arch/powerpc/platforms/pseries/lpar.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * pSeries_lpar.c
   4  * Copyright (C) 2001 Todd Inglett, IBM Corporation
   5  *
   6  * pSeries LPAR support.
   7  */
   8
   9 /* Enables debugging of low-level hash table routines - careful! */
  10 #undef DEBUG
  11 #define pr_fmt(fmt) "lpar: " fmt
  12
  13 #include <linux/kernel.h>
  14 #include <linux/dma-mapping.h>
  15 #include <linux/console.h>
  16 #include <linux/export.h>
  17 #include <linux/jump_label.h>
  18 #include <linux/delay.h>
  19 #include <linux/stop_machine.h>
  20 #include <linux/spinlock.h>
  21 #include <linux/cpuhotplug.h>
  22 #include <linux/workqueue.h>
  23 #include <linux/proc_fs.h>
  24 #include <linux/pgtable.h>
  25 #include <asm/processor.h>
  26 #include <asm/mmu.h>
  27 #include <asm/page.h>
  28 #include <asm/machdep.h>
  29 #include <asm/mmu_context.h>
  30 #include <asm/iommu.h>
  31 #include <asm/tlb.h>
  32 #include <asm/prom.h>
  33 #include <asm/cputable.h>
  34 #include <asm/udbg.h>
  35 #include <asm/smp.h>
  36 #include <asm/trace.h>
  37 #include <asm/firmware.h>
  38 #include <asm/plpar_wrappers.h>
  39 #include <asm/kexec.h>
  40 #include <asm/fadump.h>
  41 #include <asm/asm-prototypes.h>
  42 #include <asm/debugfs.h>
  43 #include <asm/dtl.h>
  44
  45 #include "pseries.h"
  46
  47 /* Flag bits for H_BULK_REMOVE */
  48 #define HBR_REQUEST     0x4000000000000000UL
  49 #define HBR_RESPONSE    0x8000000000000000UL
  50 #define HBR_END         0xc000000000000000UL
  51 #define HBR_AVPN        0x0200000000000000UL
  52 #define HBR_ANDCOND     0x0100000000000000UL
  53
  54
  55 /* in hvCall.S */
  56 EXPORT_SYMBOL(plpar_hcall);
  57 EXPORT_SYMBOL(plpar_hcall9);
  58 EXPORT_SYMBOL(plpar_hcall_norets);
  59
  60 /*
  61  * H_BLOCK_REMOVE supported block size for this page size in segment who's base
  62  * page size is that page size.
  63  *
  64  * The first index is the segment base page size, the second one is the actual
  65  * page size.
  66  */
  67 static int hblkrm_size[MMU_PAGE_COUNT][MMU_PAGE_COUNT] __ro_after_init;
  68
  69 /*
  70  * Due to the involved complexity, and that the current hypervisor is only
  71  * returning this value or 0, we are limiting the support of the H_BLOCK_REMOVE
  72  * buffer size to 8 size block.
  73  */
  74 #define HBLKRM_SUPPORTED_BLOCK_SIZE 8
  75
  76 #ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
  77 static u8 dtl_mask = DTL_LOG_PREEMPT;
  78 #else
  79 static u8 dtl_mask;
  80 #endif
  81
  82 void alloc_dtl_buffers(unsigned long *time_limit)
  83 {
  84         int cpu;
  85         struct paca_struct *pp;
  86         struct dtl_entry *dtl;
  87
  88         for_each_possible_cpu(cpu) {
  89                 pp = paca_ptrs[cpu];
  90                 if (pp->dispatch_log)
  91                         continue;
  92                 dtl = kmem_cache_alloc(dtl_cache, GFP_KERNEL);
  93                 if (!dtl) {
  94                         pr_warn("Failed to allocate dispatch trace log for cpu %d\n",
  95                                 cpu);
  96 #ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
  97                         pr_warn("Stolen time statistics will be unreliable\n");
  98 #endif
  99                         break;
 100                 }
 101
 102                 pp->dtl_ridx = 0;
 103                 pp->dispatch_log = dtl;
 104                 pp->dispatch_log_end = dtl + N_DISPATCH_LOG;
 105                 pp->dtl_curr = dtl;
 106
 107                 if (time_limit && time_after(jiffies, *time_limit)) {
 108                         cond_resched();
 109                         *time_limit = jiffies + HZ;
 110                 }
 111         }
 112 }
 113
 114 void register_dtl_buffer(int cpu)
 115 {
 116         long ret;
 117         struct paca_struct *pp;
 118         struct dtl_entry *dtl;
 119         int hwcpu = get_hard_smp_processor_id(cpu);
 120
 121         pp = paca_ptrs[cpu];
 122         dtl = pp->dispatch_log;
 123         if (dtl && dtl_mask) {
 124                 pp->dtl_ridx = 0;
 125                 pp->dtl_curr = dtl;
 126                 lppaca_of(cpu).dtl_idx = 0;
 127
 128                 /* hypervisor reads buffer length from this field */
 129                 dtl->enqueue_to_dispatch_time = cpu_to_be32(DISPATCH_LOG_BYTES);
 130                 ret = register_dtl(hwcpu, __pa(dtl));
 131                 if (ret)
 132                         pr_err("WARNING: DTL registration of cpu %d (hw %d) failed with %ld\n",
 133                                cpu, hwcpu, ret);
 134
 135                 lppaca_of(cpu).dtl_enable_mask = dtl_mask;
 136         }
 137 }
 138
 139 #ifdef CONFIG_PPC_SPLPAR
 140 struct dtl_worker {
 141         struct delayed_work work;
 142         int cpu;
 143 };
 144
 145 struct vcpu_dispatch_data {
 146         int last_disp_cpu;
 147
 148         int total_disp;
 149
 150         int same_cpu_disp;
 151         int same_chip_disp;
 152         int diff_chip_disp;
 153         int far_chip_disp;
 154
 155         int numa_home_disp;
 156         int numa_remote_disp;
 157         int numa_far_disp;
 158 };
 159
 160 /*
 161  * This represents the number of cpus in the hypervisor. Since there is no
 162  * architected way to discover the number of processors in the host, we
 163  * provision for dealing with NR_CPUS. This is currently 2048 by default, and
 164  * is sufficient for our purposes. This will need to be tweaked if
 165  * CONFIG_NR_CPUS is changed.
 166  */
 167 #define NR_CPUS_H       NR_CPUS
 168
 169 DEFINE_RWLOCK(dtl_access_lock);
 170 static DEFINE_PER_CPU(struct vcpu_dispatch_data, vcpu_disp_data);
 171 static DEFINE_PER_CPU(u64, dtl_entry_ridx);
 172 static DEFINE_PER_CPU(struct dtl_worker, dtl_workers);
 173 static enum cpuhp_state dtl_worker_state;
 174 static DEFINE_MUTEX(dtl_enable_mutex);
 175 static int vcpudispatch_stats_on __read_mostly;
 176 static int vcpudispatch_stats_freq = 50;
 177 static __be32 *vcpu_associativity, *pcpu_associativity;
 178
 179
 180 static void free_dtl_buffers(unsigned long *time_limit)
 181 {
 182 #ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
 183         int cpu;
 184         struct paca_struct *pp;
 185
 186         for_each_possible_cpu(cpu) {
 187                 pp = paca_ptrs[cpu];
 188                 if (!pp->dispatch_log)
 189                         continue;
 190                 kmem_cache_free(dtl_cache, pp->dispatch_log);
 191                 pp->dtl_ridx = 0;
 192                 pp->dispatch_log = 0;
 193                 pp->dispatch_log_end = 0;
 194                 pp->dtl_curr = 0;
 195
 196                 if (time_limit && time_after(jiffies, *time_limit)) {
 197                         cond_resched();
 198                         *time_limit = jiffies + HZ;
 199                 }
 200         }
 201 #endif
 202 }
 203
 204 static int init_cpu_associativity(void)
 205 {
 206         vcpu_associativity = kcalloc(num_possible_cpus() / threads_per_core,
 207                         VPHN_ASSOC_BUFSIZE * sizeof(__be32), GFP_KERNEL);
 208         pcpu_associativity = kcalloc(NR_CPUS_H / threads_per_core,
 209                         VPHN_ASSOC_BUFSIZE * sizeof(__be32), GFP_KERNEL);
 210
 211         if (!vcpu_associativity || !pcpu_associativity) {
 212                 pr_err("error allocating memory for associativity information\n");
 213                 return -ENOMEM;
 214         }
 215
 216         return 0;
 217 }
 218
 219 static void destroy_cpu_associativity(void)
 220 {
 221         kfree(vcpu_associativity);
 222         kfree(pcpu_associativity);
 223         vcpu_associativity = pcpu_associativity = 0;
 224 }
 225
 226 static __be32 *__get_cpu_associativity(int cpu, __be32 *cpu_assoc, int flag)
 227 {
 228         __be32 *assoc;
 229         int rc = 0;
 230
 231         assoc = &cpu_assoc[(int)(cpu / threads_per_core) * VPHN_ASSOC_BUFSIZE];
 232         if (!assoc[0]) {
 233                 rc = hcall_vphn(cpu, flag, &assoc[0]);
 234                 if (rc)
 235                         return NULL;
 236         }
 237
 238         return assoc;
 239 }
 240
 241 static __be32 *get_pcpu_associativity(int cpu)
 242 {
 243         return __get_cpu_associativity(cpu, pcpu_associativity, VPHN_FLAG_PCPU);
 244 }
 245
 246 static __be32 *get_vcpu_associativity(int cpu)
 247 {
 248         return __get_cpu_associativity(cpu, vcpu_associativity, VPHN_FLAG_VCPU);
 249 }
 250
 251 static int cpu_relative_dispatch_distance(int last_disp_cpu, int cur_disp_cpu)
 252 {
 253         __be32 *last_disp_cpu_assoc, *cur_disp_cpu_assoc;
 254
 255         if (last_disp_cpu >= NR_CPUS_H || cur_disp_cpu >= NR_CPUS_H)
 256                 return -EINVAL;
 257
 258         last_disp_cpu_assoc = get_pcpu_associativity(last_disp_cpu);
 259         cur_disp_cpu_assoc = get_pcpu_associativity(cur_disp_cpu);
 260
 261         if (!last_disp_cpu_assoc || !cur_disp_cpu_assoc)
 262                 return -EIO;
 263
 264         return cpu_distance(last_disp_cpu_assoc, cur_disp_cpu_assoc);
 265 }
 266
 267 static int cpu_home_node_dispatch_distance(int disp_cpu)
 268 {
 269         __be32 *disp_cpu_assoc, *vcpu_assoc;
 270         int vcpu_id = smp_processor_id();
 271
 272         if (disp_cpu >= NR_CPUS_H) {
