arch/powerpc/platforms/pseries/hotplug-cpu.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * pseries CPU Hotplug infrastructure.
   4  *
   5  * Split out from arch/powerpc/platforms/pseries/setup.c
   6  *  arch/powerpc/kernel/rtas.c, and arch/powerpc/platforms/pseries/smp.c
   7  *
   8  * Peter Bergner, IBM   March 2001.
   9  * Copyright (C) 2001 IBM.
  10  * Dave Engebretsen, Peter Bergner, and
  11  * Mike Corrigan {engebret|bergner|mikec}@us.ibm.com
  12  * Plus various changes from other IBM teams...
  13  *
  14  * Copyright (C) 2006 Michael Ellerman, IBM Corporation
  15  */
  16
  17 #define pr_fmt(fmt)     "pseries-hotplug-cpu: " fmt
  18
  19 #include <linux/kernel.h>
  20 #include <linux/interrupt.h>
  21 #include <linux/delay.h>
  22 #include <linux/sched.h>        /* for idle_task_exit */
  23 #include <linux/sched/hotplug.h>
  24 #include <linux/cpu.h>
  25 #include <linux/of.h>
  26 #include <linux/slab.h>
  27 #include <asm/prom.h>
  28 #include <asm/rtas.h>
  29 #include <asm/firmware.h>
  30 #include <asm/machdep.h>
  31 #include <asm/vdso_datapage.h>
  32 #include <asm/xics.h>
  33 #include <asm/xive.h>
  34 #include <asm/plpar_wrappers.h>
  35 #include <asm/topology.h>
  36
  37 #include "pseries.h"
  38
  39 /* This version can't take the spinlock, because it never returns */
  40 static int rtas_stop_self_token = RTAS_UNKNOWN_SERVICE;
  41
  42 /*
  43  * Record the CPU ids used on each nodes.
  44  * Protected by cpu_add_remove_lock.
  45  */
  46 static cpumask_var_t node_recorded_ids_map[MAX_NUMNODES];
  47
  48 static void rtas_stop_self(void)
  49 {
  50         static struct rtas_args args;
  51
  52         local_irq_disable();
  53
  54         BUG_ON(rtas_stop_self_token == RTAS_UNKNOWN_SERVICE);
  55
  56         rtas_call_unlocked(&args, rtas_stop_self_token, 0, 1, NULL);
  57
  58         panic("Alas, I survived.\n");
  59 }
  60
  61 static void pseries_cpu_offline_self(void)
  62 {
  63         unsigned int hwcpu = hard_smp_processor_id();
  64
  65         local_irq_disable();
  66         idle_task_exit();
  67         if (xive_enabled())
  68                 xive_teardown_cpu();
  69         else
  70                 xics_teardown_cpu();
  71
  72         unregister_slb_shadow(hwcpu);
  73         rtas_stop_self();
  74
  75         /* Should never get here... */
  76         BUG();
  77         for(;;);
  78 }
  79
  80 static int pseries_cpu_disable(void)
  81 {
  82         int cpu = smp_processor_id();
  83
  84         set_cpu_online(cpu, false);
  85         vdso_data->processorCount--;
  86
  87         /*fix boot_cpuid here*/
  88         if (cpu == boot_cpuid)
  89                 boot_cpuid = cpumask_any(cpu_online_mask);
  90
  91         /* FIXME: abstract this to not be platform specific later on */
  92         if (xive_enabled())
  93                 xive_smp_disable_cpu();
  94         else
  95                 xics_migrate_irqs_away();
  96
  97         cleanup_cpu_mmu_context();
  98
  99         return 0;
 100 }
 101
 102 /*
 103  * pseries_cpu_die: Wait for the cpu to die.
 104  * @cpu: logical processor id of the CPU whose death we're awaiting.
 105  *
 106  * This function is called from the context of the thread which is performing
 107  * the cpu-offline. Here we wait for long enough to allow the cpu in question
 108  * to self-destroy so that the cpu-offline thread can send the CPU_DEAD
 109  * notifications.
 110  *
 111  * OTOH, pseries_cpu_offline_self() is called by the @cpu when it wants to
 112  * self-destruct.
