libtopo/topo.c

#include <pthread.h>
#include <stdio.h>
#include <stdlib.h>
#include <errno.h>
#include <string.h>

#include <sys/types.h>
#include <sys/_cpuset.h>
#include <sys/cpuset.h>
#include <sys/sysctl.h>
#include <stdatomic.h>


#include "libxml/parser.h"
#include "libxml/tree.h"
#include "libxml/xmlmemory.h"
#include "libxml/xmlversion.h"

#include "topo.h"
#include "topop.h"

#define TOPO_INVALID (-1)

#define TOPO_CACHE_L2 (2)
#define TOPO_CACHE_L3 (3)
#define TOPO_CACHE_NONE (0)

#define TOPO_FLAG_NONE (0)
#define TOPO_FLAG_NUMA (0x4)
#define TOPO_FLAG_HTT (0x1)
#define TOPO_FLAG_SMT (0x2)
#define TOPO_FLAG_THREAD (TOPO_FLAG_HTT | TOPO_FLAG_SMT)

struct topo_obj {
	int cache_level;
	cpuset_t mask;
	int flags;
    int level;
	int num_children;
    struct topo_obj * parent;
	struct topo_obj * children[TOPO_MAX_CHILDREN];
};

static struct topo_desc g_default_desc = {0};
static _Atomic(int) initialized = 0;
static volatile int init_rc = 0;

static void 
mask_to_cpuset(const char * mask, cpuset_t * cset)
{
    char * last;
    char * copy = strdup(mask);
    char * first = strtok_r(copy, ",", &last);
    char tmp[2] = {0};

    int qword = 0;
    while (first != NULL) {
        int shift = 0;
        for (int i = (int)strlen(first) - 1; i >= 0; i--) {
            tmp[0] = first[i];
            int val = (int)strtol(tmp, NULL, 16);
            int count = 0;
            while (val > 0) {
                int bit = val & 1;
                if (bit == 1) {
                    CPU_SET(qword * 64 + shift + count, cset);
                    //printf("Setting bit %d to 1!\n", qword * 64 + shift + count);
                }
                val = val >> 1;
                count++;
            }
            shift += 4;
        }

        first = strtok_r(NULL, ",", &last);
        qword++;
    }

    free(copy);
}

static void 
tobj_preorder_dump(struct topo_obj * root, int indent) {
    cpuset_t tmpset;
    CPU_COPY(&root->mask, &tmpset);
    const char * flag_str;
    
    switch(root->flags) {
        case TOPO_FLAG_NUMA:
            flag_str = "NUMA";
            break;
        case TOPO_FLAG_THREAD:
            flag_str = "THREAD";
            break;
        case TOPO_FLAG_HTT:
            flag_str = "HTT";
            break;
        case TOPO_FLAG_SMT:
            flag_str = "SMT";
            break;
        default:
            flag_str = "NONE";
    }

    printf("%*sgroup level: %d, cache-level: %d, flag: %s, cores: ", indent * 4, "", root->level, root->cache_level, flag_str);
    while (CPU_FFS(&tmpset) != 0) {
        int u = (int)CPU_FFS(&tmpset);
        printf("%d ", u - 1);
        CPU_CLR(u -1 , &tmpset);
    }
    printf("\n");
    for (int i = 0; i < root->num_children; i++) {
        tobj_preorder_dump(root->children[i], indent + 1);
    }
}

static xmlAttr *
xml_find_attr(xmlNode * root, const char * attr_name)
{
    xmlAttr * attr = root->properties;
    while (attr != NULL) {
        if (strcmp((const char *)attr->name, attr_name) == 0) {
            break;
        }
        attr = attr->next;
    }

    return attr;
}

static void
tobj_free_root(struct topo_obj * root)
{
    if (root == NULL) {
        return;
    }

    if (root->num_children > 0) {
        for (int i = 0; i < root->num_children; i++) {
            tobj_free_root(root->children[i]);
        }
    }

    free(root);
}

static int
tobj_populate_root(xmlNode * root, struct topo_obj * parent, struct topo_obj ** out)
{
    int rc = 0;

    if (root->type == XML_ELEMENT_NODE) {
        if (strcmp((const char*)root->name, "group") == 0) {
            // this is a group, then check cache-level
            xmlAttr * attr = xml_find_attr(root, "cache-level");

