Re: [PATCH v15 06/12] numa: Extend CLI to provide memory latency and bandwidth information

[Igor Mammedov <imammedo@redhat.com>]

On Thu,  7 Nov 2019 15:45:05 +0800
Tao Xu <tao3.xu@intel.com> wrote:

> From: Liu Jingqi <jingqi.liu@intel.com>
> 
> Add -numa hmat-lb option to provide System Locality Latency and
> Bandwidth Information. These memory attributes help to build
> System Locality Latency and Bandwidth Information Structure(s)
> in ACPI Heterogeneous Memory Attribute Table (HMAT).
> 
> Signed-off-by: Liu Jingqi <jingqi.liu@intel.com>
> Signed-off-by: Tao Xu <tao3.xu@intel.com>
> ---
> 
> Changes in v15:
>     - Change the QAPI version tag to 5.0 (Eric)
> 
> Changes in v14:
>     - Use qemu ctz64 and clz64 instead of builtin function
>     - Improve help message in qemu-options.hx
> 
> Changes in v13:
>     - Reuse Garray to store the raw bandwidth and bandwidth data
>     - Calculate common base unit using range bitmap (Igor)
> ---
>  hw/core/numa.c        | 136 ++++++++++++++++++++++++++++++++++++++++++
>  include/sysemu/numa.h |  68 +++++++++++++++++++++
>  qapi/machine.json     |  94 ++++++++++++++++++++++++++++-
>  qemu-options.hx       |  49 ++++++++++++++-
>  4 files changed, 344 insertions(+), 3 deletions(-)
> 
> diff --git a/hw/core/numa.c b/hw/core/numa.c
> index eba66ab768..f391760c20 100644
> --- a/hw/core/numa.c
> +++ b/hw/core/numa.c
> @@ -23,6 +23,7 @@
>   */
>  
>  #include "qemu/osdep.h"
> +#include "qemu/units.h"
>  #include "sysemu/hostmem.h"
>  #include "sysemu/numa.h"
>  #include "sysemu/sysemu.h"
> @@ -198,6 +199,128 @@ void parse_numa_distance(MachineState *ms, NumaDistOptions *dist, Error **errp)
>      ms->numa_state->have_numa_distance = true;
>  }
>  
> +void parse_numa_hmat_lb(NumaState *numa_state, NumaHmatLBOptions *node,
> +                        Error **errp)
> +{
> +    int first_bit, last_bit;
> +    uint64_t max_latency, temp_base_la;
> +    NodeInfo *numa_info = numa_state->nodes;
> +    HMAT_LB_Info *hmat_lb =
> +        numa_state->hmat_lb[node->hierarchy][node->data_type];
> +    HMAT_LB_Data lb_data;
maybe
  HMAT_LB_Data lb_data = {};
to make sure that no members are left un-initialized 

> +
> +    /* Error checking */
> +    if (node->initiator >= numa_state->num_nodes) {
> +        error_setg(errp, "Invalid initiator=%d, it should be less than %d.",
error_setg(), shouldn't have "." at the end
see include/qapi/error.h:error_setg() description

this applies to all such uses with this series.

> +                   node->initiator, numa_state->num_nodes);
> +        return;
> +    }
> +    if (node->target >= numa_state->num_nodes) {
> +        error_setg(errp, "Invalid target=%d, it should be less than %d.",
> +                   node->target, numa_state->num_nodes);
> +        return;
> +    }
> +    if (!numa_info[node->initiator].has_cpu) {
> +        error_setg(errp, "Invalid initiator=%d, it isn't an "
> +                   "initiator proximity domain.", node->initiator);
> +        return;
> +    }
> +    if (!numa_info[node->target].present) {
> +        error_setg(errp, "Invalid target=%d, it hasn't a valid NUMA node.",
s/\..*\./the target should point to an existing node"/

