source: libcfa/src/device/cpu.cfa @ b6f2e7ab

//
// Cforall Version 1.0.0 Copyright (C) 2021 University of Waterloo
//
// The contents of this file are covered under the licence agreement in the
// file "LICENCE" distributed with Cforall.
//
// topology.cfa -- read the data structure
//
// Author           : Thierry Delisle
// Created On       : Thu Jun 10 16:13:07 2021
// Last Modified By :
// Last Modified On :
// Update Count     :
//

#include "device/cpu.hfa"

#include <math.hfa>
#include <stdlib.hfa>

#include <errno.h>
#include <stdio.h>
#include <string.h>
#include <unistd.h>

extern "C" {
        #include <dirent.h>
        #include <sys/types.h>
        #include <sys/stat.h>
        #include <fcntl.h>
}

#include "bits/defs.hfa"
#include "algorithms/range_iterator.hfa"

// search a string for character 'character' but looking atmost at len
// chars
static const char * strnchr(const char * str, int character, size_t len) {
        return (const char *)memchr(str, character, strnlen(str, len));
}

// Check if have string matches the want string
// ignoring any characters that are longer than the want string
static bool strmatch(const char * want, char * have) {
        size_t w = strlen(want);
        return strncmp(want, have, w) == 0;
}

typedef const char * idx_range_t;

// read the value of a string and evaluate it
// get the end pointer and make sure it is all evaluated
static unsigned read_value(idx_range_t map, size_t len, const char ** end) {
        unsigned long val = strtoul(map, (char**)end, 10);
        /* paranoid */ __attribute__((unused)) size_t read = (*end - map);
        /* paranoid */ verifyf(read <= len, "String '%s' passed with inconsistent length %zu", map, len);
        /* paranoid */ verifyf(read == len, "String %.*s not entirely a number, %zu chars left", (int)len, map, len - read);
        return val;
}

// Evaluate the width of a comma seperated list of idx
// for example 'A-B,C-D,E,F' has a width of '(B-A) + (D-C) + 1 + 1'
// Also has an (non-optional) end ptr like strtoul and friends
//
// FIXME : the current implementation only supports 1 comma
static unsigned read_width(idx_range_t map, size_t len, const char ** end) {
        // Do we have a comma
        const char * comma = strnchr(map, ',', len);
        if(comma != 0p) {
                // We do! recurse and sum the widths
                const char * _;
                size_t split = comma - map;
                unsigned lhs = read_width(map, split, &_);
                unsigned rhs = read_width(comma + 1, len - split - 1, end);
                return lhs + rhs;
        }

        // No commas, check for a range
        const char * dash = strnchr(map, '-', len);
        if(dash != 0p) {
                const char * _;
                size_t split = dash - map;
                unsigned lhs = read_value(map, split, &_);
                unsigned rhs = read_value(dash + 1, len - split - 1, end);
                return rhs - lhs + 1;
        }

        // No range, no comma, just a single value
        // It's width is 1 and we can consume everything
        /* paranoid */ verifyf( ({strtoul(map, (char**)end, 10); *end == (map + len); }), "Value in range '%.*s' not a number", (int)len, map);
        *end = map + len;
        return 1;
}

// go through a directory calling fn on each file
static int iterate_dir( const char * path, void (*fn)(struct dirent * ent) ) {
        // open the directory
        DIR *dir = opendir(path);
        if(dir == 0p) { return ENOTDIR; }

        // call fn for each
        struct dirent * ent;
        while ((ent = readdir(dir)) != 0p) {
                fn( ent );
        }

        // no longer need this
        closedir(dir);
        return 0;
}

// count the number of directories with the specified prefix
// the directories counted have the form '[prefix]N' where prefix is the parameter
// and N is an base 10 integer.
static int count_prefix_dirs(const char * path, const char * prefix) {
        // read the directory and find the cpu count
        // and make sure everything is as expected
        int max = -1;
        int count = 0;
        void lambda(struct dirent * ent) {
                // were are looking for prefixX, where X is a number
                // check that it starts with 'cpu
                char * s = strstr(ent->d_name, prefix);
                if(s == 0p) { return; }
                if(s != ent->d_name) { return; }

