arch/s390/kernel/topology.c

   1 /*
   2  *    Copyright IBM Corp. 2007, 2011
   3  *    Author(s): Heiko Carstens <heiko.carstens@de.ibm.com>
   4  */
   5
   6 #define KMSG_COMPONENT "cpu"
   7 #define pr_fmt(fmt) KMSG_COMPONENT ": " fmt
   8
   9 #include <linux/workqueue.h>
  10 #include <linux/cpuset.h>
  11 #include <linux/device.h>
  12 #include <linux/export.h>
  13 #include <linux/kernel.h>
  14 #include <linux/sched.h>
  15 #include <linux/delay.h>
  16 #include <linux/init.h>
  17 #include <linux/slab.h>
  18 #include <linux/cpu.h>
  19 #include <linux/smp.h>
  20 #include <linux/mm.h>
  21 #include <linux/nodemask.h>
  22 #include <linux/node.h>
  23 #include <asm/sysinfo.h>
  24 #include <asm/numa.h>
  25
  26 #define PTF_HORIZONTAL  (0UL)
  27 #define PTF_VERTICAL    (1UL)
  28 #define PTF_CHECK       (2UL)
  29
  30 struct mask_info {
  31         struct mask_info *next;
  32         unsigned char id;
  33         cpumask_t mask;
  34 };
  35
  36 static void set_topology_timer(void);
  37 static void topology_work_fn(struct work_struct *work);
  38 static struct sysinfo_15_1_x *tl_info;
  39
  40 static DECLARE_WORK(topology_work, topology_work_fn);
  41
  42 /*
  43  * Socket/Book linked lists and per_cpu(cpu_topology) updates are
  44  * protected by "sched_domains_mutex".
  45  */
  46 static struct mask_info socket_info;
  47 static struct mask_info book_info;
  48 static struct mask_info drawer_info;
  49
  50 DEFINE_PER_CPU(struct cpu_topology_s390, cpu_topology);
  51 EXPORT_PER_CPU_SYMBOL_GPL(cpu_topology);
  52
  53 static cpumask_t cpu_group_map(struct mask_info *info, unsigned int cpu)
  54 {
  55         cpumask_t mask;
  56
  57         cpumask_copy(&mask, cpumask_of(cpu));
  58         if (!MACHINE_HAS_TOPOLOGY)
  59                 return mask;
  60         for (; info; info = info->next) {
  61                 if (cpumask_test_cpu(cpu, &info->mask))
  62                         return info->mask;
  63         }
  64         return mask;
  65 }
  66
  67 static cpumask_t cpu_thread_map(unsigned int cpu)
  68 {
  69         cpumask_t mask;
  70         int i;
  71
  72         cpumask_copy(&mask, cpumask_of(cpu));
  73         if (!MACHINE_HAS_TOPOLOGY)
  74                 return mask;
  75         cpu -= cpu % (smp_cpu_mtid + 1);
  76         for (i = 0; i <= smp_cpu_mtid; i++)
  77                 if (cpu_present(cpu + i))
  78                         cpumask_set_cpu(cpu + i, &mask);
  79         return mask;
  80 }
  81
  82 static void add_cpus_to_mask(struct topology_core *tl_core,
  83                              struct mask_info *drawer,
  84                              struct mask_info *book,
  85                              struct mask_info *socket)
  86 {
  87         struct cpu_topology_s390 *topo;
  88         unsigned int core;
  89
  90         for_each_set_bit(core, &tl_core->mask[0], TOPOLOGY_CORE_BITS) {
  91                 unsigned int rcore;
  92                 int lcpu, i;
  93
  94                 rcore = TOPOLOGY_CORE_BITS - 1 - core + tl_core->origin;
  95                 lcpu = smp_find_processor_id(rcore << smp_cpu_mt_shift);
  96                 if (lcpu < 0)
  97                         continue;
  98                 for (i = 0; i <= smp_cpu_mtid; i++) {
  99                         topo = &per_cpu(cpu_topology, lcpu + i);
 100                         topo->drawer_id = drawer->id;
 101                         topo->book_id = book->id;
 102                         topo->socket_id = socket->id;
 103                         topo->core_id = rcore;
 104                         topo->thread_id = lcpu + i;
 105                         cpumask_set_cpu(lcpu + i, &drawer->mask);
 106                         cpumask_set_cpu(lcpu + i, &book->mask);
 107                         cpumask_set_cpu(lcpu + i, &socket->mask);
 108                         smp_cpu_set_polarization(lcpu + i, tl_core->pp);
 109                 }
 110         }
 111 }
 112
 113 static void clear_masks(void)
 114 {
 115         struct mask_info *info;
 116
