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                         topo->dedicated = tl_core->d;
 137                         cpumask_set_cpu(lcpu + i, &drawer->mask);
 138                         cpumask_set_cpu(lcpu + i, &book->mask);
 139                         cpumask_set_cpu(lcpu + i, &socket->mask);
 140                         cpumask_set_cpu(lcpu + i, &cpus_with_topology);
 141                         smp_cpu_set_polarization(lcpu + i, tl_core->pp);
 142                 }
 143         }
 144 }
 145
 146 static void clear_masks(void)
 147 {
 148         struct mask_info *info;
 149
 150         info = &socket_info;
 151         while (info) {
 152                 cpumask_clear(&info->mask);
 153                 info = info->next;
 154         }
 155         info = &book_info;
 156         while (info) {
 157                 cpumask_clear(&info->mask);
 158                 info = info->next;
 159         }
 160         info = &drawer_info;
 161         while (info) {
 162                 cpumask_clear(&info->mask);
 163                 info = info->next;
 164         }
 165 }
 166
 167 static union topology_entry *next_tle(union topology_entry *tle)
 168 {
 169         if (!tle->nl)
 170                 return (union topology_entry *)((struct topology_core *)tle + 1);
 171         return (union topology_entry *)((struct topology_container *)tle + 1);
 172 }
 173
 174 static void tl_to_masks(struct sysinfo_15_1_x *info)
 175 {
 176         struct mask_info *socket = &socket_info;
 177         struct mask_info *book = &book_info;
 178         struct mask_info *drawer = &drawer_info;
 179         union topology_entry *tle, *end;
 180
 181         clear_masks();
 182         tle = info->tle;
 183         end = (union topology_entry *)((unsigned long)info + info->length);
 184         while (tle < end) {
 185                 switch (tle->nl) {
 186                 case 3:
 187                         drawer = drawer->next;
 188                         drawer->id = tle->container.id;
 189                         break;
 190                 case 2:
 191                         book = book->next;
 192                         book->id = tle->container.id;
 193                         break;
 194                 case 1:
 195                         socket = socket->next;
 196                         socket->id = tle->container.id;
 197                         break;
 198                 case 0:
 199                         add_cpus_to_mask(&tle->cpu, drawer, book, socket);
 200                         break;
 201                 default:
 202                         clear_masks();
 203                         return;
 204                 }
 205                 tle = next_tle(tle);
 206         }
 207 }
 208
 209 static void topology_update_polarization_simple(void)
 210 {
 211         int cpu;
 212
 213         for_each_possible_cpu(cpu)
 214                 smp_cpu_set_polarization(cpu, POLARIZATION_HRZ);
 215 }
 216
 217 static int ptf(unsigned long fc)
 218 {
 219         int rc;
 220
 221         asm volatile(
 222                 "       .insn   rre,0xb9a20000,%1,%1\n"
 223                 "       ipm     %0\n"
 224                 "       srl     %0,28\n"
 225                 : "=d" (rc)
 226                 : "d" (fc)  : "cc");
 227         return rc;
 228 }
 229
 230 int topology_set_cpu_management(int fc)
 231 {
 232         int cpu, rc;
 233
 234         if (!MACHINE_HAS_TOPOLOGY)
 235                 return -EOPNOTSUPP;
 236         if (fc)
 237                 rc = ptf(PTF_VERTICAL);
 238         else
 239                 rc = ptf(PTF_HORIZONTAL);
 240         if (rc)
 241                 return -EBUSY;
 242         for_each_possible_cpu(cpu)
 243                 smp_cpu_set_polarization(cpu, POLARIZATION_UNKNOWN);
 244         return rc;
 245 }
 246
 247 static void update_cpu_masks(void)
 248 {
 249         struct cpu_topology_s390 *topo;
 250         int cpu, id;
 251
 252         for_each_possible_cpu(cpu) {
 253                 topo = &cpu_topology[cpu];
 254                 topo->thread_mask = cpu_thread_map(cpu);
 255                 topo->core_mask = cpu_group_map(&socket_info, cpu);
 256                 topo->book_mask = cpu_group_map(&book_info, cpu);
 257                 topo->drawer_mask = cpu_group_map(&drawer_info, cpu);
 258                 if (topology_mode != TOPOLOGY_MODE_HW) {