 273                 pr_debug_ratelimited("vcpu dispatch cpu %d > %d\n",
 274                                                 disp_cpu, NR_CPUS_H);
 275                 return -EINVAL;
 276         }
 277
 278         disp_cpu_assoc = get_pcpu_associativity(disp_cpu);
 279         vcpu_assoc = get_vcpu_associativity(vcpu_id);
 280
 281         if (!disp_cpu_assoc || !vcpu_assoc)
 282                 return -EIO;
 283
 284         return cpu_distance(disp_cpu_assoc, vcpu_assoc);
 285 }
 286
 287 static void update_vcpu_disp_stat(int disp_cpu)
 288 {
 289         struct vcpu_dispatch_data *disp;
 290         int distance;
 291
 292         disp = this_cpu_ptr(&vcpu_disp_data);
 293         if (disp->last_disp_cpu == -1) {
 294                 disp->last_disp_cpu = disp_cpu;
 295                 return;
 296         }
 297
 298         disp->total_disp++;
 299
 300         if (disp->last_disp_cpu == disp_cpu ||
 301                 (cpu_first_thread_sibling(disp->last_disp_cpu) ==
 302                                         cpu_first_thread_sibling(disp_cpu)))
 303                 disp->same_cpu_disp++;
 304         else {
 305                 distance = cpu_relative_dispatch_distance(disp->last_disp_cpu,
 306                                                                 disp_cpu);
 307                 if (distance < 0)
 308                         pr_debug_ratelimited("vcpudispatch_stats: cpu %d: error determining associativity\n",
 309                                         smp_processor_id());
 310                 else {
 311                         switch (distance) {
 312                         case 0:
 313                                 disp->same_chip_disp++;
 314                                 break;
 315                         case 1:
 316                                 disp->diff_chip_disp++;
 317                                 break;
 318                         case 2:
 319                                 disp->far_chip_disp++;
 320                                 break;
 321                         default:
 322                                 pr_debug_ratelimited("vcpudispatch_stats: cpu %d (%d -> %d): unexpected relative dispatch distance %d\n",
 323                                                  smp_processor_id(),
 324                                                  disp->last_disp_cpu,
 325                                                  disp_cpu,
 326                                                  distance);
 327                         }
 328                 }
 329         }
 330
 331         distance = cpu_home_node_dispatch_distance(disp_cpu);
 332         if (distance < 0)
 333                 pr_debug_ratelimited("vcpudispatch_stats: cpu %d: error determining associativity\n",
 334                                 smp_processor_id());
 335         else {
 336                 switch (distance) {
 337                 case 0:
 338                         disp->numa_home_disp++;
 339                         break;
 340                 case 1:
 341                         disp->numa_remote_disp++;
 342                         break;
 343                 case 2:
 344                         disp->numa_far_disp++;
 345                         break;
 346                 default:
 347                         pr_debug_ratelimited("vcpudispatch_stats: cpu %d on %d: unexpected numa dispatch distance %d\n",
 348                                                  smp_processor_id(),
 349                                                  disp_cpu,
 350                                                  distance);
 351                 }
 352         }
 353
 354         disp->last_disp_cpu = disp_cpu;
 355 }
 356
 357 static void process_dtl_buffer(struct work_struct *work)
 358 {
 359         struct dtl_entry dtle;
 360         u64 i = __this_cpu_read(dtl_entry_ridx);
 361         struct dtl_entry *dtl = local_paca->dispatch_log + (i % N_DISPATCH_LOG);
 362         struct dtl_entry *dtl_end = local_paca->dispatch_log_end;
 363         struct lppaca *vpa = local_paca->lppaca_ptr;
 364         struct dtl_worker *d = container_of(work, struct dtl_worker, work.work);
 365
 366         if (!local_paca->dispatch_log)
 367                 return;
 368
 369         /* if we have been migrated away, we cancel ourself */
 370         if (d->cpu != smp_processor_id()) {
 371                 pr_debug("vcpudispatch_stats: cpu %d worker migrated -- canceling worker\n",
 372                                                 smp_processor_id());
 373                 return;
 374         }
 375
 376         if (i == be64_to_cpu(vpa->dtl_idx))
 377                 goto out;
 378
 379         while (i < be64_to_cpu(vpa->dtl_idx)) {
 380                 dtle = *dtl;
 381                 barrier();
 382                 if (i + N_DISPATCH_LOG < be64_to_cpu(vpa->dtl_idx)) {
 383                         /* buffer has overflowed */
 384                         pr_debug_ratelimited("vcpudispatch_stats: cpu %d lost %lld DTL samples\n",
 385                                 d->cpu,
 386                                 be64_to_cpu(vpa->dtl_idx) - N_DISPATCH_LOG - i);
 387                         i = be64_to_cpu(vpa->dtl_idx) - N_DISPATCH_LOG;
 388                         dtl = local_paca->dispatch_log + (i % N_DISPATCH_LOG);
 389                         continue;
 390                 }
 391                 update_vcpu_disp_stat(be16_to_cpu(dtle.processor_id));
 392                 ++i;
 393                 ++dtl;
 394                 if (dtl == dtl_end)
 395                         dtl = local_paca->dispatch_log;
 396         }
 397
 398         __this_cpu_write(dtl_entry_ridx, i);
 399
 400 out:
 401         schedule_delayed_work_on(d->cpu, to_delayed_work(work),
 402                                         HZ / vcpudispatch_stats_freq);
 403 }
 404
 405 static int dtl_worker_online(unsigned int cpu)
 406 {
 407         struct dtl_worker *d = &per_cpu(dtl_workers, cpu);
 408
 409         memset(d, 0, sizeof(*d));
 410         INIT_DELAYED_WORK(&d->work, process_dtl_buffer);
 411         d->cpu = cpu;
 412
 413 #ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
 414         per_cpu(dtl_entry_ridx, cpu) = 0;
 415         register_dtl_buffer(cpu);
 416 #else
 417         per_cpu(dtl_entry_ridx, cpu) = be64_to_cpu(lppaca_of(cpu).dtl_idx);
 418 #endif
 419
 420         schedule_delayed_work_on(cpu, &d->work, HZ / vcpudispatch_stats_freq);
 421         return 0;
 422 }
 423
 424 static int dtl_worker_offline(unsigned int cpu)
 425 {
 426         struct dtl_worker *d = &per_cpu(dtl_workers, cpu);
 427
 428         cancel_delayed_work_sync(&d->work);
 429
 430 #ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
 431         unregister_dtl(get_hard_smp_processor_id(cpu));
 432 #endif
 433
 434         return 0;
 435 }
 436
 437 static void set_global_dtl_mask(u8 mask)
 438 {
 439         int cpu;
 440
 441         dtl_mask = mask;
 442         for_each_present_cpu(cpu)
 443                 lppaca_of(cpu).dtl_enable_mask = dtl_mask;
 444 }
 445
 446 static void reset_global_dtl_mask(void)
 447 {
 448         int cpu;
 449
 450 #ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
 451         dtl_mask = DTL_LOG_PREEMPT;
 452 #else
 453         dtl_mask = 0;
 454 #endif
 455         for_each_present_cpu(cpu)
 456                 lppaca_of(cpu).dtl_enable_mask = dtl_mask;
 457 }
 458
 459 static int dtl_worker_enable(unsigned long *time_limit)
 460 {
 461         int rc = 0, state;
 462
 463         if (!write_trylock(&dtl_access_lock)) {
 464                 rc = -EBUSY;
 465                 goto out;
 466         }
 467
 468         set_global_dtl_mask(DTL_LOG_ALL);
 469
 470         /* Setup dtl buffers and register those */
 471         alloc_dtl_buffers(time_limit);
 472
 473         state = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "powerpc/dtl:online",
 474                                         dtl_worker_online, dtl_worker_offline);
 475         if (state < 0) {
 476                 pr_err("vcpudispatch_stats: unable to setup workqueue for DTL processing\n");
 477                 free_dtl_buffers(time_limit);
 478                 reset_global_dtl_mask();
 479                 write_unlock(&dtl_access_lock);
 480                 rc = -EINVAL;
 481                 goto out;
 482         }
 483         dtl_worker_state = state;
 484
 485 out:
 486         return rc;
 487 }
 488
 489 static void dtl_worker_disable(unsigned long *time_limit)
 490 {
 491         cpuhp_remove_state(dtl_worker_state);
 492         free_dtl_buffers(time_limit);
 493         reset_global_dtl_mask();
 494         write_unlock(&dtl_access_lock);
 495 }
 496
 497 static ssize_t vcpudispatch_stats_write(struct file *file, const char __user *p,
 498                 size_t count, loff_t *ppos)
 499 {
 500         unsigned long time_limit = jiffies + HZ;
 501         struct vcpu_dispatch_data *disp;
 502         int rc, cmd, cpu;
 503         char buf[16];
 504
 505         if (count > 15)
 506                 return -EINVAL;
 507
 508         if (copy_from_user(buf, p, count))
 509                 return -EFAULT;
 510
 511         buf[count] = 0;
 512         rc = kstrtoint(buf, 0, &cmd);
 513         if (rc || cmd < 0 || cmd > 1) {
