arch/s390/kernel/topology.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  *    Copyright IBM Corp. 2007, 2011
   4  *    Author(s): Heiko Carstens <heiko.carstens@de.ibm.com>
   5  */
   6
   7 #define KMSG_COMPONENT "cpu"
   8 #define pr_fmt(fmt) KMSG_COMPONENT ": " fmt
   9
  10 #include <linux/workqueue.h>
  11 #include <linux/memblock.h>
  12 #include <linux/uaccess.h>
  13 #include <linux/sysctl.h>
  14 #include <linux/cpuset.h>
  15 #include <linux/device.h>
  16 #include <linux/export.h>
  17 #include <linux/kernel.h>
  18 #include <linux/sched.h>
  19 #include <linux/sched/topology.h>
  20 #include <linux/delay.h>
  21 #include <linux/init.h>
  22 #include <linux/slab.h>
  23 #include <linux/cpu.h>
  24 #include <linux/smp.h>
  25 #include <linux/mm.h>
  26 #include <linux/nodemask.h>
  27 #include <linux/node.h>
  28 #include <asm/sysinfo.h>
  29 #include <asm/numa.h>
  30
  31 #define PTF_HORIZONTAL  (0UL)
  32 #define PTF_VERTICAL    (1UL)
  33 #define PTF_CHECK       (2UL)
  34
  35 enum {
  36         TOPOLOGY_MODE_HW,
  37         TOPOLOGY_MODE_SINGLE,
  38         TOPOLOGY_MODE_PACKAGE,
  39         TOPOLOGY_MODE_UNINITIALIZED
  40 };
  41
  42 struct mask_info {
  43         struct mask_info *next;
  44         unsigned char id;
  45         cpumask_t mask;
  46 };
  47
  48 static int topology_mode = TOPOLOGY_MODE_UNINITIALIZED;
  49 static void set_topology_timer(void);
  50 static void topology_work_fn(struct work_struct *work);
  51 static struct sysinfo_15_1_x *tl_info;
  52
  53 static DECLARE_WORK(topology_work, topology_work_fn);
  54
  55 /*
  56  * Socket/Book linked lists and cpu_topology updates are
  57  * protected by "sched_domains_mutex".
  58  */
  59 static struct mask_info socket_info;
  60 static struct mask_info book_info;
  61 static struct mask_info drawer_info;
  62
  63 struct cpu_topology_s390 cpu_topology[NR_CPUS];
  64 EXPORT_SYMBOL_GPL(cpu_topology);
  65
  66 cpumask_t cpus_with_topology;
  67
  68 int __cpu_to_node(int cpu)
  69 {
  70         return cpu_topology[cpu].node_id;
  71 }
  72
  73 EXPORT_SYMBOL(__cpu_to_node);
  74
  75 static cpumask_t cpu_group_map(struct mask_info *info, unsigned int cpu)
  76 {
  77         cpumask_t mask;
  78
  79         cpumask_copy(&mask, cpumask_of(cpu));
  80         switch (topology_mode) {
  81         case TOPOLOGY_MODE_HW:
  82                 while (info) {
  83                         if (cpumask_test_cpu(cpu, &info->mask)) {
  84                                 mask = info->mask;
  85                                 break;
  86                         }
  87                         info = info->next;
  88                 }
  89                 if (cpumask_empty(&mask))
  90                         cpumask_copy(&mask, cpumask_of(cpu));
  91                 break;
  92         case TOPOLOGY_MODE_PACKAGE:
  93                 cpumask_copy(&mask, cpu_present_mask);
  94                 break;
  95         default:
  96                 /* fallthrough */
  97         case TOPOLOGY_MODE_SINGLE:
  98                 cpumask_copy(&mask, cpumask_of(cpu));
  99                 break;
 100         }
 101         return mask;
 102 }
 103
 104 static cpumask_t cpu_thread_map(unsigned int cpu)
 105 {
 106         cpumask_t mask;
 107         int i;
 108
 109         cpumask_copy(&mask, cpumask_of(cpu));
 110         if (topology_mode != TOPOLOGY_MODE_HW)
 111                 return mask;
 112         cpu -= cpu % (smp_cpu_mtid + 1);
 113         for (i = 0; i <= smp_cpu_mtid; i++)
 114                 if (cpu_present(cpu + i))
 115                         cpumask_set_cpu(cpu + i, &mask);
 116         return mask;
 117 }
 118
 119 #define TOPOLOGY_CORE_BITS      64
 120
 121 static void add_cpus_to_mask(struct topology_core *tl_core,
 122                              struct mask_info *drawer,
 123                              struct mask_info *book,
 124                              struct mask_info *socket)
 125 {
 126         struct cpu_topology_s390 *topo;
 127         unsigned int core;
 128
 129         for_each_set_bit(core, &tl_core->mask, TOPOLOGY_CORE_BITS) {