                // check that the next part is a number
                s += strlen(prefix);
                char * end;
                long int val = strtol(s, &end, 10);
                if(*end != '\0' || val < 0) { return; }

                // check that it's a directory
                if(ent->d_type != DT_DIR) { return; }

                // it's a match!
                max = max(val, max);
                count++;
        }
        int ret = iterate_dir(path, lambda);
        if(ret == ENOTDIR) return 0;

        /* paranoid */ verifyf(count == max + 1, "Inconsistent %s count, counted %d, but max %s was %d", prefix, count, prefix, (int)max);

        return count;
}

// Count number of cpus in the system
static [int, const char *] count_cpus(void) {
        const char * fpath = "/sys/devices/system/cpu/online";
        int fd = open(fpath, 0, O_RDONLY);
        /* paranoid */ verifyf(fd >= 0, "Could not open file %s", fpath);

        char buff[128];
        ssize_t r = read(fd, buff, 128);
        /* paranoid */ verifyf(r > 0, "Could not read file %s", fpath);
        /* paranoid */ verify( buff[r-1] == '\n' );
        buff[r-1] = '\0';

        /* paranoid */ __attribute__((unused)) int ret =
        close(fd);
        /* paranoid */ verifyf(ret == 0, "Could not close file %s", fpath);

        const char * _;
        return [read_width(buff, r - 1, &_), strndup(buff, r - 1)];
}

// Count number of cache *indexes* in the system
// cache indexes are distinct from cache level as Data or Instruction cache
// can share a level but not an index
// PITFALL: assumes all cpus have the same indexes as cpu0
static int count_cache_indexes(void) {
        return count_prefix_dirs("/sys/devices/system/cpu/cpu0/cache", "index");
}


// read information about a spcficic cache index/cpu file into the output buffer
static size_t read_cpuidxinfo_into(unsigned cpu, unsigned idx, const char * file, char * out, size_t out_len) {
        // Pick the file we want and read it
        char buf[128];
        /* paranoid */ __attribute__((unused)) int len =
        snprintf(buf, 128, "/sys/devices/system/cpu/cpu%u/cache/index%u/%s", cpu, idx, file);
        /* paranoid */ verifyf(len > 0, "Could not generate '%s' filename for cpu %u, index %u", file, cpu, idx);

        int fd = open(buf, 0, O_RDONLY);
        /* paranoid */ verifyf(fd > 0, "Could not open file '%s'", buf);

        ssize_t r = read(fd, out, out_len);
        /* paranoid */ verifyf(r > 0, "Could not read file '%s'", buf);

        /* paranoid */ __attribute__((unused)) int ret =
        close(fd);
        /* paranoid */ verifyf(ret == 0, "Could not close file '%s'", buf);
        return r;
}

// Iterate over the cache indexes of a given cpu
typedef void (*handle_func_t)(unsigned idx, unsigned char level, idx_range_t range, size_t len);
static void foreach_cacheidx(unsigned cpu, unsigned idxs, handle_func_t handle) {
        for(i; idxs) {
                unsigned idx = idxs - 1 - i;
                char buf[32];

                // Type says what kind of cache this is,
                // Options are: Unified, Data, Instruction
                read_cpuidxinfo_into(cpu, idx, "type", buf, 32);
                if((!strmatch("Unified", buf)) && (!strmatch("Data", buf))) {
                        // We don't care about instruction caches
                        continue;
                }

                // Level is the cache level: higher means bigger and slower
                read_cpuidxinfo_into(cpu, idx, "level", buf, 32);
                char * end;
                unsigned long level = strtoul(buf, &end, 10);
                /* paranoid */ verifyf(level <= 250, "Cpu %u has more than 250 levels of cache, this is not supported", cpu);

                // shared_cpu_list is a range of cpus that share this particular cache
                size_t n = read_cpuidxinfo_into(cpu, idx, "shared_cpu_list", buf, 32);
                /* paranoid */ verify( buf[n-1] == '\n' );
                buf[n-1] = '\0';

