arch/sparc/kernel/cpumap.c

   1 /* cpumap.c: used for optimizing CPU assignment
   2  *
   3  * Copyright (C) 2009 Hong H. Pham <hong.pham@windriver.com>
   4  */
   5
   6 #include <linux/module.h>
   7 #include <linux/slab.h>
   8 #include <linux/kernel.h>
   9 #include <linux/init.h>
  10 #include <linux/cpumask.h>
  11 #include <linux/spinlock.h>
  12 #include <asm/cpudata.h>
  13 #include "cpumap.h"
  14
  15
  16 enum {
  17         CPUINFO_LVL_ROOT = 0,
  18         CPUINFO_LVL_NODE,
  19         CPUINFO_LVL_CORE,
  20         CPUINFO_LVL_PROC,
  21         CPUINFO_LVL_MAX,
  22 };
  23
  24 enum {
  25         ROVER_NO_OP              = 0,
  26         /* Increment rover every time level is visited */
  27         ROVER_INC_ON_VISIT       = 1 << 0,
  28         /* Increment parent's rover every time rover wraps around */
  29         ROVER_INC_PARENT_ON_LOOP = 1 << 1,
  30 };
  31
  32 struct cpuinfo_node {
  33         int id;
  34         int level;
  35         int num_cpus;    /* Number of CPUs in this hierarchy */
  36         int parent_index;
  37         int child_start; /* Array index of the first child node */
  38         int child_end;   /* Array index of the last child node */
  39         int rover;       /* Child node iterator */
  40 };
  41
  42 struct cpuinfo_level {
  43         int start_index; /* Index of first node of a level in a cpuinfo tree */
  44         int end_index;   /* Index of last node of a level in a cpuinfo tree */
  45         int num_nodes;   /* Number of nodes in a level in a cpuinfo tree */
  46 };
  47
  48 struct cpuinfo_tree {
  49         int total_nodes;
  50
  51         /* Offsets into nodes[] for each level of the tree */
  52         struct cpuinfo_level level[CPUINFO_LVL_MAX];
  53         struct cpuinfo_node  nodes[0];
  54 };
  55
  56
  57 static struct cpuinfo_tree *cpuinfo_tree;
  58
  59 static u16 cpu_distribution_map[NR_CPUS];
  60 static DEFINE_SPINLOCK(cpu_map_lock);
  61
  62
  63 /* Niagara optimized cpuinfo tree traversal. */
  64 static const int niagara_iterate_method[] = {
  65         [CPUINFO_LVL_ROOT] = ROVER_NO_OP,
  66
  67         /* Strands (or virtual CPUs) within a core may not run concurrently
  68          * on the Niagara, as instruction pipeline(s) are shared.  Distribute
  69          * work to strands in different cores first for better concurrency.
  70          * Go to next NUMA node when all cores are used.
  71          */
  72         [CPUINFO_LVL_NODE] = ROVER_INC_ON_VISIT|ROVER_INC_PARENT_ON_LOOP,
  73
  74         /* Strands are grouped together by proc_id in cpuinfo_sparc, i.e.
  75          * a proc_id represents an instruction pipeline.  Distribute work to
  76          * strands in different proc_id groups if the core has multiple
  77          * instruction pipelines (e.g. the Niagara 2/2+ has two).
  78          */
  79         [CPUINFO_LVL_CORE] = ROVER_INC_ON_VISIT,
  80
  81         /* Pick the next strand in the proc_id group. */
  82         [CPUINFO_LVL_PROC] = ROVER_INC_ON_VISIT,
  83 };
  84
  85 /* Generic cpuinfo tree traversal.  Distribute work round robin across NUMA
  86  * nodes.
  87  */
  88 static const int generic_iterate_method[] = {
  89         [CPUINFO_LVL_ROOT] = ROVER_INC_ON_VISIT,
  90         [CPUINFO_LVL_NODE] = ROVER_NO_OP,
  91         [CPUINFO_LVL_CORE] = ROVER_INC_PARENT_ON_LOOP,
  92         [CPUINFO_LVL_PROC] = ROVER_INC_ON_VISIT|ROVER_INC_PARENT_ON_LOOP,
  93 };
  94
  95
  96 static int cpuinfo_id(int cpu, int level)
  97 {
  98         int id;
  99
 100         switch (level) {
 101         case CPUINFO_LVL_ROOT:
 102                 id = 0;
 103                 break;
 104         case CPUINFO_LVL_NODE:
 105                 id = cpu_to_node(cpu);
 106                 break;
 107         case CPUINFO_LVL_CORE:
 108                 id = cpu_data(cpu).core_id;
 109                 break;
 110         case CPUINFO_LVL_PROC:
 111                 id = cpu_data(cpu).proc_id;
 112                 break;
 113         default:
 114                 id = -EINVAL;
 115         }
 116         return id;
 117 }
 118
 119 /*
 120  * Enumerate the CPU information in __cpu_data to determine the start index,
 121  * end index, and number of nodes for each level in the cpuinfo tree.  The
 122  * total number of cpuinfo nodes required to build the tree is returned.