 130                 unsigned int rcore;
 131                 int lcpu, i;
 132
 133                 rcore = TOPOLOGY_CORE_BITS - 1 - core + tl_core->origin;
 134                 lcpu = smp_find_processor_id(rcore << smp_cpu_mt_shift);
 135                 if (lcpu < 0)
 136                         continue;
 137                 for (i = 0; i <= smp_cpu_mtid; i++) {
 138                         topo = &cpu_topology[lcpu + i];
 139                         topo->drawer_id = drawer->id;
 140                         topo->book_id = book->id;
 141                         topo->socket_id = socket->id;
 142                         topo->core_id = rcore;
 143                         topo->thread_id = lcpu + i;
 144                         topo->dedicated = tl_core->d;
 145                         cpumask_set_cpu(lcpu + i, &drawer->mask);
 146                         cpumask_set_cpu(lcpu + i, &book->mask);
 147                         cpumask_set_cpu(lcpu + i, &socket->mask);
 148                         cpumask_set_cpu(lcpu + i, &cpus_with_topology);
 149                         smp_cpu_set_polarization(lcpu + i, tl_core->pp);
 150                 }
 151         }
 152 }
 153
 154 static void clear_masks(void)
 155 {
 156         struct mask_info *info;
 157
 158         info = &socket_info;
 159         while (info) {
 160                 cpumask_clear(&info->mask);
 161                 info = info->next;
 162         }
 163         info = &book_info;
 164         while (info) {
 165                 cpumask_clear(&info->mask);
 166                 info = info->next;
 167         }
 168         info = &drawer_info;
 169         while (info) {
 170                 cpumask_clear(&info->mask);
 171                 info = info->next;
 172         }
 173 }
 174
 175 static union topology_entry *next_tle(union topology_entry *tle)
 176 {
 177         if (!tle->nl)
 178                 return (union topology_entry *)((struct topology_core *)tle + 1);
 179         return (union topology_entry *)((struct topology_container *)tle + 1);
 180 }
 181
 182 static void tl_to_masks(struct sysinfo_15_1_x *info)
 183 {
 184         struct mask_info *socket = &socket_info;
 185         struct mask_info *book = &book_info;
 186         struct mask_info *drawer = &drawer_info;
 187         union topology_entry *tle, *end;
 188
 189         clear_masks();
 190         tle = info->tle;
 191         end = (union topology_entry *)((unsigned long)info + info->length);
 192         while (tle < end) {
 193                 switch (tle->nl) {
 194                 case 3:
 195                         drawer = drawer->next;
 196                         drawer->id = tle->container.id;
 197                         break;
 198                 case 2:
 199                         book = book->next;
 200                         book->id = tle->container.id;
 201                         break;
 202                 case 1:
 203                         socket = socket->next;
 204                         socket->id = tle->container.id;
 205                         break;
 206                 case 0:
 207                         add_cpus_to_mask(&tle->cpu, drawer, book, socket);
 208                         break;
 209                 default:
 210                         clear_masks();
 211                         return;
 212                 }
 213                 tle = next_tle(tle);
 214         }
 215 }
 216
 217 static void topology_update_polarization_simple(void)
 218 {
 219         int cpu;
 220
 221         for_each_possible_cpu(cpu)
 222                 smp_cpu_set_polarization(cpu, POLARIZATION_HRZ);
 223 }
 224
 225 static int ptf(unsigned long fc)
 226 {
 227         int rc;
 228
 229         asm volatile(
 230                 "       .insn   rre,0xb9a20000,%1,%1\n"
 231                 "       ipm     %0\n"
 232                 "       srl     %0,28\n"
 233                 : "=d" (rc)
 234                 : "d" (fc)  : "cc");
 235         return rc;
 236 }
 237
 238 int topology_set_cpu_management(int fc)
 239 {