 514                 pr_err("vcpudispatch_stats: please use 0 to disable or 1 to enable dispatch statistics\n");
 515                 return rc ? rc : -EINVAL;
 516         }
 517
 518         mutex_lock(&dtl_enable_mutex);
 519
 520         if ((cmd == 0 && !vcpudispatch_stats_on) ||
 521                         (cmd == 1 && vcpudispatch_stats_on))
 522                 goto out;
 523
 524         if (cmd) {
 525                 rc = init_cpu_associativity();
 526                 if (rc)
 527                         goto out;
 528
 529                 for_each_possible_cpu(cpu) {
 530                         disp = per_cpu_ptr(&vcpu_disp_data, cpu);
 531                         memset(disp, 0, sizeof(*disp));
 532                         disp->last_disp_cpu = -1;
 533                 }
 534
 535                 rc = dtl_worker_enable(&time_limit);
 536                 if (rc) {
 537                         destroy_cpu_associativity();
 538                         goto out;
 539                 }
 540         } else {
 541                 dtl_worker_disable(&time_limit);
 542                 destroy_cpu_associativity();
 543         }
 544
 545         vcpudispatch_stats_on = cmd;
 546
 547 out:
 548         mutex_unlock(&dtl_enable_mutex);
 549         if (rc)
 550                 return rc;
 551         return count;
 552 }
 553
 554 static int vcpudispatch_stats_display(struct seq_file *p, void *v)
 555 {
 556         int cpu;
 557         struct vcpu_dispatch_data *disp;
 558
 559         if (!vcpudispatch_stats_on) {
 560                 seq_puts(p, "off\n");
 561                 return 0;
 562         }
 563
 564         for_each_online_cpu(cpu) {
 565                 disp = per_cpu_ptr(&vcpu_disp_data, cpu);
 566                 seq_printf(p, "cpu%d", cpu);
 567                 seq_put_decimal_ull(p, " ", disp->total_disp);
 568                 seq_put_decimal_ull(p, " ", disp->same_cpu_disp);
 569                 seq_put_decimal_ull(p, " ", disp->same_chip_disp);
 570                 seq_put_decimal_ull(p, " ", disp->diff_chip_disp);
 571                 seq_put_decimal_ull(p, " ", disp->far_chip_disp);
 572                 seq_put_decimal_ull(p, " ", disp->numa_home_disp);
 573                 seq_put_decimal_ull(p, " ", disp->numa_remote_disp);
 574                 seq_put_decimal_ull(p, " ", disp->numa_far_disp);
 575                 seq_puts(p, "\n");
 576         }
 577
 578         return 0;
 579 }
 580
 581 static int vcpudispatch_stats_open(struct inode *inode, struct file *file)
 582 {
 583         return single_open(file, vcpudispatch_stats_display, NULL);
 584 }
 585
 586 static const struct proc_ops vcpudispatch_stats_proc_ops = {
 587         .proc_open      = vcpudispatch_stats_open,
 588         .proc_read      = seq_read,
 589         .proc_write     = vcpudispatch_stats_write,
 590         .proc_lseek     = seq_lseek,
 591         .proc_release   = single_release,
 592 };
 593
 594 static ssize_t vcpudispatch_stats_freq_write(struct file *file,
 595                 const char __user *p, size_t count, loff_t *ppos)
 596 {
 597         int rc, freq;
 598         char buf[16];
 599
 600         if (count > 15)
 601                 return -EINVAL;
 602
 603         if (copy_from_user(buf, p, count))
 604                 return -EFAULT;
 605
 606         buf[count] = 0;
 607         rc = kstrtoint(buf, 0, &freq);
 608         if (rc || freq < 1 || freq > HZ) {
 609                 pr_err("vcpudispatch_stats_freq: please specify a frequency between 1 and %d\n",
 610                                 HZ);
 611                 return rc ? rc : -EINVAL;
 612         }
 613
 614         vcpudispatch_stats_freq = freq;
 615
 616         return count;
 617 }
 618
 619 static int vcpudispatch_stats_freq_display(struct seq_file *p, void *v)
 620 {
 621         seq_printf(p, "%d\n", vcpudispatch_stats_freq);
 622         return 0;
 623 }
 624
 625 static int vcpudispatch_stats_freq_open(struct inode *inode, struct file *file)
 626 {
 627         return single_open(file, vcpudispatch_stats_freq_display, NULL);
 628 }
 629
 630 static const struct proc_ops vcpudispatch_stats_freq_proc_ops = {
 631         .proc_open      = vcpudispatch_stats_freq_open,
 632         .proc_read      = seq_read,
 633         .proc_write     = vcpudispatch_stats_freq_write,
 634         .proc_lseek     = seq_lseek,
 635         .proc_release   = single_release,
 636 };
 637
 638 static int __init vcpudispatch_stats_procfs_init(void)
 639 {
 640         /*
 641          * Avoid smp_processor_id while preemptible. All CPUs should have
 642          * the same value for lppaca_shared_proc.
 643          */
 644         preempt_disable();
 645         if (!lppaca_shared_proc(get_lppaca())) {
 646                 preempt_enable();
 647                 return 0;
 648         }
 649         preempt_enable();
 650
 651         if (!proc_create("powerpc/vcpudispatch_stats", 0600, NULL,
 652                                         &vcpudispatch_stats_proc_ops))
 653                 pr_err("vcpudispatch_stats: error creating procfs file\n");
 654         else if (!proc_create("powerpc/vcpudispatch_stats_freq", 0600, NULL,
 655                                         &vcpudispatch_stats_freq_proc_ops))
 656                 pr_err("vcpudispatch_stats_freq: error creating procfs file\n");
 657
 658         return 0;
 659 }
 660
 661 machine_device_initcall(pseries, vcpudispatch_stats_procfs_init);
 662 #endif /* CONFIG_PPC_SPLPAR */
 663
 664 void vpa_init(int cpu)
 665 {
 666         int hwcpu = get_hard_smp_processor_id(cpu);
 667         unsigned long addr;
 668         long ret;
 669
 670         /*
 671          * The spec says it "may be problematic" if CPU x registers the VPA of
 672          * CPU y. We should never do that, but wail if we ever do.
 673          */
 674         WARN_ON(cpu != smp_processor_id());
 675
 676         if (cpu_has_feature(CPU_FTR_ALTIVEC))
 677                 lppaca_of(cpu).vmxregs_in_use = 1;
 678
 679         if (cpu_has_feature(CPU_FTR_ARCH_207S))
 680                 lppaca_of(cpu).ebb_regs_in_use = 1;
 681
 682         addr = __pa(&lppaca_of(cpu));
 683         ret = register_vpa(hwcpu, addr);
 684
 685         if (ret) {
 686                 pr_err("WARNING: VPA registration for cpu %d (hw %d) of area "
 687                        "%lx failed with %ld\n", cpu, hwcpu, addr, ret);
 688                 return;
 689         }
 690
 691 #ifdef CONFIG_PPC_BOOK3S_64
 692         /*
 693          * PAPR says this feature is SLB-Buffer but firmware never
 694          * reports that.  All SPLPAR support SLB shadow buffer.
 695          */
 696         if (!radix_enabled() && firmware_has_feature(FW_FEATURE_SPLPAR)) {
 697                 addr = __pa(paca_ptrs[cpu]->slb_shadow_ptr);
 698                 ret = register_slb_shadow(hwcpu, addr);
 699                 if (ret)
 700                         pr_err("WARNING: SLB shadow buffer registration for "
 701                                "cpu %d (hw %d) of area %lx failed with %ld\n",
 702                                cpu, hwcpu, addr, ret);
 703         }
 704 #endif /* CONFIG_PPC_BOOK3S_64 */
 705
 706         /*
 707          * Register dispatch trace log, if one has been allocated.
 708          */
 709         register_dtl_buffer(cpu);
 710 }
 711
 712 #ifdef CONFIG_PPC_BOOK3S_64
 713
 714 static long pSeries_lpar_hpte_insert(unsigned long hpte_group,
 715                                      unsigned long vpn, unsigned long pa,
 716                                      unsigned long rflags, unsigned long vflags,
 717                                      int psize, int apsize, int ssize)
 718 {
 719         unsigned long lpar_rc;
 720         unsigned long flags;
 721         unsigned long slot;
 722         unsigned long hpte_v, hpte_r;
 723
 724         if (!(vflags & HPTE_V_BOLTED))
 725                 pr_devel("hpte_insert(group=%lx, vpn=%016lx, "
 726                          "pa=%016lx, rflags=%lx, vflags=%lx, psize=%d)\n",
 727                          hpte_group, vpn,  pa, rflags, vflags, psize);
 728
 729         hpte_v = hpte_encode_v(vpn, psize, apsize, ssize) | vflags | HPTE_V_VALID;
 730         hpte_r = hpte_encode_r(pa, psize, apsize) | rflags;
 731
 732         if (!(vflags & HPTE_V_BOLTED))
 733                 pr_devel(" hpte_v=%016lx, hpte_r=%016lx\n", hpte_v, hpte_r);
 734
 735         /* Now fill in the actual HPTE */
 736         /* Set CEC cookie to 0         */
 737         /* Zero page = 0               */
 738         /* I-cache Invalidate = 0      */
 739         /* I-cache synchronize = 0     */
 740         /* Exact = 0                   */
 741         flags = 0;
 742
 743         if (firmware_has_feature(FW_FEATURE_XCMO) && !(hpte_r & HPTE_R_N))
 744                 flags |= H_COALESCE_CAND;
 745
 746         lpar_rc = plpar_pte_enter(flags, hpte_group, hpte_v, hpte_r, &slot);
 747         if (unlikely(lpar_rc == H_PTEG_FULL)) {
 748                 pr_devel("Hash table group is full\n");
 749                 return -1;
 750         }
 751
 752         /*
 753          * Since we try and ioremap PHBs we don't own, the pte insert
 754          * will fail. However we must catch the failure in hash_page
 755          * or we will loop forever, so return -2 in this case.
