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/bootmem.h>
  11 #include <linux/uaccess.h>
  12 #include <linux/sysctl.h>
  13 #include <linux/cpuset.h>
  14 #include <linux/device.h>
  15 #include <linux/export.h>
  16 #include <linux/kernel.h>
  17 #include <linux/sched.h>
  18 #include <linux/sched/topology.h>
  19 #include <linux/delay.h>
  20 #include <linux/init.h>
  21 #include <linux/slab.h>
  22 #include <linux/cpu.h>
  23 #include <linux/smp.h>
  24 #include <linux/mm.h>
  25 #include <linux/nodemask.h>
  26 #include <linux/node.h>
  27 #include <asm/sysinfo.h>
  28 #include <asm/numa.h>
  29
  30 #define PTF_HORIZONTAL  (0UL)
  31 #define PTF_VERTICAL    (1UL)
  32 #define PTF_CHECK       (2UL)
  33
  34 enum {
  35         TOPOLOGY_MODE_HW,
  36         TOPOLOGY_MODE_SINGLE,
  37         TOPOLOGY_MODE_PACKAGE,
  38         TOPOLOGY_MODE_UNINITIALIZED
  39 };
  40
  41 struct mask_info {
  42         struct mask_info *next;
  43         unsigned char id;
  44         cpumask_t mask;
  45 };
  46
  47 static int topology_mode = TOPOLOGY_MODE_UNINITIALIZED;
  48 static void set_topology_timer(void);
  49 static void topology_work_fn(struct work_struct *work);
  50 static struct sysinfo_15_1_x *tl_info;
  51
  52 static DECLARE_WORK(topology_work, topology_work_fn);
  53
  54 /*
  55  * Socket/Book linked lists and cpu_topology updates are
  56  * protected by "sched_domains_mutex".
  57  */
  58 static struct mask_info socket_info;
  59 static struct mask_info book_info;
  60 static struct mask_info drawer_info;
  61
  62 struct cpu_topology_s390 cpu_topology[NR_CPUS];
  63 EXPORT_SYMBOL_GPL(cpu_topology);
  64
  65 cpumask_t cpus_with_topology;
  66
  67 static cpumask_t cpu_group_map(struct mask_info *info, unsigned int cpu)
  68 {
  69         cpumask_t mask;
  70
  71         cpumask_copy(&mask, cpumask_of(cpu));
  72         switch (topology_mode) {
  73         case TOPOLOGY_MODE_HW:
  74                 while (info) {
  75                         if (cpumask_test_cpu(cpu, &info->mask)) {
  76                                 mask = info->mask;
  77                                 break;
  78                         }
  79                         info = info->next;
  80                 }
  81                 if (cpumask_empty(&mask))
  82                         cpumask_copy(&mask, cpumask_of(cpu));
  83                 break;
  84         case TOPOLOGY_MODE_PACKAGE:
  85                 cpumask_copy(&mask, cpu_present_mask);
  86                 break;
  87         default:
  88                 /* fallthrough */
  89         case TOPOLOGY_MODE_SINGLE:
  90                 cpumask_copy(&mask, cpumask_of(cpu));
  91                 break;
  92         }
  93         return mask;
  94 }
  95
  96 static cpumask_t cpu_thread_map(unsigned int cpu)
  97 {
  98         cpumask_t mask;
  99         int i;
 100
 101         cpumask_copy(&mask, cpumask_of(cpu));
 102         if (topology_mode != TOPOLOGY_MODE_HW)
 103                 return mask;
 104         cpu -= cpu % (smp_cpu_mtid + 1);
 105         for (i = 0; i <= smp_cpu_mtid; i++)
 106                 if (cpu_present(cpu + i))
 107                         cpumask_set_cpu(cpu + i, &mask);
 108         return mask;
 109 }
 110
 111 #define TOPOLOGY_CORE_BITS      64
 112
 113 static void add_cpus_to_mask(struct topology_core *tl_core,
 114                              struct mask_info *drawer,
 115                              struct mask_info *book,
 116                              struct mask_info *socket)
 117 {
 118         struct cpu_topology_s390 *topo;
 119         unsigned int core;
 120
 121         for_each_set_bit(core, &tl_core->mask, TOPOLOGY_CORE_BITS) {
 122                 unsigned int rcore;
 123                 int lcpu, i;
 124
 125                 rcore = TOPOLOGY_CORE_BITS - 1 - core + tl_core->origin;