 240         int cpu, rc;
 241
 242         if (!MACHINE_HAS_TOPOLOGY)
 243                 return -EOPNOTSUPP;
 244         if (fc)
 245                 rc = ptf(PTF_VERTICAL);
 246         else
 247                 rc = ptf(PTF_HORIZONTAL);
 248         if (rc)
 249                 return -EBUSY;
 250         for_each_possible_cpu(cpu)
 251                 smp_cpu_set_polarization(cpu, POLARIZATION_UNKNOWN);
 252         return rc;
 253 }
 254
 255 static void update_cpu_masks(void)
 256 {
 257         struct cpu_topology_s390 *topo;
 258         int cpu, id;
 259
 260         for_each_possible_cpu(cpu) {
 261                 topo = &cpu_topology[cpu];
 262                 topo->thread_mask = cpu_thread_map(cpu);
 263                 topo->core_mask = cpu_group_map(&socket_info, cpu);
 264                 topo->book_mask = cpu_group_map(&book_info, cpu);
 265                 topo->drawer_mask = cpu_group_map(&drawer_info, cpu);
 266                 if (topology_mode != TOPOLOGY_MODE_HW) {
 267                         id = topology_mode == TOPOLOGY_MODE_PACKAGE ? 0 : cpu;
 268                         topo->thread_id = cpu;
 269                         topo->core_id = cpu;
 270                         topo->socket_id = id;
 271                         topo->book_id = id;
 272                         topo->drawer_id = id;
 273                         if (cpu_present(cpu))
 274                                 cpumask_set_cpu(cpu, &cpus_with_topology);
 275                 }
 276         }
 277         numa_update_cpu_topology();
 278 }
 279
 280 void store_topology(struct sysinfo_15_1_x *info)
 281 {
 282         stsi(info, 15, 1, topology_mnest_limit());
 283 }
 284
 285 static void __arch_update_dedicated_flag(void *arg)
 286 {
 287         if (topology_cpu_dedicated(smp_processor_id()))
 288                 set_cpu_flag(CIF_DEDICATED_CPU);
 289         else
 290                 clear_cpu_flag(CIF_DEDICATED_CPU);
 291 }
 292
 293 static int __arch_update_cpu_topology(void)
 294 {
 295         struct sysinfo_15_1_x *info = tl_info;
 296         int rc = 0;
 297
 298         mutex_lock(&smp_cpu_state_mutex);
 299         cpumask_clear(&cpus_with_topology);
 300         if (MACHINE_HAS_TOPOLOGY) {
 301                 rc = 1;
 302                 store_topology(info);
 303                 tl_to_masks(info);
 304         }
 305         update_cpu_masks();
 306         if (!MACHINE_HAS_TOPOLOGY)
 307                 topology_update_polarization_simple();
 308         mutex_unlock(&smp_cpu_state_mutex);
 309         return rc;
 310 }
 311
 312 int arch_update_cpu_topology(void)
 313 {
 314         struct device *dev;
 315         int cpu, rc;
 316
 317         rc = __arch_update_cpu_topology();
 318         on_each_cpu(__arch_update_dedicated_flag, NULL, 0);
 319         for_each_online_cpu(cpu) {
 320                 dev = get_cpu_device(cpu);
 321                 kobject_uevent(&dev->kobj, KOBJ_CHANGE);
 322         }
 323         return rc;
 324 }
 325
 326 static void topology_work_fn(struct work_struct *work)
 327 {
 328         rebuild_sched_domains();
 329 }
 330
 331 void topology_schedule_update(void)
 332 {
 333         schedule_work(&topology_work);
 334 }
 335
 336 static void topology_flush_work(void)
 337 {
 338         flush_work(&topology_work);
 339 }
 340
 341 static void topology_timer_fn(struct timer_list *unused)
 342 {
 343         if (ptf(PTF_CHECK))
 344                 topology_schedule_update();
 345         set_topology_timer();
 346 }
 347
 348 static struct timer_list topology_timer;
 349
 350 static atomic_t topology_poll = ATOMIC_INIT(0);
 351
 352 static void set_topology_timer(void)
 353 {
 354         if (atomic_add_unless(&topology_poll, -1, 0))
 355                 mod_timer(&topology_timer, jiffies + HZ / 10);
 356         else
 357                 mod_timer(&topology_timer, jiffies + HZ * 60);
 358 }
 359
 360 void topology_expect_change(void)
 361 {
 362         if (!MACHINE_HAS_TOPOLOGY)
 363                 return;
 364         /* This is racy, but it doesn't matter since it is just a heuristic.
