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

   1 /*
   2  * pseries CPU Hotplug infrastructure.
   3  *
   4  * Split out from arch/powerpc/platforms/pseries/setup.c
   5  *  arch/powerpc/kernel/rtas.c, and arch/powerpc/platforms/pseries/smp.c
   6  *
   7  * Peter Bergner, IBM   March 2001.
   8  * Copyright (C) 2001 IBM.
   9  * Dave Engebretsen, Peter Bergner, and
  10  * Mike Corrigan {engebret|bergner|mikec}@us.ibm.com
  11  * Plus various changes from other IBM teams...
  12  *
  13  * Copyright (C) 2006 Michael Ellerman, IBM Corporation
  14  *
  15  *      This program is free software; you can redistribute it and/or
  16  *      modify it under the terms of the GNU General Public License
  17  *      as published by the Free Software Foundation; either version
  18  *      2 of the License, or (at your option) any later version.
  19  */
  20
  21 #define pr_fmt(fmt)     "pseries-hotplug-cpu: " fmt
  22
  23 #include <linux/kernel.h>
  24 #include <linux/interrupt.h>
  25 #include <linux/delay.h>
  26 #include <linux/sched.h>        /* for idle_task_exit */
  27 #include <linux/sched/hotplug.h>
  28 #include <linux/cpu.h>
  29 #include <linux/of.h>
  30 #include <linux/slab.h>
  31 #include <asm/prom.h>
  32 #include <asm/rtas.h>
  33 #include <asm/firmware.h>
  34 #include <asm/machdep.h>
  35 #include <asm/vdso_datapage.h>
  36 #include <asm/xics.h>
  37 #include <asm/xive.h>
  38 #include <asm/plpar_wrappers.h>
  39 #include <asm/topology.h>
  40
  41 #include "pseries.h"
  42 #include "offline_states.h"
  43
  44 /* This version can't take the spinlock, because it never returns */
  45 static int rtas_stop_self_token = RTAS_UNKNOWN_SERVICE;
  46
  47 static DEFINE_PER_CPU(enum cpu_state_vals, preferred_offline_state) =
  48                                                         CPU_STATE_OFFLINE;
  49 static DEFINE_PER_CPU(enum cpu_state_vals, current_state) = CPU_STATE_OFFLINE;
  50
  51 static enum cpu_state_vals default_offline_state = CPU_STATE_OFFLINE;
  52
  53 static bool cede_offline_enabled __read_mostly = true;
  54
  55 /*
  56  * Enable/disable cede_offline when available.
  57  */
  58 static int __init setup_cede_offline(char *str)
  59 {
  60         return (kstrtobool(str, &cede_offline_enabled) == 0);
  61 }
  62
  63 __setup("cede_offline=", setup_cede_offline);
  64
  65 enum cpu_state_vals get_cpu_current_state(int cpu)
  66 {
  67         return per_cpu(current_state, cpu);
  68 }
  69
  70 void set_cpu_current_state(int cpu, enum cpu_state_vals state)
  71 {
  72         per_cpu(current_state, cpu) = state;
  73 }
  74
  75 enum cpu_state_vals get_preferred_offline_state(int cpu)
  76 {
  77         return per_cpu(preferred_offline_state, cpu);
  78 }
  79
  80 void set_preferred_offline_state(int cpu, enum cpu_state_vals state)
  81 {
  82         per_cpu(preferred_offline_state, cpu) = state;
  83 }
  84
  85 void set_default_offline_state(int cpu)
  86 {
  87         per_cpu(preferred_offline_state, cpu) = default_offline_state;
  88 }
  89
  90 static void rtas_stop_self(void)
  91 {
  92         static struct rtas_args args;
  93
  94         local_irq_disable();
  95
  96         BUG_ON(rtas_stop_self_token == RTAS_UNKNOWN_SERVICE);
  97
  98         rtas_call_unlocked(&args, rtas_stop_self_token, 0, 1, NULL);
  99
 100         panic("Alas, I survived.\n");
 101 }
 102
 103 static void pseries_mach_cpu_die(void)
 104 {
 105         unsigned int cpu = smp_processor_id();
 106         unsigned int hwcpu = hard_smp_processor_id();
 107         u8 cede_latency_hint = 0;
 108
 109         local_irq_disable();
 110         idle_task_exit();
 111         if (xive_enabled())
 112                 xive_teardown_cpu();
 113         else
 114                 xics_teardown_cpu();
 115
 116         if (get_preferred_offline_state(cpu) == CPU_STATE_INACTIVE) {
 117                 set_cpu_current_state(cpu, CPU_STATE_INACTIVE);
 118                 if (ppc_md.suspend_disable_cpu)
 119                         ppc_md.suspend_disable_cpu();
 120
 121                 cede_latency_hint = 2;
 122
 123                 get_lppaca()->idle = 1;
 124                 if (!lppaca_shared_proc(get_lppaca()))
 125                         get_lppaca()->donate_dedicated_cpu = 1;
 126
 127                 while (get_preferred_offline_state(cpu) == CPU_STATE_INACTIVE) {
 128                         while (!prep_irq_for_idle()) {
 129                                 local_irq_enable();
 130                                 local_irq_disable();
 131                         }
 132
 133                         extended_cede_processor(cede_latency_hint);
 134                 }
 135
 136                 local_irq_disable();
 137
 138                 if (!lppaca_shared_proc(get_lppaca()))
 139                         get_lppaca()->donate_dedicated_cpu = 0;
 140                 get_lppaca()->idle = 0;
 141
 142                 if (get_preferred_offline_state(cpu) == CPU_STATE_ONLINE) {
 143                         unregister_slb_shadow(hwcpu);
 144
 145                         hard_irq_disable();
 146                         /*
 147                          * Call to start_secondary_resume() will not return.