 756          */
 757         if (unlikely(lpar_rc != H_SUCCESS)) {
 758                 pr_err("Failed hash pte insert with error %ld\n", lpar_rc);
 759                 return -2;
 760         }
 761         if (!(vflags & HPTE_V_BOLTED))
 762                 pr_devel(" -> slot: %lu\n", slot & 7);
 763
 764         /* Because of iSeries, we have to pass down the secondary
 765          * bucket bit here as well
 766          */
 767         return (slot & 7) | (!!(vflags & HPTE_V_SECONDARY) << 3);
 768 }
 769
 770 static DEFINE_SPINLOCK(pSeries_lpar_tlbie_lock);
 771
 772 static long pSeries_lpar_hpte_remove(unsigned long hpte_group)
 773 {
 774         unsigned long slot_offset;
 775         unsigned long lpar_rc;
 776         int i;
 777         unsigned long dummy1, dummy2;
 778
 779         /* pick a random slot to start at */
 780         slot_offset = mftb() & 0x7;
 781
 782         for (i = 0; i < HPTES_PER_GROUP; i++) {
 783
 784                 /* don't remove a bolted entry */
 785                 lpar_rc = plpar_pte_remove(H_ANDCOND, hpte_group + slot_offset,
 786                                            HPTE_V_BOLTED, &dummy1, &dummy2);
 787                 if (lpar_rc == H_SUCCESS)
 788                         return i;
 789
 790                 /*
 791                  * The test for adjunct partition is performed before the
 792                  * ANDCOND test.  H_RESOURCE may be returned, so we need to
 793                  * check for that as well.
 794                  */
 795                 BUG_ON(lpar_rc != H_NOT_FOUND && lpar_rc != H_RESOURCE);
 796
 797                 slot_offset++;
 798                 slot_offset &= 0x7;
 799         }
 800
 801         return -1;
 802 }
 803
 804 static void manual_hpte_clear_all(void)
 805 {
 806         unsigned long size_bytes = 1UL << ppc64_pft_size;
 807         unsigned long hpte_count = size_bytes >> 4;
 808         struct {
 809                 unsigned long pteh;
 810                 unsigned long ptel;
 811         } ptes[4];
 812         long lpar_rc;
 813         unsigned long i, j;
 814
 815         /* Read in batches of 4,
 816          * invalidate only valid entries not in the VRMA
 817          * hpte_count will be a multiple of 4
 818          */
 819         for (i = 0; i < hpte_count; i += 4) {
 820                 lpar_rc = plpar_pte_read_4_raw(0, i, (void *)ptes);
 821                 if (lpar_rc != H_SUCCESS) {
 822                         pr_info("Failed to read hash page table at %ld err %ld\n",
 823                                 i, lpar_rc);
 824                         continue;
 825                 }
 826                 for (j = 0; j < 4; j++){
 827                         if ((ptes[j].pteh & HPTE_V_VRMA_MASK) ==
 828                                 HPTE_V_VRMA_MASK)
 829                                 continue;
 830                         if (ptes[j].pteh & HPTE_V_VALID)
 831                                 plpar_pte_remove_raw(0, i + j, 0,
 832                                         &(ptes[j].pteh), &(ptes[j].ptel));
 833                 }
 834         }
 835 }
 836
 837 static int hcall_hpte_clear_all(void)
 838 {
 839         int rc;
 840
 841         do {
 842                 rc = plpar_hcall_norets(H_CLEAR_HPT);
 843         } while (rc == H_CONTINUE);
 844
 845         return rc;
 846 }
 847
 848 static void pseries_hpte_clear_all(void)
 849 {
 850         int rc;
 851
 852         rc = hcall_hpte_clear_all();
 853         if (rc != H_SUCCESS)
 854                 manual_hpte_clear_all();
 855
 856 #ifdef __LITTLE_ENDIAN__
 857         /*
 858          * Reset exceptions to big endian.
 859          *
 860          * FIXME this is a hack for kexec, we need to reset the exception
 861          * endian before starting the new kernel and this is a convenient place
 862          * to do it.
 863          *
 864          * This is also called on boot when a fadump happens. In that case we
 865          * must not change the exception endian mode.
 866          */
 867         if (firmware_has_feature(FW_FEATURE_SET_MODE) && !is_fadump_active())
 868                 pseries_big_endian_exceptions();
 869 #endif
 870 }
 871
 872 /*
 873  * NOTE: for updatepp ops we are fortunate that the linux "newpp" bits and
 874  * the low 3 bits of flags happen to line up.  So no transform is needed.
 875  * We can probably optimize here and assume the high bits of newpp are
 876  * already zero.  For now I am paranoid.
 877  */
 878 static long pSeries_lpar_hpte_updatepp(unsigned long slot,
 879                                        unsigned long newpp,
 880                                        unsigned long vpn,
 881                                        int psize, int apsize,
 882                                        int ssize, unsigned long inv_flags)
 883 {
 884         unsigned long lpar_rc;
 885         unsigned long flags;
 886         unsigned long want_v;
 887
 888         want_v = hpte_encode_avpn(vpn, psize, ssize);
 889
 890         flags = (newpp & (HPTE_R_PP | HPTE_R_N | HPTE_R_KEY_LO)) | H_AVPN;
 891         flags |= (newpp & HPTE_R_KEY_HI) >> 48;
 892         if (mmu_has_feature(MMU_FTR_KERNEL_RO))
 893                 /* Move pp0 into bit 8 (IBM 55) */
 894                 flags |= (newpp & HPTE_R_PP0) >> 55;
 895
 896         pr_devel("    update: avpnv=%016lx, hash=%016lx, f=%lx, psize: %d ...",
 897                  want_v, slot, flags, psize);
 898
 899         lpar_rc = plpar_pte_protect(flags, slot, want_v);
 900
 901         if (lpar_rc == H_NOT_FOUND) {
 902                 pr_devel("not found !\n");
 903                 return -1;
 904         }
 905
 906         pr_devel("ok\n");
 907
 908         BUG_ON(lpar_rc != H_SUCCESS);
 909
 910         return 0;
 911 }
 912
 913 static long __pSeries_lpar_hpte_find(unsigned long want_v, unsigned long hpte_group)
 914 {
 915         long lpar_rc;
 916         unsigned long i, j;
 917         struct {
 918                 unsigned long pteh;
 919                 unsigned long ptel;
 920         } ptes[4];
 921
 922         for (i = 0; i < HPTES_PER_GROUP; i += 4, hpte_group += 4) {
 923
 924                 lpar_rc = plpar_pte_read_4(0, hpte_group, (void *)ptes);
 925                 if (lpar_rc != H_SUCCESS) {
 926                         pr_info("Failed to read hash page table at %ld err %ld\n",
 927                                 hpte_group, lpar_rc);
 928                         continue;
 929                 }
 930
 931                 for (j = 0; j < 4; j++) {
 932                         if (HPTE_V_COMPARE(ptes[j].pteh, want_v) &&
 933                             (ptes[j].pteh & HPTE_V_VALID))
 934                                 return i + j;
 935                 }
 936         }
 937
 938         return -1;
 939 }
 940
 941 static long pSeries_lpar_hpte_find(unsigned long vpn, int psize, int ssize)
 942 {
 943         long slot;
 944         unsigned long hash;
 945         unsigned long want_v;
 946         unsigned long hpte_group;
 947
 948         hash = hpt_hash(vpn, mmu_psize_defs[psize].shift, ssize);
 949         want_v = hpte_encode_avpn(vpn, psize, ssize);
 950
 951         /*
 952          * We try to keep bolted entries always in primary hash
 953          * But in some case we can find them in secondary too.
 954          */
 955         hpte_group = (hash & htab_hash_mask) * HPTES_PER_GROUP;
 956         slot = __pSeries_lpar_hpte_find(want_v, hpte_group);
 957         if (slot < 0) {
 958                 /* Try in secondary */
 959                 hpte_group = (~hash & htab_hash_mask) * HPTES_PER_GROUP;
 960                 slot = __pSeries_lpar_hpte_find(want_v, hpte_group);
 961                 if (slot < 0)
 962                         return -1;
 963         }
 964         return hpte_group + slot;
 965 }
 966
 967 static void pSeries_lpar_hpte_updateboltedpp(unsigned long newpp,
 968                                              unsigned long ea,
 969                                              int psize, int ssize)
 970 {
 971         unsigned long vpn;
 972         unsigned long lpar_rc, slot, vsid, flags;
 973
 974         vsid = get_kernel_vsid(ea, ssize);
 975         vpn = hpt_vpn(ea, vsid, ssize);
 976
 977         slot = pSeries_lpar_hpte_find(vpn, psize, ssize);
 978         BUG_ON(slot == -1);
 979
 980         flags = newpp & (HPTE_R_PP | HPTE_R_N);
 981         if (mmu_has_feature(MMU_FTR_KERNEL_RO))
 982                 /* Move pp0 into bit 8 (IBM 55) */
 983                 flags |= (newpp & HPTE_R_PP0) >> 55;
 984
 985         flags |= ((newpp & HPTE_R_KEY_HI) >> 48) | (newpp & HPTE_R_KEY_LO);
 986
 987         lpar_rc = plpar_pte_protect(flags, slot, 0);
 988
 989         BUG_ON(lpar_rc != H_SUCCESS);
 990 }
 991
 992 static void pSeries_lpar_hpte_invalidate(unsigned long slot, unsigned long vpn,
 993                                          int psize, int apsize,
 994                                          int ssize, int local)
 995 {
 996         unsigned long want_v;
 997         unsigned long lpar_rc;
 998         unsigned long dummy1, dummy2;
 999
1000         pr_devel("    inval : slot=%lx, vpn=%016lx, psize: %d, local: %d\n",
1001                  slot, vpn, psize, local);
1002
1003         want_v = hpte_encode_avpn(vpn, psize, ssize);
1004         lpar_rc = plpar_pte_remove(H_AVPN, slot, want_v, &dummy1, &dummy2);
1005         if (lpar_rc == H_NOT_FOUND)
1006                 return;
1007
1008         BUG_ON(lpar_rc != H_SUCCESS);
1009 }
1010
1011
1012 /*
1013  * As defined in the PAPR's section 14.5.4.1.8
1014  * The control mask doesn't include the returned reference and change bit from
1015  * the processed PTE.
1016  */
1017 #define HBLKR_AVPN              0x0100000000000000UL
1018 #define HBLKR_CTRL_MASK         0xf800000000000000UL
1019 #define HBLKR_CTRL_SUCCESS      0x8000000000000000UL
1020 #define HBLKR_CTRL_ERRNOTFOUND  0x8800000000000000UL
1021 #define HBLKR_CTRL_ERRBUSY      0xa000000000000000UL
1022
1023 /*
1024  * Returned true if we are supporting this block size for the specified segment
1025  * base page size and actual page size.
1026  *
1027  * Currently, we only support 8 size block.