> +                   node->target);
> +        return;
> +    }
> +
> +    if (!hmat_lb) {
> +        hmat_lb = g_malloc0(sizeof(*hmat_lb));
> +        numa_state->hmat_lb[node->hierarchy][node->data_type] = hmat_lb;
> +        hmat_lb->latency = g_array_new(false, true, sizeof(HMAT_LB_Data));
> +        hmat_lb->bandwidth = g_array_new(false, true, sizeof(HMAT_LB_Data));
> +    }
> +    hmat_lb->hierarchy = node->hierarchy;
> +    hmat_lb->data_type = node->data_type;
> +    lb_data.initiator = node->initiator;
> +    lb_data.target = node->target;
> +
> +    /* Input latency data */
> +    if (node->data_type <= HMATLB_DATA_TYPE_WRITE_LATENCY) {
> +        if (!node->has_latency) {
> +            error_setg(errp, "Missing 'latency' option.");
> +            return;
> +        }
> +        if (node->has_bandwidth) {
> +            error_setg(errp, "Invalid option 'bandwidth' since "
> +                       "the data type is latency.");
> +            return;
> +        }
> +
> +        if (hmat_lb->base_latency == 0) {
> +            hmat_lb->base_latency = UINT64_MAX;
> +        }
> +
> +        /* Calculate the temporary base and compressed latency */
> +        max_latency = node->latency;
> +        temp_base_la = 1;
> +        while (QEMU_IS_ALIGNED(max_latency, 10)) {
> +            max_latency /= 10;
> +            temp_base_la *= 10;
> +        }
> +
> +        /* Calculate the max compressed latency */
> +        hmat_lb->base_latency = MIN(hmat_lb->base_latency, temp_base_la);
> +        max_latency = node->latency / hmat_lb->base_latency;
> +        hmat_lb->max_entry_la = MAX(hmat_lb->max_entry_la, max_latency);
> +
> +        if (hmat_lb->max_entry_la >= UINT16_MAX) {
> +            error_setg(errp, "Latency %" PRIu64 " between initiator=%d and "
> +                       "target=%d should not differ from previously entered "
> +                       "values on more than %d.", node->latency,
s/values/min/max values/

> +                       node->initiator, node->target, UINT16_MAX - 1);
> +            return;
> +        }
> +
> +        lb_data.rawdata = node->latency;
> +        g_array_append_val(hmat_lb->latency, lb_data);
> +    }
I'd use ... else if ... here and then below

> +
> +    /* Input bandwidth data */
> +    if (node->data_type >= HMATLB_DATA_TYPE_ACCESS_BANDWIDTH) {
> +        if (!node->has_bandwidth) {
> +            error_setg(errp, "Missing 'bandwidth' option.");
> +            return;
> +        }
> +        if (node->has_latency) {
> +            error_setg(errp, "Invalid option 'latency' since "
> +                       "the data type is bandwidth.");
> +            return;
> +        }
> +        if (!QEMU_IS_ALIGNED(node->bandwidth, MiB)) {
> +            error_setg(errp, "Bandwidth %" PRIu64 " between initiator=%d and "
> +                       "target=%d should be 1MB aligned.", node->bandwidth,
> +                       node->initiator, node->target);
> +            return;
> +        }
> +
> +        hmat_lb->range_bitmap_bw |= node->bandwidth;
> +
> +        first_bit = ctz64(hmat_lb->range_bitmap_bw);
> +        hmat_lb->base_bandwidth = UINT64_C(1) << first_bit;
> +        last_bit = 64 - clz64(hmat_lb->range_bitmap_bw);
> +        if ((last_bit - first_bit) > UINT16_BITS ||
> +            (MAKE_64BIT_MASK(first_bit, UINT16_BITS) == node->bandwidth)) {
> +            error_setg(errp, "Bandwidth %" PRIu64 " between initiator=%d and "
> +                       "target=%d should not differ from previously entered "
> +                       "values on more than %d.", node->bandwidth,
> +                       node->initiator, node->target, UINT16_MAX - 1);
> +            return;
> +        }
> +
> +        hmat_lb->base_bandwidth = UINT64_C(1) << first_bit;
> +        lb_data.rawdata = node->bandwidth;
> +        g_array_append_val(hmat_lb->bandwidth, lb_data);
> +    }
else {
   assert()
}


for each case we also should check that compressed into 16bits lb_data.rawdata
value won't be 0xFFFF.