 148                          * Kernel stack will be reset and start_secondary()
 149                          * will be called to continue the online operation.
 150                          */
 151                         start_secondary_resume();
 152                 }
 153         }
 154
 155         /* Requested state is CPU_STATE_OFFLINE at this point */
 156         WARN_ON(get_preferred_offline_state(cpu) != CPU_STATE_OFFLINE);
 157
 158         set_cpu_current_state(cpu, CPU_STATE_OFFLINE);
 159         unregister_slb_shadow(hwcpu);
 160         rtas_stop_self();
 161
 162         /* Should never get here... */
 163         BUG();
 164         for(;;);
 165 }
 166
 167 static int pseries_cpu_disable(void)
 168 {
 169         int cpu = smp_processor_id();
 170
 171         set_cpu_online(cpu, false);
 172         vdso_data->processorCount--;
 173
 174         /*fix boot_cpuid here*/
 175         if (cpu == boot_cpuid)
 176                 boot_cpuid = cpumask_any(cpu_online_mask);
 177
 178         /* FIXME: abstract this to not be platform specific later on */
 179         if (xive_enabled())
 180                 xive_smp_disable_cpu();
 181         else
 182                 xics_migrate_irqs_away();
 183         return 0;
 184 }
 185
 186 /*
 187  * pseries_cpu_die: Wait for the cpu to die.
 188  * @cpu: logical processor id of the CPU whose death we're awaiting.
 189  *
 190  * This function is called from the context of the thread which is performing
 191  * the cpu-offline. Here we wait for long enough to allow the cpu in question
 192  * to self-destroy so that the cpu-offline thread can send the CPU_DEAD
 193  * notifications.
 194  *
 195  * OTOH, pseries_mach_cpu_die() is called by the @cpu when it wants to
 196  * self-destruct.
 197  */
 198 static void pseries_cpu_die(unsigned int cpu)
 199 {
 200         int tries;
 201         int cpu_status = 1;
 202         unsigned int pcpu = get_hard_smp_processor_id(cpu);
 203
 204         if (get_preferred_offline_state(cpu) == CPU_STATE_INACTIVE) {
 205                 cpu_status = 1;
 206                 for (tries = 0; tries < 5000; tries++) {
 207                         if (get_cpu_current_state(cpu) == CPU_STATE_INACTIVE) {
 208                                 cpu_status = 0;
 209                                 break;
 210                         }
 211                         msleep(1);
 212                 }
 213         } else if (get_preferred_offline_state(cpu) == CPU_STATE_OFFLINE) {
 214
 215                 for (tries = 0; tries < 25; tries++) {
 216                         cpu_status = smp_query_cpu_stopped(pcpu);
 217                         if (cpu_status == QCSS_STOPPED ||
 218                             cpu_status == QCSS_HARDWARE_ERROR)
 219                                 break;
 220                         cpu_relax();
 221                 }
 222         }
 223
 224         if (cpu_status != 0) {
 225                 printk("Querying DEAD? cpu %i (%i) shows %i\n",
 226                        cpu, pcpu, cpu_status);
 227         }
 228
 229         /* Isolation and deallocation are definitely done by
 230          * drslot_chrp_cpu.  If they were not they would be
 231          * done here.  Change isolate state to Isolate and
 232          * change allocation-state to Unusable.