 259                         id = topology_mode == TOPOLOGY_MODE_PACKAGE ? 0 : cpu;
 260                         topo->thread_id = cpu;
 261                         topo->core_id = cpu;
 262                         topo->socket_id = id;
 263                         topo->book_id = id;
 264                         topo->drawer_id = id;
 265                         if (cpu_present(cpu))
 266                                 cpumask_set_cpu(cpu, &cpus_with_topology);
 267                 }
 268         }
 269         numa_update_cpu_topology();
 270 }
 271
 272 void store_topology(struct sysinfo_15_1_x *info)
 273 {
 274         stsi(info, 15, 1, topology_mnest_limit());
 275 }
 276
 277 static void __arch_update_dedicated_flag(void *arg)
 278 {
 279         if (topology_cpu_dedicated(smp_processor_id()))
 280                 set_cpu_flag(CIF_DEDICATED_CPU);
 281         else
 282                 clear_cpu_flag(CIF_DEDICATED_CPU);
 283 }
 284
 285 static int __arch_update_cpu_topology(void)
 286 {
 287         struct sysinfo_15_1_x *info = tl_info;
 288         int rc = 0;
 289
 290         mutex_lock(&smp_cpu_state_mutex);
 291         cpumask_clear(&cpus_with_topology);
 292         if (MACHINE_HAS_TOPOLOGY) {
 293                 rc = 1;
 294                 store_topology(info);
 295                 tl_to_masks(info);
 296         }
 297         update_cpu_masks();
 298         if (!MACHINE_HAS_TOPOLOGY)
 299                 topology_update_polarization_simple();
 300         mutex_unlock(&smp_cpu_state_mutex);
 301         return rc;
 302 }
 303
 304 int arch_update_cpu_topology(void)
 305 {
 306         struct device *dev;
 307         int cpu, rc;
 308
 309         rc = __arch_update_cpu_topology();
 310         on_each_cpu(__arch_update_dedicated_flag, NULL, 0);
 311         for_each_online_cpu(cpu) {
 312                 dev = get_cpu_device(cpu);
 313                 kobject_uevent(&dev->kobj, KOBJ_CHANGE);
 314         }
 315         return rc;
 316 }
 317
 318 static void topology_work_fn(struct work_struct *work)
 319 {
 320         rebuild_sched_domains();
 321 }
 322
 323 void topology_schedule_update(void)
 324 {
 325         schedule_work(&topology_work);
 326 }
 327
 328 static void topology_flush_work(void)
 329 {
 330         flush_work(&topology_work);
 331 }
 332
 333 static void topology_timer_fn(struct timer_list *unused)
 334 {
 335         if (ptf(PTF_CHECK))
 336                 topology_schedule_update();
 337         set_topology_timer();
 338 }
 339
 340 static struct timer_list topology_timer;
 341
 342 static atomic_t topology_poll = ATOMIC_INIT(0);
 343
 344 static void set_topology_timer(void)
 345 {
 346         if (atomic_add_unless(&topology_poll, -1, 0))
 347                 mod_timer(&topology_timer, jiffies + HZ / 10);
 348         else
 349                 mod_timer(&topology_timer, jiffies + HZ * 60);
 350 }
 351
 352 void topology_expect_change(void)
 353 {
 354         if (!MACHINE_HAS_TOPOLOGY)
 355                 return;
 356         /* This is racy, but it doesn't matter since it is just a heuristic.
 357          * Worst case is that we poll in a higher frequency for a bit longer.
 358          */
 359         if (atomic_read(&topology_poll) > 60)
 360                 return;
 361         atomic_add(60, &topology_poll);
 362         set_topology_timer();
 363 }
 364
 365 static int cpu_management;
 366
 367 static ssize_t dispatching_show(struct device *dev,
 368                                 struct device_attribute *attr,
 369                                 char *buf)
 370 {
 371         ssize_t count;
 372
 373         mutex_lock(&smp_cpu_state_mutex);
 374         count = sprintf(buf, "%d\n", cpu_management);
 375         mutex_unlock(&smp_cpu_state_mutex);
 376         return count;
 377 }
 378
 379 static ssize_t dispatching_store(struct device *dev,
 380                                  struct device_attribute *attr,
 381                                  const char *buf,
 382                                  size_t count)
 383 {
 384         int val, rc;
 385         char delim;
 386
 387         if (sscanf(buf, "%d %c", &val, &delim) != 1)
 388                 return -EINVAL;
 389         if (val != 0 && val != 1)
 390                 return -EINVAL;
 391         rc = 0;
 392         get_online_cpus();
 393         mutex_lock(&smp_cpu_state_mutex);
 394         if (cpu_management == val)
 395                 goto out;