 365          * Worst case is that we poll in a higher frequency for a bit longer.
 366          */
 367         if (atomic_read(&topology_poll) > 60)
 368                 return;
 369         atomic_add(60, &topology_poll);
 370         set_topology_timer();
 371 }
 372
 373 static int cpu_management;
 374
 375 static ssize_t dispatching_show(struct device *dev,
 376                                 struct device_attribute *attr,
 377                                 char *buf)
 378 {
 379         ssize_t count;
 380
 381         mutex_lock(&smp_cpu_state_mutex);
 382         count = sprintf(buf, "%d\n", cpu_management);
 383         mutex_unlock(&smp_cpu_state_mutex);
 384         return count;
 385 }
 386
 387 static ssize_t dispatching_store(struct device *dev,
 388                                  struct device_attribute *attr,
 389                                  const char *buf,
 390                                  size_t count)
 391 {
 392         int val, rc;
 393         char delim;
 394
 395         if (sscanf(buf, "%d %c", &val, &delim) != 1)
 396                 return -EINVAL;
 397         if (val != 0 && val != 1)
 398                 return -EINVAL;
 399         rc = 0;
 400         get_online_cpus();
 401         mutex_lock(&smp_cpu_state_mutex);
 402         if (cpu_management == val)
 403                 goto out;
 404         rc = topology_set_cpu_management(val);
 405         if (rc)
 406                 goto out;
 407         cpu_management = val;
 408         topology_expect_change();
 409 out:
 410         mutex_unlock(&smp_cpu_state_mutex);
 411         put_online_cpus();
 412         return rc ? rc : count;
 413 }
 414 static DEVICE_ATTR_RW(dispatching);
 415
 416 static ssize_t cpu_polarization_show(struct device *dev,
 417                                      struct device_attribute *attr, char *buf)
 418 {
 419         int cpu = dev->id;
 420         ssize_t count;
 421
 422         mutex_lock(&smp_cpu_state_mutex);
 423         switch (smp_cpu_get_polarization(cpu)) {
 424         case POLARIZATION_HRZ:
 425                 count = sprintf(buf, "horizontal\n");
 426                 break;
 427         case POLARIZATION_VL:
 428                 count = sprintf(buf, "vertical:low\n");
 429                 break;
 430         case POLARIZATION_VM:
 431                 count = sprintf(buf, "vertical:medium\n");
 432                 break;
 433         case POLARIZATION_VH:
 434                 count = sprintf(buf, "vertical:high\n");
 435                 break;
 436         default:
 437                 count = sprintf(buf, "unknown\n");
 438                 break;
 439         }
 440         mutex_unlock(&smp_cpu_state_mutex);
 441         return count;
 442 }
 443 static DEVICE_ATTR(polarization, 0444, cpu_polarization_show, NULL);
 444
 445 static struct attribute *topology_cpu_attrs[] = {
 446         &dev_attr_polarization.attr,
 447         NULL,
 448 };
 449
 450 static struct attribute_group topology_cpu_attr_group = {
 451         .attrs = topology_cpu_attrs,
 452 };
 453
 454 static ssize_t cpu_dedicated_show(struct device *dev,
 455                                   struct device_attribute *attr, char *buf)
 456 {
 457         int cpu = dev->id;
 458         ssize_t count;
 459
 460         mutex_lock(&smp_cpu_state_mutex);
 461         count = sprintf(buf, "%d\n", topology_cpu_dedicated(cpu));