 117         info = &socket_info;
 118         while (info) {
 119                 cpumask_clear(&info->mask);
 120                 info = info->next;
 121         }
 122         info = &book_info;
 123         while (info) {
 124                 cpumask_clear(&info->mask);
 125                 info = info->next;
 126         }
 127         info = &drawer_info;
 128         while (info) {
 129                 cpumask_clear(&info->mask);
 130                 info = info->next;
 131         }
 132 }
 133
 134 static union topology_entry *next_tle(union topology_entry *tle)
 135 {
 136         if (!tle->nl)
 137                 return (union topology_entry *)((struct topology_core *)tle + 1);
 138         return (union topology_entry *)((struct topology_container *)tle + 1);
 139 }
 140
 141 static void tl_to_masks(struct sysinfo_15_1_x *info)
 142 {
 143         struct mask_info *socket = &socket_info;
 144         struct mask_info *book = &book_info;
 145         struct mask_info *drawer = &drawer_info;
 146         union topology_entry *tle, *end;
 147
 148         clear_masks();
 149         tle = info->tle;
 150         end = (union topology_entry *)((unsigned long)info + info->length);
 151         while (tle < end) {
 152                 switch (tle->nl) {
 153                 case 3:
 154                         drawer = drawer->next;
 155                         drawer->id = tle->container.id;
 156                         break;
 157                 case 2:
 158                         book = book->next;
 159                         book->id = tle->container.id;
 160                         break;
 161                 case 1:
 162                         socket = socket->next;
 163                         socket->id = tle->container.id;
 164                         break;
 165                 case 0:
 166                         add_cpus_to_mask(&tle->cpu, drawer, book, socket);
 167                         break;
 168                 default:
 169                         clear_masks();
 170                         return;
 171                 }
 172                 tle = next_tle(tle);
 173         }
 174 }
 175
 176 static void topology_update_polarization_simple(void)
 177 {
 178         int cpu;
 179
 180         mutex_lock(&smp_cpu_state_mutex);
 181         for_each_possible_cpu(cpu)
 182                 smp_cpu_set_polarization(cpu, POLARIZATION_HRZ);
 183         mutex_unlock(&smp_cpu_state_mutex);
 184 }
 185
 186 static int ptf(unsigned long fc)
 187 {
 188         int rc;
 189
 190         asm volatile(
 191                 "       .insn   rre,0xb9a20000,%1,%1\n"
 192                 "       ipm     %0\n"
 193                 "       srl     %0,28\n"
 194                 : "=d" (rc)
 195                 : "d" (fc)  : "cc");
 196         return rc;
 197 }
 198
 199 int topology_set_cpu_management(int fc)
 200 {
 201         int cpu, rc;
 202
 203         if (!MACHINE_HAS_TOPOLOGY)
 204                 return -EOPNOTSUPP;
 205         if (fc)
 206                 rc = ptf(PTF_VERTICAL);
 207         else
 208                 rc = ptf(PTF_HORIZONTAL);
 209         if (rc)
 210                 return -EBUSY;
 211         for_each_possible_cpu(cpu)
 212                 smp_cpu_set_polarization(cpu, POLARIZATION_UNKNOWN);
 213         return rc;
 214 }
 215
 216 static void update_cpu_masks(void)
 217 {
 218         struct cpu_topology_s390 *topo;
 219         int cpu;
 220
 221         for_each_possible_cpu(cpu) {
 222                 topo = &per_cpu(cpu_topology, cpu);
 223                 topo->thread_mask = cpu_thread_map(cpu);
 224                 topo->core_mask = cpu_group_map(&socket_info, cpu);
 225                 topo->book_mask = cpu_group_map(&book_info, cpu);
 226                 topo->drawer_mask = cpu_group_map(&drawer_info, cpu);
 227                 if (!MACHINE_HAS_TOPOLOGY) {
 228                         topo->thread_id = cpu;
 229                         topo->core_id = cpu;
 230                         topo->socket_id = cpu;
