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 bool topology_enabled = true;
  41 static DECLARE_WORK(topology_work, topology_work_fn);
  42
  43 /*
  44  * Socket/Book linked lists and per_cpu(cpu_topology) updates are
  45  * protected by "sched_domains_mutex".
  46  */
  47 static struct mask_info socket_info;
  48 static struct mask_info book_info;
  49 static struct mask_info drawer_info;
  50
  51 DEFINE_PER_CPU(struct cpu_topology_s390, cpu_topology);
  52 EXPORT_PER_CPU_SYMBOL_GPL(cpu_topology);
  53
  54 static cpumask_t cpu_group_map(struct mask_info *info, unsigned int cpu)
  55 {
  56         cpumask_t mask;
  57
  58         cpumask_copy(&mask, cpumask_of(cpu));
  59         if (!topology_enabled || !MACHINE_HAS_TOPOLOGY)
  60                 return mask;
  61         for (; info; info = info->next) {
  62                 if (cpumask_test_cpu(cpu, &info->mask))
  63                         return info->mask;
  64         }
  65         return mask;
  66 }
  67
  68 static cpumask_t cpu_thread_map(unsigned int cpu)
  69 {
  70         cpumask_t mask;
  71         int i;
  72
  73         cpumask_copy(&mask, cpumask_of(cpu));
  74         if (!topology_enabled || !MACHINE_HAS_TOPOLOGY)
  75                 return mask;
  76         cpu -= cpu % (smp_cpu_mtid + 1);
  77         for (i = 0; i <= smp_cpu_mtid; i++)
  78                 if (cpu_present(cpu + i))
  79                         cpumask_set_cpu(cpu + i, &mask);
  80         return mask;
  81 }
  82
  83 static struct mask_info *add_cpus_to_mask(struct topology_core *tl_core,
  84                                           struct mask_info *drawer,
  85                                           struct mask_info *book,
  86                                           struct mask_info *socket,
  87                                           int one_socket_per_cpu)
  88 {
  89         struct cpu_topology_s390 *topo;
  90         unsigned int core;
  91
  92         for_each_set_bit(core, &tl_core->mask[0], TOPOLOGY_CORE_BITS) {
  93                 unsigned int rcore;
  94                 int lcpu, i;
  95
  96                 rcore = TOPOLOGY_CORE_BITS - 1 - core + tl_core->origin;
  97                 lcpu = smp_find_processor_id(rcore << smp_cpu_mt_shift);
  98                 if (lcpu < 0)
  99                         continue;
 100                 for (i = 0; i <= smp_cpu_mtid; i++) {
 101                         topo = &per_cpu(cpu_topology, lcpu + i);
 102                         topo->drawer_id = drawer->id;
 103                         topo->book_id = book->id;
 104                         topo->core_id = rcore;
 105                         topo->thread_id = lcpu + i;
 106                         cpumask_set_cpu(lcpu + i, &drawer->mask);
 107                         cpumask_set_cpu(lcpu + i, &book->mask);
 108                         cpumask_set_cpu(lcpu + i, &socket->mask);
 109                         if (one_socket_per_cpu)
 110                                 topo->socket_id = rcore;
 111                         else
 112                                 topo->socket_id = socket->id;
 113                         smp_cpu_set_polarization(lcpu + i, tl_core->pp);
 114                 }
 115                 if (one_socket_per_cpu)
 116                         socket = socket->next;
 117         }
 118         return socket;
 119 }
 120
 121 static void clear_masks(void)
 122 {
 123         struct mask_info *info;
 124
 125         info = &socket_info;
 126         while (info) {
 127                 cpumask_clear(&info->mask);
 128                 info = info->next;
 129         }
 130         info = &book_info;
 131         while (info) {
 132                 cpumask_clear(&info->mask);
 133                 info = info->next;
 134         }
 135         info = &drawer_info;
 136         while (info) {
 137                 cpumask_clear(&info->mask);
 138                 info = info->next;
 139         }
 140 }
 141
 142 static union topology_entry *next_tle(union topology_entry *tle)
 143 {
 144         if (!tle->nl)
 145                 return (union topology_entry *)((struct topology_core *)tle + 1);
 146         return (union topology_entry *)((struct topology_container *)tle + 1);
 147 }
 148
 149 static void __tl_to_masks_generic(struct sysinfo_15_1_x *info)
