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 <asm/processor.h>
  25 #include <asm/mmu.h>
  26 #include <asm/page.h>
  27 #include <asm/pgtable.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
  44 #include "pseries.h"
  45
  46 /* Flag bits for H_BULK_REMOVE */
  47 #define HBR_REQUEST     0x4000000000000000UL
  48 #define HBR_RESPONSE    0x8000000000000000UL
  49 #define HBR_END         0xc000000000000000UL
  50 #define HBR_AVPN        0x0200000000000000UL
  51 #define HBR_ANDCOND     0x0100000000000000UL
  52
  53
  54 /* in hvCall.S */
  55 EXPORT_SYMBOL(plpar_hcall);
  56 EXPORT_SYMBOL(plpar_hcall9);
  57 EXPORT_SYMBOL(plpar_hcall_norets);
  58
  59 /*
  60  * H_BLOCK_REMOVE supported block size for this page size in segment who's base
  61  * page size is that page size.
  62  *
  63  * The first index is the segment base page size, the second one is the actual
  64  * page size.
  65  */
  66 static int hblkrm_size[MMU_PAGE_COUNT][MMU_PAGE_COUNT] __ro_after_init;
  67
  68 /*
  69  * Due to the involved complexity, and that the current hypervisor is only
  70  * returning this value or 0, we are limiting the support of the H_BLOCK_REMOVE
  71  * buffer size to 8 size block.
  72  */
  73 #define HBLKRM_SUPPORTED_BLOCK_SIZE 8
  74
  75 #ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
  76 static u8 dtl_mask = DTL_LOG_PREEMPT;
  77 #else
  78 static u8 dtl_mask;
  79 #endif
  80
  81 void alloc_dtl_buffers(unsigned long *time_limit)
  82 {
  83         int cpu;
  84         struct paca_struct *pp;
  85         struct dtl_entry *dtl;
  86
  87         for_each_possible_cpu(cpu) {
  88                 pp = paca_ptrs[cpu];
  89                 if (pp->dispatch_log)
  90                         continue;
  91                 dtl = kmem_cache_alloc(dtl_cache, GFP_KERNEL);
  92                 if (!dtl) {
  93                         pr_warn("Failed to allocate dispatch trace log for cpu %d\n",
  94                                 cpu);
  95 #ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
  96                         pr_warn("Stolen time statistics will be unreliable\n");
  97 #endif
  98                         break;
  99                 }
 100
 101                 pp->dtl_ridx = 0;
 102                 pp->dispatch_log = dtl;
 103                 pp->dispatch_log_end = dtl + N_DISPATCH_LOG;
 104                 pp->dtl_curr = dtl;
 105
 106                 if (time_limit && time_after(jiffies, *time_limit)) {
 107                         cond_resched();
 108                         *time_limit = jiffies + HZ;
 109                 }
 110         }
 111 }
 112
 113 void register_dtl_buffer(int cpu)
 114 {
 115         long ret;
 116         struct paca_struct *pp;
 117         struct dtl_entry *dtl;
 118         int hwcpu = get_hard_smp_processor_id(cpu);
 119
 120         pp = paca_ptrs[cpu];
 121         dtl = pp->dispatch_log;
 122         if (dtl && dtl_mask) {
 123                 pp->dtl_ridx = 0;
 124                 pp->dtl_curr = dtl;
 125                 lppaca_of(cpu).dtl_idx = 0;
 126
 127                 /* hypervisor reads buffer length from this field */
 128                 dtl->enqueue_to_dispatch_time = cpu_to_be32(DISPATCH_LOG_BYTES);
 129                 ret = register_dtl(hwcpu, __pa(dtl));
 130                 if (ret)
 131                         pr_err("WARNING: DTL registration of cpu %d (hw %d) failed with %ld\n",
 132                                cpu, hwcpu, ret);
 133
 134                 lppaca_of(cpu).dtl_enable_mask = dtl_mask;
 135         }
 136 }
 137
 138 #ifdef CONFIG_PPC_SPLPAR
 139 struct dtl_worker {
 140         struct delayed_work work;
 141         int cpu;
 142 };
 143
 144 struct vcpu_dispatch_data {
 145         int last_disp_cpu;
 146
 147         int total_disp;
 148
 149         int same_cpu_disp;
 150         int same_chip_disp;
 151         int diff_chip_disp;
 152         int far_chip_disp;
 153
 154         int numa_home_disp;
 155         int numa_remote_disp;
 156         int numa_far_disp;
 157 };
 158
 159 /*
 160  * This represents the number of cpus in the hypervisor. Since there is no
 161  * architected way to discover the number of processors in the host, we
 162  * provision for dealing with NR_CPUS. This is currently 2048 by default, and
 163  * is sufficient for our purposes. This will need to be tweaked if
 164  * CONFIG_NR_CPUS is changed.