 462         mutex_unlock(&smp_cpu_state_mutex);
 463         return count;
 464 }
 465 static DEVICE_ATTR(dedicated, 0444, cpu_dedicated_show, NULL);
 466
 467 static struct attribute *topology_extra_cpu_attrs[] = {
 468         &dev_attr_dedicated.attr,
 469         NULL,
 470 };
 471
 472 static struct attribute_group topology_extra_cpu_attr_group = {
 473         .attrs = topology_extra_cpu_attrs,
 474 };
 475
 476 int topology_cpu_init(struct cpu *cpu)
 477 {
 478         int rc;
 479
 480         rc = sysfs_create_group(&cpu->dev.kobj, &topology_cpu_attr_group);
 481         if (rc || !MACHINE_HAS_TOPOLOGY)
 482                 return rc;
 483         rc = sysfs_create_group(&cpu->dev.kobj, &topology_extra_cpu_attr_group);
 484         if (rc)
 485                 sysfs_remove_group(&cpu->dev.kobj, &topology_cpu_attr_group);
 486         return rc;
 487 }
 488
 489 static const struct cpumask *cpu_thread_mask(int cpu)
 490 {
 491         return &cpu_topology[cpu].thread_mask;
 492 }
 493
 494
 495 const struct cpumask *cpu_coregroup_mask(int cpu)
 496 {
 497         return &cpu_topology[cpu].core_mask;
 498 }
 499
 500 static const struct cpumask *cpu_book_mask(int cpu)
 501 {
 502         return &cpu_topology[cpu].book_mask;
 503 }
 504
 505 static const struct cpumask *cpu_drawer_mask(int cpu)
 506 {
 507         return &cpu_topology[cpu].drawer_mask;
 508 }
 509
 510 static struct sched_domain_topology_level s390_topology[] = {
 511         { cpu_thread_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
 512         { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
 513         { cpu_book_mask, SD_INIT_NAME(BOOK) },
 514         { cpu_drawer_mask, SD_INIT_NAME(DRAWER) },
 515         { cpu_cpu_mask, SD_INIT_NAME(DIE) },
 516         { NULL, },
 517 };
 518
 519 static void __init alloc_masks(struct sysinfo_15_1_x *info,
 520                                struct mask_info *mask, int offset)
 521 {
 522         int i, nr_masks;
 523
 524         nr_masks = info->mag[TOPOLOGY_NR_MAG - offset];
 525         for (i = 0; i < info->mnest - offset; i++)
 526                 nr_masks *= info->mag[TOPOLOGY_NR_MAG - offset - 1 - i];
 527         nr_masks = max(nr_masks, 1);
 528         for (i = 0; i < nr_masks; i++) {
 529                 mask->next = memblock_alloc(sizeof(*mask->next), 8);
 530                 if (!mask->next)
 531                         panic("%s: Failed to allocate %zu bytes align=0x%x\n",
 532                               __func__, sizeof(*mask->next), 8);
 533                 mask = mask->next;
 534         }
 535 }
 536
 537 void __init topology_init_early(void)
 538 {
 539         struct sysinfo_15_1_x *info;
 540
 541         set_sched_topology(s390_topology);
 542         if (topology_mode == TOPOLOGY_MODE_UNINITIALIZED) {
 543                 if (MACHINE_HAS_TOPOLOGY)
 544                         topology_mode = TOPOLOGY_MODE_HW;
 545                 else
 546                         topology_mode = TOPOLOGY_MODE_SINGLE;
 547         }
 548         if (!MACHINE_HAS_TOPOLOGY)
 549                 goto out;
 550         tl_info = memblock_alloc(PAGE_SIZE, PAGE_SIZE);
 551         if (!tl_info)
 552                 panic("%s: Failed to allocate %lu bytes align=0x%lx\n",
 553                       __func__, PAGE_SIZE, PAGE_SIZE);
 554         info = tl_info;
 555         store_topology(info);
