx, &c->edx);
> > +        switch (i) {
> > +        case 0x8000001d:
> > +            /* Query for all AMD cache information leaves */
> > +            for (j = 0; ; j++) {
> > +                c->function = i;
> > +                c->flags = KVM_CPUID_FLAG_SIGNIFCANT_INDEX;
> > +                c->index = j;
> > +                cpu_x86_cpuid(env, i, j, &c->eax, &c->ebx, &c->ecx, 
> > &c->edx);
> > +
> > +                if (c->eax == 0) {
> > +                    break;
> > +                }
> > +                if (cpuid_i == KVM_MAX_CPUID_ENTRIES) {
> > +                    fprintf(stderr, "cpuid_data is full, no space for "
> > +                            "cpuid(eax:0x%x,ecx:0x%x)\n", i, j);
> > +                    abort();
> > +                }
> > +                c = &cpuid_data.entries[cpuid_i++];
> > +            }
> > +            break;
> > +        default:
> > +            c->function = i;
> > +            c->flags = 0;
> > +            cpu_x86_cpuid(env, i, 0, &c->eax, &c->ebx, &c->ecx, &c->edx);
> > +            break;
> > +        }
> >      }
> >
> >      /* Call Centaur's CPUID instructions they are supported. */
> > --
> > 1.8.3.1
> >
> 
> --
> Eduardo