 150 {
 151         struct mask_info *socket = &socket_info;
 152         struct mask_info *book = &book_info;
 153         struct mask_info *drawer = &drawer_info;
 154         union topology_entry *tle, *end;
 155
 156         tle = info->tle;
 157         end = (union topology_entry *)((unsigned long)info + info->length);
 158         while (tle < end) {
 159                 switch (tle->nl) {
 160                 case 3:
 161                         drawer = drawer->next;
 162                         drawer->id = tle->container.id;
 163                         break;
 164                 case 2:
 165                         book = book->next;
 166                         book->id = tle->container.id;
 167                         break;
 168                 case 1:
 169                         socket = socket->next;
 170                         socket->id = tle->container.id;
 171                         break;
 172                 case 0:
 173                         add_cpus_to_mask(&tle->cpu, drawer, book, socket, 0);
 174                         break;
 175                 default:
 176                         clear_masks();
 177                         return;
 178                 }
 179                 tle = next_tle(tle);
 180         }
 181 }
 182
 183 static void __tl_to_masks_z10(struct sysinfo_15_1_x *info)
 184 {
 185         struct mask_info *socket = &socket_info;
 186         struct mask_info *book = &book_info;
 187         struct mask_info *drawer = &drawer_info;
 188         union topology_entry *tle, *end;
 189
 190         tle = info->tle;
 191         end = (union topology_entry *)((unsigned long)info + info->length);
 192         while (tle < end) {
 193                 switch (tle->nl) {
 194                 case 1:
 195                         book = book->next;
 196                         book->id = tle->container.id;
 197                         break;
 198                 case 0:
 199                         socket = add_cpus_to_mask(&tle->cpu, drawer, book, socket, 1);
 200                         break;
 201                 default:
 202                         clear_masks();
 203                         return;
 204                 }
 205                 tle = next_tle(tle);
 206         }
 207 }
 208
 209 static void tl_to_masks(struct sysinfo_15_1_x *info)
 210 {
 211         struct cpuid cpu_id;
 212
 213         get_cpu_id(&cpu_id);
 214         clear_masks();
 215         switch (cpu_id.machine) {
 216         case 0x2097:
 217         case 0x2098:
 218                 __tl_to_masks_z10(info);
 219                 break;
 220         default:
 221                 __tl_to_masks_generic(info);
 222         }
 223 }
 224
 225 static void topology_update_polarization_simple(void)
 226 {
 227         int cpu;
 228
 229         mutex_lock(&smp_cpu_state_mutex);
 230         for_each_possible_cpu(cpu)
 231                 smp_cpu_set_polarization(cpu, POLARIZATION_HRZ);
 232         mutex_unlock(&smp_cpu_state_mutex);
 233 }
 234
 235 static int ptf(unsigned long fc)
 236 {
 237         int rc;
 238
 239         asm volatile(
 240                 "       .insn   rre,0xb9a20000,%1,%1\n"
 241                 "       ipm     %0\n"
 242                 "       srl     %0,28\n"
 243                 : "=d" (rc)
 244                 : "d" (fc)  : "cc");
 245         return rc;
 246 }
 247
 248 int topology_set_cpu_management(int fc)
 249 {
 250         int cpu, rc;
 251
 252         if (!MACHINE_HAS_TOPOLOGY)
 253                 return -EOPNOTSUPP;
 254         if (fc)
 255                 rc = ptf(PTF_VERTICAL);
 256         else
 257                 rc = ptf(PTF_HORIZONTAL);
 258         if (rc)
 259                 return -EBUSY;
 260         for_each_possible_cpu(cpu)
 261                 smp_cpu_set_polarization(cpu, POLARIZATION_UNKNOWN);
 262         return rc;
 263 }
 264
 265 static void update_cpu_masks(void)
 266 {
 267         struct cpu_topology_s390 *topo;
 268         int cpu;
 269
 270         for_each_possible_cpu(cpu) {
 271                 topo = &per_cpu(cpu_topology, cpu);
 272                 topo->thread_mask = cpu_thread_map(cpu);
 273                 topo->core_mask = cpu_group_map(&socket_info, cpu);