 233          */
 234         paca[cpu].cpu_start = 0;
 235 }
 236
 237 /*
 238  * Update cpu_present_mask and paca(s) for a new cpu node.  The wrinkle
 239  * here is that a cpu device node may represent up to two logical cpus
 240  * in the SMT case.  We must honor the assumption in other code that
 241  * the logical ids for sibling SMT threads x and y are adjacent, such
 242  * that x^1 == y and y^1 == x.
 243  */
 244 static int pseries_add_processor(struct device_node *np)
 245 {
 246         unsigned int cpu;
 247         cpumask_var_t candidate_mask, tmp;
 248         int err = -ENOSPC, len, nthreads, i;
 249         const __be32 *intserv;
 250
 251         intserv = of_get_property(np, "ibm,ppc-interrupt-server#s", &len);
 252         if (!intserv)
 253                 return 0;
 254
 255         zalloc_cpumask_var(&candidate_mask, GFP_KERNEL);
 256         zalloc_cpumask_var(&tmp, GFP_KERNEL);
 257
 258         nthreads = len / sizeof(u32);
 259         for (i = 0; i < nthreads; i++)
 260                 cpumask_set_cpu(i, tmp);
 261
 262         cpu_maps_update_begin();
 263
 264         BUG_ON(!cpumask_subset(cpu_present_mask, cpu_possible_mask));
 265
 266         /* Get a bitmap of unoccupied slots. */
 267         cpumask_xor(candidate_mask, cpu_possible_mask, cpu_present_mask);
 268         if (cpumask_empty(candidate_mask)) {
 269                 /* If we get here, it most likely means that NR_CPUS is
 270                  * less than the partition's max processors setting.
 271                  */
 272                 printk(KERN_ERR "Cannot add cpu %pOF; this system configuration"
 273                        " supports %d logical cpus.\n", np,
 274                        num_possible_cpus());
 275                 goto out_unlock;
 276         }
 277
 278         while (!cpumask_empty(tmp))
 279                 if (cpumask_subset(tmp, candidate_mask))
 280                         /* Found a range where we can insert the new cpu(s) */
 281                         break;
 282                 else
 283                         cpumask_shift_left(tmp, tmp, nthreads);
 284
 285         if (cpumask_empty(tmp)) {
 286                 printk(KERN_ERR "Unable to find space in cpu_present_mask for"
 287                        " processor %s with %d thread(s)\n", np->name,
 288                        nthreads);
 289                 goto out_unlock;
 290         }
 291
 292         for_each_cpu(cpu, tmp) {
 293                 BUG_ON(cpu_present(cpu));
 294                 set_cpu_present(cpu, true);
 295                 set_hard_smp_processor_id(cpu, be32_to_cpu(*intserv++));
 296         }
 297         err = 0;
 298 out_unlock:
 299         cpu_maps_update_done();
 300         free_cpumask_var(candidate_mask);
 301         free_cpumask_var(tmp);
 302         return err;
 303 }
 304
 305 /*
 306  * Update the present map for a cpu node which is going away, and set
 307  * the hard id in the paca(s) to -1 to be consistent with boot time
 308  * convention for non-present cpus.
 309  */
 310 static void pseries_remove_processor(struct device_node *np)
 311 {
 312         unsigned int cpu;
 313         int len, nthreads, i;
 314         const __be32 *intserv;
 315         u32 thread;
 316
 317         intserv = of_get_property(np, "ibm,ppc-interrupt-server#s", &len);
 318         if (!intserv)
 319                 return;
 320
 321         nthreads = len / sizeof(u32);
 322
 323         cpu_maps_update_begin();
 324         for (i = 0; i < nthreads; i++) {
 325                 thread = be32_to_cpu(intserv[i]);
 326                 for_each_present_cpu(cpu) {
 327                         if (get_hard_smp_processor_id(cpu) != thread)
 328                                 continue;
 329                         BUG_ON(cpu_online(cpu));
 330                         set_cpu_present(cpu, false);
 331                         set_hard_smp_processor_id(cpu, -1);
 332                         update_numa_cpu_lookup_table(cpu, -1);
 333                         break;
 334                 }
 335                 if (cpu >= nr_cpu_ids)
 336                         printk(KERN_WARNING "Could not find cpu to remove "
 337                                "with physical id 0x%x\n", thread);
 338         }
 339         cpu_maps_update_done();
 340 }
 341
 342 static int dlpar_online_cpu(struct device_node *dn)
 343 {
 344         int rc = 0;
 345         unsigned int cpu;
 346         int len, nthreads, i;
 347         const __be32 *intserv;
 348         u32 thread;
 349
 350         intserv = of_get_property(dn, "ibm,ppc-interrupt-server#s", &len);
 351         if (!intserv)
 352                 return -EINVAL;
 353
 354         nthreads = len / sizeof(u32);
 355
 356         cpu_maps_update_begin();
 357         for (i = 0; i < nthreads; i++) {
 358                 thread = be32_to_cpu(intserv[i]);
 359                 for_each_present_cpu(cpu) {