1028  */
1029 static inline bool is_supported_hlbkrm(int bpsize, int psize)
1030 {
1031         return (hblkrm_size[bpsize][psize] == HBLKRM_SUPPORTED_BLOCK_SIZE);
1032 }
1033
1034 /**
1035  * H_BLOCK_REMOVE caller.
1036  * @idx should point to the latest @param entry set with a PTEX.
1037  * If PTE cannot be processed because another CPUs has already locked that
1038  * group, those entries are put back in @param starting at index 1.
1039  * If entries has to be retried and @retry_busy is set to true, these entries
1040  * are retried until success. If @retry_busy is set to false, the returned
1041  * is the number of entries yet to process.
1042  */
1043 static unsigned long call_block_remove(unsigned long idx, unsigned long *param,
1044                                        bool retry_busy)
1045 {
1046         unsigned long i, rc, new_idx;
1047         unsigned long retbuf[PLPAR_HCALL9_BUFSIZE];
1048
1049         if (idx < 2) {
1050                 pr_warn("Unexpected empty call to H_BLOCK_REMOVE");
1051                 return 0;
1052         }
1053 again:
1054         new_idx = 0;
1055         if (idx > PLPAR_HCALL9_BUFSIZE) {
1056                 pr_err("Too many PTEs (%lu) for H_BLOCK_REMOVE", idx);
1057                 idx = PLPAR_HCALL9_BUFSIZE;
1058         } else if (idx < PLPAR_HCALL9_BUFSIZE)
1059                 param[idx] = HBR_END;
1060
1061         rc = plpar_hcall9(H_BLOCK_REMOVE, retbuf,
1062                           param[0], /* AVA */
1063                           param[1],  param[2],  param[3],  param[4], /* TS0-7 */
1064                           param[5],  param[6],  param[7],  param[8]);
1065         if (rc == H_SUCCESS)
1066                 return 0;
1067
1068         BUG_ON(rc != H_PARTIAL);
1069
1070         /* Check that the unprocessed entries were 'not found' or 'busy' */
1071         for (i = 0; i < idx-1; i++) {
1072                 unsigned long ctrl = retbuf[i] & HBLKR_CTRL_MASK;
1073
1074                 if (ctrl == HBLKR_CTRL_ERRBUSY) {
1075                         param[++new_idx] = param[i+1];
1076                         continue;
1077                 }
1078
1079                 BUG_ON(ctrl != HBLKR_CTRL_SUCCESS
1080                        && ctrl != HBLKR_CTRL_ERRNOTFOUND);
1081         }
1082
1083         /*
1084          * If there were entries found busy, retry these entries if requested,
1085          * of if all the entries have to be retried.
1086          */
1087         if (new_idx && (retry_busy || new_idx == (PLPAR_HCALL9_BUFSIZE-1))) {
1088                 idx = new_idx + 1;
1089                 goto again;
1090         }
1091
1092         return new_idx;
1093 }
1094
1095 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1096 /*
1097  * Limit iterations holding pSeries_lpar_tlbie_lock to 3. We also need
1098  * to make sure that we avoid bouncing the hypervisor tlbie lock.
1099  */
1100 #define PPC64_HUGE_HPTE_BATCH 12
1101
1102 static void hugepage_block_invalidate(unsigned long *slot, unsigned long *vpn,
1103                                       int count, int psize, int ssize)
1104 {
1105         unsigned long param[PLPAR_HCALL9_BUFSIZE];
1106         unsigned long shift, current_vpgb, vpgb;
1107         int i, pix = 0;
1108
1109         shift = mmu_psize_defs[psize].shift;
1110
1111         for (i = 0; i < count; i++) {
1112                 /*
1113                  * Shifting 3 bits more on the right to get a
1114                  * 8 pages aligned virtual addresse.
1115                  */
1116                 vpgb = (vpn[i] >> (shift - VPN_SHIFT + 3));
1117                 if (!pix || vpgb != current_vpgb) {
1118                         /*
1119                          * Need to start a new 8 pages block, flush
1120                          * the current one if needed.
1121                          */
1122                         if (pix)
1123                                 (void)call_block_remove(pix, param, true);
1124                         current_vpgb = vpgb;
1125                         param[0] = hpte_encode_avpn(vpn[i], psize, ssize);
1126                         pix = 1;
1127                 }
1128
1129                 param[pix++] = HBR_REQUEST | HBLKR_AVPN | slot[i];
1130                 if (pix == PLPAR_HCALL9_BUFSIZE) {
1131                         pix = call_block_remove(pix, param, false);
1132                         /*
1133                          * pix = 0 means that all the entries were
1134                          * removed, we can start a new block.
1135                          * Otherwise, this means that there are entries
1136                          * to retry, and pix points to latest one, so
1137                          * we should increment it and try to continue
1138                          * the same block.
1139                          */
1140                         if (pix)
1141                                 pix++;
1142                 }
1143         }
1144         if (pix)
1145                 (void)call_block_remove(pix, param, true);
1146 }
1147
1148 static void hugepage_bulk_invalidate(unsigned long *slot, unsigned long *vpn,
1149                                      int count, int psize, int ssize)
1150 {
1151         unsigned long param[PLPAR_HCALL9_BUFSIZE];
1152         int i = 0, pix = 0, rc;
1153
1154         for (i = 0; i < count; i++) {
1155
1156                 if (!firmware_has_feature(FW_FEATURE_BULK_REMOVE)) {
1157                         pSeries_lpar_hpte_invalidate(slot[i], vpn[i], psize, 0,
1158                                                      ssize, 0);
1159                 } else {
1160                         param[pix] = HBR_REQUEST | HBR_AVPN | slot[i];
1161                         param[pix+1] = hpte_encode_avpn(vpn[i], psize, ssize);
1162                         pix += 2;
1163                         if (pix == 8) {
1164                                 rc = plpar_hcall9(H_BULK_REMOVE, param,
1165                                                   param[0], param[1], param[2],
1166                                                   param[3], param[4], param[5],
1167                                                   param[6], param[7]);
1168                                 BUG_ON(rc != H_SUCCESS);
1169                                 pix = 0;
1170                         }
1171                 }
1172         }
1173         if (pix) {
1174                 param[pix] = HBR_END;
1175                 rc = plpar_hcall9(H_BULK_REMOVE, param, param[0], param[1],
1176                                   param[2], param[3], param[4], param[5],
1177                                   param[6], param[7]);
1178                 BUG_ON(rc != H_SUCCESS);
1179         }
1180 }
1181
1182 static inline void __pSeries_lpar_hugepage_invalidate(unsigned long *slot,
1183                                                       unsigned long *vpn,
1184                                                       int count, int psize,
1185                                                       int ssize)
1186 {
1187         unsigned long flags = 0;
1188         int lock_tlbie = !mmu_has_feature(MMU_FTR_LOCKLESS_TLBIE);
1189
1190         if (lock_tlbie)
1191                 spin_lock_irqsave(&pSeries_lpar_tlbie_lock, flags);
1192
1193         /* Assuming THP size is 16M */
1194         if (is_supported_hlbkrm(psize, MMU_PAGE_16M))
1195                 hugepage_block_invalidate(slot, vpn, count, psize, ssize);
1196         else
1197                 hugepage_bulk_invalidate(slot, vpn, count, psize, ssize);
1198
1199         if (lock_tlbie)
1200                 spin_unlock_irqrestore(&pSeries_lpar_tlbie_lock, flags);
1201 }
1202
1203 static void pSeries_lpar_hugepage_invalidate(unsigned long vsid,
1204                                              unsigned long addr,
1205                                              unsigned char *hpte_slot_array,
1206                                              int psize, int ssize, int local)
1207 {
1208         int i, index = 0;
1209         unsigned long s_addr = addr;
1210         unsigned int max_hpte_count, valid;
1211         unsigned long vpn_array[PPC64_HUGE_HPTE_BATCH];
1212         unsigned long slot_array[PPC64_HUGE_HPTE_BATCH];
1213         unsigned long shift, hidx, vpn = 0, hash, slot;
1214
1215         shift = mmu_psize_defs[psize].shift;
1216         max_hpte_count = 1U << (PMD_SHIFT - shift);
1217
1218         for (i = 0; i < max_hpte_count; i++) {
1219                 valid = hpte_valid(hpte_slot_array, i);
1220                 if (!valid)
1221                         continue;
1222                 hidx =  hpte_hash_index(hpte_slot_array, i);
1223
1224                 /* get the vpn */
1225                 addr = s_addr + (i * (1ul << shift));
1226                 vpn = hpt_vpn(addr, vsid, ssize);
1227                 hash = hpt_hash(vpn, shift, ssize);
1228                 if (hidx & _PTEIDX_SECONDARY)
1229                         hash = ~hash;
1230
1231                 slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
1232                 slot += hidx & _PTEIDX_GROUP_IX;
1233
1234                 slot_array[index] = slot;
1235                 vpn_array[index] = vpn;
1236                 if (index == PPC64_HUGE_HPTE_BATCH - 1) {
1237                         /*
1238                          * Now do a bluk invalidate
1239                          */
1240                         __pSeries_lpar_hugepage_invalidate(slot_array,
1241                                                            vpn_array,
1242                                                            PPC64_HUGE_HPTE_BATCH,
1243                                                            psize, ssize);
1244                         index = 0;
1245                 } else
1246                         index++;
1247         }
1248         if (index)
1249                 __pSeries_lpar_hugepage_invalidate(slot_array, vpn_array,
1250                                                    index, psize, ssize);
1251 }
1252 #else
1253 static void pSeries_lpar_hugepage_invalidate(unsigned long vsid,
1254                                              unsigned long addr,
1255                                              unsigned char *hpte_slot_array,
1256                                              int psize, int ssize, int local)
1257 {
1258         WARN(1, "%s called without THP support\n", __func__);
1259 }
1260 #endif
1261
1262 static int pSeries_lpar_hpte_removebolted(unsigned long ea,
1263                                           int psize, int ssize)
1264 {
1265         unsigned long vpn;
1266         unsigned long slot, vsid;
1267
1268         vsid = get_kernel_vsid(ea, ssize);
1269         vpn = hpt_vpn(ea, vsid, ssize);
1270
1271         slot = pSeries_lpar_hpte_find(vpn, psize, ssize);
1272         if (slot == -1)
1273                 return -ENOENT;
1274
1275         /*
1276          * lpar doesn't use the passed actual page size
1277          */
1278         pSeries_lpar_hpte_invalidate(slot, vpn, psize, 0, ssize, 0);
1279         return 0;
1280 }
1281
1282
1283 static inline unsigned long compute_slot(real_pte_t pte,
1284                                          unsigned long vpn,
1285                                          unsigned long index,
1286                                          unsigned long shift,
1287                                          int ssize)
1288 {
1289         unsigned long slot, hash, hidx;
1290
1291         hash = hpt_hash(vpn, shift, ssize);
1292         hidx = __rpte_to_hidx(pte, index);
1293         if (hidx & _PTEIDX_SECONDARY)
1294                 hash = ~hash;
1295         slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
1296         slot += hidx & _PTEIDX_GROUP_IX;
1297         return slot;
1298 }
1299
1300 /**
1301  * The hcall H_BLOCK_REMOVE implies that the virtual pages to processed are
1302  * "all within the same naturally aligned 8 page virtual address block".