 165  */
 166 #define NR_CPUS_H       NR_CPUS
 167
 168 DEFINE_RWLOCK(dtl_access_lock);
 169 static DEFINE_PER_CPU(struct vcpu_dispatch_data, vcpu_disp_data);
 170 static DEFINE_PER_CPU(u64, dtl_entry_ridx);
 171 static DEFINE_PER_CPU(struct dtl_worker, dtl_workers);
 172 static enum cpuhp_state dtl_worker_state;
 173 static DEFINE_MUTEX(dtl_enable_mutex);
 174 static int vcpudispatch_stats_on __read_mostly;
 175 static int vcpudispatch_stats_freq = 50;
 176 static __be32 *vcpu_associativity, *pcpu_associativity;
 177
 178
 179 static void free_dtl_buffers(unsigned long *time_limit)
 180 {
 181 #ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
 182         int cpu;
 183         struct paca_struct *pp;
 184
 185         for_each_possible_cpu(cpu) {
 186                 pp = paca_ptrs[cpu];
 187                 if (!pp->dispatch_log)
 188                         continue;
 189                 kmem_cache_free(dtl_cache, pp->dispatch_log);
 190                 pp->dtl_ridx = 0;
 191                 pp->dispatch_log = 0;
 192                 pp->dispatch_log_end = 0;
 193                 pp->dtl_curr = 0;
 194
 195                 if (time_limit && time_after(jiffies, *time_limit)) {
 196                         cond_resched();
 197                         *time_limit = jiffies + HZ;
 198                 }
 199         }
 200 #endif
 201 }
 202
 203 static int init_cpu_associativity(void)
 204 {
 205         vcpu_associativity = kcalloc(num_possible_cpus() / threads_per_core,
 206                         VPHN_ASSOC_BUFSIZE * sizeof(__be32), GFP_KERNEL);
 207         pcpu_associativity = kcalloc(NR_CPUS_H / threads_per_core,
 208                         VPHN_ASSOC_BUFSIZE * sizeof(__be32), GFP_KERNEL);
 209
 210         if (!vcpu_associativity || !pcpu_associativity) {
 211                 pr_err("error allocating memory for associativity information\n");
 212                 return -ENOMEM;
 213         }
 214
 215         return 0;
 216 }
 217
 218 static void destroy_cpu_associativity(void)
 219 {
 220         kfree(vcpu_associativity);
 221         kfree(pcpu_associativity);
 222         vcpu_associativity = pcpu_associativity = 0;
 223 }
 224
 225 static __be32 *__get_cpu_associativity(int cpu, __be32 *cpu_assoc, int flag)
 226 {
 227         __be32 *assoc;
 228         int rc = 0;
 229
 230         assoc = &cpu_assoc[(int)(cpu / threads_per_core) * VPHN_ASSOC_BUFSIZE];
 231         if (!assoc[0]) {
 232                 rc = hcall_vphn(cpu, flag, &assoc[0]);
 233                 if (rc)
 234                         return NULL;
 235         }
 236
 237         return assoc;
 238 }
 239
 240 static __be32 *get_pcpu_associativity(int cpu)
 241 {
 242         return __get_cpu_associativity(cpu, pcpu_associativity, VPHN_FLAG_PCPU);
 243 }
 244
 245 static __be32 *get_vcpu_associativity(int cpu)
 246 {
 247         return __get_cpu_associativity(cpu, vcpu_associativity, VPHN_FLAG_VCPU);
 248 }
 249
 250 static int cpu_relative_dispatch_distance(int last_disp_cpu, int cur_disp_cpu)
 251 {
 252         __be32 *last_disp_cpu_assoc, *cur_disp_cpu_assoc;
 253
 254         if (last_disp_cpu >= NR_CPUS_H || cur_disp_cpu >= NR_CPUS_H)
 255                 return -EINVAL;
 256
 257         last_disp_cpu_assoc = get_pcpu_associativity(last_disp_cpu);
 258         cur_disp_cpu_assoc = get_pcpu_associativity(cur_disp_cpu);
 259
 260         if (!last_disp_cpu_assoc || !cur_disp_cpu_assoc)
 261                 return -EIO;
 262
 263         return cpu_distance(last_disp_cpu_assoc, cur_disp_cpu_assoc);
 264 }
 265
 266 static int cpu_home_node_dispatch_distance(int disp_cpu)
 267 {
 268         __be32 *disp_cpu_assoc, *vcpu_assoc;
 269         int vcpu_id = smp_processor_id();
 270
 271         if (disp_cpu >= NR_CPUS_H) {
 272                 pr_debug_ratelimited("vcpu dispatch cpu %d > %d\n",