 360                         if (get_hard_smp_processor_id(cpu) != thread)
 361                                 continue;
 362                         BUG_ON(get_cpu_current_state(cpu)
 363                                         != CPU_STATE_OFFLINE);
 364                         cpu_maps_update_done();
 365                         rc = device_online(get_cpu_device(cpu));
 366                         if (rc)
 367                                 goto out;
 368                         cpu_maps_update_begin();
 369
 370                         break;
 371                 }
 372                 if (cpu == num_possible_cpus())
 373                         printk(KERN_WARNING "Could not find cpu to online "
 374                                "with physical id 0x%x\n", thread);
 375         }
 376         cpu_maps_update_done();
 377
 378 out:
 379         return rc;
 380
 381 }
 382
 383 static bool dlpar_cpu_exists(struct device_node *parent, u32 drc_index)
 384 {
 385         struct device_node *child = NULL;
 386         u32 my_drc_index;
 387         bool found;
 388         int rc;
 389
 390         /* Assume cpu doesn't exist */
 391         found = false;
 392
 393         for_each_child_of_node(parent, child) {
 394                 rc = of_property_read_u32(child, "ibm,my-drc-index",
 395                                           &my_drc_index);
 396                 if (rc)
 397                         continue;
 398
 399                 if (my_drc_index == drc_index) {
 400                         of_node_put(child);
 401                         found = true;
 402                         break;
 403                 }
 404         }
 405
 406         return found;
 407 }
 408
 409 static bool valid_cpu_drc_index(struct device_node *parent, u32 drc_index)
 410 {
 411         bool found = false;
 412         int rc, index;
 413
 414         index = 0;
 415         while (!found) {
 416                 u32 drc;
 417
 418                 rc = of_property_read_u32_index(parent, "ibm,drc-indexes",
 419                                                 index++, &drc);
 420                 if (rc)
 421                         break;
 422
 423                 if (drc == drc_index)
 424                         found = true;
 425         }
 426
 427         return found;
 428 }
 429
 430 static ssize_t dlpar_cpu_add(u32 drc_index)
 431 {
 432         struct device_node *dn, *parent;
 433         int rc, saved_rc;
 434
 435         pr_debug("Attempting to add CPU, drc index: %x\n", drc_index);
 436
 437         parent = of_find_node_by_path("/cpus");
 438         if (!parent) {
 439                 pr_warn("Failed to find CPU root node \"/cpus\"\n");
 440                 return -ENODEV;
 441         }
 442
 443         if (dlpar_cpu_exists(parent, drc_index)) {
 444                 of_node_put(parent);
 445                 pr_warn("CPU with drc index %x already exists\n", drc_index);
 446                 return -EINVAL;
 447         }
 448
 449         if (!valid_cpu_drc_index(parent, drc_index)) {
 450                 of_node_put(parent);
 451                 pr_warn("Cannot find CPU (drc index %x) to add.\n", drc_index);
 452                 return -EINVAL;
 453         }
 454
 455         rc = dlpar_acquire_drc(drc_index);
 456         if (rc) {
 457                 pr_warn("Failed to acquire DRC, rc: %d, drc index: %x\n",
 458                         rc, drc_index);
 459                 of_node_put(parent);
 460                 return -EINVAL;
 461         }
 462
 463         dn = dlpar_configure_connector(cpu_to_be32(drc_index), parent);
 464         if (!dn) {
 465                 pr_warn("Failed call to configure-connector, drc index: %x\n",
 466                         drc_index);
 467                 dlpar_release_drc(drc_index);
 468                 of_node_put(parent);
 469                 return -EINVAL;
 470         }
 471
 472         rc = dlpar_attach_node(dn, parent);
 473
 474         /* Regardless we are done with parent now */
 475         of_node_put(parent);
 476
 477         if (rc) {
 478                 saved_rc = rc;
 479                 pr_warn("Failed to attach node %s, rc: %d, drc index: %x\n",
 480                         dn->name, rc, drc_index);
 481
 482                 rc = dlpar_release_drc(drc_index);
 483                 if (!rc)
 484                         dlpar_free_cc_nodes(dn);
 485
 486                 return saved_rc;
 487         }
 488
 489         rc = dlpar_online_cpu(dn);
 490         if (rc) {
 491                 saved_rc = rc;
 492                 pr_warn("Failed to online cpu %s, rc: %d, drc index: %x\n",
 493                         dn->name, rc, drc_index);
 494
 495                 rc = dlpar_detach_node(dn);
 496                 if (!rc)
 497                         dlpar_release_drc(drc_index);
 498
 499                 return saved_rc;
 500         }
 501
 502         pr_debug("Successfully added CPU %s, drc index: %x\n", dn->name,