 126                 lcpu = smp_find_processor_id(rcore << smp_cpu_mt_shift);
 127                 if (lcpu < 0)
 128                         continue;
 129                 for (i = 0; i <= smp_cpu_mtid; i++) {
 130                         topo = &cpu_topology[lcpu + i];
 131                         topo->drawer_id = drawer->id;
 132                         topo->book_id = book->id;
 133                         topo->socket_id = socket->id;
 134                         topo->core_id = rcore;
 135                         topo->thread_id = lcpu + i;
 136                         cpumask_set_cpu(lcpu + i, &drawer->mask);
 137                         cpumask_set_cpu(lcpu + i, &book->mask);
 138                         cpumask_set_cpu(lcpu + i, &socket->mask);
 139                         cpumask_set_cpu(lcpu + i, &cpus_with_topology);
 140                         smp_cpu_set_polarization(lcpu + i, tl_core->pp);
 141                 }
 142         }
 143 }
 144
 145 static void clear_masks(void)
 146 {
 147         struct mask_info *info;
 148
 149         info = &socket_info;
 150         while (info) {
 151                 cpumask_clear(&info->mask);
 152                 info = info->next;
 153         }
 154         info = &book_info;
 155         while (info) {
 156                 cpumask_clear(&info->mask);
 157                 info = info->next;
 158         }
 159         info = &drawer_info;
 160         while (info) {
 161                 cpumask_clear(&info->mask);
 162                 info = info->next;
 163         }
 164 }
 165
 166 static union topology_entry *next_tle(union topology_entry *tle)
 167 {
 168         if (!tle->nl)
 169                 return (union topology_entry *)((struct topology_core *)tle + 1);
 170         return (union topology_entry *)((struct topology_container *)tle + 1);
 171 }
 172
 173 static void tl_to_masks(struct sysinfo_15_1_x *info)
 174 {
 175         struct mask_info *socket = &socket_info;
 176         struct mask_info *book = &book_info;
 177         struct mask_info *drawer = &drawer_info;
 178         union topology_entry *tle, *end;
 179
 180         clear_masks();
 181         tle = info->tle;
 182         end = (union topology_entry *)((unsigned long)info + info->length);
 183         while (tle < end) {
 184                 switch (tle->nl) {
 185                 case 3:
 186                         drawer = drawer->next;
 187                         drawer->id = tle->container.id;
 188                         break;
 189                 case 2:
 190                         book = book->next;
 191                         book->id = tle->container.id;
 192                         break;
 193                 case 1:
 194                         socket = socket->next;
 195                         socket->id = tle->container.id;
 196                         break;
 197                 case 0:
 198                         add_cpus_to_mask(&tle->cpu, drawer, book, socket);
 199                         break;
 200                 default:
 201                         clear_masks();
 202                         return;
 203                 }
 204                 tle = next_tle(tle);
 205         }
 206 }
 207
 208 static void topology_update_polarization_simple(void)
 209 {
 210         int cpu;
 211
 212         for_each_possible_cpu(cpu)
 213                 smp_cpu_set_polarization(cpu, POLARIZATION_HRZ);
 214 }
 215
 216 static int ptf(unsigned long fc)
 217 {
 218         int rc;
 219
 220         asm volatile(
 221                 "       .insn   rre,0xb9a20000,%1,%1\n"
 222                 "       ipm     %0\n"
 223                 "       srl     %0,28\n"
 224                 : "=d" (rc)
 225                 : "d" (fc)  : "cc");
 226         return rc;
 227 }
 228
 229 int topology_set_cpu_management(int fc)
 230 {
 231         int cpu, rc;
 232
 233         if (!MACHINE_HAS_TOPOLOGY)
