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 void 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 {
  88         struct cpu_topology_s390 *topo;
  89         unsigned int core;
  90
  91         for_each_set_bit(core, &tl_core->mask[0], TOPOLOGY_CORE_BITS) {
  92                 unsigned int rcore;
  93                 int lcpu, i;
  94
  95                 rcore = TOPOLOGY_CORE_BITS - 1 - core + tl_core->origin;
  96                 lcpu = smp_find_processor_id(rcore << smp_cpu_mt_shift);
  97                 if (lcpu < 0)
  98                         continue;
  99                 for (i = 0; i <= smp_cpu_mtid; i++) {
 100                         topo = &per_cpu(cpu_topology, lcpu + i);
 101                         topo->drawer_id = drawer->id;
 102                         topo->book_id = book->id;
 103                         topo->socket_id = socket->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                         smp_cpu_set_polarization(lcpu + i, tl_core->pp);
 110                 }
 111         }
 112 }
 113
 114 static void clear_masks(void)
 115 {
 116         struct mask_info *info;
 117
 118         info = &socket_info;
 119         while (info) {
 120                 cpumask_clear(&info->mask);
 121                 info = info->next;
 122         }
 123         info = &book_info;
 124         while (info) {
 125                 cpumask_clear(&info->mask);
 126                 info = info->next;
 127         }
 128         info = &drawer_info;
 129         while (info) {
 130                 cpumask_clear(&info->mask);
 131                 info = info->next;
 132         }
 133 }
 134
 135 static union topology_entry *next_tle(union topology_entry *tle)
 136 {
 137         if (!tle->nl)
 138                 return (union topology_entry *)((struct topology_core *)tle + 1);
 139         return (union topology_entry *)((struct topology_container *)tle + 1);
 140 }
 141
 142 static void tl_to_masks(struct sysinfo_15_1_x *info)
 143 {
 144         struct mask_info *socket = &socket_info;
 145         struct mask_info *book = &book_info;
 146         struct mask_info *drawer = &drawer_info;
 147         union topology_entry *tle, *end;
 148
 149         clear_masks();
 150         tle = info->tle;
 151         end = (union topology_entry *)((unsigned long)info + info->length);
 152         while (tle < end) {
 153                 switch (tle->nl) {
 154                 case 3:
 155                         drawer = drawer->next;
 156                         drawer->id = tle->container.id;
 157                         break;
 158                 case 2:
 159                         book = book->next;
 160                         book->id = tle->container.id;
 161                         break;
 162                 case 1:
 163                         socket = socket->next;
 164                         socket->id = tle->container.id;
 165                         break;
 166                 case 0:
 167                         add_cpus_to_mask(&tle->cpu, drawer, book, socket);
 168                         break;
 169                 default:
 170                         clear_masks();
 171                         return;
 172                 }
 173                 tle = next_tle(tle);
 174         }
 175 }
 176
 177 static void topology_update_polarization_simple(void)
 178 {
 179         int cpu;
 180
 181         mutex_lock(&smp_cpu_state_mutex);
 182         for_each_possible_cpu(cpu)
 183                 smp_cpu_set_polarization(cpu, POLARIZATION_HRZ);
 184         mutex_unlock(&smp_cpu_state_mutex);
 185 }
 186
 187 static int ptf(unsigned long fc)
 188 {
 189         int rc;
 190
 191         asm volatile(
 192                 "       .insn   rre,0xb9a20000,%1,%1\n"
 193                 "       ipm     %0\n"
 194                 "       srl     %0,28\n"
 195                 : "=d" (rc)
 196                 : "d" (fc)  : "cc");
 197         return rc;
 198 }
 199
 200 int topology_set_cpu_management(int fc)
 201 {
 202         int cpu, rc;
 203
 204         if (!MACHINE_HAS_TOPOLOGY)
 205                 return -EOPNOTSUPP;
 206         if (fc)
 207                 rc = ptf(PTF_VERTICAL);
 208         else