 274                 topo->book_mask = cpu_group_map(&book_info, cpu);
 275                 topo->drawer_mask = cpu_group_map(&drawer_info, cpu);
 276                 if (!MACHINE_HAS_TOPOLOGY) {
 277                         topo->thread_id = cpu;
 278                         topo->core_id = cpu;
 279                         topo->socket_id = cpu;
 280                         topo->book_id = cpu;
 281                         topo->drawer_id = cpu;
 282                 }
 283         }
 284         numa_update_cpu_topology();
 285 }
 286
 287 void store_topology(struct sysinfo_15_1_x *info)
 288 {
 289         stsi(info, 15, 1, min(topology_max_mnest, 4));
 290 }
 291
 292 int arch_update_cpu_topology(void)
 293 {
 294         struct sysinfo_15_1_x *info = tl_info;
 295         struct device *dev;
 296         int cpu, rc = 0;
 297
 298         if (MACHINE_HAS_TOPOLOGY) {
 299                 rc = 1;
 300                 store_topology(info);
 301                 tl_to_masks(info);
 302         }
 303         update_cpu_masks();
 304         if (!MACHINE_HAS_TOPOLOGY)
 305                 topology_update_polarization_simple();
 306         for_each_online_cpu(cpu) {
 307                 dev = get_cpu_device(cpu);
 308                 kobject_uevent(&dev->kobj, KOBJ_CHANGE);
 309         }
 310         return rc;
 311 }
 312
 313 static void topology_work_fn(struct work_struct *work)
 314 {
 315         rebuild_sched_domains();
 316 }
 317
 318 void topology_schedule_update(void)
 319 {
 320         schedule_work(&topology_work);
 321 }
 322
 323 static void topology_timer_fn(unsigned long ignored)
 324 {
 325         if (ptf(PTF_CHECK))
 326                 topology_schedule_update();
 327         set_topology_timer();
 328 }
 329
 330 static struct timer_list topology_timer =
 331         TIMER_DEFERRED_INITIALIZER(topology_timer_fn, 0, 0);
 332
 333 static atomic_t topology_poll = ATOMIC_INIT(0);
 334
 335 static void set_topology_timer(void)
 336 {
 337         if (atomic_add_unless(&topology_poll, -1, 0))
 338                 mod_timer(&topology_timer, jiffies + HZ / 10);
 339         else
 340                 mod_timer(&topology_timer, jiffies + HZ * 60);
 341 }
 342
 343 void topology_expect_change(void)
 344 {
 345         if (!MACHINE_HAS_TOPOLOGY)
 346                 return;
 347         /* This is racy, but it doesn't matter since it is just a heuristic.
 348          * Worst case is that we poll in a higher frequency for a bit longer.
 349          */
 350         if (atomic_read(&topology_poll) > 60)
 351                 return;
 352         atomic_add(60, &topology_poll);
 353         set_topology_timer();
 354 }
 355
 356 static int cpu_management;
 357
 358 static ssize_t dispatching_show(struct device *dev,
 359                                 struct device_attribute *attr,
 360                                 char *buf)
 361 {
 362         ssize_t count;
 363
 364         mutex_lock(&smp_cpu_state_mutex);
 365         count = sprintf(buf, "%d\n", cpu_management);
 366         mutex_unlock(&smp_cpu_state_mutex);
 367         return count;
 368 }
 369
 370 static ssize_t dispatching_store(struct device *dev,
 371                                  struct device_attribute *attr,
 372                                  const char *buf,
 373                                  size_t count)
 374 {
 375         int val, rc;
 376         char delim;
 377
 378         if (sscanf(buf, "%d %c", &val, &delim) != 1)
 379                 return -EINVAL;
 380         if (val != 0 && val != 1)
 381                 return -EINVAL;
 382         rc = 0;
 383         get_online_cpus();
 384         mutex_lock(&smp_cpu_state_mutex);
 385         if (cpu_management == val)
 386                 goto out;
 387         rc = topology_set_cpu_management(val);
 388         if (rc)
 389                 goto out;
 390         cpu_management = val;
 391         topology_expect_change();
 392 out:
 393         mutex_unlock(&smp_cpu_state_mutex);
 394         put_online_cpus();
 395         return rc ? rc : count;
 396 }
 397 static DEVICE_ATTR(dispatching, 0644, dispatching_show,
 398                          dispatching_store);
 399
 400 static ssize_t cpu_polarization_show(struct device *dev,
 401                                      struct device_attribute *attr, char *buf)