1303  */
1304 static void do_block_remove(unsigned long number, struct ppc64_tlb_batch *batch,
1305                             unsigned long *param)
1306 {
1307         unsigned long vpn;
1308         unsigned long i, pix = 0;
1309         unsigned long index, shift, slot, current_vpgb, vpgb;
1310         real_pte_t pte;
1311         int psize, ssize;
1312
1313         psize = batch->psize;
1314         ssize = batch->ssize;
1315
1316         for (i = 0; i < number; i++) {
1317                 vpn = batch->vpn[i];
1318                 pte = batch->pte[i];
1319                 pte_iterate_hashed_subpages(pte, psize, vpn, index, shift) {
1320                         /*
1321                          * Shifting 3 bits more on the right to get a
1322                          * 8 pages aligned virtual addresse.
1323                          */
1324                         vpgb = (vpn >> (shift - VPN_SHIFT + 3));
1325                         if (!pix || vpgb != current_vpgb) {
1326                                 /*
1327                                  * Need to start a new 8 pages block, flush
1328                                  * the current one if needed.
1329                                  */
1330                                 if (pix)
1331                                         (void)call_block_remove(pix, param,
1332                                                                 true);
1333                                 current_vpgb = vpgb;
1334                                 param[0] = hpte_encode_avpn(vpn, psize,
1335                                                             ssize);
1336                                 pix = 1;
1337                         }
1338
1339                         slot = compute_slot(pte, vpn, index, shift, ssize);
1340                         param[pix++] = HBR_REQUEST | HBLKR_AVPN | slot;
1341
1342                         if (pix == PLPAR_HCALL9_BUFSIZE) {
1343                                 pix = call_block_remove(pix, param, false);
1344                                 /*
1345                                  * pix = 0 means that all the entries were
1346                                  * removed, we can start a new block.
1347                                  * Otherwise, this means that there are entries
1348                                  * to retry, and pix points to latest one, so
1349                                  * we should increment it and try to continue
1350                                  * the same block.
1351                                  */
1352                                 if (pix)
1353                                         pix++;
1354                         }
1355                 } pte_iterate_hashed_end();
1356         }
1357
1358         if (pix)
1359                 (void)call_block_remove(pix, param, true);
1360 }
1361
1362 /*
1363  * TLB Block Invalidate Characteristics
1364  *
1365  * These characteristics define the size of the block the hcall H_BLOCK_REMOVE
1366  * is able to process for each couple segment base page size, actual page size.
1367  *
1368  * The ibm,get-system-parameter properties is returning a buffer with the
1369  * following layout:
1370  *
1371  * [ 2 bytes size of the RTAS buffer (excluding these 2 bytes) ]
1372  * -----------------
1373  * TLB Block Invalidate Specifiers:
1374  * [ 1 byte LOG base 2 of the TLB invalidate block size being specified ]
1375  * [ 1 byte Number of page sizes (N) that are supported for the specified
1376  *          TLB invalidate block size ]
1377  * [ 1 byte Encoded segment base page size and actual page size
1378  *          MSB=0 means 4k segment base page size and actual page size
1379  *          MSB=1 the penc value in mmu_psize_def ]
1380  * ...
1381  * -----------------
1382  * Next TLB Block Invalidate Specifiers...
1383  * -----------------
1384  * [ 0 ]
1385  */
1386 static inline void set_hblkrm_bloc_size(int bpsize, int psize,
1387                                         unsigned int block_size)
1388 {
1389         if (block_size > hblkrm_size[bpsize][psize])
1390                 hblkrm_size[bpsize][psize] = block_size;
1391 }
1392
1393 /*
1394  * Decode the Encoded segment base page size and actual page size.
1395  * PAPR specifies:
1396  *   - bit 7 is the L bit
1397  *   - bits 0-5 are the penc value
1398  * If the L bit is 0, this means 4K segment base page size and actual page size
1399  * otherwise the penc value should be read.
1400  */
1401 #define HBLKRM_L_MASK           0x80
1402 #define HBLKRM_PENC_MASK        0x3f
1403 static inline void __init check_lp_set_hblkrm(unsigned int lp,
1404                                               unsigned int block_size)
1405 {
1406         unsigned int bpsize, psize;
1407
1408         /* First, check the L bit, if not set, this means 4K */
1409         if ((lp & HBLKRM_L_MASK) == 0) {
1410                 set_hblkrm_bloc_size(MMU_PAGE_4K, MMU_PAGE_4K, block_size);
1411                 return;
1412         }
1413
1414         lp &= HBLKRM_PENC_MASK;
1415         for (bpsize = 0; bpsize < MMU_PAGE_COUNT; bpsize++) {
1416                 struct mmu_psize_def *def = &mmu_psize_defs[bpsize];
1417
1418                 for (psize = 0; psize < MMU_PAGE_COUNT; psize++) {
1419                         if (def->penc[psize] == lp) {
1420                                 set_hblkrm_bloc_size(bpsize, psize, block_size);
1421                                 return;
1422                         }
1423                 }
1424         }
1425 }
1426
1427 #define SPLPAR_TLB_BIC_TOKEN            50
1428
1429 /*
1430  * The size of the TLB Block Invalidate Characteristics is variable. But at the
1431  * maximum it will be the number of possible page sizes *2 + 10 bytes.
1432  * Currently MMU_PAGE_COUNT is 16, which means 42 bytes. Use a cache line size
1433  * (128 bytes) for the buffer to get plenty of space.
1434  */
1435 #define SPLPAR_TLB_BIC_MAXLENGTH        128
1436
1437 void __init pseries_lpar_read_hblkrm_characteristics(void)
1438 {
1439         unsigned char local_buffer[SPLPAR_TLB_BIC_MAXLENGTH];
1440         int call_status, len, idx, bpsize;
1441
1442         if (!firmware_has_feature(FW_FEATURE_BLOCK_REMOVE))
1443                 return;
1444
1445         spin_lock(&rtas_data_buf_lock);
1446         memset(rtas_data_buf, 0, RTAS_DATA_BUF_SIZE);
1447         call_status = rtas_call(rtas_token("ibm,get-system-parameter"), 3, 1,
1448                                 NULL,
1449                                 SPLPAR_TLB_BIC_TOKEN,
1450                                 __pa(rtas_data_buf),
1451                                 RTAS_DATA_BUF_SIZE);
1452         memcpy(local_buffer, rtas_data_buf, SPLPAR_TLB_BIC_MAXLENGTH);
1453         local_buffer[SPLPAR_TLB_BIC_MAXLENGTH - 1] = '\0';
1454         spin_unlock(&rtas_data_buf_lock);
1455
1456         if (call_status != 0) {
1457                 pr_warn("%s %s Error calling get-system-parameter (0x%x)\n",
1458                         __FILE__, __func__, call_status);
1459                 return;
1460         }
1461
1462         /*
1463          * The first two (2) bytes of the data in the buffer are the length of
1464          * the returned data, not counting these first two (2) bytes.
1465          */
1466         len = be16_to_cpu(*((u16 *)local_buffer)) + 2;
1467         if (len > SPLPAR_TLB_BIC_MAXLENGTH) {
1468                 pr_warn("%s too large returned buffer %d", __func__, len);
1469                 return;
1470         }
1471
1472         idx = 2;
1473         while (idx < len) {
1474                 u8 block_shift = local_buffer[idx++];
1475                 u32 block_size;
1476                 unsigned int npsize;
1477
1478                 if (!block_shift)
1479                         break;
1480
1481                 block_size = 1 << block_shift;
1482
1483                 for (npsize = local_buffer[idx++];
1484                      npsize > 0 && idx < len; npsize--)
1485                         check_lp_set_hblkrm((unsigned int) local_buffer[idx++],
1486                                             block_size);
1487         }
1488
1489         for (bpsize = 0; bpsize < MMU_PAGE_COUNT; bpsize++)
1490                 for (idx = 0; idx < MMU_PAGE_COUNT; idx++)
1491                         if (hblkrm_size[bpsize][idx])
1492                                 pr_info("H_BLOCK_REMOVE supports base psize:%d psize:%d block size:%d",
1493                                         bpsize, idx, hblkrm_size[bpsize][idx]);
1494 }
1495
1496 /*
1497  * Take a spinlock around flushes to avoid bouncing the hypervisor tlbie
1498  * lock.