 503                  drc_index);
 504         return rc;
 505 }
 506
 507 static int dlpar_offline_cpu(struct device_node *dn)
 508 {
 509         int rc = 0;
 510         unsigned int cpu;
 511         int len, nthreads, i;
 512         const __be32 *intserv;
 513         u32 thread;
 514
 515         intserv = of_get_property(dn, "ibm,ppc-interrupt-server#s", &len);
 516         if (!intserv)
 517                 return -EINVAL;
 518
 519         nthreads = len / sizeof(u32);
 520
 521         cpu_maps_update_begin();
 522         for (i = 0; i < nthreads; i++) {
 523                 thread = be32_to_cpu(intserv[i]);
 524                 for_each_present_cpu(cpu) {
 525                         if (get_hard_smp_processor_id(cpu) != thread)
 526                                 continue;
 527
 528                         if (get_cpu_current_state(cpu) == CPU_STATE_OFFLINE)
 529                                 break;
 530
 531                         if (get_cpu_current_state(cpu) == CPU_STATE_ONLINE) {
 532                                 set_preferred_offline_state(cpu,
 533                                                             CPU_STATE_OFFLINE);
 534                                 cpu_maps_update_done();
 535                                 rc = device_offline(get_cpu_device(cpu));
 536                                 if (rc)
 537                                         goto out;
 538                                 cpu_maps_update_begin();
 539                                 break;
 540
 541                         }
 542
 543                         /*
 544                          * The cpu is in CPU_STATE_INACTIVE.
 545                          * Upgrade it's state to CPU_STATE_OFFLINE.
 546                          */
 547                         set_preferred_offline_state(cpu, CPU_STATE_OFFLINE);
 548                         BUG_ON(plpar_hcall_norets(H_PROD, thread)
 549                                                                 != H_SUCCESS);
 550                         __cpu_die(cpu);
 551                         break;
 552                 }
 553                 if (cpu == num_possible_cpus())
 554                         printk(KERN_WARNING "Could not find cpu to offline with physical id 0x%x\n", thread);
 555         }
 556         cpu_maps_update_done();
 557
 558 out:
 559         return rc;
 560
 561 }
 562
 563 static ssize_t dlpar_cpu_remove(struct device_node *dn, u32 drc_index)
 564 {
 565         int rc;
 566
 567         pr_debug("Attempting to remove CPU %s, drc index: %x\n",
 568                  dn->name, drc_index);
 569
 570         rc = dlpar_offline_cpu(dn);
 571         if (rc) {
 572                 pr_warn("Failed to offline CPU %s, rc: %d\n", dn->name, rc);
 573                 return -EINVAL;
 574         }
 575
 576         rc = dlpar_release_drc(drc_index);
 577         if (rc) {
 578                 pr_warn("Failed to release drc (%x) for CPU %s, rc: %d\n",
 579                         drc_index, dn->name, rc);
 580                 dlpar_online_cpu(dn);
 581                 return rc;
 582         }
 583
 584         rc = dlpar_detach_node(dn);
 585         if (rc) {
 586                 int saved_rc = rc;
 587
 588                 pr_warn("Failed to detach CPU %s, rc: %d", dn->name, rc);
 589
 590                 rc = dlpar_acquire_drc(drc_index);
 591                 if (!rc)
 592                         dlpar_online_cpu(dn);
 593
 594                 return saved_rc;
 595         }
 596
 597         pr_debug("Successfully removed CPU, drc index: %x\n", drc_index);
 598         return 0;
 599 }
 600
 601 static struct device_node *cpu_drc_index_to_dn(u32 drc_index)
 602 {
 603         struct device_node *dn;
 604         u32 my_index;
 605         int rc;
 606
 607         for_each_node_by_type(dn, "cpu") {
 608                 rc = of_property_read_u32(dn, "ibm,my-drc-index", &my_index);
 609                 if (rc)
 610                         continue;
 611
 612                 if (my_index == drc_index)
 613                         break;
 614         }
 615
 616         return dn;
 617 }
 618
 619 static int dlpar_cpu_remove_by_index(u32 drc_index)
 620 {
 621         struct device_node *dn;
 622         int rc;
 623
 624         dn = cpu_drc_index_to_dn(drc_index);
 625         if (!dn) {
 626                 pr_warn("Cannot find CPU (drc index %x) to remove\n",
 627                         drc_index);
 628                 return -ENODEV;
 629         }
 630
 631         rc = dlpar_cpu_remove(dn, drc_index);
 632         of_node_put(dn);
 633         return rc;
 634 }
 635
 636 static int find_dlpar_cpus_to_remove(u32 *cpu_drcs, int cpus_to_remove)
 637 {
 638         struct device_node *dn;
 639         int cpus_found = 0;
 640         int rc;
 641
 642         /* We want to find cpus_to_remove + 1 CPUs to ensure we do not
 643          * remove the last CPU.