 113  */
 114 static void pseries_cpu_die(unsigned int cpu)
 115 {
 116         int cpu_status = 1;
 117         unsigned int pcpu = get_hard_smp_processor_id(cpu);
 118         unsigned long timeout = jiffies + msecs_to_jiffies(120000);
 119
 120         while (true) {
 121                 cpu_status = smp_query_cpu_stopped(pcpu);
 122                 if (cpu_status == QCSS_STOPPED ||
 123                     cpu_status == QCSS_HARDWARE_ERROR)
 124                         break;
 125
 126                 if (time_after(jiffies, timeout)) {
 127                         pr_warn("CPU %i (hwid %i) didn't die after 120 seconds\n",
 128                                 cpu, pcpu);
 129                         timeout = jiffies + msecs_to_jiffies(120000);
 130                 }
 131
 132                 cond_resched();
 133         }
 134
 135         if (cpu_status == QCSS_HARDWARE_ERROR) {
 136                 pr_warn("CPU %i (hwid %i) reported error while dying\n",
 137                         cpu, pcpu);
 138         }
 139
 140         /* Isolation and deallocation are definitely done by
 141          * drslot_chrp_cpu.  If they were not they would be
 142          * done here.  Change isolate state to Isolate and
 143          * change allocation-state to Unusable.
 144          */
 145         paca_ptrs[cpu]->cpu_start = 0;
 146 }
 147
 148 /**
 149  * find_cpu_id_range - found a linear ranger of @nthreads free CPU ids.
 150  * @nthreads : the number of threads (cpu ids)
 151  * @assigned_node : the node it belongs to or NUMA_NO_NODE if free ids from any
 152  *                  node can be peek.
 153  * @cpu_mask: the returned CPU mask.
 154  *
 155  * Returns 0 on success.
 156  */
 157 static int find_cpu_id_range(unsigned int nthreads, int assigned_node,
 158                              cpumask_var_t *cpu_mask)
 159 {
 160         cpumask_var_t candidate_mask;
 161         unsigned int cpu, node;
 162         int rc = -ENOSPC;
 163
 164         if (!zalloc_cpumask_var(&candidate_mask, GFP_KERNEL))
 165                 return -ENOMEM;
 166
 167         cpumask_clear(*cpu_mask);
 168         for (cpu = 0; cpu < nthreads; cpu++)
 169                 cpumask_set_cpu(cpu, *cpu_mask);
 170
 171         BUG_ON(!cpumask_subset(cpu_present_mask, cpu_possible_mask));
 172
 173         /* Get a bitmap of unoccupied slots. */
 174         cpumask_xor(candidate_mask, cpu_possible_mask, cpu_present_mask);
 175
 176         if (assigned_node != NUMA_NO_NODE) {
 177                 /*
 178                  * Remove free ids previously assigned on the other nodes. We
 179                  * can walk only online nodes because once a node became online
 180                  * it is not turned offlined back.
 181                  */
 182                 for_each_online_node(node) {
 183                         if (node == assigned_node)
 184                                 continue;
 185                         cpumask_andnot(candidate_mask, candidate_mask,
 186                                        node_recorded_ids_map[node]);
 187                 }
 188         }
 189
 190         if (cpumask_empty(candidate_mask))
 191                 goto out;
 192
 193         while (!cpumask_empty(*cpu_mask)) {
 194                 if (cpumask_subset(*cpu_mask, candidate_mask))
 195                         /* Found a range where we can insert the new cpu(s) */
 196                         break;
 197                 cpumask_shift_left(*cpu_mask, *cpu_mask, nthreads);
 198         }
 199
 200         if (!cpumask_empty(*cpu_mask))
 201                 rc = 0;
 202
 203 out:
 204         free_cpumask_var(candidate_mask);
 205         return rc;
 206 }
 207
 208 /*
 209  * Update cpu_present_mask and paca(s) for a new cpu node.  The wrinkle
 210  * here is that a cpu device node may represent multiple logical cpus
 211  * in the SMT case.  We must honor the assumption in other code that
 212  * the logical ids for sibling SMT threads x and y are adjacent, such
 213  * that x^1 == y and y^1 == x.
 214  */
 215 static int pseries_add_processor(struct device_node *np)
 216 {
 217         int len, nthreads, node, cpu, assigned_node;
 218         int rc = 0;
 219         cpumask_var_t cpu_mask;
 220         const __be32 *intserv;
 221
 222         intserv = of_get_property(np, "ibm,ppc-interrupt-server#s", &len);
 223         if (!intserv)
 224                 return 0;
 225
 226         nthreads = len / sizeof(u32);
 227
 228         if (!alloc_cpumask_var(&cpu_mask, GFP_KERNEL))
 229                 return -ENOMEM;
 230
 231         /*
 232          * Fetch from the DT nodes read by dlpar_configure_connector() the NUMA
 233          * node id the added CPU belongs to.
 234          */
 235         node = of_node_to_nid(np);
 236         if (node < 0 || !node_possible(node))
 237                 node = first_online_node;
 238
 239         BUG_ON(node == NUMA_NO_NODE);
 240         assigned_node = node;
 241
 242         cpu_maps_update_begin();
 243
 244         rc = find_cpu_id_range(nthreads, node, &cpu_mask);
 245         if (rc && nr_node_ids > 1) {
 246                 /*
 247                  * Try again, considering the free CPU ids from the other node.