                // Simply call the functor
                handle(idx, level, buf, n - 1);
        }
}


struct raw_cache_instance {
        idx_range_t range;      // A text description of the cpus covered
        unsigned width;         // The number of cpus covered
        unsigned char level;    // the cache level
        // FIXME add at least size and type
};

static void  ?{}(raw_cache_instance & this) { this.range = 0p;}
static void ^?{}(raw_cache_instance & this) { free(this.range);}

// Returns a 2D array of instances of size [cpu count][cache levels]
// where cache level doesn't include instruction caches
raw_cache_instance ** build_raw_cache_table(unsigned cpus_c, idx_range_t cpus, unsigned idxs, unsigned cache_levels) {
        raw_cache_instance ** raw = alloc(cpus_c, '\0'`fill);

        RangeIter rc = { cpus };
        while(moveNext(rc)) {
                unsigned i = rc.com;
                raw[i] = alloc(cache_levels);
                void addcache(unsigned fidx, unsigned char level, idx_range_t range, size_t len) {
                        /* paranoid */ verifyf(level <= cache_levels, "Unexpected cache level %d on cpu %u index %u", (int)level, i, fidx);

                        unsigned idx = cache_levels - level;
                        raw_cache_instance & r = raw[i][idx];
                        r.range = strndup(range, len);
                        r.level = level;
                        const char * end;
                        r.width = read_width(range, len, &end);
                }
                foreach_cacheidx(i, idxs, addcache);
        }

        return raw;
}

struct llc_map_t {
        raw_cache_instance * raw;
        unsigned count;
        unsigned start;
};

// returns an allocate list of all the different distinct last level caches
static [*llc_map_t, size_t cnt] distinct_llcs(idx_range_t cpus, unsigned llc_idx, raw_cache_instance ** raw) {
        // Allocate at least one element
        llc_map_t* ranges = alloc();
        size_t range_cnt = 1;

        RangeIter rc = { cpus };
        __attribute__((unused)) bool ret =
        moveNext(rc);
        /* paranoid */ verify( ret );
        /* paranoid */ verify( rc.com >= 0 );

        // Initialize with element 0
        ranges->raw = &raw[rc.com][llc_idx];
        ranges->count = 0;
        ranges->start = -1u;

        // Go over all other cpus
        CPU_LOOP: while(moveNext(rc)) {
                unsigned i = rc.com;
                // Check if the range is already there
                raw_cache_instance * candidate = &raw[i][llc_idx];
                for(j; range_cnt) {
                        llc_map_t & exist = ranges[j];
                        // If the range is already there just jump to the next cpu
                        if(0 == strcmp(candidate->range, exist.raw->range)) continue CPU_LOOP;
                }

                // The range wasn't there, added to the list
                ranges = alloc(range_cnt + 1, ranges`realloc);
                ranges[range_cnt].raw = candidate;
                ranges[range_cnt].count = 0;
                ranges[range_cnt].start = -1u;
                range_cnt++;
        }

        // return what we have
        return [ranges, range_cnt];
}

struct cpu_pairing_t {
        unsigned cpu;
        unsigned id;
};

int ?<?( cpu_pairing_t lhs, cpu_pairing_t rhs ) {
        return lhs.id < rhs.id;
}

static [[]cpu_pairing_t] get_cpu_pairings(unsigned cpus_c, idx_range_t cpus, raw_cache_instance ** raw, llc_map_t * maps, size_t map_cnt) {
        cpu_pairing_t * pairings = alloc(cpus_c);

        RangeIter rc = { cpus };
        CPU_LOOP: while(moveNext(rc)) {
                unsigned i = rc.com;
                pairings[i].cpu = i;
                idx_range_t want = raw[i][0].range;
                MAP_LOOP: for(j; map_cnt) {
                        if(0 != strcmp(want, maps[j].raw->range)) continue MAP_LOOP;

                        pairings[i].id = j;
                        continue CPU_LOOP;
                }

                /* paranoid */ verifyf( false, "Cpu %u map doesn't match", i );
        }

        return pairings;
}

#include <fstream.hfa>

extern "C" {
        void __cfaabi_device_startup( void ) {
                int cpus_c;
                const char * cpus;
                [cpus_c, cpus] = count_cpus();
                #if defined(__CFA_WITH_VERIFY__)
                // Verify that the mapping is self consistant.
                {
                        RangeIter rc = { cpus };
                        while(moveNext(rc)) {
                                unsigned i = rc.com;
                                verify(cpus_c > i);
                        }
                }
                #endif

                int idxs = count_cache_indexes();