 234                 return -EOPNOTSUPP;
 235         if (fc)
 236                 rc = ptf(PTF_VERTICAL);
 237         else
 238                 rc = ptf(PTF_HORIZONTAL);
 239         if (rc)
 240                 return -EBUSY;
 241         for_each_possible_cpu(cpu)
 242                 smp_cpu_set_polarization(cpu, POLARIZATION_UNKNOWN);
 243         return rc;
 244 }
 245
 246 static void update_cpu_masks(void)
 247 {
 248         struct cpu_topology_s390 *topo;
 249         int cpu, id;
 250
 251         for_each_possible_cpu(cpu) {
 252                 topo = &cpu_topology[cpu];
 253                 topo->thread_mask = cpu_thread_map(cpu);
 254                 topo->core_mask = cpu_group_map(&socket_info, cpu);
 255                 topo->book_mask = cpu_group_map(&book_info, cpu);
 256                 topo->drawer_mask = cpu_group_map(&drawer_info, cpu);
 257                 if (topology_mode != TOPOLOGY_MODE_HW) {
 258                         id = topology_mode == TOPOLOGY_MODE_PACKAGE ? 0 : cpu;
 259                         topo->thread_id = cpu;
 260                         topo->core_id = cpu;
 261                         topo->socket_id = id;
 262                         topo->book_id = id;
 263                         topo->drawer_id = id;
 264                         if (cpu_present(cpu))
 265                                 cpumask_set_cpu(cpu, &cpus_with_topology);
 266                 }
 267         }
 268         numa_update_cpu_topology();
 269 }
 270
 271 void store_topology(struct sysinfo_15_1_x *info)
 272 {
 273         stsi(info, 15, 1, topology_mnest_limit());
 274 }
 275
 276 static int __arch_update_cpu_topology(void)
 277 {
 278         struct sysinfo_15_1_x *info = tl_info;
 279         int rc = 0;
 280
 281         mutex_lock(&smp_cpu_state_mutex);
 282         cpumask_clear(&cpus_with_topology);
 283         if (MACHINE_HAS_TOPOLOGY) {
 284                 rc = 1;
 285                 store_topology(info);
 286                 tl_to_masks(info);
 287         }
 288         update_cpu_masks();
 289         if (!MACHINE_HAS_TOPOLOGY)
 290                 topology_update_polarization_simple();
 291         mutex_unlock(&smp_cpu_state_mutex);
 292         return rc;
 293 }
 294
 295 int arch_update_cpu_topology(void)
 296 {
 297         struct device *dev;
 298         int cpu, rc;
 299
 300         rc = __arch_update_cpu_topology();
 301         for_each_online_cpu(cpu) {
 302                 dev = get_cpu_device(cpu);
 303                 if (dev)
 304                         kobject_uevent(&dev->kobj, KOBJ_CHANGE);
 305         }
 306         return rc;
 307 }
 308
 309 static void topology_work_fn(struct work_struct *work)
 310 {
 311         rebuild_sched_domains();
 312 }
 313
 314 void topology_schedule_update(void)
 315 {
 316         schedule_work(&topology_work);
 317 }
 318
 319 static void topology_flush_work(void)
 320 {
 321         flush_work(&topology_work);
 322 }
 323
 324 static void topology_timer_fn(unsigned long ignored)
 325 {
 326         if (ptf(PTF_CHECK))
 327                 topology_schedule_update();
 328         set_topology_timer();
 329 }
 330
 331 static struct timer_list topology_timer =
 332         TIMER_DEFERRED_INITIALIZER(topology_timer_fn, 0, 0);
 333
 334 static atomic_t topology_poll = ATOMIC_INIT(0);
 335
 336 static void set_topology_timer(void)
 337 {
 338         if (atomic_add_unless(&topology_poll, -1, 0))
 339                 mod_timer(&topology_timer, jiffies + HZ / 10);
 340         else
 341                 mod_timer(&topology_timer, jiffies + HZ * 60);
 342 }
 343
 344 void topology_expect_change(void)
 345 {
 346         if (!MACHINE_HAS_TOPOLOGY)
 347                 return;
 348         /* This is racy, but it doesn't matter since it is just a heuristic.
 349          * Worst case is that we poll in a higher frequency for a bit longer.