 123  */
 124 static int enumerate_cpuinfo_nodes(struct cpuinfo_level *tree_level)
 125 {
 126         int prev_id[CPUINFO_LVL_MAX];
 127         int i, n, num_nodes;
 128
 129         for (i = CPUINFO_LVL_ROOT; i < CPUINFO_LVL_MAX; i++) {
 130                 struct cpuinfo_level *lv = &tree_level[i];
 131
 132                 prev_id[i] = -1;
 133                 lv->start_index = lv->end_index = lv->num_nodes = 0;
 134         }
 135
 136         num_nodes = 1; /* Include the root node */
 137
 138         for (i = 0; i < num_possible_cpus(); i++) {
 139                 if (!cpu_online(i))
 140                         continue;
 141
 142                 n = cpuinfo_id(i, CPUINFO_LVL_NODE);
 143                 if (n > prev_id[CPUINFO_LVL_NODE]) {
 144                         tree_level[CPUINFO_LVL_NODE].num_nodes++;
 145                         prev_id[CPUINFO_LVL_NODE] = n;
 146                         num_nodes++;
 147                 }
 148                 n = cpuinfo_id(i, CPUINFO_LVL_CORE);
 149                 if (n > prev_id[CPUINFO_LVL_CORE]) {
 150                         tree_level[CPUINFO_LVL_CORE].num_nodes++;
 151                         prev_id[CPUINFO_LVL_CORE] = n;
 152                         num_nodes++;
 153                 }
 154                 n = cpuinfo_id(i, CPUINFO_LVL_PROC);
 155                 if (n > prev_id[CPUINFO_LVL_PROC]) {
 156                         tree_level[CPUINFO_LVL_PROC].num_nodes++;
 157                         prev_id[CPUINFO_LVL_PROC] = n;
 158                         num_nodes++;
 159                 }
 160         }
 161
 162         tree_level[CPUINFO_LVL_ROOT].num_nodes = 1;
 163
 164         n = tree_level[CPUINFO_LVL_NODE].num_nodes;
 165         tree_level[CPUINFO_LVL_NODE].start_index = 1;
 166         tree_level[CPUINFO_LVL_NODE].end_index   = n;
 167
 168         n++;
 169         tree_level[CPUINFO_LVL_CORE].start_index = n;
 170         n += tree_level[CPUINFO_LVL_CORE].num_nodes;
 171         tree_level[CPUINFO_LVL_CORE].end_index   = n - 1;
 172
 173         tree_level[CPUINFO_LVL_PROC].start_index = n;
 174         n += tree_level[CPUINFO_LVL_PROC].num_nodes;
 175         tree_level[CPUINFO_LVL_PROC].end_index   = n - 1;
 176
 177         return num_nodes;
 178 }
 179
 180 /* Build a tree representation of the CPU hierarchy using the per CPU
 181  * information in __cpu_data.  Entries in __cpu_data[0..NR_CPUS] are
 182  * assumed to be sorted in ascending order based on node, core_id, and
 183  * proc_id (in order of significance).