            if (attr == NULL) {
                fprintf(stderr, "libtopo: could not find attr cache-level in group.\n");
                return -1;
            }

            xmlAttr * lvl_attr = xml_find_attr(root, "level");
            if (lvl_attr == NULL) {
                fprintf(stderr, "libtopo: could not find attr level in group.\n");
                return -1;
            }

            int level = atoi((const char *)lvl_attr->children->content);
            int cache_level = atoi((const char *)attr->children->content);

            struct topo_obj * tobj = malloc(sizeof(struct topo_obj));
            tobj->cache_level = cache_level;
            tobj->num_children = 0;
            tobj->level = level;
            tobj->parent = parent;
            tobj->flags = TOPO_FLAG_NONE;
            CPU_ZERO(&tobj->mask);
            if (level == 1) {
                if (*out != NULL) {
                    fprintf(stderr, "libtopo: multiple level 1 group detected.\n");
                    free(tobj);
                    return -1;
                }
                *out = tobj;
            } else {
                if (parent == NULL) {
                    fprintf(stderr, "libtopo: level 1 group does not appear to be the outermost\n");
                    free(tobj);
                    return -1;
                }
                parent->children[parent->num_children] = tobj;
                parent->num_children++;
            }

            parent = tobj;
        } else if (strcmp((const char*)root->name, "cpu") == 0) {
            xmlAttr * attr = xml_find_attr(root, "mask");

            if (attr == NULL) {
                fprintf(stderr, "libtopo: could not find attr mask in cpu.\n");
                return -1;
            }

            mask_to_cpuset((const char *)attr->children->content, &parent->mask);
        } else if (strcmp((const char *)root->name, "flag") == 0) {
            xmlAttr * attr = xml_find_attr(root, "name");

            if (attr == NULL) {
                fprintf(stderr, "libtopo: could not find attr name in flag.\n");
                return -1;
            }

            if (strcmp((const char *)attr->children->content, "THREAD") == 0) {
                parent->flags |= TOPO_FLAG_THREAD;
            } else if (strcmp((const char *)attr->children->content, "SMT") == 0) {
                parent->flags |= TOPO_FLAG_SMT;
            } else if (strcmp((const char *)attr->children->content, "HTT") == 0) {
                parent->flags |= TOPO_FLAG_HTT;
            } else if (strcmp((const char *)attr->children->content, "NODE") == 0) {
                parent->flags |= TOPO_FLAG_NUMA;
            }
        }

        xmlNode * child = root->children;
        while (child != NULL) {
            if ((rc = tobj_populate_root(child, parent, out)) != 0) {
                break;
            }
            child = child->next;
        }
    }
    return rc;
}

static int
tobj_populate(struct topo_obj **out)
{
    int rc;
    size_t sz;

    LIBXML_TEST_VERSION;

    rc = sysctlbyname("kern.sched.topology_spec", NULL, &sz, NULL, 0);
    if (rc != 0) {
        return rc;
    }

    char * buf = malloc(sz);
    if (buf == NULL) {
        errno = ENOMEM;
        return -1;
    }

    rc = sysctlbyname("kern.sched.topology_spec", buf, &sz, NULL, 0);
    if (rc != 0) {
        return rc;
    }

    //printf("xml:\n%s",buf);
    
    xmlDoc * doc = xmlReadMemory(buf, (int)sz, NULL, NULL, 0);
    if (doc == NULL) {
        errno = EINVAL;
        return -1;
    }
    rc = tobj_populate_root(xmlDocGetRootElement(doc), NULL, out) != 0;
    free(buf);
    xmlFreeDoc(doc);

    if (rc != 0) {
        tobj_free_root(*out);
        errno = EINVAL;
        return rc;
    }

    return rc;
}

static struct topo_obj *
tobj_find_leaf_node_by_mask(struct topo_obj * root, cpuset_t * set)
{
    if (!CPU_SUBSET(&root->mask, set)) {
        return NULL;
    }

    struct topo_obj * best_fit = root;
    for (int i = 0; i < root->num_children; i++) {
        struct topo_obj * child_fit = tobj_find_leaf_node_by_mask(root->children[i], set);
        if (child_fit != NULL) {
            best_fit = child_fit;
            break;
        }
    }