 556         pr_info("The CPU configuration topology of the machine is: %d %d %d %d %d %d / %d\n",
 557                 info->mag[0], info->mag[1], info->mag[2], info->mag[3],
 558                 info->mag[4], info->mag[5], info->mnest);
 559         alloc_masks(info, &socket_info, 1);
 560         alloc_masks(info, &book_info, 2);
 561         alloc_masks(info, &drawer_info, 3);
 562 out:
 563         __arch_update_cpu_topology();
 564         __arch_update_dedicated_flag(NULL);
 565 }
 566
 567 static inline int topology_get_mode(int enabled)
 568 {
 569         if (!enabled)
 570                 return TOPOLOGY_MODE_SINGLE;
 571         return MACHINE_HAS_TOPOLOGY ? TOPOLOGY_MODE_HW : TOPOLOGY_MODE_PACKAGE;
 572 }
 573
 574 static inline int topology_is_enabled(void)
 575 {
 576         return topology_mode != TOPOLOGY_MODE_SINGLE;
 577 }
 578
 579 static int __init topology_setup(char *str)
 580 {
 581         bool enabled;
 582         int rc;
 583
 584         rc = kstrtobool(str, &enabled);
 585         if (rc)
 586                 return rc;
 587         topology_mode = topology_get_mode(enabled);
 588         return 0;
 589 }
 590 early_param("topology", topology_setup);
 591
 592 static int topology_ctl_handler(struct ctl_table *ctl, int write,
 593                                 void __user *buffer, size_t *lenp, loff_t *ppos)
 594 {
 595         int enabled = topology_is_enabled();
 596         int new_mode;
 597         int rc;
 598         struct ctl_table ctl_entry = {
 599                 .procname       = ctl->procname,
 600                 .data           = &enabled,
 601                 .maxlen         = sizeof(int),
 602                 .extra1         = SYSCTL_ZERO,
 603                 .extra2         = SYSCTL_ONE,
 604         };
 605
 606         rc = proc_douintvec_minmax(&ctl_entry, write, buffer, lenp, ppos);
 607         if (rc < 0 || !write)
 608                 return rc;
 609
 610         mutex_lock(&smp_cpu_state_mutex);
 611         new_mode = topology_get_mode(enabled);
 612         if (topology_mode != new_mode) {
 613                 topology_mode = new_mode;
 614                 topology_schedule_update();
 615         }
 616         mutex_unlock(&smp_cpu_state_mutex);
 617         topology_flush_work();
 618
 619         return rc;
 620 }
 621
 622 static struct ctl_table topology_ctl_table[] = {
 623         {
 624                 .procname       = "topology",
 625                 .mode           = 0644,
 626                 .proc_handler   = topology_ctl_handler,
 627         },
 628         { },
 629 };
 630
 631 static struct ctl_table topology_dir_table[] = {
 632         {
 633                 .procname       = "s390",
 634                 .maxlen         = 0,
 635                 .mode           = 0555,
 636                 .child          = topology_ctl_table,
 637         },
 638         { },
 639 };
 640
 641 static int __init topology_init(void)
 642 {
 643         timer_setup(&topology_timer, topology_timer_fn, TIMER_DEFERRABLE);
 644         if (MACHINE_HAS_TOPOLOGY)
 645                 set_topology_timer();
 646         else
 647                 topology_update_polarization_simple();
 648         register_sysctl_table(topology_dir_table);
 649         return device_create_file(cpu_subsys.dev_root, &dev_attr_dispatching);
 650 }
 651 device_initcall(topology_init);