 184  */
 185 static struct cpuinfo_tree *build_cpuinfo_tree(void)
 186 {
 187         struct cpuinfo_tree *new_tree;
 188         struct cpuinfo_node *node;
 189         struct cpuinfo_level tmp_level[CPUINFO_LVL_MAX];
 190         int num_cpus[CPUINFO_LVL_MAX];
 191         int level_rover[CPUINFO_LVL_MAX];
 192         int prev_id[CPUINFO_LVL_MAX];
 193         int n, id, cpu, prev_cpu, last_cpu, level;
 194
 195         n = enumerate_cpuinfo_nodes(tmp_level);
 196
 197         new_tree = kzalloc(sizeof(struct cpuinfo_tree) +
 198                            (sizeof(struct cpuinfo_node) * n), GFP_ATOMIC);
 199         if (!new_tree)
 200                 return NULL;
 201
 202         new_tree->total_nodes = n;
 203         memcpy(&new_tree->level, tmp_level, sizeof(tmp_level));
 204
 205         prev_cpu = cpu = cpumask_first(cpu_online_mask);
 206
 207         /* Initialize all levels in the tree with the first CPU */
 208         for (level = CPUINFO_LVL_PROC; level >= CPUINFO_LVL_ROOT; level--) {
 209                 n = new_tree->level[level].start_index;
 210
 211                 level_rover[level] = n;
 212                 node = &new_tree->nodes[n];
 213
 214                 id = cpuinfo_id(cpu, level);
 215                 if (unlikely(id < 0)) {
 216                         kfree(new_tree);
 217                         return NULL;
 218                 }
 219                 node->id = id;
 220                 node->level = level;
 221                 node->num_cpus = 1;
 222
 223                 node->parent_index = (level > CPUINFO_LVL_ROOT)
 224                     ? new_tree->level[level - 1].start_index : -1;
 225
 226                 node->child_start = node->child_end = node->rover =
 227                     (level == CPUINFO_LVL_PROC)
 228                     ? cpu : new_tree->level[level + 1].start_index;
 229
 230                 prev_id[level] = node->id;
 231                 num_cpus[level] = 1;
 232         }
 233
 234         for (last_cpu = (num_possible_cpus() - 1); last_cpu >= 0; last_cpu--) {
 235                 if (cpu_online(last_cpu))
 236                         break;
 237         }
 238
 239         while (++cpu <= last_cpu) {
 240                 if (!cpu_online(cpu))
 241                         continue;
 242
 243                 for (level = CPUINFO_LVL_PROC; level >= CPUINFO_LVL_ROOT;
 244                      level--) {
 245                         id = cpuinfo_id(cpu, level);
 246                         if (unlikely(id < 0)) {
 247                                 kfree(new_tree);
 248                                 return NULL;
 249                         }
 250
 251                         if ((id != prev_id[level]) || (cpu == last_cpu)) {
 252                                 prev_id[level] = id;
 253                                 node = &new_tree->nodes[level_rover[level]];
 254                                 node->num_cpus = num_cpus[level];
 255                                 num_cpus[level] = 1;
 256
 257                                 if (cpu == last_cpu)
 258                                         node->num_cpus++;
 259
 260                                 /* Connect tree node to parent */
 261                                 if (level == CPUINFO_LVL_ROOT)
 262                                         node->parent_index = -1;
 263                                 else
 264                                         node->parent_index =
 265                                             level_rover[level - 1];
 266
 267                                 if (level == CPUINFO_LVL_PROC) {
 268                                         node->child_end =
 269                                             (cpu == last_cpu) ? cpu : prev_cpu;
 270                                 } else {
 271                                         node->child_end =
 272                                             level_rover[level + 1] - 1;
 273                                 }
 274
 275                                 /* Initialize the next node in the same level */
 276                                 n = ++level_rover[level];
 277                                 if (n <= new_tree->level[level].end_index) {
 278                                         node = &new_tree->nodes[n];
 279                                         node->id = id;
 280                                         node->level = level;
 281
 282                                         /* Connect node to child */
 283                                         node->child_start = node->child_end =
 284                                         node->rover =
 285                                             (level == CPUINFO_LVL_PROC)
 286                                             ? cpu : level_rover[level + 1];
 287                                 }
 288                         } else
 289                                 num_cpus[level]++;
 290                 }