 396         rc = topology_set_cpu_management(val);
 397         if (rc)
 398                 goto out;
 399         cpu_management = val;
 400         topology_expect_change();
 401 out:
 402         mutex_unlock(&smp_cpu_state_mutex);
 403         put_online_cpus();
 404         return rc ? rc : count;
 405 }
 406 static DEVICE_ATTR(dispatching, 0644, dispatching_show,
 407                          dispatching_store);
 408
 409 static ssize_t cpu_polarization_show(struct device *dev,
 410                                      struct device_attribute *attr, char *buf)
 411 {
 412         int cpu = dev->id;
 413         ssize_t count;
 414
 415         mutex_lock(&smp_cpu_state_mutex);
 416         switch (smp_cpu_get_polarization(cpu)) {
 417         case POLARIZATION_HRZ:
 418                 count = sprintf(buf, "horizontal\n");
 419                 break;
 420         case POLARIZATION_VL:
 421                 count = sprintf(buf, "vertical:low\n");
 422                 break;
 423         case POLARIZATION_VM:
 424                 count = sprintf(buf, "vertical:medium\n");
 425                 break;
 426         case POLARIZATION_VH:
 427                 count = sprintf(buf, "vertical:high\n");
 428                 break;
 429         default:
 430                 count = sprintf(buf, "unknown\n");
 431                 break;
 432         }
 433         mutex_unlock(&smp_cpu_state_mutex);
 434         return count;
 435 }
 436 static DEVICE_ATTR(polarization, 0444, cpu_polarization_show, NULL);
 437
 438 static struct attribute *topology_cpu_attrs[] = {
 439         &dev_attr_polarization.attr,
 440         NULL,
 441 };
 442
 443 static struct attribute_group topology_cpu_attr_group = {
 444         .attrs = topology_cpu_attrs,
 445 };
 446
 447 static ssize_t cpu_dedicated_show(struct device *dev,
 448                                   struct device_attribute *attr, char *buf)
 449 {
 450         int cpu = dev->id;
 451         ssize_t count;
 452
 453         mutex_lock(&smp_cpu_state_mutex);
 454         count = sprintf(buf, "%d\n", topology_cpu_dedicated(cpu));
 455         mutex_unlock(&smp_cpu_state_mutex);
 456         return count;
 457 }
 458 static DEVICE_ATTR(dedicated, 0444, cpu_dedicated_show, NULL);
 459
 460 static struct attribute *topology_extra_cpu_attrs[] = {
 461         &dev_attr_dedicated.attr,
 462         NULL,
 463 };
 464
 465 static struct attribute_group topology_extra_cpu_attr_group = {
 466         .attrs = topology_extra_cpu_attrs,
 467 };
 468
 469 int topology_cpu_init(struct cpu *cpu)
 470 {
 471         int rc;
 472
 473         rc = sysfs_create_group(&cpu->dev.kobj, &topology_cpu_attr_group);
 474         if (rc || !MACHINE_HAS_TOPOLOGY)
 475                 return rc;
 476         rc = sysfs_create_group(&cpu->dev.kobj, &topology_extra_cpu_attr_group);
 477         if (rc)
 478                 sysfs_remove_group(&cpu->dev.kobj, &topology_cpu_attr_group);
 479         return rc;
 480 }
 481
 482 static const struct cpumask *cpu_thread_mask(int cpu)
 483 {
 484         return &cpu_topology[cpu].thread_mask;
 485 }
 486
 487
 488 const struct cpumask *cpu_coregroup_mask(int cpu)
 489 {
 490         return &cpu_topology[cpu].core_mask;
 491 }
 492
 493 static const struct cpumask *cpu_book_mask(int cpu)
 494 {
 495         return &cpu_topology[cpu].book_mask;
 496 }
 497
 498 static const struct cpumask *cpu_drawer_mask(int cpu)
 499 {
 500         return &cpu_topology[cpu].drawer_mask;
 501 }
 502
 503 static struct sched_domain_topology_level s390_topology[] = {
 504         { cpu_thread_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
 505         { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
 506         { cpu_book_mask, SD_INIT_NAME(BOOK) },
 507         { cpu_drawer_mask, SD_INIT_NAME(DRAWER) },
 508         { cpu_cpu_mask, SD_INIT_NAME(DIE) },
 509         { NULL, },
 510 };
 511
 512 static void __init alloc_masks(struct sysinfo_15_1_x *info,
 513                                struct mask_info *mask, int offset)
 514 {
 515         int i, nr_masks;
 516
 517         nr_masks = info->mag[TOPOLOGY_NR_MAG - offset];
 518         for (i = 0; i < info->mnest - offset; i++)
 519                 nr_masks *= info->mag[TOPOLOGY_NR_MAG - offset - 1 - i];
 520         nr_masks = max(nr_masks, 1);
 521         for (i = 0; i < nr_masks; i++) {
 522                 mask->next = memblock_virt_alloc(sizeof(*mask->next), 8);