1499  */
1500 static void pSeries_lpar_flush_hash_range(unsigned long number, int local)
1501 {
1502         unsigned long vpn;
1503         unsigned long i, pix, rc;
1504         unsigned long flags = 0;
1505         struct ppc64_tlb_batch *batch = this_cpu_ptr(&ppc64_tlb_batch);
1506         int lock_tlbie = !mmu_has_feature(MMU_FTR_LOCKLESS_TLBIE);
1507         unsigned long param[PLPAR_HCALL9_BUFSIZE];
1508         unsigned long index, shift, slot;
1509         real_pte_t pte;
1510         int psize, ssize;
1511
1512         if (lock_tlbie)
1513                 spin_lock_irqsave(&pSeries_lpar_tlbie_lock, flags);
1514
1515         if (is_supported_hlbkrm(batch->psize, batch->psize)) {
1516                 do_block_remove(number, batch, param);
1517                 goto out;
1518         }
1519
1520         psize = batch->psize;
1521         ssize = batch->ssize;
1522         pix = 0;
1523         for (i = 0; i < number; i++) {
1524                 vpn = batch->vpn[i];
1525                 pte = batch->pte[i];
1526                 pte_iterate_hashed_subpages(pte, psize, vpn, index, shift) {
1527                         slot = compute_slot(pte, vpn, index, shift, ssize);
1528                         if (!firmware_has_feature(FW_FEATURE_BULK_REMOVE)) {
1529                                 /*
1530                                  * lpar doesn't use the passed actual page size
1531                                  */
1532                                 pSeries_lpar_hpte_invalidate(slot, vpn, psize,
1533                                                              0, ssize, local);
1534                         } else {
1535                                 param[pix] = HBR_REQUEST | HBR_AVPN | slot;
1536                                 param[pix+1] = hpte_encode_avpn(vpn, psize,
1537                                                                 ssize);
1538                                 pix += 2;
1539                                 if (pix == 8) {
1540                                         rc = plpar_hcall9(H_BULK_REMOVE, param,
1541                                                 param[0], param[1], param[2],
1542                                                 param[3], param[4], param[5],
1543                                                 param[6], param[7]);
1544                                         BUG_ON(rc != H_SUCCESS);
1545                                         pix = 0;
1546                                 }
1547                         }
1548                 } pte_iterate_hashed_end();
1549         }
1550         if (pix) {
1551                 param[pix] = HBR_END;
1552                 rc = plpar_hcall9(H_BULK_REMOVE, param, param[0], param[1],
1553                                   param[2], param[3], param[4], param[5],
1554                                   param[6], param[7]);
1555                 BUG_ON(rc != H_SUCCESS);
1556         }
1557
1558 out:
1559         if (lock_tlbie)
1560                 spin_unlock_irqrestore(&pSeries_lpar_tlbie_lock, flags);
1561 }
1562
1563 static int __init disable_bulk_remove(char *str)
1564 {
1565         if (strcmp(str, "off") == 0 &&
1566             firmware_has_feature(FW_FEATURE_BULK_REMOVE)) {
1567                 pr_info("Disabling BULK_REMOVE firmware feature");
1568                 powerpc_firmware_features &= ~FW_FEATURE_BULK_REMOVE;
1569         }
1570         return 1;
1571 }
1572
1573 __setup("bulk_remove=", disable_bulk_remove);
1574
1575 #define HPT_RESIZE_TIMEOUT      10000 /* ms */
1576
1577 struct hpt_resize_state {
1578         unsigned long shift;
1579         int commit_rc;
1580 };
1581
1582 static int pseries_lpar_resize_hpt_commit(void *data)
1583 {
1584         struct hpt_resize_state *state = data;
1585
1586         state->commit_rc = plpar_resize_hpt_commit(0, state->shift);
1587         if (state->commit_rc != H_SUCCESS)
1588                 return -EIO;
1589
1590         /* Hypervisor has transitioned the HTAB, update our globals */
1591         ppc64_pft_size = state->shift;
1592         htab_size_bytes = 1UL << ppc64_pft_size;
1593         htab_hash_mask = (htab_size_bytes >> 7) - 1;
1594
1595         return 0;
1596 }
1597
1598 /*
1599  * Must be called in process context. The caller must hold the
1600  * cpus_lock.
1601  */
1602 static int pseries_lpar_resize_hpt(unsigned long shift)
1603 {
1604         struct hpt_resize_state state = {
1605                 .shift = shift,
1606                 .commit_rc = H_FUNCTION,
1607         };
1608         unsigned int delay, total_delay = 0;
1609         int rc;
1610         ktime_t t0, t1, t2;
1611
1612         might_sleep();
1613
1614         if (!firmware_has_feature(FW_FEATURE_HPT_RESIZE))
1615                 return -ENODEV;
1616
1617         pr_info("Attempting to resize HPT to shift %lu\n", shift);
1618
1619         t0 = ktime_get();
1620
1621         rc = plpar_resize_hpt_prepare(0, shift);
1622         while (H_IS_LONG_BUSY(rc)) {
1623                 delay = get_longbusy_msecs(rc);
1624                 total_delay += delay;
1625                 if (total_delay > HPT_RESIZE_TIMEOUT) {
1626                         /* prepare with shift==0 cancels an in-progress resize */
1627                         rc = plpar_resize_hpt_prepare(0, 0);
1628                         if (rc != H_SUCCESS)
1629                                 pr_warn("Unexpected error %d cancelling timed out HPT resize\n",
1630                                        rc);
1631                         return -ETIMEDOUT;
1632                 }
1633                 msleep(delay);
1634                 rc = plpar_resize_hpt_prepare(0, shift);
1635         }
1636
1637         switch (rc) {
1638         case H_SUCCESS:
1639                 /* Continue on */
1640                 break;
1641
1642         case H_PARAMETER:
1643                 pr_warn("Invalid argument from H_RESIZE_HPT_PREPARE\n");
1644                 return -EINVAL;
1645         case H_RESOURCE:
1646                 pr_warn("Operation not permitted from H_RESIZE_HPT_PREPARE\n");
1647                 return -EPERM;
1648         default:
1649                 pr_warn("Unexpected error %d from H_RESIZE_HPT_PREPARE\n", rc);
1650                 return -EIO;
1651         }
1652
1653         t1 = ktime_get();
1654
1655         rc = stop_machine_cpuslocked(pseries_lpar_resize_hpt_commit,
1656                                      &state, NULL);
1657
1658         t2 = ktime_get();
1659
1660         if (rc != 0) {
1661                 switch (state.commit_rc) {
1662                 case H_PTEG_FULL:
1663                         return -ENOSPC;
1664
1665                 default:
1666                         pr_warn("Unexpected error %d from H_RESIZE_HPT_COMMIT\n",
1667                                 state.commit_rc);
1668                         return -EIO;
1669                 };
1670         }
1671
1672         pr_info("HPT resize to shift %lu complete (%lld ms / %lld ms)\n",
1673                 shift, (long long) ktime_ms_delta(t1, t0),
1674                 (long long) ktime_ms_delta(t2, t1));
1675
1676         return 0;
1677 }
1678
1679 static int pseries_lpar_register_process_table(unsigned long base,
1680                         unsigned long page_size, unsigned long table_size)
1681 {
1682         long rc;
1683         unsigned long flags = 0;
1684
1685         if (table_size)
1686                 flags |= PROC_TABLE_NEW;
1687         if (radix_enabled()) {
1688                 flags |= PROC_TABLE_RADIX;
1689                 if (mmu_has_feature(MMU_FTR_GTSE))
1690                         flags |= PROC_TABLE_GTSE;
1691         } else
1692                 flags |= PROC_TABLE_HPT_SLB;
1693         for (;;) {
1694                 rc = plpar_hcall_norets(H_REGISTER_PROC_TBL, flags, base,
1695                                         page_size, table_size);
1696                 if (!H_IS_LONG_BUSY(rc))
1697                         break;
1698                 mdelay(get_longbusy_msecs(rc));
1699         }
1700         if (rc != H_SUCCESS) {
1701                 pr_err("Failed to register process table (rc=%ld)\n", rc);
1702                 BUG();
1703         }
1704         return rc;
1705 }
1706
1707 void __init hpte_init_pseries(void)
1708 {
1709         mmu_hash_ops.hpte_invalidate     = pSeries_lpar_hpte_invalidate;
1710         mmu_hash_ops.hpte_updatepp       = pSeries_lpar_hpte_updatepp;
1711         mmu_hash_ops.hpte_updateboltedpp = pSeries_lpar_hpte_updateboltedpp;
1712         mmu_hash_ops.hpte_insert         = pSeries_lpar_hpte_insert;
1713         mmu_hash_ops.hpte_remove         = pSeries_lpar_hpte_remove;
1714         mmu_hash_ops.hpte_removebolted   = pSeries_lpar_hpte_removebolted;
1715         mmu_hash_ops.flush_hash_range    = pSeries_lpar_flush_hash_range;
1716         mmu_hash_ops.hpte_clear_all      = pseries_hpte_clear_all;
1717         mmu_hash_ops.hugepage_invalidate = pSeries_lpar_hugepage_invalidate;
1718
1719         if (firmware_has_feature(FW_FEATURE_HPT_RESIZE))
1720                 mmu_hash_ops.resize_hpt = pseries_lpar_resize_hpt;
1721
1722         /*
1723          * On POWER9, we need to do a H_REGISTER_PROC_TBL hcall
1724          * to inform the hypervisor that we wish to use the HPT.