 644          */
 645         for_each_node_by_type(dn, "cpu") {
 646                 cpus_found++;
 647
 648                 if (cpus_found > cpus_to_remove) {
 649                         of_node_put(dn);
 650                         break;
 651                 }
 652
 653                 /* Note that cpus_found is always 1 ahead of the index
 654                  * into the cpu_drcs array, so we use cpus_found - 1
 655                  */
 656                 rc = of_property_read_u32(dn, "ibm,my-drc-index",
 657                                           &cpu_drcs[cpus_found - 1]);
 658                 if (rc) {
 659                         pr_warn("Error occurred getting drc-index for %s\n",
 660                                 dn->name);
 661                         of_node_put(dn);
 662                         return -1;
 663                 }
 664         }
 665
 666         if (cpus_found < cpus_to_remove) {
 667                 pr_warn("Failed to find enough CPUs (%d of %d) to remove\n",
 668                         cpus_found, cpus_to_remove);
 669         } else if (cpus_found == cpus_to_remove) {
 670                 pr_warn("Cannot remove all CPUs\n");
 671         }
 672
 673         return cpus_found;
 674 }
 675
 676 static int dlpar_cpu_remove_by_count(u32 cpus_to_remove)
 677 {
 678         u32 *cpu_drcs;
 679         int cpus_found;
 680         int cpus_removed = 0;
 681         int i, rc;
 682
 683         pr_debug("Attempting to hot-remove %d CPUs\n", cpus_to_remove);
 684
 685         cpu_drcs = kcalloc(cpus_to_remove, sizeof(*cpu_drcs), GFP_KERNEL);
 686         if (!cpu_drcs)
 687                 return -EINVAL;
 688
 689         cpus_found = find_dlpar_cpus_to_remove(cpu_drcs, cpus_to_remove);
 690         if (cpus_found <= cpus_to_remove) {
 691                 kfree(cpu_drcs);
 692                 return -EINVAL;
 693         }
 694
 695         for (i = 0; i < cpus_to_remove; i++) {
 696                 rc = dlpar_cpu_remove_by_index(cpu_drcs[i]);
 697                 if (rc)
 698                         break;
 699
 700                 cpus_removed++;
 701         }
 702
 703         if (cpus_removed != cpus_to_remove) {
 704                 pr_warn("CPU hot-remove failed, adding back removed CPUs\n");
 705
 706                 for (i = 0; i < cpus_removed; i++)
 707                         dlpar_cpu_add(cpu_drcs[i]);
 708
 709                 rc = -EINVAL;
 710         } else {
 711                 rc = 0;
 712         }
 713
 714         kfree(cpu_drcs);
 715         return rc;
 716 }
 717
 718 static int find_dlpar_cpus_to_add(u32 *cpu_drcs, u32 cpus_to_add)
 719 {
 720         struct device_node *parent;
 721         int cpus_found = 0;
 722         int index, rc;
 723
 724         parent = of_find_node_by_path("/cpus");
 725         if (!parent) {
 726                 pr_warn("Could not find CPU root node in device tree\n");
 727                 kfree(cpu_drcs);
 728                 return -1;
 729         }
 730
 731         /* Search the ibm,drc-indexes array for possible CPU drcs to
 732          * add. Note that the format of the ibm,drc-indexes array is
 733          * the number of entries in the array followed by the array
 734          * of drc values so we start looking at index = 1.