 248                  */
 249                 node = NUMA_NO_NODE;
 250                 rc = find_cpu_id_range(nthreads, NUMA_NO_NODE, &cpu_mask);
 251         }
 252
 253         if (rc) {
 254                 pr_err("Cannot add cpu %pOF; this system configuration"
 255                        " supports %d logical cpus.\n", np, num_possible_cpus());
 256                 goto out;
 257         }
 258
 259         for_each_cpu(cpu, cpu_mask) {
 260                 BUG_ON(cpu_present(cpu));
 261                 set_cpu_present(cpu, true);
 262                 set_hard_smp_processor_id(cpu, be32_to_cpu(*intserv++));
 263         }
 264
 265         /* Record the newly used CPU ids for the associate node. */
 266         cpumask_or(node_recorded_ids_map[assigned_node],
 267                    node_recorded_ids_map[assigned_node], cpu_mask);
 268
 269         /*
 270          * If node is set to NUMA_NO_NODE, CPU ids have be reused from
 271          * another node, remove them from its mask.
 272          */
 273         if (node == NUMA_NO_NODE) {
 274                 cpu = cpumask_first(cpu_mask);
 275                 pr_warn("Reusing free CPU ids %d-%d from another node\n",
 276                         cpu, cpu + nthreads - 1);
 277                 for_each_online_node(node) {
 278                         if (node == assigned_node)
 279                                 continue;
 280                         cpumask_andnot(node_recorded_ids_map[node],
 281                                        node_recorded_ids_map[node],
 282                                        cpu_mask);
 283                 }
 284         }
 285
 286 out:
 287         cpu_maps_update_done();
 288         free_cpumask_var(cpu_mask);
 289         return rc;
 290 }
 291
 292 /*
 293  * Update the present map for a cpu node which is going away, and set
 294  * the hard id in the paca(s) to -1 to be consistent with boot time
 295  * convention for non-present cpus.
 296  */
 297 static void pseries_remove_processor(struct device_node *np)
 298 {
 299         unsigned int cpu;
 300         int len, nthreads, i;
 301         const __be32 *intserv;
 302         u32 thread;
 303
 304         intserv = of_get_property(np, "ibm,ppc-interrupt-server#s", &len);
 305         if (!intserv)
 306                 return;
 307
 308         nthreads = len / sizeof(u32);
 309
 310         cpu_maps_update_begin();
 311         for (i = 0; i < nthreads; i++) {
 312                 thread = be32_to_cpu(intserv[i]);
 313                 for_each_present_cpu(cpu) {
 314                         if (get_hard_smp_processor_id(cpu) != thread)
 315                                 continue;
 316                         BUG_ON(cpu_online(cpu));
 317                         set_cpu_present(cpu, false);
 318                         set_hard_smp_processor_id(cpu, -1);
 319                         update_numa_cpu_lookup_table(cpu, -1);
 320                         break;
 321                 }
 322                 if (cpu >= nr_cpu_ids)
 323                         printk(KERN_WARNING "Could not find cpu to remove "
 324                                "with physical id 0x%x\n", thread);
 325         }
 326         cpu_maps_update_done();
 327 }
 328
 329 static int dlpar_offline_cpu(struct device_node *dn)
 330 {
 331         int rc = 0;
 332         unsigned int cpu;
 333         int len, nthreads, i;
 334         const __be32 *intserv;
 335         u32 thread;
 336
 337         intserv = of_get_property(dn, "ibm,ppc-interrupt-server#s", &len);
 338         if (!intserv)
 339                 return -EINVAL;
 340
 341         nthreads = len / sizeof(u32);
 342
 343         cpu_maps_update_begin();
 344         for (i = 0; i < nthreads; i++) {
 345                 thread = be32_to_cpu(intserv[i]);
 346                 for_each_present_cpu(cpu) {
 347                         if (get_hard_smp_processor_id(cpu) != thread)
 348                                 continue;
 349
 350                         if (!cpu_online(cpu))
 351                                 break;
 352
 353                         /*
 354                          * device_offline() will return -EBUSY (via cpu_down()) if there
 355                          * is only one CPU left. Check it here to fail earlier and with a
 356                          * more informative error message, while also retaining the
 357                          * cpu_add_remove_lock to be sure that no CPUs are being
 358                          * online/offlined during this check.
