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