 231                         topo->book_id = cpu;
 232                         topo->drawer_id = cpu;
 233                 }
 234         }
 235         numa_update_cpu_topology();
 236 }
 237
 238 void store_topology(struct sysinfo_15_1_x *info)
 239 {
 240         stsi(info, 15, 1, min(topology_max_mnest, 4));
 241 }
 242
 243 int arch_update_cpu_topology(void)
 244 {
 245         struct sysinfo_15_1_x *info = tl_info;
 246         struct device *dev;
 247         int cpu, rc = 0;
 248
 249         if (MACHINE_HAS_TOPOLOGY) {
 250                 rc = 1;
 251                 store_topology(info);
 252                 tl_to_masks(info);
 253         }
 254         update_cpu_masks();
 255         if (!MACHINE_HAS_TOPOLOGY)
 256                 topology_update_polarization_simple();
 257         for_each_online_cpu(cpu) {
 258                 dev = get_cpu_device(cpu);
 259                 if (dev)
 260                         kobject_uevent(&dev->kobj, KOBJ_CHANGE);
 261         }
 262         return rc;
 263 }
 264
 265 static void topology_work_fn(struct work_struct *work)
 266 {
 267         rebuild_sched_domains();
 268 }
 269
 270 void topology_schedule_update(void)
 271 {
 272         schedule_work(&topology_work);
 273 }
 274
 275 static void topology_timer_fn(unsigned long ignored)
 276 {
 277         if (ptf(PTF_CHECK))
 278                 topology_schedule_update();
 279         set_topology_timer();
 280 }
 281
 282 static struct timer_list topology_timer =
 283         TIMER_DEFERRED_INITIALIZER(topology_timer_fn, 0, 0);
 284
 285 static atomic_t topology_poll = ATOMIC_INIT(0);
 286
 287 static void set_topology_timer(void)
 288 {
 289         if (atomic_add_unless(&topology_poll, -1, 0))
 290                 mod_timer(&topology_timer, jiffies + HZ / 10);
 291         else
 292                 mod_timer(&topology_timer, jiffies + HZ * 60);
 293 }
 294
 295 void topology_expect_change(void)
 296 {
 297         if (!MACHINE_HAS_TOPOLOGY)
 298                 return;
 299         /* This is racy, but it doesn't matter since it is just a heuristic.
 300          * Worst case is that we poll in a higher frequency for a bit longer.
 301          */
 302         if (atomic_read(&topology_poll) > 60)
 303                 return;
 304         atomic_add(60, &topology_poll);
 305         set_topology_timer();
 306 }
 307
 308 static int cpu_management;
 309
 310 static ssize_t dispatching_show(struct device *dev,
 311                                 struct device_attribute *attr,
 312                                 char *buf)
 313 {
 314         ssize_t count;
 315
 316         mutex_lock(&smp_cpu_state_mutex);
 317         count = sprintf(buf, "%d\n", cpu_management);
 318         mutex_unlock(&smp_cpu_state_mutex);
 319         return count;
 320 }
 321
 322 static ssize_t dispatching_store(struct device *dev,
 323                                  struct device_attribute *attr,
 324                                  const char *buf,
 325                                  size_t count)
 326 {
 327         int val, rc;
 328         char delim;
 329
 330         if (sscanf(buf, "%d %c", &val, &delim) != 1)
 331                 return -EINVAL;
 332         if (val != 0 && val != 1)
 333                 return -EINVAL;
 334         rc = 0;
 335         get_online_cpus();
 336         mutex_lock(&smp_cpu_state_mutex);
 337         if (cpu_management == val)
 338                 goto out;
 339         rc = topology_set_cpu_management(val);
 340         if (rc)
 341                 goto out;
 342         cpu_management = val;
 343         topology_expect_change();
 344 out:
 345         mutex_unlock(&smp_cpu_state_mutex);
 346         put_online_cpus();
 347         return rc ? rc : count;
 348 }
 349 static DEVICE_ATTR(dispatching, 0644, dispatching_show,
 350                          dispatching_store);
 351
 352 static ssize_t cpu_polarization_show(struct device *dev,
 353                                      struct device_attribute *attr, char *buf)
 354 {
 355         int cpu = dev->id;