> +}
> +
>  void set_numa_options(MachineState *ms, NumaOptions *object, Error **errp)
>  {
>      Error *err = NULL;
> @@ -236,6 +359,19 @@ void set_numa_options(MachineState *ms, NumaOptions *object, Error **errp)
>          machine_set_cpu_numa_node(ms, qapi_NumaCpuOptions_base(&object->u.cpu),
>                                    &err);
>          break;
> +    case NUMA_OPTIONS_TYPE_HMAT_LB:
> +        if (!ms->numa_state->hmat_enabled) {
> +            error_setg(errp, "ACPI Heterogeneous Memory Attribute Table "
> +                       "(HMAT) is disabled, enable it with -machine hmat=on "
> +                       "before using any of hmat specific options.");
> +            return;
> +        }
> +
> +        parse_numa_hmat_lb(ms->numa_state, &object->u.hmat_lb, &err);
> +        if (err) {
> +            goto end;
> +        }
> +        break;
>      default:
>          abort();
>      }
> diff --git a/include/sysemu/numa.h b/include/sysemu/numa.h
> index 788cbec7a2..36e1b4dece 100644
> --- a/include/sysemu/numa.h
> +++ b/include/sysemu/numa.h
> @@ -14,6 +14,29 @@ struct CPUArchId;
>  #define NUMA_DISTANCE_MAX         254
>  #define NUMA_DISTANCE_UNREACHABLE 255
>  
> +/* the value of AcpiHmatLBInfo flags */
> +enum {
> +    HMAT_LB_MEM_MEMORY           = 0,
> +    HMAT_LB_MEM_CACHE_1ST_LEVEL  = 1,
> +    HMAT_LB_MEM_CACHE_2ND_LEVEL  = 2,
> +    HMAT_LB_MEM_CACHE_3RD_LEVEL  = 3,
> +};
> +
> +/* the value of AcpiHmatLBInfo data type */
> +enum {
> +    HMAT_LB_DATA_ACCESS_LATENCY   = 0,
> +    HMAT_LB_DATA_READ_LATENCY     = 1,
> +    HMAT_LB_DATA_WRITE_LATENCY    = 2,
> +    HMAT_LB_DATA_ACCESS_BANDWIDTH = 3,
> +    HMAT_LB_DATA_READ_BANDWIDTH   = 4,
> +    HMAT_LB_DATA_WRITE_BANDWIDTH  = 5,
> +};
> +
> +#define UINT16_BITS       16
> +
> +#define HMAT_LB_LEVELS    (HMAT_LB_MEM_CACHE_3RD_LEVEL + 1)
> +#define HMAT_LB_TYPES     (HMAT_LB_DATA_WRITE_BANDWIDTH + 1)
such separate defines tend to become un-synced from enum definitions,
leading to out of bound access if enums are grown later.

one can use following trick

enum {
  ...,
  HMAT_LB_TYPES /* must be the last entry */
}


>  struct NodeInfo {
>      uint64_t node_mem;
>      struct HostMemoryBackend *node_memdev;
> @@ -28,6 +51,46 @@ struct NumaNodeMem {
>      uint64_t node_plugged_mem;
>  };
>  
> +struct HMAT_LB_Data {
> +    uint8_t     initiator;
> +    uint8_t     target;
> +    uint64_t    rawdata;
just 'data' is sufficient,
raw is confusing as it rises question what is non-raw data and there is no answer in this patch to it.