 273                                                 disp_cpu, NR_CPUS_H);
 274                 return -EINVAL;
 275         }
 276
 277         disp_cpu_assoc = get_pcpu_associativity(disp_cpu);
 278         vcpu_assoc = get_vcpu_associativity(vcpu_id);
 279
 280         if (!disp_cpu_assoc || !vcpu_assoc)
 281                 return -EIO;
 282
 283         return cpu_distance(disp_cpu_assoc, vcpu_assoc);
 284 }
 285
 286 static void update_vcpu_disp_stat(int disp_cpu)
 287 {
 288         struct vcpu_dispatch_data *disp;
 289         int distance;
 290
 291         disp = this_cpu_ptr(&vcpu_disp_data);
 292         if (disp->last_disp_cpu == -1) {
 293                 disp->last_disp_cpu = disp_cpu;
 294                 return;
 295         }
 296
 297         disp->total_disp++;
 298
 299         if (disp->last_disp_cpu == disp_cpu ||
 300                 (cpu_first_thread_sibling(disp->last_disp_cpu) ==
 301                                         cpu_first_thread_sibling(disp_cpu)))
 302                 disp->same_cpu_disp++;
 303         else {
 304                 distance = cpu_relative_dispatch_distance(disp->last_disp_cpu,
 305                                                                 disp_cpu);
 306                 if (distance < 0)
 307                         pr_debug_ratelimited("vcpudispatch_stats: cpu %d: error determining associativity\n",
 308                                         smp_processor_id());
 309                 else {
 310                         switch (distance) {
 311                         case 0:
 312                                 disp->same_chip_disp++;
 313                                 break;
 314                         case 1:
 315                                 disp->diff_chip_disp++;
 316                                 break;
 317                         case 2:
 318                                 disp->far_chip_disp++;
 319                                 break;
 320                         default:
 321                                 pr_debug_ratelimited("vcpudispatch_stats: cpu %d (%d -> %d): unexpected relative dispatch distance %d\n",
 322                                                  smp_processor_id(),
 323                                                  disp->last_disp_cpu,
 324                                                  disp_cpu,
 325                                                  distance);
 326                         }
 327                 }
 328         }
 329
 330         distance = cpu_home_node_dispatch_distance(disp_cpu);
 331         if (distance < 0)
 332                 pr_debug_ratelimited("vcpudispatch_stats: cpu %d: error determining associativity\n",
 333                                 smp_processor_id());
 334         else {
 335                 switch (distance) {
 336                 case 0:
 337                         disp->numa_home_disp++;
 338                         break;
 339                 case 1:
 340                         disp->numa_remote_disp++;
 341                         break;
 342                 case 2:
 343                         disp->numa_far_disp++;
 344                         break;
 345                 default:
 346                         pr_debug_ratelimited("vcpudispatch_stats: cpu %d on %d: unexpected numa dispatch distance %d\n",
 347                                                  smp_processor_id(),
 348                                                  disp_cpu,
 349                                                  distance);
 350                 }
 351         }
 352
 353         disp->last_disp_cpu = disp_cpu;
 354 }
 355
 356 static void process_dtl_buffer(struct work_struct *work)
 357 {
 358         struct dtl_entry dtle;
 359         u64 i = __this_cpu_read(dtl_entry_ridx);
 360         struct dtl_entry *dtl = local_paca->dispatch_log + (i % N_DISPATCH_LOG);
 361         struct dtl_entry *dtl_end = local_paca->dispatch_log_end;