 359                          */
 360                         if (num_online_cpus() == 1) {
 361                                 pr_warn("Unable to remove last online CPU %pOFn\n", dn);
 362                                 rc = -EBUSY;
 363                                 goto out_unlock;
 364                         }
 365
 366                         cpu_maps_update_done();
 367                         rc = device_offline(get_cpu_device(cpu));
 368                         if (rc)
 369                                 goto out;
 370                         cpu_maps_update_begin();
 371                         break;
 372                 }
 373                 if (cpu == num_possible_cpus()) {
 374                         pr_warn("Could not find cpu to offline with physical id 0x%x\n",
 375                                 thread);
 376                 }
 377         }
 378 out_unlock:
 379         cpu_maps_update_done();
 380
 381 out:
 382         return rc;
 383 }
 384
 385 static int dlpar_online_cpu(struct device_node *dn)
 386 {
 387         int rc = 0;
 388         unsigned int cpu;
 389         int len, nthreads, i;
 390         const __be32 *intserv;
 391         u32 thread;
 392
 393         intserv = of_get_property(dn, "ibm,ppc-interrupt-server#s", &len);
 394         if (!intserv)
 395                 return -EINVAL;
 396
 397         nthreads = len / sizeof(u32);
 398
 399         cpu_maps_update_begin();
 400         for (i = 0; i < nthreads; i++) {
 401                 thread = be32_to_cpu(intserv[i]);
 402                 for_each_present_cpu(cpu) {
 403                         if (get_hard_smp_processor_id(cpu) != thread)
 404                                 continue;
 405                         cpu_maps_update_done();
 406                         find_and_online_cpu_nid(cpu);
 407                         rc = device_online(get_cpu_device(cpu));
 408                         if (rc) {
 409                                 dlpar_offline_cpu(dn);
 410                                 goto out;
 411                         }
 412                         cpu_maps_update_begin();
 413
 414                         break;
 415                 }
 416                 if (cpu == num_possible_cpus())
 417                         printk(KERN_WARNING "Could not find cpu to online "
 418                                "with physical id 0x%x\n", thread);
 419         }
 420         cpu_maps_update_done();
 421
 422 out:
 423         return rc;
 424
 425 }
 426
 427 static bool dlpar_cpu_exists(struct device_node *parent, u32 drc_index)
 428 {
 429         struct device_node *child = NULL;
 430         u32 my_drc_index;
 431         bool found;
 432         int rc;
 433
 434         /* Assume cpu doesn't exist */
 435         found = false;
 436
 437         for_each_child_of_node(parent, child) {
 438                 rc = of_property_read_u32(child, "ibm,my-drc-index",
 439                                           &my_drc_index);
 440                 if (rc)
 441                         continue;
 442
 443                 if (my_drc_index == drc_index) {
 444                         of_node_put(child);
 445                         found = true;
 446                         break;
 447                 }
 448         }
 449
 450         return found;
 451 }
 452
 453 static bool drc_info_valid_index(struct device_node *parent, u32 drc_index)
 454 {
 455         struct property *info;
 456         struct of_drc_info drc;
 457         const __be32 *value;
 458         u32 index;
 459         int count, i, j;
 460
 461         info = of_find_property(parent, "ibm,drc-info", NULL);
 462         if (!info)
 463                 return false;
 464
 465         value = of_prop_next_u32(info, NULL, &count);
 466
 467         /* First value of ibm,drc-info is number of drc-info records */
 468         if (value)
 469                 value++;
 470         else
 471                 return false;
 472
 473         for (i = 0; i < count; i++) {
 474                 if (of_read_drc_info_cell(&info, &value, &drc))
 475                         return false;
 476
 477                 if (strncmp(drc.drc_type, "CPU", 3))
 478                         break;
 479
 480                 if (drc_index > drc.last_drc_index)
 481                         continue;
 482
 483                 index = drc.drc_index_start;
 484                 for (j = 0; j < drc.num_sequential_elems; j++) {
 485                         if (drc_index == index)
 486                                 return true;
 487
 488                         index += drc.sequential_inc;
 489                 }
 490         }
 491
 492         return false;
 493 }
 494
 495 static bool valid_cpu_drc_index(struct device_node *parent, u32 drc_index)
 496 {
 497         bool found = false;
 498         int rc, index;
 499
 500         if (of_find_property(parent, "ibm,drc-info", NULL))
 501                 return drc_info_valid_index(parent, drc_index);
 502
 503         /* Note that the format of the ibm,drc-indexes array is
 504          * the number of entries in the array followed by the array
 505          * of drc values so we start looking at index = 1.