> +};
> +typedef struct HMAT_LB_Data HMAT_LB_Data;
> +
> +struct HMAT_LB_Info {
> +    /* Indicates it's memory or the specified level memory side cache. */
> +    uint8_t     hierarchy;
> +
> +    /* Present the type of data, access/read/write latency or bandwidth. */
> +    uint8_t     data_type;
> +
> +    /* The max compressed latency for calculating common latency base */
> +    uint64_t    max_entry_la;
> +
> +    /* The range bitmap of bandwidth for calculating common bandwidth base */
> +    uint64_t    range_bitmap_bw;
> +
> +    /* The common base unit for latencies */
> +    uint64_t    base_latency;
> +
> +    /* The common base unit for bandwidths */
> +    uint64_t    base_bandwidth;
> +
> +    /* Array to store the compressed latencies */
> +    uint16_t    *entry_latency;
> +
> +    /* Array to store the compressed latencies */
> +    uint16_t    *entry_bandwidth;
> +
> +    /* Array to store the latencies */
> +    GArray      *latency;
> +
> +    /* Array to store the bandwidthes */
> +    GArray      *bandwidth;
> +};
> +typedef struct HMAT_LB_Info HMAT_LB_Info;
> +
>  struct NumaState {
>      /* Number of NUMA nodes */
>      int num_nodes;
> @@ -40,11 +103,16 @@ struct NumaState {
>  
>      /* NUMA nodes information */
>      NodeInfo nodes[MAX_NODES];
> +
> +    /* NUMA nodes HMAT Locality Latency and Bandwidth Information */
> +    HMAT_LB_Info *hmat_lb[HMAT_LB_LEVELS][HMAT_LB_TYPES];
>  };
>  typedef struct NumaState NumaState;
>  
>  void set_numa_options(MachineState *ms, NumaOptions *object, Error **errp);
>  void parse_numa_opts(MachineState *ms);
> +void parse_numa_hmat_lb(NumaState *numa_state, NumaHmatLBOptions *node,
> +                        Error **errp);
>  void numa_complete_configuration(MachineState *ms);
>  void query_numa_node_mem(NumaNodeMem node_mem[], MachineState *ms);
>  extern QemuOptsList qemu_numa_opts;
> diff --git a/qapi/machine.json b/qapi/machine.json
> index 27d0e37534..da4d2c2cfc 100644
> --- a/qapi/machine.json
> +++ b/qapi/machine.json
> @@ -426,10 +426,12 @@
>  #
>  # @cpu: property based CPU(s) to node mapping (Since: 2.10)
>  #
> +# @hmat-lb: memory latency and bandwidth information (Since: 5.0)
> +#
>  # Since: 2.1
>  ##
>  { 'enum': 'NumaOptionsType',
> -  'data': [ 'node', 'dist', 'cpu' ] }
> +  'data': [ 'node', 'dist', 'cpu', 'hmat-lb' ] }
>  
>  ##
>  # @NumaOptions:
> @@ -444,7 +446,8 @@
>    'data': {
>      'node': 'NumaNodeOptions',
>      'dist': 'NumaDistOptions',
> -    'cpu': 'NumaCpuOptions' }}
> +    'cpu': 'NumaCpuOptions',
> +    'hmat-lb': 'NumaHmatLBOptions' }}
>  
>  ##
>  # @NumaNodeOptions:
> @@ -557,6 +560,93 @@
>     'base': 'CpuInstanceProperties',
>     'data' : {} }
>  
> +##
> +# @HmatLBMemoryHierarchy:
> +#
> +# The memory hierarchy in the System Locality Latency
> +# and Bandwidth Information Structure of HMAT (Heterogeneous
> +# Memory Attribute Table)
> +#
> +# For more information of @HmatLBMemoryHierarchy see
s/of/about/

> +# the chapter 5.2.27.4: Table 5-142: Field "Flags" of ACPI 6.3 spec.
> +#
> +# @memory: the structure represents the memory performance
> +#
> +# @first-level: first level memory of memory side cached memory
maybe it should say
  "first level of memory side cache"

and similar for other cache levels?

> +#
> +# @second-level: second level memory of memory side cached memory
> +#
> +# @third-level: third level memory of memory side cached memory
> +#
> +# Since: 5.0
> +##
> +{ 'enum': 'HmatLBMemoryHierarchy',
> +  'data': [ 'memory', 'first-level', 'second-level', 'third-level' ] }
> +
> +##
> +# @HmatLBDataType:
> +#
> +# Data type in the System Locality Latency
> +# and Bandwidth Information Structure of HMAT (Heterogeneous
> +# Memory Attribute Table)
> +#
> +# For more information of @HmatLBDataType see
s/of/about/