1725          */
1726         if (cpu_has_feature(CPU_FTR_ARCH_300))
1727                 pseries_lpar_register_process_table(0, 0, 0);
1728 }
1729
1730 #ifdef CONFIG_PPC_RADIX_MMU
1731 void radix_init_pseries(void)
1732 {
1733         pr_info("Using radix MMU under hypervisor\n");
1734
1735         pseries_lpar_register_process_table(__pa(process_tb),
1736                                                 0, PRTB_SIZE_SHIFT - 12);
1737 }
1738 #endif
1739
1740 #ifdef CONFIG_PPC_SMLPAR
1741 #define CMO_FREE_HINT_DEFAULT 1
1742 static int cmo_free_hint_flag = CMO_FREE_HINT_DEFAULT;
1743
1744 static int __init cmo_free_hint(char *str)
1745 {
1746         char *parm;
1747         parm = strstrip(str);
1748
1749         if (strcasecmp(parm, "no") == 0 || strcasecmp(parm, "off") == 0) {
1750                 pr_info("%s: CMO free page hinting is not active.\n", __func__);
1751                 cmo_free_hint_flag = 0;
1752                 return 1;
1753         }
1754
1755         cmo_free_hint_flag = 1;
1756         pr_info("%s: CMO free page hinting is active.\n", __func__);
1757
1758         if (strcasecmp(parm, "yes") == 0 || strcasecmp(parm, "on") == 0)
1759                 return 1;
1760
1761         return 0;
1762 }
1763
1764 __setup("cmo_free_hint=", cmo_free_hint);
1765
1766 static void pSeries_set_page_state(struct page *page, int order,
1767                                    unsigned long state)
1768 {
1769         int i, j;
1770         unsigned long cmo_page_sz, addr;
1771
1772         cmo_page_sz = cmo_get_page_size();
1773         addr = __pa((unsigned long)page_address(page));
1774
1775         for (i = 0; i < (1 << order); i++, addr += PAGE_SIZE) {
1776                 for (j = 0; j < PAGE_SIZE; j += cmo_page_sz)
1777                         plpar_hcall_norets(H_PAGE_INIT, state, addr + j, 0);
1778         }
1779 }
1780
1781 void arch_free_page(struct page *page, int order)
1782 {
1783         if (radix_enabled())
1784                 return;
1785         if (!cmo_free_hint_flag || !firmware_has_feature(FW_FEATURE_CMO))
1786                 return;
1787
1788         pSeries_set_page_state(page, order, H_PAGE_SET_UNUSED);
1789 }
1790 EXPORT_SYMBOL(arch_free_page);
1791
1792 #endif /* CONFIG_PPC_SMLPAR */
1793 #endif /* CONFIG_PPC_BOOK3S_64 */
1794
1795 #ifdef CONFIG_TRACEPOINTS
1796 #ifdef CONFIG_JUMP_LABEL
1797 struct static_key hcall_tracepoint_key = STATIC_KEY_INIT;
1798
1799 int hcall_tracepoint_regfunc(void)
1800 {
1801         static_key_slow_inc(&hcall_tracepoint_key);
1802         return 0;
1803 }
1804
1805 void hcall_tracepoint_unregfunc(void)
1806 {
1807         static_key_slow_dec(&hcall_tracepoint_key);
1808 }
1809 #else
1810 /*
1811  * We optimise our hcall path by placing hcall_tracepoint_refcount
1812  * directly in the TOC so we can check if the hcall tracepoints are
1813  * enabled via a single load.
1814  */
1815
1816 /* NB: reg/unreg are called while guarded with the tracepoints_mutex */
1817 extern long hcall_tracepoint_refcount;
1818
1819 int hcall_tracepoint_regfunc(void)
1820 {
1821         hcall_tracepoint_refcount++;
1822         return 0;
1823 }
1824
1825 void hcall_tracepoint_unregfunc(void)
1826 {
1827         hcall_tracepoint_refcount--;
1828 }
1829 #endif
1830
1831 /*
1832  * Since the tracing code might execute hcalls we need to guard against
1833  * recursion. One example of this are spinlocks calling H_YIELD on
1834  * shared processor partitions.
1835  */
1836 static DEFINE_PER_CPU(unsigned int, hcall_trace_depth);
1837
1838
1839 void __trace_hcall_entry(unsigned long opcode, unsigned long *args)
1840 {
1841         unsigned long flags;
1842         unsigned int *depth;
1843
1844         /*
1845          * We cannot call tracepoints inside RCU idle regions which
1846          * means we must not trace H_CEDE.
1847          */
1848         if (opcode == H_CEDE)
1849                 return;
1850
1851         local_irq_save(flags);
1852
1853         depth = this_cpu_ptr(&hcall_trace_depth);
1854
1855         if (*depth)
1856                 goto out;
1857
1858         (*depth)++;
1859         preempt_disable();
1860         trace_hcall_entry(opcode, args);
1861         (*depth)--;
1862
1863 out:
1864         local_irq_restore(flags);
1865 }
1866
1867 void __trace_hcall_exit(long opcode, long retval, unsigned long *retbuf)
1868 {
1869         unsigned long flags;
1870         unsigned int *depth;
1871
1872         if (opcode == H_CEDE)
1873                 return;
1874
1875         local_irq_save(flags);
1876
1877         depth = this_cpu_ptr(&hcall_trace_depth);
1878
1879         if (*depth)
1880                 goto out;
1881
1882         (*depth)++;
1883         trace_hcall_exit(opcode, retval, retbuf);
1884         preempt_enable();
1885         (*depth)--;
1886
1887 out:
1888         local_irq_restore(flags);
1889 }
1890 #endif
1891
1892 /**
1893  * h_get_mpp
1894  * H_GET_MPP hcall returns info in 7 parms
1895  */
1896 int h_get_mpp(struct hvcall_mpp_data *mpp_data)
1897 {
1898         int rc;
1899         unsigned long retbuf[PLPAR_HCALL9_BUFSIZE];
1900
1901         rc = plpar_hcall9(H_GET_MPP, retbuf);
1902
1903         mpp_data->entitled_mem = retbuf[0];
1904         mpp_data->mapped_mem = retbuf[1];
1905
1906         mpp_data->group_num = (retbuf[2] >> 2 * 8) & 0xffff;
1907         mpp_data->pool_num = retbuf[2] & 0xffff;
1908
1909         mpp_data->mem_weight = (retbuf[3] >> 7 * 8) & 0xff;
1910         mpp_data->unallocated_mem_weight = (retbuf[3] >> 6 * 8) & 0xff;
1911         mpp_data->unallocated_entitlement = retbuf[3] & 0xffffffffffffUL;
1912
1913         mpp_data->pool_size = retbuf[4];
1914         mpp_data->loan_request = retbuf[5];
1915         mpp_data->backing_mem = retbuf[6];
1916
1917         return rc;
1918 }
1919 EXPORT_SYMBOL(h_get_mpp);
1920
1921 int h_get_mpp_x(struct hvcall_mpp_x_data *mpp_x_data)
1922 {
1923         int rc;
1924         unsigned long retbuf[PLPAR_HCALL9_BUFSIZE] = { 0 };
1925
1926         rc = plpar_hcall9(H_GET_MPP_X, retbuf);
1927
1928         mpp_x_data->coalesced_bytes = retbuf[0];
1929         mpp_x_data->pool_coalesced_bytes = retbuf[1];
1930         mpp_x_data->pool_purr_cycles = retbuf[2];
1931         mpp_x_data->pool_spurr_cycles = retbuf[3];
1932
1933         return rc;
1934 }
1935
1936 static unsigned long vsid_unscramble(unsigned long vsid, int ssize)
1937 {
1938         unsigned long protovsid;
1939         unsigned long va_bits = VA_BITS;
1940         unsigned long modinv, vsid_modulus;
1941         unsigned long max_mod_inv, tmp_modinv;
1942
1943         if (!mmu_has_feature(MMU_FTR_68_BIT_VA))
1944                 va_bits = 65;
1945
1946         if (ssize == MMU_SEGSIZE_256M) {
1947                 modinv = VSID_MULINV_256M;
1948                 vsid_modulus = ((1UL << (va_bits - SID_SHIFT)) - 1);
1949         } else {
1950                 modinv = VSID_MULINV_1T;
1951                 vsid_modulus = ((1UL << (va_bits - SID_SHIFT_1T)) - 1);
1952         }
1953
1954         /*
1955          * vsid outside our range.
1956          */
1957         if (vsid >= vsid_modulus)
1958                 return 0;
1959
1960         /*
1961          * If modinv is the modular multiplicate inverse of (x % vsid_modulus)
1962          * and vsid = (protovsid * x) % vsid_modulus, then we say:
1963          *   protovsid = (vsid * modinv) % vsid_modulus
1964          */
1965
1966         /* Check if (vsid * modinv) overflow (63 bits) */
1967         max_mod_inv = 0x7fffffffffffffffull / vsid;
1968         if (modinv < max_mod_inv)
1969                 return (vsid * modinv) % vsid_modulus;
1970
1971         tmp_modinv = modinv/max_mod_inv;
1972         modinv %= max_mod_inv;
1973
1974         protovsid = (((vsid * max_mod_inv) % vsid_modulus) * tmp_modinv) % vsid_modulus;
1975         protovsid = (protovsid + vsid * modinv) % vsid_modulus;
1976
1977         return protovsid;
1978 }
1979
1980 static int __init reserve_vrma_context_id(void)
1981 {
1982         unsigned long protovsid;
1983
1984         /*
1985          * Reserve context ids which map to reserved virtual addresses. For now
1986          * we only reserve the context id which maps to the VRMA VSID. We ignore
1987          * the addresses in "ibm,adjunct-virtual-addresses" because we don't
1988          * enable adjunct support via the "ibm,client-architecture-support"
1989          * interface.
1990          */
1991         protovsid = vsid_unscramble(VRMA_VSID, MMU_SEGSIZE_1T);
1992         hash__reserve_context_id(protovsid >> ESID_BITS_1T);
1993         return 0;
1994 }
1995 machine_device_initcall(pseries, reserve_vrma_context_id);
1996
1997 #ifdef CONFIG_DEBUG_FS
1998 /* debugfs file interface for vpa data */
1999 static ssize_t vpa_file_read(struct file *filp, char __user *buf, size_t len,
2000                               loff_t *pos)
2001 {
2002         int cpu = (long)filp->private_data;
2003         struct lppaca *lppaca = &lppaca_of(cpu);
2004
2005         return simple_read_from_buffer(buf, len, pos, lppaca,
2006                                 sizeof(struct lppaca));
2007 }
2008
2009 static const struct file_operations vpa_fops = {
2010         .open           = simple_open,
2011         .read           = vpa_file_read,
2012         .llseek         = default_llseek,
2013 };
2014
2015 static int __init vpa_debugfs_init(void)
2016 {
2017         char name[16];
2018         long i;
2019         struct dentry *vpa_dir;
2020
2021         if (!firmware_has_feature(FW_FEATURE_SPLPAR))
2022                 return 0;
2023
2024         vpa_dir = debugfs_create_dir("vpa", powerpc_debugfs_root);
2025
2026         /* set up the per-cpu vpa file*/
2027         for_each_possible_cpu(i) {
2028                 sprintf(name, "cpu-%ld", i);
2029                 debugfs_create_file(name, 0400, vpa_dir, (void *)i, &vpa_fops);
2030         }
2031
2032         return 0;
2033 }
2034 machine_arch_initcall(pseries, vpa_debugfs_init);
2035 #endif /* CONFIG_DEBUG_FS */