 362         struct lppaca *vpa = local_paca->lppaca_ptr;
 363         struct dtl_worker *d = container_of(work, struct dtl_worker, work.work);
 364
 365         if (!local_paca->dispatch_log)
 366                 return;
 367
 368         /* if we have been migrated away, we cancel ourself */
 369         if (d->cpu != smp_processor_id()) {
 370                 pr_debug("vcpudispatch_stats: cpu %d worker migrated -- canceling worker\n",
 371                                                 smp_processor_id());
 372                 return;
 373         }
 374
 375         if (i == be64_to_cpu(vpa->dtl_idx))
 376                 goto out;
 377
 378         while (i < be64_to_cpu(vpa->dtl_idx)) {
 379                 dtle = *dtl;
 380                 barrier();
 381                 if (i + N_DISPATCH_LOG < be64_to_cpu(vpa->dtl_idx)) {
 382                         /* buffer has overflowed */
 383                         pr_debug_ratelimited("vcpudispatch_stats: cpu %d lost %lld DTL samples\n",
 384                                 d->cpu,
 385                                 be64_to_cpu(vpa->dtl_idx) - N_DISPATCH_LOG - i);
 386                         i = be64_to_cpu(vpa->dtl_idx) - N_DISPATCH_LOG;
 387                         dtl = local_paca->dispatch_log + (i % N_DISPATCH_LOG);
 388                         continue;
 389                 }
 390                 update_vcpu_disp_stat(be16_to_cpu(dtle.processor_id));
 391                 ++i;
 392                 ++dtl;
 393                 if (dtl == dtl_end)
 394                         dtl = local_paca->dispatch_log;
 395         }
 396
 397         __this_cpu_write(dtl_entry_ridx, i);
 398
 399 out:
 400         schedule_delayed_work_on(d->cpu, to_delayed_work(work),
 401                                         HZ / vcpudispatch_stats_freq);
 402 }
 403
 404 static int dtl_worker_online(unsigned int cpu)
 405 {
 406         struct dtl_worker *d = &per_cpu(dtl_workers, cpu);
 407
 408         memset(d, 0, sizeof(*d));
 409         INIT_DELAYED_WORK(&d->work, process_dtl_buffer);
 410         d->cpu = cpu;
 411
 412 #ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
 413         per_cpu(dtl_entry_ridx, cpu) = 0;
 414         register_dtl_buffer(cpu);
 415 #else
 416         per_cpu(dtl_entry_ridx, cpu) = be64_to_cpu(lppaca_of(cpu).dtl_idx);
 417 #endif
 418
 419         schedule_delayed_work_on(cpu, &d->work, HZ / vcpudispatch_stats_freq);
 420         return 0;
 421 }
 422
 423 static int dtl_worker_offline(unsigned int cpu)
 424 {
 425         struct dtl_worker *d = &per_cpu(dtl_workers, cpu);
 426
 427         cancel_delayed_work_sync(&d->work);
 428
 429 #ifndef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
 430         unregister_dtl(get_hard_smp_processor_id(cpu));
 431 #endif
 432
 433         return 0;
 434 }
 435
 436 static void set_global_dtl_mask(u8 mask)
 437 {
 438         int cpu;
 439
 440         dtl_mask = mask;
 441         for_each_present_cpu(cpu)
 442                 lppaca_of(cpu).dtl_enable_mask = dtl_mask;
 443 }
 444
 445 static void reset_global_dtl_mask(void)
 446 {
 447         int cpu;
 448
 449 #ifdef CONFIG_VIRT_CPU_ACCOUNTING_NATIVE
 450         dtl_mask = DTL_LOG_PREEMPT;
 451 #else
 452         dtl_mask = 0;
 453 #endif
 454         for_each_present_cpu(cpu)
 455                 lppaca_of(cpu).dtl_enable_mask = dtl_mask;
 456 }
 457
 458 static int dtl_worker_enable(unsigned long *time_limit)
 459 {
 460         int rc = 0, state;
 461
 462         if (!write_trylock(&dtl_access_lock)) {
 463                 rc = -EBUSY;
 464                 goto out;
 465         }
 466
 467         set_global_dtl_mask(DTL_LOG_ALL);
 468
 469         /* Setup dtl buffers and register those */
 470         alloc_dtl_buffers(time_limit);
 471
 472         state = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "powerpc/dtl:online",