 506          */
 507         index = 1;
 508         while (!found) {
 509                 u32 drc;
 510
 511                 rc = of_property_read_u32_index(parent, "ibm,drc-indexes",
 512                                                 index++, &drc);
 513
 514                 if (rc)
 515                         break;
 516
 517                 if (drc == drc_index)
 518                         found = true;
 519         }
 520
 521         return found;
 522 }
 523
 524 static ssize_t dlpar_cpu_add(u32 drc_index)
 525 {
 526         struct device_node *dn, *parent;
 527         int rc, saved_rc;
 528
 529         pr_debug("Attempting to add CPU, drc index: %x\n", drc_index);
 530
 531         parent = of_find_node_by_path("/cpus");
 532         if (!parent) {
 533                 pr_warn("Failed to find CPU root node \"/cpus\"\n");
 534                 return -ENODEV;
 535         }
 536
 537         if (dlpar_cpu_exists(parent, drc_index)) {
 538                 of_node_put(parent);
 539                 pr_warn("CPU with drc index %x already exists\n", drc_index);
 540                 return -EINVAL;
 541         }
 542
 543         if (!valid_cpu_drc_index(parent, drc_index)) {
 544                 of_node_put(parent);
 545                 pr_warn("Cannot find CPU (drc index %x) to add.\n", drc_index);
 546                 return -EINVAL;
 547         }
 548
 549         rc = dlpar_acquire_drc(drc_index);
 550         if (rc) {
 551                 pr_warn("Failed to acquire DRC, rc: %d, drc index: %x\n",
 552                         rc, drc_index);
 553                 of_node_put(parent);
 554                 return -EINVAL;
 555         }
 556
 557         dn = dlpar_configure_connector(cpu_to_be32(drc_index), parent);
 558         if (!dn) {
 559                 pr_warn("Failed call to configure-connector, drc index: %x\n",
 560                         drc_index);
 561                 dlpar_release_drc(drc_index);
 562                 of_node_put(parent);
 563                 return -EINVAL;
 564         }
 565
 566         rc = dlpar_attach_node(dn, parent);
 567
 568         /* Regardless we are done with parent now */
 569         of_node_put(parent);
 570
 571         if (rc) {
 572                 saved_rc = rc;
 573                 pr_warn("Failed to attach node %pOFn, rc: %d, drc index: %x\n",
 574                         dn, rc, drc_index);
 575
 576                 rc = dlpar_release_drc(drc_index);
 577                 if (!rc)
 578                         dlpar_free_cc_nodes(dn);
 579
 580                 return saved_rc;
 581         }
 582
 583         update_numa_distance(dn);
 584
 585         rc = dlpar_online_cpu(dn);
 586         if (rc) {
 587                 saved_rc = rc;
 588                 pr_warn("Failed to online cpu %pOFn, rc: %d, drc index: %x\n",
 589                         dn, rc, drc_index);
 590
 591                 rc = dlpar_detach_node(dn);
 592                 if (!rc)
 593                         dlpar_release_drc(drc_index);
 594
 595                 return saved_rc;
 596         }
 597
 598         pr_debug("Successfully added CPU %pOFn, drc index: %x\n", dn,
 599                  drc_index);
 600         return rc;
 601 }
 602
 603 static ssize_t dlpar_cpu_remove(struct device_node *dn, u32 drc_index)
 604 {
 605         int rc;
 606
 607         pr_debug("Attempting to remove CPU %pOFn, drc index: %x\n",
 608                  dn, drc_index);
 609
 610         rc = dlpar_offline_cpu(dn);
 611         if (rc) {
 612                 pr_warn("Failed to offline CPU %pOFn, rc: %d\n", dn, rc);
 613                 return -EINVAL;
 614         }
 615
 616         rc = dlpar_release_drc(drc_index);
 617         if (rc) {
 618                 pr_warn("Failed to release drc (%x) for CPU %pOFn, rc: %d\n",
 619                         drc_index, dn, rc);
 620                 dlpar_online_cpu(dn);
 621                 return rc;
 622         }
 623
 624         rc = dlpar_detach_node(dn);
 625         if (rc) {
 626                 int saved_rc = rc;
 627
 628                 pr_warn("Failed to detach CPU %pOFn, rc: %d", dn, rc);
 629
 630                 rc = dlpar_acquire_drc(drc_index);
 631                 if (!rc)
 632                         dlpar_online_cpu(dn);
 633
 634                 return saved_rc;
 635         }
 636
 637         pr_debug("Successfully removed CPU, drc index: %x\n", drc_index);
 638         return 0;
 639 }
 640
 641 static struct device_node *cpu_drc_index_to_dn(u32 drc_index)
 642 {
 643         struct device_node *dn;
 644         u32 my_index;
 645         int rc;
 646
 647         for_each_node_by_type(dn, "cpu") {
 648                 rc = of_property_read_u32(dn, "ibm,my-drc-index", &my_index);
 649                 if (rc)
 650                         continue;
 651
 652                 if (my_index == drc_index)
 653                         break;
 654         }
 655
 656         return dn;
 657 }
 658
 659 static int dlpar_cpu_remove_by_index(u32 drc_index)
 660 {
 661         struct device_node *dn;
 662         int rc;
 663
 664         dn = cpu_drc_index_to_dn(drc_index);
 665         if (!dn) {
 666                 pr_warn("Cannot find CPU (drc index %x) to remove\n",
 667                         drc_index);
 668                 return -ENODEV;
 669         }
 670
 671         rc = dlpar_cpu_remove(dn, drc_index);
 672         of_node_put(dn);
 673         return rc;
 674 }
 675
 676 static int find_dlpar_cpus_to_remove(u32 *cpu_drcs, int cpus_to_remove)
 677 {
 678         struct device_node *dn;
 679         int cpus_found = 0;
 680         int rc;
 681
 682         /* We want to find cpus_to_remove + 1 CPUs to ensure we do not
 683          * remove the last CPU.