 209                 rc = ptf(PTF_HORIZONTAL);
 210         if (rc)
 211                 return -EBUSY;
 212         for_each_possible_cpu(cpu)
 213                 smp_cpu_set_polarization(cpu, POLARIZATION_UNKNOWN);
 214         return rc;
 215 }
 216
 217 static void update_cpu_masks(void)
 218 {
 219         struct cpu_topology_s390 *topo;
 220         int cpu;
 221
 222         for_each_possible_cpu(cpu) {
 223                 topo = &per_cpu(cpu_topology, cpu);
 224                 topo->thread_mask = cpu_thread_map(cpu);
 225                 topo->core_mask = cpu_group_map(&socket_info, cpu);
 226                 topo->book_mask = cpu_group_map(&book_info, cpu);
 227                 topo->drawer_mask = cpu_group_map(&drawer_info, cpu);
 228                 if (!MACHINE_HAS_TOPOLOGY) {
 229                         topo->thread_id = cpu;
 230                         topo->core_id = cpu;
 231                         topo->socket_id = cpu;
 232                         topo->book_id = cpu;
 233                         topo->drawer_id = cpu;
 234                 }
 235         }
 236         numa_update_cpu_topology();
 237 }
 238
 239 void store_topology(struct sysinfo_15_1_x *info)
 240 {
 241         stsi(info, 15, 1, min(topology_max_mnest, 4));
 242 }
 243
 244 int arch_update_cpu_topology(void)
 245 {
 246         struct sysinfo_15_1_x *info = tl_info;
 247         struct device *dev;
 248         int cpu, rc = 0;
 249
 250         if (MACHINE_HAS_TOPOLOGY) {
 251                 rc = 1;
 252                 store_topology(info);
 253                 tl_to_masks(info);
 254         }
 255         update_cpu_masks();
 256         if (!MACHINE_HAS_TOPOLOGY)
 257                 topology_update_polarization_simple();
 258         for_each_online_cpu(cpu) {
 259                 dev = get_cpu_device(cpu);
 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 int __init early_parse_topology(char *p)
 417 {
 418         return kstrtobool(p, &topology_enabled);
 419 }
 420 early_param("topology", early_parse_topology);
 421
 422 static struct sched_domain_topology_level s390_topology[] = {
 423         { cpu_thread_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
 424         { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
 425         { cpu_book_mask, SD_INIT_NAME(BOOK) },
 426         { cpu_drawer_mask, SD_INIT_NAME(DRAWER) },
 427         { cpu_cpu_mask, SD_INIT_NAME(DIE) },
 428         { NULL, },
 429 };
 430
 431 static void __init alloc_masks(struct sysinfo_15_1_x *info,
 432                                struct mask_info *mask, int offset)
 433 {
 434         int i, nr_masks;
 435
 436         nr_masks = info->mag[TOPOLOGY_NR_MAG - offset];
 437         for (i = 0; i < info->mnest - offset; i++)
 438                 nr_masks *= info->mag[TOPOLOGY_NR_MAG - offset - 1 - i];
 439         nr_masks = max(nr_masks, 1);
 440         for (i = 0; i < nr_masks; i++) {
 441                 mask->next = kzalloc(sizeof(*mask->next), GFP_KERNEL);
 442                 mask = mask->next;
 443         }
 444 }
 445
 446 static int __init s390_topology_init(void)
 447 {
 448         struct sysinfo_15_1_x *info;
 449         int i;
 450
 451         if (!MACHINE_HAS_TOPOLOGY)
 452                 return 0;
 453         tl_info = (struct sysinfo_15_1_x *)__get_free_page(GFP_KERNEL);
 454         info = tl_info;
 455         store_topology(info);
 456         pr_info("The CPU configuration topology of the machine is:");
 457         for (i = 0; i < TOPOLOGY_NR_MAG; i++)
 458                 printk(KERN_CONT " %d", info->mag[i]);
 459         printk(KERN_CONT " / %d\n", info->mnest);
 460         alloc_masks(info, &socket_info, 1);
 461         alloc_masks(info, &book_info, 2);
 462         alloc_masks(info, &drawer_info, 3);
 463         set_sched_topology(s390_topology);
 464         return 0;
 465 }
 466 early_initcall(s390_topology_init);
 467
 468 static int __init topology_init(void)
 469 {
 470         if (MACHINE_HAS_TOPOLOGY)
 471                 set_topology_timer();
 472         else
 473                 topology_update_polarization_simple();
 474         return device_create_file(cpu_subsys.dev_root, &dev_attr_dispatching);
 475 }
 476 device_initcall(topology_init);