 473                                         dtl_worker_online, dtl_worker_offline);
 474         if (state < 0) {
 475                 pr_err("vcpudispatch_stats: unable to setup workqueue for DTL processing\n");
 476                 free_dtl_buffers(time_limit);
 477                 reset_global_dtl_mask();
 478                 write_unlock(&dtl_access_lock);
 479                 rc = -EINVAL;
 480                 goto out;
 481         }
 482         dtl_worker_state = state;
 483
 484 out:
 485         return rc;
 486 }
 487
 488 static void dtl_worker_disable(unsigned long *time_limit)
 489 {
 490         cpuhp_remove_state(dtl_worker_state);
 491         free_dtl_buffers(time_limit);
 492         reset_global_dtl_mask();
 493         write_unlock(&dtl_access_lock);
 494 }
 495
 496 static ssize_t vcpudispatch_stats_write(struct file *file, const char __user *p,
 497                 size_t count, loff_t *ppos)
 498 {
 499         unsigned long time_limit = jiffies + HZ;
 500         struct vcpu_dispatch_data *disp;
 501         int rc, cmd, cpu;
 502         char buf[16];
 503
 504         if (count > 15)
 505                 return -EINVAL;
 506
 507         if (copy_from_user(buf, p, count))
 508                 return -EFAULT;
 509
 510         buf[count] = 0;
 511         rc = kstrtoint(buf, 0, &cmd);
 512         if (rc || cmd < 0 || cmd > 1) {
 513                 pr_err("vcpudispatch_stats: please use 0 to disable or 1 to enable dispatch statistics\n");
 514                 return rc ? rc : -EINVAL;
 515         }
 516
 517         mutex_lock(&dtl_enable_mutex);
 518
 519         if ((cmd == 0 && !vcpudispatch_stats_on) ||
 520                         (cmd == 1 && vcpudispatch_stats_on))
 521                 goto out;
 522
 523         if (cmd) {
 524                 rc = init_cpu_associativity();
 525                 if (rc)
 526                         goto out;
 527
 528                 for_each_possible_cpu(cpu) {
 529                         disp = per_cpu_ptr(&vcpu_disp_data, cpu);
 530                         memset(disp, 0, sizeof(*disp));
 531                         disp->last_disp_cpu = -1;
 532                 }
 533
 534                 rc = dtl_worker_enable(&time_limit);
 535                 if (rc) {
 536                         destroy_cpu_associativity();
 537                         goto out;
 538                 }
 539         } else {
 540                 dtl_worker_disable(&time_limit);
 541                 destroy_cpu_associativity();
 542         }
 543
 544         vcpudispatch_stats_on = cmd;
 545
 546 out:
 547         mutex_unlock(&dtl_enable_mutex);
 548         if (rc)
 549                 return rc;
 550         return count;
 551 }
 552
 553 static int vcpudispatch_stats_display(struct seq_file *p, void *v)
 554 {
 555         int cpu;
 556         struct vcpu_dispatch_data *disp;
 557
 558         if (!vcpudispatch_stats_on) {
 559                 seq_puts(p, "off\n");
 560                 return 0;
 561         }
 562
 563         for_each_online_cpu(cpu) {
 564                 disp = per_cpu_ptr(&vcpu_disp_data, cpu);
 565                 seq_printf(p, "cpu%d", cpu);
 566                 seq_put_decimal_ull(p, " ", disp->total_disp);
 567                 seq_put_decimal_ull(p, " ", disp->same_cpu_disp);
 568                 seq_put_decimal_ull(p, " ", disp->same_chip_disp);
 569                 seq_put_decimal_ull(p, " ", disp->diff_chip_disp);
 570                 seq_put_decimal_ull(p, " ", disp->far_chip_disp);
 571                 seq_put_decimal_ull(p, " ", disp->numa_home_disp);
 572                 seq_put_decimal_ull(p, " ", disp->numa_remote_disp);
 573                 seq_put_decimal_ull(p, " ", disp->numa_far_disp);
 574                 seq_puts(p, "\n");
 575         }
 576
 577         return 0;
 578 }
 579