 350          */
 351         if (atomic_read(&topology_poll) > 60)
 352                 return;
 353         atomic_add(60, &topology_poll);
 354         set_topology_timer();
 355 }
 356
 357 static int cpu_management;
 358
 359 static ssize_t dispatching_show(struct device *dev,
 360                                 struct device_attribute *attr,
 361                                 char *buf)
 362 {
 363         ssize_t count;
 364
 365         mutex_lock(&smp_cpu_state_mutex);
 366         count = sprintf(buf, "%d\n", cpu_management);
 367         mutex_unlock(&smp_cpu_state_mutex);
 368         return count;
 369 }
 370
 371 static ssize_t dispatching_store(struct device *dev,
 372                                  struct device_attribute *attr,
 373                                  const char *buf,
 374                                  size_t count)
 375 {
 376         int val, rc;
 377         char delim;
 378
 379         if (sscanf(buf, "%d %c", &val, &delim) != 1)
 380                 return -EINVAL;
 381         if (val != 0 && val != 1)
 382                 return -EINVAL;
 383         rc = 0;
 384         get_online_cpus();
 385         mutex_lock(&smp_cpu_state_mutex);
 386         if (cpu_management == val)
 387                 goto out;
 388         rc = topology_set_cpu_management(val);
 389         if (rc)
 390                 goto out;
 391         cpu_management = val;
 392         topology_expect_change();
 393 out:
 394         mutex_unlock(&smp_cpu_state_mutex);
 395         put_online_cpus();
 396         return rc ? rc : count;
 397 }
 398 static DEVICE_ATTR(dispatching, 0644, dispatching_show,
 399                          dispatching_store);
 400
 401 static ssize_t cpu_polarization_show(struct device *dev,
 402                                      struct device_attribute *attr, char *buf)
 403 {
 404         int cpu = dev->id;
 405         ssize_t count;
 406
 407         mutex_lock(&smp_cpu_state_mutex);
 408         switch (smp_cpu_get_polarization(cpu)) {
 409         case POLARIZATION_HRZ:
 410                 count = sprintf(buf, "horizontal\n");
 411                 break;
 412         case POLARIZATION_VL:
 413                 count = sprintf(buf, "vertical:low\n");
 414                 break;
 415         case POLARIZATION_VM:
 416                 count = sprintf(buf, "vertical:medium\n");
 417                 break;
 418         case POLARIZATION_VH:
 419                 count = sprintf(buf, "vertical:high\n");
 420                 break;
 421         default:
 422                 count = sprintf(buf, "unknown\n");
 423                 break;
 424         }
 425         mutex_unlock(&smp_cpu_state_mutex);
 426         return count;
 427 }
 428 static DEVICE_ATTR(polarization, 0444, cpu_polarization_show, NULL);
 429
 430 static struct attribute *topology_cpu_attrs[] = {
 431         &dev_attr_polarization.attr,
 432         NULL,
 433 };
 434
 435 static struct attribute_group topology_cpu_attr_group = {
 436         .attrs = topology_cpu_attrs,
 437 };
 438
 439 int topology_cpu_init(struct cpu *cpu)
 440 {
 441         return sysfs_create_group(&cpu->dev.kobj, &topology_cpu_attr_group);
 442 }
 443
 444 static const struct cpumask *cpu_thread_mask(int cpu)
 445 {
 446         return &cpu_topology[cpu].thread_mask;
 447 }
 448
 449
 450 const struct cpumask *cpu_coregroup_mask(int cpu)
 451 {
 452         return &cpu_topology[cpu].core_mask;
 453 }
 454
 455 static const struct cpumask *cpu_book_mask(int cpu)
 456 {
 457         return &cpu_topology[cpu].book_mask;
 458 }
 459
 460 static const struct cpumask *cpu_drawer_mask(int cpu)
 461 {
 462         return &cpu_topology[cpu].drawer_mask;
 463 }
 464
 465 static struct sched_domain_topology_level s390_topology[] = {
 466         { cpu_thread_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
 467         { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
 468         { cpu_book_mask, SD_INIT_NAME(BOOK) },
 469         { cpu_drawer_mask, SD_INIT_NAME(DRAWER) },
 470         { cpu_cpu_mask, SD_INIT_NAME(DIE) },
 471         { NULL, },
 472 };
 473
 474 static void __init alloc_masks(struct sysinfo_15_1_x *info,
 475                                struct mask_info *mask, int offset)
 476 {
 477         int i, nr_masks;
 478
 479         nr_masks = info->mag[TOPOLOGY_NR_MAG - offset];
 480         for (i = 0; i < info->mnest - offset; i++)
 481                 nr_masks *= info->mag[TOPOLOGY_NR_MAG - offset - 1 - i];
