Re: [PATCH 4/6] plugins/cache: Supported multicore cache modelling

["Alex Bennée" <alex.bennee@linaro.org>]

Mahmoud Mandour <ma.mandourr@gmail.com> writes:

> Multicore L1 cache modelling is introduced and is supported for both
> full system emulation and linux-user.
>
> For full-system emulation, L1 icache and dcache are maintained for each
> available core, since this information is exposed to the plugin through
> `qemu_plugin_n_vcpus()`.
>
> For linux-user, a static number of cores is assumed (default 1 core, and
> can be provided as a plugin argument `cores=N`). Every memory access
> goes through one of these caches, this approach is taken as it's
> somewhat akin to what happens on real setup, where a program that
> dispatches more threads than the available cores, they'll thrash
> each other
>
> Signed-off-by: Mahmoud Mandour <ma.mandourr@gmail.com>
> ---
>  contrib/plugins/cache.c | 156 ++++++++++++++++++++++++++++++----------
>  1 file changed, 117 insertions(+), 39 deletions(-)
>
> diff --git a/contrib/plugins/cache.c b/contrib/plugins/cache.c
> index a452aba01c..60f7be208b 100644
> --- a/contrib/plugins/cache.c
> +++ b/contrib/plugins/cache.c
> @@ -23,11 +23,6 @@ static GRand *rng;
>  static int limit;
>  static bool sys;
>  
> -static uint64_t dmem_accesses;
> -static uint64_t dmisses;
> -
> -static uint64_t imem_accesses;
> -static uint64_t imisses;
>  
>  enum EvictionPolicy {
>      LRU,
> @@ -90,13 +85,22 @@ typedef struct {
>      uint64_t imisses;
>  } InsnData;
>  
> +typedef struct {
> +    uint64_t dmem_accesses;
> +    uint64_t dmisses;
> +    uint64_t imem_accesses;
> +    uint64_t imisses;
> +} CoreStats;
> +
>  void (*update_hit)(Cache *cache, int set, int blk);
>  void (*update_miss)(Cache *cache, int set, int blk);
>  
>  void (*metadata_init)(Cache *cache);
>  void (*metadata_destroy)(Cache *cache);
>  
> -Cache *dcache, *icache;
> +static int cores;
> +CoreStats *stats;
> +Cache **dcaches, **icaches;
>  
>  static int pow_of_two(int num)
>  {
> @@ -233,10 +237,6 @@ static bool bad_cache_params(int blksize, int assoc, int cachesize)
>  
>  static Cache *cache_init(int blksize, int assoc, int cachesize)
>  {
> -    if (bad_cache_params(blksize, assoc, cachesize)) {
> -        return NULL;
> -    }
> -
>      Cache *cache;
>      int i;
>      uint64_t blk_mask;
> @@ -263,6 +263,24 @@ static Cache *cache_init(int blksize, int assoc, int cachesize)
>      return cache;
>  }
>  
> +static Cache **caches_init(int blksize, int assoc, int cachesize)
> +{
> +    Cache **caches;
> +    int i;
> +
> +    if (bad_cache_params(blksize, assoc, cachesize)) {
> +        return NULL;
> +    }
> +
> +    caches = g_new(Cache *, cores);
> +
> +    for (i = 0; i < cores; i++) {
> +        caches[i] = cache_init(blksize, assoc, cachesize);
> +    }
> +
> +    return caches;
> +}
> +
>  static int get_invalid_block(Cache *cache, uint64_t set)
>  {
>      int i;
> @@ -353,6 +371,7 @@ static void vcpu_mem_access(unsigned int vcpu_index, qemu_plugin_meminfo_t info,
>  {
>      uint64_t effective_addr;
>      struct qemu_plugin_hwaddr *hwaddr;
> +    int cache_idx;
>      InsnData *insn;
>  
>      hwaddr = qemu_plugin_get_hwaddr(info, vaddr);
> @@ -363,17 +382,24 @@ static void vcpu_mem_access(unsigned int vcpu_index, qemu_plugin_meminfo_t info,
>      effective_addr = hwaddr ? qemu_plugin_hwaddr_phys_addr(hwaddr) : vaddr;
>  
>      g_mutex_lock(&mtx);
> -    if (dcache == NULL) {
> +    cache_idx = vcpu_index % cores;
> +    if (dcaches[cache_idx] == NULL) {
>          g_mutex_unlock(&mtx);
>          return;
>      }

I was hoping we could avoid this with some careful atomic/memory barrier
use given dcaches is setup on init. However it's the exit case that gets
us which further increases my desire to fix the exit race in linux-user
and make sure callbacks don't occur in this case.