 580 static int vcpudispatch_stats_open(struct inode *inode, struct file *file)
 581 {
 582         return single_open(file, vcpudispatch_stats_display, NULL);
 583 }
 584
 585 static const struct file_operations vcpudispatch_stats_proc_ops = {
 586         .open           = vcpudispatch_stats_open,
 587         .read           = seq_read,
 588         .write          = vcpudispatch_stats_write,
 589         .llseek         = seq_lseek,
 590         .release        = single_release,
 591 };
 592
 593 static ssize_t vcpudispatch_stats_freq_write(struct file *file,
 594                 const char __user *p, size_t count, loff_t *ppos)
 595 {
 596         int rc, freq;
 597         char buf[16];
 598
 599         if (count > 15)
 600                 return -EINVAL;
 601
 602         if (copy_from_user(buf, p, count))
 603                 return -EFAULT;
 604
 605         buf[count] = 0;
 606         rc = kstrtoint(buf, 0, &freq);
 607         if (rc || freq < 1 || freq > HZ) {
 608                 pr_err("vcpudispatch_stats_freq: please specify a frequency between 1 and %d\n",
 609                                 HZ);
 610                 return rc ? rc : -EINVAL;
 611         }
 612
 613         vcpudispatch_stats_freq = freq;
 614
 615         return count;
 616 }
 617
 618 static int vcpudispatch_stats_freq_display(struct seq_file *p, void *v)
 619 {
 620         seq_printf(p, "%d\n", vcpudispatch_stats_freq);
 621         return 0;
 622 }
 623
 624 static int vcpudispatch_stats_freq_open(struct inode *inode, struct file *file)
 625 {
 626         return single_open(file, vcpudispatch_stats_freq_display, NULL);
 627 }
 628
 629 static const struct file_operations vcpudispatch_stats_freq_proc_ops = {
 630         .open           = vcpudispatch_stats_freq_open,
 631         .read           = seq_read,
 632         .write          = vcpudispatch_stats_freq_write,
 633         .llseek         = seq_lseek,
 634         .release        = single_release,
 635 };
 636
 637 static int __init vcpudispatch_stats_procfs_init(void)
 638 {
 639         if (!lppaca_shared_proc(get_lppaca()))
 640                 return 0;
 641
 642         if (!proc_create("powerpc/vcpudispatch_stats", 0600, NULL,
 643                                         &vcpudispatch_stats_proc_ops))
 644                 pr_err("vcpudispatch_stats: error creating procfs file\n");
 645         else if (!proc_create("powerpc/vcpudispatch_stats_freq", 0600, NULL,
 646                                         &vcpudispatch_stats_freq_proc_ops))
 647                 pr_err("vcpudispatch_stats_freq: error creating procfs file\n");
 648
 649         return 0;
 650 }
 651
 652 machine_device_initcall(pseries, vcpudispatch_stats_procfs_init);
 653 #endif /* CONFIG_PPC_SPLPAR */
 654
 655 void vpa_init(int cpu)
 656 {
 657         int hwcpu = get_hard_smp_processor_id(cpu);
 658         unsigned long addr;
 659         long ret;
 660
 661         /*
 662          * The spec says it "may be problematic" if CPU x registers the VPA of
 663          * CPU y. We should never do that, but wail if we ever do.
 664          */
 665         WARN_ON(cpu != smp_processor_id());
 666
 667         if (cpu_has_feature(CPU_FTR_ALTIVEC))
 668                 lppaca_of(cpu).vmxregs_in_use = 1;
 669
 670         if (cpu_has_feature(CPU_FTR_ARCH_207S))
 671                 lppaca_of(cpu).ebb_regs_in_use = 1;
 672
 673         addr = __pa(&lppaca_of(cpu));
 674         ret = register_vpa(hwcpu, addr);
 675
 676         if (ret) {
 677                 pr_err("WARNING: VPA registration for cpu %d (hw %d) of area "
 678                        "%lx failed with %ld\n", cpu, hwcpu, addr, ret);
 679                 return;
 680         }
 681
 682 #ifdef CONFIG_PPC_BOOK3S_64
 683         /*
 684          * PAPR says this feature is SLB-Buffer but firmware never
 685          * reports that.  All SPLPAR support SLB shadow buffer.