 356         ssize_t count;
 357
 358         mutex_lock(&smp_cpu_state_mutex);
 359         switch (smp_cpu_get_polarization(cpu)) {
 360         case POLARIZATION_HRZ:
 361                 count = sprintf(buf, "horizontal\n");
 362                 break;
 363         case POLARIZATION_VL:
 364                 count = sprintf(buf, "vertical:low\n");
 365                 break;
 366         case POLARIZATION_VM:
 367                 count = sprintf(buf, "vertical:medium\n");
 368                 break;
 369         case POLARIZATION_VH:
 370                 count = sprintf(buf, "vertical:high\n");
 371                 break;
 372         default:
 373                 count = sprintf(buf, "unknown\n");
 374                 break;
 375         }
 376         mutex_unlock(&smp_cpu_state_mutex);
 377         return count;
 378 }
 379 static DEVICE_ATTR(polarization, 0444, cpu_polarization_show, NULL);
 380
 381 static struct attribute *topology_cpu_attrs[] = {
 382         &dev_attr_polarization.attr,
 383         NULL,
 384 };
 385
 386 static struct attribute_group topology_cpu_attr_group = {
 387         .attrs = topology_cpu_attrs,
 388 };
 389
 390 int topology_cpu_init(struct cpu *cpu)
 391 {
 392         return sysfs_create_group(&cpu->dev.kobj, &topology_cpu_attr_group);
 393 }
 394
 395 static const struct cpumask *cpu_thread_mask(int cpu)
 396 {
 397         return &per_cpu(cpu_topology, cpu).thread_mask;
 398 }
 399
 400
 401 const struct cpumask *cpu_coregroup_mask(int cpu)
 402 {
 403         return &per_cpu(cpu_topology, cpu).core_mask;
 404 }
 405
 406 static const struct cpumask *cpu_book_mask(int cpu)
 407 {
 408         return &per_cpu(cpu_topology, cpu).book_mask;
 409 }
 410
 411 static const struct cpumask *cpu_drawer_mask(int cpu)
 412 {
 413         return &per_cpu(cpu_topology, cpu).drawer_mask;
 414 }
 415
 416 static struct sched_domain_topology_level s390_topology[] = {
 417         { cpu_thread_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
 418         { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
 419         { cpu_book_mask, SD_INIT_NAME(BOOK) },
 420         { cpu_drawer_mask, SD_INIT_NAME(DRAWER) },
 421         { cpu_cpu_mask, SD_INIT_NAME(DIE) },
 422         { NULL, },
 423 };
 424
 425 static void __init alloc_masks(struct sysinfo_15_1_x *info,
 426                                struct mask_info *mask, int offset)
 427 {
 428         int i, nr_masks;
 429
 430         nr_masks = info->mag[TOPOLOGY_NR_MAG - offset];
 431         for (i = 0; i < info->mnest - offset; i++)
 432                 nr_masks *= info->mag[TOPOLOGY_NR_MAG - offset - 1 - i];
 433         nr_masks = max(nr_masks, 1);
 434         for (i = 0; i < nr_masks; i++) {
 435                 mask->next = kzalloc(sizeof(*mask->next), GFP_KERNEL);
 436                 mask = mask->next;
 437         }
 438 }
 439
 440 static int __init s390_topology_init(void)
 441 {
 442         struct sysinfo_15_1_x *info;
 443         int i;
 444
 445         set_sched_topology(s390_topology);
 446         if (!MACHINE_HAS_TOPOLOGY)
 447                 return 0;
 448         tl_info = (struct sysinfo_15_1_x *)__get_free_page(GFP_KERNEL);
 449         info = tl_info;
 450         store_topology(info);
 451         pr_info("The CPU configuration topology of the machine is:");
 452         for (i = 0; i < TOPOLOGY_NR_MAG; i++)
 453                 printk(KERN_CONT " %d", info->mag[i]);
 454         printk(KERN_CONT " / %d\n", info->mnest);
 455         alloc_masks(info, &socket_info, 1);
 456         alloc_masks(info, &book_info, 2);
 457         alloc_masks(info, &drawer_info, 3);
 458         return 0;
 459 }
 460 early_initcall(s390_topology_init);
 461
 462 static int __init topology_init(void)
 463 {
 464         if (MACHINE_HAS_TOPOLOGY)
 465                 set_topology_timer();
 466         else
 467                 topology_update_polarization_simple();
 468         return device_create_file(cpu_subsys.dev_root, &dev_attr_dispatching);
 469 }
 470 device_initcall(topology_init);