 684          */
 685         for_each_node_by_type(dn, "cpu") {
 686                 cpus_found++;
 687
 688                 if (cpus_found > cpus_to_remove) {
 689                         of_node_put(dn);
 690                         break;
 691                 }
 692
 693                 /* Note that cpus_found is always 1 ahead of the index
 694                  * into the cpu_drcs array, so we use cpus_found - 1
 695                  */
 696                 rc = of_property_read_u32(dn, "ibm,my-drc-index",
 697                                           &cpu_drcs[cpus_found - 1]);
 698                 if (rc) {
 699                         pr_warn("Error occurred getting drc-index for %pOFn\n",
 700                                 dn);
 701                         of_node_put(dn);
 702                         return -1;
 703                 }
 704         }
 705
 706         if (cpus_found < cpus_to_remove) {
 707                 pr_warn("Failed to find enough CPUs (%d of %d) to remove\n",
 708                         cpus_found, cpus_to_remove);
 709         } else if (cpus_found == cpus_to_remove) {
 710                 pr_warn("Cannot remove all CPUs\n");
 711         }
 712
 713         return cpus_found;
 714 }
 715
 716 static int dlpar_cpu_remove_by_count(u32 cpus_to_remove)
 717 {
 718         u32 *cpu_drcs;
 719         int cpus_found;
 720         int cpus_removed = 0;
 721         int i, rc;
 722
 723         pr_debug("Attempting to hot-remove %d CPUs\n", cpus_to_remove);
 724
 725         cpu_drcs = kcalloc(cpus_to_remove, sizeof(*cpu_drcs), GFP_KERNEL);
 726         if (!cpu_drcs)
 727                 return -EINVAL;
 728
 729         cpus_found = find_dlpar_cpus_to_remove(cpu_drcs, cpus_to_remove);
 730         if (cpus_found <= cpus_to_remove) {
 731                 kfree(cpu_drcs);
 732                 return -EINVAL;
 733         }
 734
 735         for (i = 0; i < cpus_to_remove; i++) {
 736                 rc = dlpar_cpu_remove_by_index(cpu_drcs[i]);
 737                 if (rc)
 738                         break;
 739
 740                 cpus_removed++;
 741         }
 742
 743         if (cpus_removed != cpus_to_remove) {
 744                 pr_warn("CPU hot-remove failed, adding back removed CPUs\n");
 745
 746                 for (i = 0; i < cpus_removed; i++)
 747                         dlpar_cpu_add(cpu_drcs[i]);
 748
 749                 rc = -EINVAL;
 750         } else {
 751                 rc = 0;
 752         }
 753
 754         kfree(cpu_drcs);
 755         return rc;
 756 }
 757
 758 static int find_drc_info_cpus_to_add(struct device_node *cpus,
 759                                      struct property *info,
 760                                      u32 *cpu_drcs, u32 cpus_to_add)
 761 {
 762         struct of_drc_info drc;
 763         const __be32 *value;
 764         u32 count, drc_index;
 765         int cpus_found = 0;
 766         int i, j;
 767
 768         if (!info)
 769                 return -1;
 770
 771         value = of_prop_next_u32(info, NULL, &count);
 772         if (value)
 773                 value++;
 774
 775         for (i = 0; i < count; i++) {
 776                 of_read_drc_info_cell(&info, &value, &drc);
 777                 if (strncmp(drc.drc_type, "CPU", 3))
 778                         break;
 779
 780                 drc_index = drc.drc_index_start;
 781                 for (j = 0; j < drc.num_sequential_elems; j++) {
 782                         if (dlpar_cpu_exists(cpus, drc_index))
 783                                 continue;
 784
 785                         cpu_drcs[cpus_found++] = drc_index;
 786
 787                         if (cpus_found == cpus_to_add)
 788                                 return cpus_found;
 789
 790                         drc_index += drc.sequential_inc;
 791                 }
 792         }
 793
 794         return cpus_found;
 795 }
 796
 797 static int find_drc_index_cpus_to_add(struct device_node *cpus,
 798                                       u32 *cpu_drcs, u32 cpus_to_add)
 799 {
 800         int cpus_found = 0;
 801         int index, rc;
 802         u32 drc_index;
 803
 804         /* Search the ibm,drc-indexes array for possible CPU drcs to
 805          * add. Note that the format of the ibm,drc-indexes array is
 806          * the number of entries in the array followed by the array
 807          * of drc values so we start looking at index = 1.