 735          */
 736         index = 1;
 737         while (cpus_found < cpus_to_add) {
 738                 u32 drc;
 739
 740                 rc = of_property_read_u32_index(parent, "ibm,drc-indexes",
 741                                                 index++, &drc);
 742                 if (rc)
 743                         break;
 744
 745                 if (dlpar_cpu_exists(parent, drc))
 746                         continue;
 747
 748                 cpu_drcs[cpus_found++] = drc;
 749         }
 750
 751         of_node_put(parent);
 752         return cpus_found;
 753 }
 754
 755 static int dlpar_cpu_add_by_count(u32 cpus_to_add)
 756 {
 757         u32 *cpu_drcs;
 758         int cpus_added = 0;
 759         int cpus_found;
 760         int i, rc;
 761
 762         pr_debug("Attempting to hot-add %d CPUs\n", cpus_to_add);
 763
 764         cpu_drcs = kcalloc(cpus_to_add, sizeof(*cpu_drcs), GFP_KERNEL);
 765         if (!cpu_drcs)
 766                 return -EINVAL;
 767
 768         cpus_found = find_dlpar_cpus_to_add(cpu_drcs, cpus_to_add);
 769         if (cpus_found < cpus_to_add) {
 770                 pr_warn("Failed to find enough CPUs (%d of %d) to add\n",
 771                         cpus_found, cpus_to_add);
 772                 kfree(cpu_drcs);
 773                 return -EINVAL;
 774         }
 775
 776         for (i = 0; i < cpus_to_add; i++) {
 777                 rc = dlpar_cpu_add(cpu_drcs[i]);
 778                 if (rc)
 779                         break;
 780
 781                 cpus_added++;
 782         }
 783
 784         if (cpus_added < cpus_to_add) {
 785                 pr_warn("CPU hot-add failed, removing any added CPUs\n");
 786
 787                 for (i = 0; i < cpus_added; i++)
 788                         dlpar_cpu_remove_by_index(cpu_drcs[i]);
 789
 790                 rc = -EINVAL;
 791         } else {
 792                 rc = 0;
 793         }
 794
 795         kfree(cpu_drcs);
 796         return rc;
 797 }
 798
 799 int dlpar_cpu_readd(int cpu)
 800 {
 801         struct device_node *dn;
 802         struct device *dev;
 803         u32 drc_index;
 804         int rc;
 805
 806         dev = get_cpu_device(cpu);
 807         dn = dev->of_node;
 808
 809         rc = of_property_read_u32(dn, "ibm,my-drc-index", &drc_index);
 810
 811         rc = dlpar_cpu_remove_by_index(drc_index);
 812         if (!rc)
 813                 rc = dlpar_cpu_add(drc_index);
 814
 815         return rc;
 816 }
 817
 818 int dlpar_cpu(struct pseries_hp_errorlog *hp_elog)
 819 {
 820         u32 count, drc_index;
 821         int rc;
 822
 823         count = hp_elog->_drc_u.drc_count;
 824         drc_index = hp_elog->_drc_u.drc_index;
 825
 826         lock_device_hotplug();
 827
 828         switch (hp_elog->action) {
 829         case PSERIES_HP_ELOG_ACTION_REMOVE:
 830                 if (hp_elog->id_type == PSERIES_HP_ELOG_ID_DRC_COUNT)
 831                         rc = dlpar_cpu_remove_by_count(count);
 832                 else if (hp_elog->id_type == PSERIES_HP_ELOG_ID_DRC_INDEX)
 833                         rc = dlpar_cpu_remove_by_index(drc_index);
 834                 else
 835                         rc = -EINVAL;
 836                 break;
 837         case PSERIES_HP_ELOG_ACTION_ADD:
 838                 if (hp_elog->id_type == PSERIES_HP_ELOG_ID_DRC_COUNT)
 839                         rc = dlpar_cpu_add_by_count(count);
 840                 else if (hp_elog->id_type == PSERIES_HP_ELOG_ID_DRC_INDEX)
 841                         rc = dlpar_cpu_add(drc_index);
 842                 else
 843                         rc = -EINVAL;
 844                 break;
 845         default:
 846                 pr_err("Invalid action (%d) specified\n", hp_elog->action);
 847                 rc = -EINVAL;
 848                 break;
 849         }
 850
 851         unlock_device_hotplug();
 852         return rc;
 853 }
 854
 855 #ifdef CONFIG_ARCH_CPU_PROBE_RELEASE
 856
 857 static ssize_t dlpar_cpu_probe(const char *buf, size_t count)