 686          */
 687         if (!radix_enabled() && firmware_has_feature(FW_FEATURE_SPLPAR)) {
 688                 addr = __pa(paca_ptrs[cpu]->slb_shadow_ptr);
 689                 ret = register_slb_shadow(hwcpu, addr);
 690                 if (ret)
 691                         pr_err("WARNING: SLB shadow buffer registration for "
 692                                "cpu %d (hw %d) of area %lx failed with %ld\n",
 693                                cpu, hwcpu, addr, ret);
 694         }
 695 #endif /* CONFIG_PPC_BOOK3S_64 */
 696
 697         /*
 698          * Register dispatch trace log, if one has been allocated.
 699          */
 700         register_dtl_buffer(cpu);
 701 }
 702
 703 #ifdef CONFIG_PPC_BOOK3S_64
 704
 705 static long pSeries_lpar_hpte_insert(unsigned long hpte_group,
 706                                      unsigned long vpn, unsigned long pa,
 707                                      unsigned long rflags, unsigned long vflags,
 708                                      int psize, int apsize, int ssize)
 709 {
 710         unsigned long lpar_rc;
 711         unsigned long flags;
 712         unsigned long slot;
 713         unsigned long hpte_v, hpte_r;
 714
 715         if (!(vflags & HPTE_V_BOLTED))
 716                 pr_devel("hpte_insert(group=%lx, vpn=%016lx, "
 717                          "pa=%016lx, rflags=%lx, vflags=%lx, psize=%d)\n",
 718                          hpte_group, vpn,  pa, rflags, vflags, psize);
 719
 720         hpte_v = hpte_encode_v(vpn, psize, apsize, ssize) | vflags | HPTE_V_VALID;
 721         hpte_r = hpte_encode_r(pa, psize, apsize) | rflags;
 722
 723         if (!(vflags & HPTE_V_BOLTED))
 724                 pr_devel(" hpte_v=%016lx, hpte_r=%016lx\n", hpte_v, hpte_r);
 725
 726         /* Now fill in the actual HPTE */
 727         /* Set CEC cookie to 0         */
 728         /* Zero page = 0               */
 729         /* I-cache Invalidate = 0      */
 730         /* I-cache synchronize = 0     */
 731         /* Exact = 0                   */
 732         flags = 0;
 733
 734         if (firmware_has_feature(FW_FEATURE_XCMO) && !(hpte_r & HPTE_R_N))
 735                 flags |= H_COALESCE_CAND;
 736
 737         lpar_rc = plpar_pte_enter(flags, hpte_group, hpte_v, hpte_r, &slot);
 738         if (unlikely(lpar_rc == H_PTEG_FULL)) {
 739                 pr_devel("Hash table group is full\n");
 740                 return -1;
 741         }
 742
 743         /*
 744          * Since we try and ioremap PHBs we don't own, the pte insert
 745          * will fail. However we must catch the failure in hash_page
 746          * or we will loop forever, so return -2 in this case.
 747          */
 748         if (unlikely(lpar_rc != H_SUCCESS)) {
 749                 pr_err("Failed hash pte insert with error %ld\n", lpar_rc);
 750                 return -2;
 751         }
 752         if (!(vflags & HPTE_V_BOLTED))
 753                 pr_devel(" -> slot: %lu\n", slot & 7);
 754
 755         /* Because of iSeries, we have to pass down the secondary
 756          * bucket bit here as well
 757          */
 758         return (slot & 7) | (!!(vflags & HPTE_V_SECONDARY) << 3);
 759 }
 760
 761 static DEFINE_SPINLOCK(pSeries_lpar_tlbie_lock);
 762
 763 static long pSeries_lpar_hpte_remove(unsigned long hpte_group)
 764 {
 765         unsigned long slot_offset;
 766         unsigned long lpar_rc;
 767         int i;
 768         unsigned long dummy1, dummy2;
 769
 770         /* pick a random slot to start at */
 771         slot_offset = mftb() & 0x7;
 772
 773         for (i = 0; i < HPTES_PER_GROUP; i++) {
 774
 775                 /* don't remove a bolted entry */
 776                 lpar_rc = plpar_pte_remove(H_ANDCOND, hpte_group + slot_offset,
 777                                            (0x1UL << 4), &dummy1, &dummy2);
 778                 if (lpar_rc == H_SUCCESS)
 779                         return i;
 780
 781                 /*
 782                  * The test for adjunct partition is performed before the
 783                  * ANDCOND test.  H_RESOURCE may be returned, so we need to
 784                  * check for that as well.