>  
> -    if (!access_cache(dcache, effective_addr)) {
> +    if (!access_cache(dcaches[cache_idx], effective_addr)) {
>          insn = (InsnData *) userdata;
>          insn->dmisses++;
> -        dmisses++;
> +        if (cores > 1) {
> +            stats[cores].dmisses++;
> +        }
> +        stats[cache_idx].dmisses++;
> +    }
> +    if (cores > 1) {
> +        stats[cores].dmem_accesses++;
>      }
> -    dmem_accesses++;
> +    stats[cache_idx].dmem_accesses++;
>      g_mutex_unlock(&mtx);
>  }
>  
> @@ -381,21 +407,29 @@ static void vcpu_insn_exec(unsigned int vcpu_index, void *userdata)
>  {
>      uint64_t insn_addr;
>      InsnData *insn;
> +    uint64_t cache_idx;
>  
>      g_mutex_lock(&mtx);
>      insn_addr = ((InsnData *) userdata)->addr;
>  
> -    if (icache == NULL) {
> +    cache_idx = vcpu_index % cores;
> +    if (icaches[cache_idx] == NULL) {
>          g_mutex_unlock(&mtx);
>          return;
>      }
>  
> -    if (!access_cache(icache, insn_addr)) {
> +    if (!access_cache(icaches[cache_idx], insn_addr)) {
>          insn = (InsnData *) userdata;
>          insn->imisses++;
> -        imisses++;
> +        if (cores > 1) {
> +            stats[cores].imisses++;
> +        }
> +        stats[cache_idx].imisses++;
>      }
> -    imem_accesses++;
> +    if (cores > 1) {
> +        stats[cores].imem_accesses++;
> +    }

I'd much rather avoid summing the stats during execution than at the end
when we can add things simply.