 482         nr_masks = max(nr_masks, 1);
 483         for (i = 0; i < nr_masks; i++) {
 484                 mask->next = memblock_virt_alloc(sizeof(*mask->next), 8);
 485                 mask = mask->next;
 486         }
 487 }
 488
 489 void __init topology_init_early(void)
 490 {
 491         struct sysinfo_15_1_x *info;
 492
 493         set_sched_topology(s390_topology);
 494         if (topology_mode == TOPOLOGY_MODE_UNINITIALIZED) {
 495                 if (MACHINE_HAS_TOPOLOGY)
 496                         topology_mode = TOPOLOGY_MODE_HW;
 497                 else
 498                         topology_mode = TOPOLOGY_MODE_SINGLE;
 499         }
 500         if (!MACHINE_HAS_TOPOLOGY)
 501                 goto out;
 502         tl_info = memblock_virt_alloc(PAGE_SIZE, PAGE_SIZE);
 503         info = tl_info;
 504         store_topology(info);
 505         pr_info("The CPU configuration topology of the machine is: %d %d %d %d %d %d / %d\n",
 506                 info->mag[0], info->mag[1], info->mag[2], info->mag[3],
 507                 info->mag[4], info->mag[5], info->mnest);
 508         alloc_masks(info, &socket_info, 1);
 509         alloc_masks(info, &book_info, 2);
 510         alloc_masks(info, &drawer_info, 3);
 511 out:
 512         __arch_update_cpu_topology();
 513 }
 514
 515 static inline int topology_get_mode(int enabled)
 516 {
 517         if (!enabled)
 518                 return TOPOLOGY_MODE_SINGLE;
 519         return MACHINE_HAS_TOPOLOGY ? TOPOLOGY_MODE_HW : TOPOLOGY_MODE_PACKAGE;
 520 }
 521
 522 static inline int topology_is_enabled(void)
 523 {
 524         return topology_mode != TOPOLOGY_MODE_SINGLE;
 525 }
 526
 527 static int __init topology_setup(char *str)
 528 {
 529         bool enabled;
 530         int rc;
 531
 532         rc = kstrtobool(str, &enabled);
 533         if (rc)
 534                 return rc;
 535         topology_mode = topology_get_mode(enabled);
 536         return 0;
 537 }
 538 early_param("topology", topology_setup);
 539
 540 static int topology_ctl_handler(struct ctl_table *ctl, int write,
 541                                 void __user *buffer, size_t *lenp, loff_t *ppos)
 542 {
 543         unsigned int len;
 544         int new_mode;
 545         char buf[2];
 546
 547         if (!*lenp || *ppos) {
 548                 *lenp = 0;
 549                 return 0;
 550         }
 551         if (!write) {
 552                 strncpy(buf, topology_is_enabled() ? "1\n" : "0\n",
 553                         ARRAY_SIZE(buf));
 554                 len = strnlen(buf, ARRAY_SIZE(buf));
 555                 if (len > *lenp)
 556                         len = *lenp;
 557                 if (copy_to_user(buffer, buf, len))
 558                         return -EFAULT;
 559                 goto out;
 560         }
 561         len = *lenp;
 562         if (copy_from_user(buf, buffer, len > sizeof(buf) ? sizeof(buf) : len))
 563                 return -EFAULT;
 564         if (buf[0] != '0' && buf[0] != '1')
 565                 return -EINVAL;
 566         mutex_lock(&smp_cpu_state_mutex);
 567         new_mode = topology_get_mode(buf[0] == '1');
 568         if (topology_mode != new_mode) {
 569                 topology_mode = new_mode;
 570                 topology_schedule_update();
 571         }
 572         mutex_unlock(&smp_cpu_state_mutex);
 573         topology_flush_work();
 574 out:
 575         *lenp = len;
 576         *ppos += len;
 577         return 0;
 578 }
 579
 580 static struct ctl_table topology_ctl_table[] = {
 581         {
 582                 .procname       = "topology",
 583                 .mode           = 0644,
 584                 .proc_handler   = topology_ctl_handler,
 585         },
 586         { },
 587 };
 588
 589 static struct ctl_table topology_dir_table[] = {
 590         {
 591                 .procname       = "s390",
 592                 .maxlen         = 0,
 593                 .mode           = 0555,
 594                 .child          = topology_ctl_table,
 595         },
 596         { },
 597 };
 598
 599 static int __init topology_init(void)
 600 {
 601         if (MACHINE_HAS_TOPOLOGY)
 602                 set_topology_timer();
 603         else
 604                 topology_update_polarization_simple();
 605         register_sysctl_table(topology_dir_table);
 606         return device_create_file(cpu_subsys.dev_root, &dev_attr_dispatching);
 607 }
 608 device_initcall(topology_init);