> +# the chapter 5.2.27.4: Table 5-142:  Field "Data Type" of ACPI 6.3 spec.
> +#
> +# @access-latency: access latency (nanoseconds)
> +#
> +# @read-latency: read latency (nanoseconds)
> +#
> +# @write-latency: write latency (nanoseconds)
> +#
> +# @access-bandwidth: access bandwidth (B/s)
> +#
> +# @read-bandwidth: read bandwidth (B/s)
> +#
> +# @write-bandwidth: write bandwidth (B/s)
> +#
> +# Since: 5.0
> +##
> +{ 'enum': 'HmatLBDataType',
> +  'data': [ 'access-latency', 'read-latency', 'write-latency',
> +            'access-bandwidth', 'read-bandwidth', 'write-bandwidth' ] }
> +
> +##
> +# @NumaHmatLBOptions:
> +#
> +# Set the system locality latency and bandwidth information
> +# between Initiator and Target proximity Domains.
> +#
> +# For more information of @NumaHmatLBOptions see
s/of/about/

> +# the chapter 5.2.27.4: Table 5-142 of ACPI 6.3 spec.
> +#
> +# @initiator: the Initiator Proximity Domain.
> +#
> +# @target: the Target Proximity Domain.
> +#
> +# @hierarchy: the Memory Hierarchy. Indicates the performance
> +#             of memory or side cache.
> +#
> +# @data-type: presents the type of data, access/read/write
> +#             latency or hit latency.
> +#
> +# @latency: the value of latency from @initiator to @target proximity domain,
> +#           the latency unit is "ns(nanosecond)".
> +#
> +# @bandwidth: the value of bandwidth between @initiator and @target proximity
> +#             domain, the bandwidth unit is "B(/s)".
> +#
> +# Since: 5.0
> +##
> +{ 'struct': 'NumaHmatLBOptions',
> +    'data': {
> +    'initiator': 'uint16',
> +    'target': 'uint16',
> +    'hierarchy': 'HmatLBMemoryHierarchy',
> +    'data-type': 'HmatLBDataType',
> +    '*latency': 'time',
> +    '*bandwidth': 'size' }}
> +
>  ##
>  # @HostMemPolicy:
>  #
> diff --git a/qemu-options.hx b/qemu-options.hx
> index cb0417b736..ec4ec37010 100644
> --- a/qemu-options.hx
> +++ b/qemu-options.hx
> @@ -168,16 +168,19 @@ DEF("numa", HAS_ARG, QEMU_OPTION_numa,
>      "-numa node[,mem=size][,cpus=firstcpu[-lastcpu]][,nodeid=node][,initiator=node]\n"
>      "-numa node[,memdev=id][,cpus=firstcpu[-lastcpu]][,nodeid=node][,initiator=node]\n"
>      "-numa dist,src=source,dst=destination,val=distance\n"
> -    "-numa cpu,node-id=node[,socket-id=x][,core-id=y][,thread-id=z]\n",
> +    "-numa cpu,node-id=node[,socket-id=x][,core-id=y][,thread-id=z]\n"
> +    "-numa hmat-lb,initiator=node,target=node,hierarchy=memory|first-level|second-level|third-level,data-type=access-latency|read-latency|write-latency[,latency=lat][,bandwidth=bw]\n",
>      QEMU_ARCH_ALL)
>  STEXI
>  @item -numa node[,mem=@var{size}][,cpus=@var{firstcpu}[-@var{lastcpu}]][,nodeid=@var{node}][,initiator=@var{initiator}]
>  @itemx -numa node[,memdev=@var{id}][,cpus=@var{firstcpu}[-@var{lastcpu}]][,nodeid=@var{node}][,initiator=@var{initiator}]
>  @itemx -numa dist,src=@var{source},dst=@var{destination},val=@var{distance}
>  @itemx -numa cpu,node-id=@var{node}[,socket-id=@var{x}][,core-id=@var{y}][,thread-id=@var{z}]
> +@itemx -numa hmat-lb,initiator=@var{node},target=@var{node},hierarchy=@var{hierarchy},data-type=@var{tpye}[,latency=@var{lat}][,bandwidth=@var{bw}]
>  @findex -numa
>  Define a NUMA node and assign RAM and VCPUs to it.
>  Set the NUMA distance from a source node to a destination node.
> +Set the ACPI Heterogeneous Memory Attributes for the given nodes.
>  
>  Legacy VCPU assignment uses @samp{cpus} option where
>  @var{firstcpu} and @var{lastcpu} are CPU indexes. Each
> @@ -256,6 +259,50 @@ specified resources, it just assigns existing resources to NUMA
>  nodes. This means that one still has to use the @option{-m},
>  @option{-smp} options to allocate RAM and VCPUs respectively.
>  
> +Use @samp{hmat-lb} to set System Locality Latency and Bandwidth Information
> +between initiator and target NUMA nodes in ACPI Heterogeneous Attribute Memory Table (HMAT).
> +Initiator NUMA node can create memory requests, usually it has one or more processors.
> +Target NUMA node contains addressable memory.
> +
> +In @samp{hmat-lb} option, @var{node} are NUMA node IDs. @var{hierarchy} is the memory
> +hierarchy of the target NUMA node: if @var{hierarchy} is 'memory', the structure
> +represents the memory performance; if @var{hierarchy} is 'first-level|second-level|third-level',
> +this structure represents aggregated performance of memory side caches for each domain.
> +@var{type} of 'data-type' is type of data represented by this structure instance:
> +if 'hierarchy' is 'memory', 'data-type' is 'access|read|write' latency or 'access|read|write'
> +bandwidth of the target memory; if 'hierarchy' is 'first-level|second-level|third-level',
> +'data-type' is 'access|read|write' hit latency or 'access|read|write' hit bandwidth of the
> +target memory side cache.
> +
> +@var{lat} of 'latency' is latency value, the possible value and units are
> +NUM[ns|us|ms] (nanosecond|microsecond|millisecond), the recommended unit is 'ns'.
> +@var{bw} is bandwidth value, the possible value and units are NUM[M|G|T], mean that
> +the bandwidth value are NUM MB/s, GB/s or TB/s. Note that for @var{lat} and @var{bw}
NUM byte (or MB/s, GB/s or TB/s depending on used suffix)
> +max NUM is 65534, if NUM is 0, means the corresponding latency or bandwidth information
not sure that 'max' statement is valid one
for once I can specify 65534M or equivalent suffix-less 68717379584
another thing is that range is a sliding window so there isn't a single max value
it's rather multiple values, where 0xFFFF-1 is shifted left or right within 64bit range.