                // Do we actually have a cache?
                if(idxs == 0) {
                        // if not just fake the data structure, it makes things easier.
                        cpu_info.hthrd_count = cpus_c;
                        cpu_info.llc_count = 0;
                        struct cpu_map_entry_t * entries = alloc(cpu_info.hthrd_count);
                        for(i; cpu_info.hthrd_count) {
                                entries[i].self  = i;
                                entries[i].start = 0;
                                entries[i].count = cpu_info.hthrd_count;
                                entries[i].cache = 0;
                        }
                        cpu_info.llc_map = entries;
                        return;
                }

                // Count actual cache levels
                unsigned cache_levels = 0;
                unsigned llc = 0;

                unsigned char prev = -1u;
                void first(unsigned idx, unsigned char level, const char * map, size_t len) {
                        /* paranoid */ verifyf(level < prev, "Index %u of cpu 0 has cache levels out of order: %u then %u", idx, (unsigned)prev, (unsigned)level);
                        llc = max(llc, level);
                        prev = level;
                        cache_levels++;
                }
                foreach_cacheidx(0, idxs, first);

                // Read in raw data
                raw_cache_instance ** raw = build_raw_cache_table(cpus_c, cpus, idxs, cache_levels);

                // Find number of distinct cache instances
                llc_map_t * maps;
                size_t map_cnt;
                [maps, map_cnt] =  distinct_llcs(cpus, cache_levels - llc, raw);

                #if defined(__CFA_WITH_VERIFY__)
                // Verify that the caches cover the all the cpus
                {
                        unsigned width1 = 0;
                        unsigned width2 = 0;
                        for(i; map_cnt) {
                                const char * _;
                                width1 += read_width(maps[i].raw->range, strlen(maps[i].raw->range), &_);
                                width2 += maps[i].raw->width;
                        }
                        verify(width1 == cpus_c);
                        verify(width2 == cpus_c);
                }
                #endif

                // Get mappings from cpu to cache instance
                cpu_pairing_t * pairings = get_cpu_pairings(cpus_c, cpus, raw, maps, map_cnt);

                // Sort by cache instance
                qsort(pairings, cpus_c);

                {
                        unsigned it = 0;
                        RangeIter rc = { cpus };
                        while(moveNext(rc)) {
                                unsigned i = rc.com;
                                unsigned llc_id = pairings[i].id;
                                if(maps[llc_id].start == -1u) {
                                        maps[llc_id].start = it;
                                        it += maps[llc_id].raw->width;
                                        /* paranoid */ verify(maps[llc_id].start < it);
                                        /* paranoid */ verify(it != -1u);
                                }
                        }
                        /* paranoid */ verify(it == cpus_c);
                }

                // From the mappings build the actual cpu map we want
                struct cpu_map_entry_t * entries = alloc(cpus_c);
                for(i; cpus_c) { entries[i].count = 0; }

                RangeIter rc = { cpus };
                while(moveNext(rc)) {
                        unsigned i = rc.com;
                        /* paranoid */ verify(pairings[i].id < map_cnt);
                        unsigned c = pairings[i].cpu;
                        unsigned llc_id = pairings[i].id;
                        unsigned start = maps[llc_id].start;
                        entries[c].count = maps[llc_id].raw->width;
                        entries[c].start = start;
                        entries[c].self  = start + (maps[llc_id].count++);
                        entries[c].cache = llc_id;
                }

                // get rid of the temporary data
                free(maps);
                free(pairings);

                for(i; cpus_c) {
                        if( raw[i] ) for(j; cache_levels) {
                                ^(raw[i][j]){};
                        }
                        free(raw[i]);
                }
                free(raw);

                cpu_info.llc_map = entries;
                cpu_info.hthrd_count = cpus_c;
                cpu_info.llc_count = map_cnt;
        }

        void __cfaabi_device_shutdown( void ) {
                free(cpu_info.llc_map);
        }
}

libcfa_public cpu_info_t cpu_info;