 808          */
 809         index = 1;
 810         while (cpus_found < cpus_to_add) {
 811                 rc = of_property_read_u32_index(cpus, "ibm,drc-indexes",
 812                                                 index++, &drc_index);
 813
 814                 if (rc)
 815                         break;
 816
 817                 if (dlpar_cpu_exists(cpus, drc_index))
 818                         continue;
 819
 820                 cpu_drcs[cpus_found++] = drc_index;
 821         }
 822
 823         return cpus_found;
 824 }
 825
 826 static int dlpar_cpu_add_by_count(u32 cpus_to_add)
 827 {
 828         struct device_node *parent;
 829         struct property *info;
 830         u32 *cpu_drcs;
 831         int cpus_added = 0;
 832         int cpus_found;
 833         int i, rc;
 834
 835         pr_debug("Attempting to hot-add %d CPUs\n", cpus_to_add);
 836
 837         cpu_drcs = kcalloc(cpus_to_add, sizeof(*cpu_drcs), GFP_KERNEL);
 838         if (!cpu_drcs)
 839                 return -EINVAL;
 840
 841         parent = of_find_node_by_path("/cpus");
 842         if (!parent) {
 843                 pr_warn("Could not find CPU root node in device tree\n");
 844                 kfree(cpu_drcs);
 845                 return -1;
 846         }
 847
 848         info = of_find_property(parent, "ibm,drc-info", NULL);
 849         if (info)
 850                 cpus_found = find_drc_info_cpus_to_add(parent, info, cpu_drcs, cpus_to_add);
 851         else
 852                 cpus_found = find_drc_index_cpus_to_add(parent, cpu_drcs, cpus_to_add);
 853
 854         of_node_put(parent);
 855
 856         if (cpus_found < cpus_to_add) {
 857                 pr_warn("Failed to find enough CPUs (%d of %d) to add\n",
 858                         cpus_found, cpus_to_add);
 859                 kfree(cpu_drcs);
 860                 return -EINVAL;
 861         }
 862
 863         for (i = 0; i < cpus_to_add; i++) {
 864                 rc = dlpar_cpu_add(cpu_drcs[i]);
 865                 if (rc)
 866                         break;
 867
 868                 cpus_added++;
 869         }
 870
 871         if (cpus_added < cpus_to_add) {
 872                 pr_warn("CPU hot-add failed, removing any added CPUs\n");
 873
 874                 for (i = 0; i < cpus_added; i++)
 875                         dlpar_cpu_remove_by_index(cpu_drcs[i]);
 876
 877                 rc = -EINVAL;
 878         } else {
 879                 rc = 0;
 880         }
 881
 882         kfree(cpu_drcs);
 883         return rc;
 884 }
 885
 886 int dlpar_cpu(struct pseries_hp_errorlog *hp_elog)
 887 {
 888         u32 count, drc_index;
 889         int rc;
 890
 891         count = hp_elog->_drc_u.drc_count;
 892         drc_index = hp_elog->_drc_u.drc_index;
 893
 894         lock_device_hotplug();
 895
 896         switch (hp_elog->action) {
 897         case PSERIES_HP_ELOG_ACTION_REMOVE:
 898                 if (hp_elog->id_type == PSERIES_HP_ELOG_ID_DRC_COUNT)
 899                         rc = dlpar_cpu_remove_by_count(count);
 900                 else if (hp_elog->id_type == PSERIES_HP_ELOG_ID_DRC_INDEX) {
 901                         rc = dlpar_cpu_remove_by_index(drc_index);
 902                         /*
 903                          * Setting the isolation state of an UNISOLATED/CONFIGURED
 904                          * device to UNISOLATE is a no-op, but the hypervisor can
 905                          * use it as a hint that the CPU removal failed.