I'd drop statement about max value.

> +is not provided. And if input numbers without any unit, the latency unit will be 'ns'
> +and the bandwidth will be MB/s.
should be B/s


> +For example, the following option assigns NUMA node 0 and 1. Node 0 has 2 cpus and
s/option assigns NUMA node 0 and 1/options describe 2 NUMA nodes/

> +a ram, node 1 has only a ram. The processors in node 0 access memory in node
> +0 with access-latency 5 nanoseconds, access-bandwidth is 200 MB/s;
> +The processors in NUMA node 0 access memory in NUMA node 1 with access-latency 10
> +nanoseconds, access-bandwidth is 100 MB/s.
> +@example
> +-machine hmat=on \
> +-m 2G \
> +-object memory-backend-ram,size=1G,id=m0 \
> +-object memory-backend-ram,size=1G,id=m1 \
> +-smp 2 \
> +-numa node,nodeid=0,memdev=m0 \
> +-numa node,nodeid=1,memdev=m1,initiator=0 \
> +-numa cpu,node-id=0,socket-id=0 \
> +-numa cpu,node-id=0,socket-id=1 \
> +-numa hmat-lb,initiator=0,target=0,hierarchy=memory,data-type=access-latency,latency=5ns \
> +-numa hmat-lb,initiator=0,target=0,hierarchy=memory,data-type=access-bandwidth,bandwidth=200M \
> +-numa hmat-lb,initiator=0,target=1,hierarchy=memory,data-type=access-latency,latency=10ns \
> +-numa hmat-lb,initiator=0,target=1,hierarchy=memory,data-type=access-bandwidth,bandwidth=100M
> +@end example
> +
>  ETEXI
>  
>  DEF("add-fd", HAS_ARG, QEMU_OPTION_add_fd,