 291                 prev_cpu = cpu;
 292         }
 293
 294         return new_tree;
 295 }
 296
 297 static void increment_rover(struct cpuinfo_tree *t, int node_index,
 298                             int root_index, const int *rover_inc_table)
 299 {
 300         struct cpuinfo_node *node = &t->nodes[node_index];
 301         int top_level, level;
 302
 303         top_level = t->nodes[root_index].level;
 304         for (level = node->level; level >= top_level; level--) {
 305                 node->rover++;
 306                 if (node->rover <= node->child_end)
 307                         return;
 308
 309                 node->rover = node->child_start;
 310                 /* If parent's rover does not need to be adjusted, stop here. */
 311                 if ((level == top_level) ||
 312                     !(rover_inc_table[level] & ROVER_INC_PARENT_ON_LOOP))
 313                         return;
 314
 315                 node = &t->nodes[node->parent_index];
 316         }
 317 }
 318
 319 static int iterate_cpu(struct cpuinfo_tree *t, unsigned int root_index)
 320 {
 321         const int *rover_inc_table;
 322         int level, new_index, index = root_index;
 323
 324         switch (sun4v_chip_type) {
 325         case SUN4V_CHIP_NIAGARA1:
 326         case SUN4V_CHIP_NIAGARA2:
 327         case SUN4V_CHIP_NIAGARA3:
 328         case SUN4V_CHIP_NIAGARA4:
 329         case SUN4V_CHIP_NIAGARA5:
 330                 rover_inc_table = niagara_iterate_method;
 331                 break;
 332         default:
 333                 rover_inc_table = generic_iterate_method;
 334         }
 335
 336         for (level = t->nodes[root_index].level; level < CPUINFO_LVL_MAX;
 337              level++) {
 338                 new_index = t->nodes[index].rover;
 339                 if (rover_inc_table[level] & ROVER_INC_ON_VISIT)
 340                         increment_rover(t, index, root_index, rover_inc_table);
 341
 342                 index = new_index;
 343         }
 344         return index;
 345 }
 346
 347 static void _cpu_map_rebuild(void)
 348 {
 349         int i;
 350
 351         if (cpuinfo_tree) {
 352                 kfree(cpuinfo_tree);
 353                 cpuinfo_tree = NULL;
 354         }
 355
 356         cpuinfo_tree = build_cpuinfo_tree();
 357         if (!cpuinfo_tree)
 358                 return;
 359
 360         /* Build CPU distribution map that spans all online CPUs.  No need
 361          * to check if the CPU is online, as that is done when the cpuinfo
 362          * tree is being built.
 363          */
 364         for (i = 0; i < cpuinfo_tree->nodes[0].num_cpus; i++)
 365                 cpu_distribution_map[i] = iterate_cpu(cpuinfo_tree, 0);
 366 }
 367
 368 /* Fallback if the cpuinfo tree could not be built.  CPU mapping is linear
 369  * round robin.
 370  */
 371 static int simple_map_to_cpu(unsigned int index)
 372 {
 373         int i, end, cpu_rover;
 374
 375         cpu_rover = 0;
 376         end = index % num_online_cpus();
 377         for (i = 0; i < num_possible_cpus(); i++) {
 378                 if (cpu_online(cpu_rover)) {
 379                         if (cpu_rover >= end)
 380                                 return cpu_rover;
 381
 382                         cpu_rover++;
 383                 }
 384         }
 385
 386         /* Impossible, since num_online_cpus() <= num_possible_cpus() */
 387         return cpumask_first(cpu_online_mask);
 388 }
 389
 390 static int _map_to_cpu(unsigned int index)
 391 {
 392         struct cpuinfo_node *root_node;
 393
 394         if (unlikely(!cpuinfo_tree)) {
 395                 _cpu_map_rebuild();
 396                 if (!cpuinfo_tree)
 397                         return simple_map_to_cpu(index);
 398         }
 399
 400         root_node = &cpuinfo_tree->nodes[0];
 401 #ifdef CONFIG_HOTPLUG_CPU
 402         if (unlikely(root_node->num_cpus != num_online_cpus())) {
 403                 _cpu_map_rebuild();
 404                 if (!cpuinfo_tree)
 405                         return simple_map_to_cpu(index);
 406         }
 407 #endif
 408         return cpu_distribution_map[index % root_node->num_cpus];
 409 }
 410
 411 int map_to_cpu(unsigned int index)
 412 {
 413         int mapped_cpu;
 414         unsigned long flag;
 415
 416         spin_lock_irqsave(&cpu_map_lock, flag);
 417         mapped_cpu = _map_to_cpu(index);
 418
 419 #ifdef CONFIG_HOTPLUG_CPU
 420         while (unlikely(!cpu_online(mapped_cpu)))
 421                 mapped_cpu = _map_to_cpu(index);
 422 #endif
 423         spin_unlock_irqrestore(&cpu_map_lock, flag);
 424         return mapped_cpu;
 425 }
 426 EXPORT_SYMBOL(map_to_cpu);
 427
 428 void cpu_map_rebuild(void)
 429 {
 430         unsigned long flag;
 431
 432         spin_lock_irqsave(&cpu_map_lock, flag);
 433         _cpu_map_rebuild();
 434         spin_unlock_irqrestore(&cpu_map_lock, flag);
 435 }