> +    stats[cache_idx].imem_accesses++;
>      g_mutex_unlock(&mtx);
>  }
>  
> @@ -475,6 +509,22 @@ static void cache_free(Cache *cache)
>      g_free(cache);
>  }
>  
> +/*
> + * caches_free(): free an array of Cache structs.
> + * @caches: caches to free
> + *
> + * Calling cache_free for each cache in @caches, and then NULL-ify them so that
> + * we mark them as freed, other async callbacks can check to see whether a cache
> + * is freed already or not by checking against NULL.
> + */
> +static void caches_free(Cache **caches)
> +{
> +    for (int i = 0; i < cores; i++) {
> +        cache_free(caches[i]);
> +        caches[i] = NULL;
> +    }
> +}
> +
>  static int dcmp(gconstpointer a, gconstpointer b)
>  {
>      InsnData *insn_a = (InsnData *) a;
> @@ -493,19 +543,38 @@ static int icmp(gconstpointer a, gconstpointer b)
>  
>  static void log_stats()
>  {
> -    g_autoptr(GString) rep = g_string_new("");
> -    g_string_append_printf(rep,
> -        "Data accesses: %lu, Misses: %lu\nMiss rate: %lf%%\n\n",
> -        dmem_accesses,
> -        dmisses,
> -        ((double) dmisses / (double) dmem_accesses) * 100.0);
> -
> -    g_string_append_printf(rep,
> -        "Instruction accesses: %lu, Misses: %lu\nMiss rate: %lf%%\n\n",
> -        imem_accesses,
> -        imisses,
> -        ((double) imisses / (double) imem_accesses) * 100.0);
> +    int i, iters;
> +    CoreStats cs;
> +    double dmiss_rate, imiss_rate;
> +
> +    g_autoptr(GString) rep = g_string_new("core #, data accesses, data misses,"
> +                                          " dmiss rate, insn accesses,"
> +                                          " insn misses, imiss rate\n");
> +
> +    /* Only iterate and print a sum row if cores > 1 */
> +    iters = cores == 1 ? 1 : cores + 1;
> +    for (i = 0; i < iters; i++) {
> +        cs = stats[i];
> +        dmiss_rate = ((double) cs.dmisses) / (cs.dmem_accesses) * 100.0;
> +        imiss_rate = ((double) cs.imisses) / (cs.imem_accesses) * 100.0;
> +
> +        if (i == cores) {
> +            g_string_append_printf(rep, "%-8s", "sum");
> +        } else {
> +            g_string_append_printf(rep, "%-8d", i);
> +        }
> +
> +        g_string_append_printf(rep, "%-14lu %-12lu %9.4lf%%  %-14lu %-12lu"
> +                               " %9.4lf%%\n",
> +                               cs.dmem_accesses,
> +                               cs.dmisses,
> +                               cs.dmem_accesses ? dmiss_rate : 0.0,
> +                               cs.imem_accesses,
> +                               cs.imisses,
> +                               cs.imem_accesses ? imiss_rate : 0.0);
> +    }
>  
> +    g_string_append(rep, "\n");
>      qemu_plugin_outs(rep->str);
>  }
>  
> @@ -553,15 +622,14 @@ static void plugin_exit(qemu_plugin_id_t id, void *p)
>      log_stats();
>      log_top_insns();
>  
> -    cache_free(dcache);
> -    dcache = NULL;
> -
> -    cache_free(icache);
> -    icache = NULL;
> +    caches_free(dcaches);
> +    caches_free(icaches);
>  
>      g_hash_table_destroy(miss_ht);
>      miss_ht = NULL;
>  
> +    g_free(stats);
> +
>      g_mutex_unlock(&mtx);
>  }
>  
> @@ -608,6 +676,8 @@ int qemu_plugin_install(qemu_plugin_id_t id, const qemu_info_t *info,
>  
>      policy = LRU;
>  
> +    cores = sys ? qemu_plugin_n_vcpus() : 1;
> +
>      for (i = 0; i < argc; i++) {
>          char *opt = argv[i];
>          if (g_str_has_prefix(opt, "iblksize=")) {
> @@ -624,6 +694,8 @@ int qemu_plugin_install(qemu_plugin_id_t id, const qemu_info_t *info,
>              dcachesize = g_ascii_strtoll(opt + 11, NULL, 10);
>          } else if (g_str_has_prefix(opt, "limit=")) {
>              limit = g_ascii_strtoll(opt + 6, NULL, 10);
> +        } else if (g_str_has_prefix(opt, "cores=")) {
> +            cores = g_ascii_strtoll(opt + 6, NULL, 10);
>          } else if (g_str_has_prefix(opt, "evict=")) {
>              gchar *p = opt + 6;
>              if (g_strcmp0(p, "rand") == 0) {
> @@ -644,22 +716,28 @@ int qemu_plugin_install(qemu_plugin_id_t id, const qemu_info_t *info,
>  
>      policy_init();
>  
> -    dcache = cache_init(dblksize, dassoc, dcachesize);
> -    if (!dcache) {
> +    dcaches = caches_init(dblksize, dassoc, dcachesize);
> +    if (!dcaches) {
>          const char *err = cache_config_error(dblksize, dassoc, dcachesize);
>          fprintf(stderr, "dcache cannot be constructed from given parameters\n");
>          fprintf(stderr, "%s\n", err);
>          return -1;
>      }
>  
> -    icache = cache_init(iblksize, iassoc, icachesize);
> -    if (!icache) {
> +    icaches = caches_init(iblksize, iassoc, icachesize);
> +    if (!icaches) {
>          const char *err = cache_config_error(iblksize, iassoc, icachesize);
>          fprintf(stderr, "icache cannot be constructed from given parameters\n");
>          fprintf(stderr, "%s\n", err);
>          return -1;
>      }
>  
> +    /*
> +     * plus one to save the sum in. If only one core is used then no need to
> +     * get an auxiliary struct.
> +     */
> +    stats = g_new0(CoreStats, cores == 1 ? 1 : cores + 1);
> +

See above, lets keep the sum as a post processing step.

>      qemu_plugin_register_vcpu_tb_trans_cb(id, vcpu_tb_trans);
>      qemu_plugin_register_atexit_cb(id, plugin_exit, NULL);


-- 
Alex Bennée