 402 {
 403         int cpu = dev->id;
 404         ssize_t count;
 405
 406         mutex_lock(&smp_cpu_state_mutex);
 407         switch (smp_cpu_get_polarization(cpu)) {
 408         case POLARIZATION_HRZ:
 409                 count = sprintf(buf, "horizontal\n");
 410                 break;
 411         case POLARIZATION_VL:
 412                 count = sprintf(buf, "vertical:low\n");
 413                 break;
 414         case POLARIZATION_VM:
 415                 count = sprintf(buf, "vertical:medium\n");
 416                 break;
 417         case POLARIZATION_VH:
 418                 count = sprintf(buf, "vertical:high\n");
 419                 break;
 420         default:
 421                 count = sprintf(buf, "unknown\n");
 422                 break;
 423         }
 424         mutex_unlock(&smp_cpu_state_mutex);
 425         return count;
 426 }
 427 static DEVICE_ATTR(polarization, 0444, cpu_polarization_show, NULL);
 428
 429 static struct attribute *topology_cpu_attrs[] = {
 430         &dev_attr_polarization.attr,
 431         NULL,
 432 };
 433
 434 static struct attribute_group topology_cpu_attr_group = {
 435         .attrs = topology_cpu_attrs,
 436 };
 437
 438 int topology_cpu_init(struct cpu *cpu)
 439 {
 440         return sysfs_create_group(&cpu->dev.kobj, &topology_cpu_attr_group);
 441 }
 442
 443 static const struct cpumask *cpu_thread_mask(int cpu)
 444 {
 445         return &per_cpu(cpu_topology, cpu).thread_mask;
 446 }
 447
 448
 449 const struct cpumask *cpu_coregroup_mask(int cpu)
 450 {
 451         return &per_cpu(cpu_topology, cpu).core_mask;
 452 }
 453
 454 static const struct cpumask *cpu_book_mask(int cpu)
 455 {
 456         return &per_cpu(cpu_topology, cpu).book_mask;
 457 }
 458
 459 static const struct cpumask *cpu_drawer_mask(int cpu)
 460 {
 461         return &per_cpu(cpu_topology, cpu).drawer_mask;
 462 }
 463
 464 static int __init early_parse_topology(char *p)
 465 {
 466         return kstrtobool(p, &topology_enabled);
 467 }
 468 early_param("topology", early_parse_topology);
 469
 470 static struct sched_domain_topology_level s390_topology[] = {
 471         { cpu_thread_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
 472         { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
 473         { cpu_book_mask, SD_INIT_NAME(BOOK) },
 474         { cpu_drawer_mask, SD_INIT_NAME(DRAWER) },
 475         { cpu_cpu_mask, SD_INIT_NAME(DIE) },
 476         { NULL, },
 477 };
 478
 479 static void __init alloc_masks(struct sysinfo_15_1_x *info,
 480                                struct mask_info *mask, int offset)
 481 {
 482         int i, nr_masks;
 483
 484         nr_masks = info->mag[TOPOLOGY_NR_MAG - offset];
 485         for (i = 0; i < info->mnest - offset; i++)
 486                 nr_masks *= info->mag[TOPOLOGY_NR_MAG - offset - 1 - i];
 487         nr_masks = max(nr_masks, 1);
 488         for (i = 0; i < nr_masks; i++) {
 489                 mask->next = kzalloc(sizeof(*mask->next), GFP_KERNEL);
 490                 mask = mask->next;
 491         }
 492 }
 493
 494 static int __init s390_topology_init(void)
 495 {
 496         struct sysinfo_15_1_x *info;
 497         int i;
 498
 499         if (!MACHINE_HAS_TOPOLOGY)
 500                 return 0;
 501         tl_info = (struct sysinfo_15_1_x *)__get_free_page(GFP_KERNEL);
 502         info = tl_info;
 503         store_topology(info);
 504         pr_info("The CPU configuration topology of the machine is:");
 505         for (i = 0; i < TOPOLOGY_NR_MAG; i++)
 506                 printk(KERN_CONT " %d", info->mag[i]);
 507         printk(KERN_CONT " / %d\n", info->mnest);
 508         alloc_masks(info, &socket_info, 1);
 509         alloc_masks(info, &book_info, 2);
 510         alloc_masks(info, &drawer_info, 3);
 511         set_sched_topology(s390_topology);
 512         return 0;
 513 }
 514 early_initcall(s390_topology_init);
 515
 516 static int __init topology_init(void)
 517 {
 518         if (MACHINE_HAS_TOPOLOGY)
 519                 set_topology_timer();
 520         else
 521                 topology_update_polarization_simple();
 522         return device_create_file(cpu_subsys.dev_root, &dev_attr_dispatching);
 523 }
 524 device_initcall(topology_init);