 858 {
 859         u32 drc_index;
 860         int rc;
 861
 862         rc = kstrtou32(buf, 0, &drc_index);
 863         if (rc)
 864                 return -EINVAL;
 865
 866         rc = dlpar_cpu_add(drc_index);
 867
 868         return rc ? rc : count;
 869 }
 870
 871 static ssize_t dlpar_cpu_release(const char *buf, size_t count)
 872 {
 873         struct device_node *dn;
 874         u32 drc_index;
 875         int rc;
 876
 877         dn = of_find_node_by_path(buf);
 878         if (!dn)
 879                 return -EINVAL;
 880
 881         rc = of_property_read_u32(dn, "ibm,my-drc-index", &drc_index);
 882         if (rc) {
 883                 of_node_put(dn);
 884                 return -EINVAL;
 885         }
 886
 887         rc = dlpar_cpu_remove(dn, drc_index);
 888         of_node_put(dn);
 889
 890         return rc ? rc : count;
 891 }
 892
 893 #endif /* CONFIG_ARCH_CPU_PROBE_RELEASE */
 894
 895 static int pseries_smp_notifier(struct notifier_block *nb,
 896                                 unsigned long action, void *data)
 897 {
 898         struct of_reconfig_data *rd = data;
 899         int err = 0;
 900
 901         switch (action) {
 902         case OF_RECONFIG_ATTACH_NODE:
 903                 err = pseries_add_processor(rd->dn);
 904                 break;
 905         case OF_RECONFIG_DETACH_NODE:
 906                 pseries_remove_processor(rd->dn);
 907                 break;
 908         }
 909         return notifier_from_errno(err);
 910 }
 911
 912 static struct notifier_block pseries_smp_nb = {
 913         .notifier_call = pseries_smp_notifier,
 914 };
 915
 916 #define MAX_CEDE_LATENCY_LEVELS         4
 917 #define CEDE_LATENCY_PARAM_LENGTH       10
 918 #define CEDE_LATENCY_PARAM_MAX_LENGTH   \
 919         (MAX_CEDE_LATENCY_LEVELS * CEDE_LATENCY_PARAM_LENGTH * sizeof(char))
 920 #define CEDE_LATENCY_TOKEN              45
 921
 922 static char cede_parameters[CEDE_LATENCY_PARAM_MAX_LENGTH];
 923
 924 static int parse_cede_parameters(void)
 925 {
 926         memset(cede_parameters, 0, CEDE_LATENCY_PARAM_MAX_LENGTH);
 927         return rtas_call(rtas_token("ibm,get-system-parameter"), 3, 1,
 928                          NULL,
 929                          CEDE_LATENCY_TOKEN,
 930                          __pa(cede_parameters),
 931                          CEDE_LATENCY_PARAM_MAX_LENGTH);
 932 }
 933
 934 static int __init pseries_cpu_hotplug_init(void)
 935 {
 936         int cpu;
 937         int qcss_tok;
 938
 939 #ifdef CONFIG_ARCH_CPU_PROBE_RELEASE
 940         ppc_md.cpu_probe = dlpar_cpu_probe;
 941         ppc_md.cpu_release = dlpar_cpu_release;
 942 #endif /* CONFIG_ARCH_CPU_PROBE_RELEASE */
 943
 944         rtas_stop_self_token = rtas_token("stop-self");
 945         qcss_tok = rtas_token("query-cpu-stopped-state");
 946
 947         if (rtas_stop_self_token == RTAS_UNKNOWN_SERVICE ||
 948                         qcss_tok == RTAS_UNKNOWN_SERVICE) {
 949                 printk(KERN_INFO "CPU Hotplug not supported by firmware "
 950                                 "- disabling.\n");
 951                 return 0;
 952         }
 953
 954         ppc_md.cpu_die = pseries_mach_cpu_die;
 955         smp_ops->cpu_disable = pseries_cpu_disable;
 956         smp_ops->cpu_die = pseries_cpu_die;
 957
 958         /* Processors can be added/removed only on LPAR */
 959         if (firmware_has_feature(FW_FEATURE_LPAR)) {
 960                 of_reconfig_notifier_register(&pseries_smp_nb);
 961                 cpu_maps_update_begin();
 962                 if (cede_offline_enabled && parse_cede_parameters() == 0) {
 963                         default_offline_state = CPU_STATE_INACTIVE;
 964                         for_each_online_cpu(cpu)
 965                                 set_default_offline_state(cpu);
 966                 }
 967                 cpu_maps_update_done();
 968         }
 969
 970         return 0;
 971 }
 972 machine_arch_initcall(pseries, pseries_cpu_hotplug_init);