 906                          */
 907                         if (rc)
 908                                 dlpar_unisolate_drc(drc_index);
 909                 }
 910                 else
 911                         rc = -EINVAL;
 912                 break;
 913         case PSERIES_HP_ELOG_ACTION_ADD:
 914                 if (hp_elog->id_type == PSERIES_HP_ELOG_ID_DRC_COUNT)
 915                         rc = dlpar_cpu_add_by_count(count);
 916                 else if (hp_elog->id_type == PSERIES_HP_ELOG_ID_DRC_INDEX)
 917                         rc = dlpar_cpu_add(drc_index);
 918                 else
 919                         rc = -EINVAL;
 920                 break;
 921         default:
 922                 pr_err("Invalid action (%d) specified\n", hp_elog->action);
 923                 rc = -EINVAL;
 924                 break;
 925         }
 926
 927         unlock_device_hotplug();
 928         return rc;
 929 }
 930
 931 #ifdef CONFIG_ARCH_CPU_PROBE_RELEASE
 932
 933 static ssize_t dlpar_cpu_probe(const char *buf, size_t count)
 934 {
 935         u32 drc_index;
 936         int rc;
 937
 938         rc = kstrtou32(buf, 0, &drc_index);
 939         if (rc)
 940                 return -EINVAL;
 941
 942         rc = dlpar_cpu_add(drc_index);
 943
 944         return rc ? rc : count;
 945 }
 946
 947 static ssize_t dlpar_cpu_release(const char *buf, size_t count)
 948 {
 949         struct device_node *dn;
 950         u32 drc_index;
 951         int rc;
 952
 953         dn = of_find_node_by_path(buf);
 954         if (!dn)
 955                 return -EINVAL;
 956
 957         rc = of_property_read_u32(dn, "ibm,my-drc-index", &drc_index);
 958         if (rc) {
 959                 of_node_put(dn);
 960                 return -EINVAL;
 961         }
 962
 963         rc = dlpar_cpu_remove(dn, drc_index);
 964         of_node_put(dn);
 965
 966         return rc ? rc : count;
 967 }
 968
 969 #endif /* CONFIG_ARCH_CPU_PROBE_RELEASE */
 970
 971 static int pseries_smp_notifier(struct notifier_block *nb,
 972                                 unsigned long action, void *data)
 973 {
 974         struct of_reconfig_data *rd = data;
 975         int err = 0;
 976
 977         switch (action) {
 978         case OF_RECONFIG_ATTACH_NODE:
 979                 err = pseries_add_processor(rd->dn);
 980                 break;
 981         case OF_RECONFIG_DETACH_NODE:
 982                 pseries_remove_processor(rd->dn);
 983                 break;
 984         }
 985         return notifier_from_errno(err);
 986 }
 987
 988 static struct notifier_block pseries_smp_nb = {
 989         .notifier_call = pseries_smp_notifier,
 990 };
 991
 992 static int __init pseries_cpu_hotplug_init(void)
 993 {
 994         int qcss_tok;
 995         unsigned int node;
 996
 997 #ifdef CONFIG_ARCH_CPU_PROBE_RELEASE
 998         ppc_md.cpu_probe = dlpar_cpu_probe;
 999         ppc_md.cpu_release = dlpar_cpu_release;
1000 #endif /* CONFIG_ARCH_CPU_PROBE_RELEASE */
1001
1002         rtas_stop_self_token = rtas_token("stop-self");
1003         qcss_tok = rtas_token("query-cpu-stopped-state");
1004
1005         if (rtas_stop_self_token == RTAS_UNKNOWN_SERVICE ||
1006                         qcss_tok == RTAS_UNKNOWN_SERVICE) {
1007                 printk(KERN_INFO "CPU Hotplug not supported by firmware "
1008                                 "- disabling.\n");
1009                 return 0;
1010         }
1011
1012         smp_ops->cpu_offline_self = pseries_cpu_offline_self;
1013         smp_ops->cpu_disable = pseries_cpu_disable;
1014         smp_ops->cpu_die = pseries_cpu_die;
1015
1016         /* Processors can be added/removed only on LPAR */
1017         if (firmware_has_feature(FW_FEATURE_LPAR)) {
1018                 for_each_node(node) {
1019                         alloc_bootmem_cpumask_var(&node_recorded_ids_map[node]);
1020
1021                         /* Record ids of CPU added at boot time */
1022                         cpumask_or(node_recorded_ids_map[node],
1023                                    node_recorded_ids_map[node],
1024                                    cpumask_of_node(node));
1025                 }
1026
1027                 of_reconfig_notifier_register(&pseries_smp_nb);
1028         }
1029
1030         return 0;
1031 }
1032 machine_arch_initcall(pseries, pseries_cpu_hotplug_init);