    return best_fit;
}

__attribute__((unused))
static struct topo_obj *
tobj_find_node_by_flag(struct topo_obj * root, int flag)
{
    if (root->flags & flag) {
        return root;
    }

    struct topo_obj * result;
    for (int i = 0; i < root->num_children; i++) {
        result = tobj_find_node_by_flag(root->children[i], flag);
        if (result != NULL) {
            return result;
        }
    }

    return NULL;
}

static inline int
tobj_is_node_numa(struct topo_obj * obj)
{
    return obj->level == 1 && obj->cache_level == 0;
}

static int
tobj_core_to_numa(struct topo_obj * root, int coreid)
{
    cpuset_t set;
    CPU_ZERO(&set);
    CPU_SET(coreid, &set);

    struct topo_obj * obj = tobj_find_leaf_node_by_mask(root, &set);
    if (obj == NULL) {
        return TOPO_INVALID;
    }
    
    while(obj->parent != NULL) {
        if (tobj_is_node_numa(obj->parent)) {
            for (int i = 0; i < obj->parent->num_children; i++) {
                if (obj->parent->children[i] == obj) {
                    return i;
                }
            }
        }
        obj = obj->parent;
    }

    return 0;
}

static int
tobj_get_num_numa(struct topo_obj * root)
{
    while (root->parent != NULL) {
        root = root->parent;
    }

    if (tobj_is_node_numa(root) && root->num_children > 0) {
        return root->num_children;
    } else {
        return 1;
    }
}

static int
tobj_get_num_core(struct topo_obj * root)
{
    return CPU_COUNT(&root->mask);
}

static void
desc_destroy(struct topo_desc * desc)
{
    tobj_free_root(desc->root);
}

static int
desc_init(struct topo_desc * desc, int verbose)
{
    desc->root = NULL;
    int rc = tobj_populate(&desc->root);

    if (rc == 0) {
        desc->num_core = tobj_get_num_core(desc->root);
        desc->num_numa = tobj_get_num_numa(desc->root);
        if (desc->num_core > TOPO_MAX_CHILDREN) {
            fprintf(stderr, "libtopo: too many cores - %d\n", desc->num_core);
            rc = ENOMEM;
        } else {  
            for (int i = 0; i < desc->num_core; i++) {
                desc->core_to_numa_lookup[i] = tobj_core_to_numa(desc->root, i);
            }
            if (verbose) {
                fprintf(stdout, "libtopo: %d cores, %d numa domains detected.\n", desc->num_core, desc->num_numa);
                for (int i = 0; i < desc->num_core; i++) {
                    fprintf(stdout, "libtopo: core #%d @ numa domain %d.\n", i, desc->core_to_numa_lookup[i]);
                }
                fprintf(stdout, "libtopo: parsed tree:\n");
                tobj_preorder_dump(desc->root, 1);
            }
        }
    } else {
        fprintf(stderr, "libtopo: failed to parse topo: %d\n", errno);
    }

    if (rc == 0) {
        rc = topo_ts_init(desc, verbose);
    }

    if (rc != 0) {
        if (desc->root != NULL) {
            tobj_free_root(desc->root);
        }
    }

    return rc;
}

//
// public APIs
//
int
topo_num_core(void)
{
    return g_default_desc.num_core;
}

int
topo_num_numa(void)
{
    return g_default_desc.num_numa;
}

int
topo_core_to_numa(int coreid)
{
    return g_default_desc.core_to_numa_lookup[coreid];
}

void
topo_destroy(void)
{
    desc_destroy(&g_default_desc);
}

uint64_t
topo_uptime_ns(void)
{
    return topo_desc_uptime_ns(&g_default_desc);
}

int
topo_init(int verbose)
{
    int expected = 0;
    if (atomic_compare_exchange_strong(&initialized, &expected, 2)) {
        init_rc = desc_init(&g_default_desc, verbose);
        atomic_store(&initialized, 1);
    } else {
        while(atomic_load(&initialized) != 1) {} // wait for init
        if (verbose) {
            fprintf(stdout, "libtopo: already initialized with rc = %d\n", init_rc);
        }
    }
    return init_rc;
}