 523                 mask = mask->next;
 524         }
 525 }
 526
 527 void __init topology_init_early(void)
 528 {
 529         struct sysinfo_15_1_x *info;
 530
 531         set_sched_topology(s390_topology);
 532         if (topology_mode == TOPOLOGY_MODE_UNINITIALIZED) {
 533                 if (MACHINE_HAS_TOPOLOGY)
 534                         topology_mode = TOPOLOGY_MODE_HW;
 535                 else
 536                         topology_mode = TOPOLOGY_MODE_SINGLE;
 537         }
 538         if (!MACHINE_HAS_TOPOLOGY)
 539                 goto out;
 540         tl_info = memblock_virt_alloc(PAGE_SIZE, PAGE_SIZE);
 541         info = tl_info;
 542         store_topology(info);
 543         pr_info("The CPU configuration topology of the machine is: %d %d %d %d %d %d / %d\n",
 544                 info->mag[0], info->mag[1], info->mag[2], info->mag[3],
 545                 info->mag[4], info->mag[5], info->mnest);
 546         alloc_masks(info, &socket_info, 1);
 547         alloc_masks(info, &book_info, 2);
 548         alloc_masks(info, &drawer_info, 3);
 549 out:
 550         __arch_update_cpu_topology();
 551         __arch_update_dedicated_flag(NULL);
 552 }
 553
 554 static inline int topology_get_mode(int enabled)
 555 {
 556         if (!enabled)
 557                 return TOPOLOGY_MODE_SINGLE;
 558         return MACHINE_HAS_TOPOLOGY ? TOPOLOGY_MODE_HW : TOPOLOGY_MODE_PACKAGE;
 559 }
 560
 561 static inline int topology_is_enabled(void)
 562 {
 563         return topology_mode != TOPOLOGY_MODE_SINGLE;
 564 }
 565
 566 static int __init topology_setup(char *str)
 567 {
 568         bool enabled;
 569         int rc;
 570
 571         rc = kstrtobool(str, &enabled);
 572         if (rc)
 573                 return rc;
 574         topology_mode = topology_get_mode(enabled);
 575         return 0;
 576 }
 577 early_param("topology", topology_setup);
 578
 579 static int topology_ctl_handler(struct ctl_table *ctl, int write,
 580                                 void __user *buffer, size_t *lenp, loff_t *ppos)
 581 {
 582         unsigned int len;
 583         int new_mode;
 584         char buf[2];
 585
 586         if (!*lenp || *ppos) {
 587                 *lenp = 0;
 588                 return 0;
 589         }
 590         if (!write) {
 591                 strncpy(buf, topology_is_enabled() ? "1\n" : "0\n",
 592                         ARRAY_SIZE(buf));
 593                 len = strnlen(buf, ARRAY_SIZE(buf));
 594                 if (len > *lenp)
 595                         len = *lenp;
 596                 if (copy_to_user(buffer, buf, len))
 597                         return -EFAULT;
 598                 goto out;
 599         }
 600         len = *lenp;
 601         if (copy_from_user(buf, buffer, len > sizeof(buf) ? sizeof(buf) : len))
 602                 return -EFAULT;
 603         if (buf[0] != '0' && buf[0] != '1')
 604                 return -EINVAL;
 605         mutex_lock(&smp_cpu_state_mutex);
 606         new_mode = topology_get_mode(buf[0] == '1');
 607         if (topology_mode != new_mode) {
 608                 topology_mode = new_mode;
 609                 topology_schedule_update();
 610         }
 611         mutex_unlock(&smp_cpu_state_mutex);
 612         topology_flush_work();
 613 out:
 614         *lenp = len;
 615         *ppos += len;
 616         return 0;
 617 }
 618
 619 static struct ctl_table topology_ctl_table[] = {
 620         {
 621                 .procname       = "topology",
 622                 .mode           = 0644,
 623                 .proc_handler   = topology_ctl_handler,
 624         },
 625         { },
 626 };
 627
 628 static struct ctl_table topology_dir_table[] = {
 629         {
 630                 .procname       = "s390",
 631                 .maxlen         = 0,
 632                 .mode           = 0555,
 633                 .child          = topology_ctl_table,
 634         },
 635         { },
 636 };
 637
 638 static int __init topology_init(void)
 639 {
 640         timer_setup(&topology_timer, topology_timer_fn, TIMER_DEFERRABLE);
 641         if (MACHINE_HAS_TOPOLOGY)
 642                 set_topology_timer();
 643         else
 644                 topology_update_polarization_simple();
 645         register_sysctl_table(topology_dir_table);
 646         return device_create_file(cpu_subsys.dev_root, &dev_attr_dispatching);
 647 }
 648 device_initcall(topology_init);