 785                  */
 786                 BUG_ON(lpar_rc != H_NOT_FOUND && lpar_rc != H_RESOURCE);
 787
 788                 slot_offset++;
 789                 slot_offset &= 0x7;
 790         }
 791
 792         return -1;
 793 }
 794
 795 static void manual_hpte_clear_all(void)
 796 {
 797         unsigned long size_bytes = 1UL << ppc64_pft_size;
 798         unsigned long hpte_count = size_bytes >> 4;
 799         struct {
 800                 unsigned long pteh;
 801                 unsigned long ptel;
 802         } ptes[4];
 803         long lpar_rc;
 804         unsigned long i, j;
 805
 806         /* Read in batches of 4,
 807          * invalidate only valid entries not in the VRMA
 808          * hpte_count will be a multiple of 4
 809          */
 810         for (i = 0; i < hpte_count; i += 4) {
 811                 lpar_rc = plpar_pte_read_4_raw(0, i, (void *)ptes);
 812                 if (lpar_rc != H_SUCCESS) {
 813                         pr_info("Failed to read hash page table at %ld err %ld\n",
 814                                 i, lpar_rc);
 815                         continue;
 816                 }
 817                 for (j = 0; j < 4; j++){
 818                         if ((ptes[j].pteh & HPTE_V_VRMA_MASK) ==
 819                                 HPTE_V_VRMA_MASK)
 820                                 continue;
 821                         if (ptes[j].pteh & HPTE_V_VALID)
 822                                 plpar_pte_remove_raw(0, i + j, 0,
 823                                         &(ptes[j].pteh), &(ptes[j].ptel));
 824                 }
 825         }
 826 }
 827
 828 static int hcall_hpte_clear_all(void)
 829 {
 830         int rc;
 831
 832         do {
 833                 rc = plpar_hcall_norets(H_CLEAR_HPT);
 834         } while (rc == H_CONTINUE);
 835
 836         return rc;
 837 }
 838
 839 static void pseries_hpte_clear_all(void)
 840 {
 841         int rc;
 842
 843         rc = hcall_hpte_clear_all();
 844         if (rc != H_SUCCESS)
 845                 manual_hpte_clear_all();
 846
 847 #ifdef __LITTLE_ENDIAN__
 848         /*
 849          * Reset exceptions to big endian.
 850          *
 851          * FIXME this is a hack for kexec, we need to reset the exception
 852          * endian before starting the new kernel and this is a convenient place
 853          * to do it.
 854          *
 855          * This is also called on boot when a fadump happens. In that case we
 856          * must not change the exception endian mode.
 857          */
 858         if (firmware_has_feature(FW_FEATURE_SET_MODE) && !is_fadump_active())
 859                 pseries_big_endian_exceptions();
 860 #endif
 861 }
 862
 863 /*
 864  * NOTE: for updatepp ops we are fortunate that the linux "newpp" bits and
 865  * the low 3 bits of flags happen to line up.  So no transform is needed.
 866  * We can probably optimize here and assume the high bits of newpp are
 867  * already zero.  For now I am paranoid.
 868  */
 869 static long pSeries_lpar_hpte_updatepp(unsigned long slot,
 870                                        unsigned long newpp,
 871                                        unsigned long vpn,
 872                                        int psize, int apsize,
 873                                        int ssize, unsigned long inv_flags)
 874 {
 875         unsigned long lpar_rc;
 876         unsigned long flags;
 877         unsigned long want_v;
 878
 879         want_v = hpte_encode_avpn(vpn, psize, ssize);
 880
 881         flags = (newpp & 7) | H_AVPN;
 882         if (mmu_has_feature(MMU_FTR_KERNEL_RO))
 883                 /* Move pp0 into bit 8 (IBM 55) */
 884                 flags |= (newpp & HPTE_R_PP0) >> 55;
 885
 886         pr_devel("    update: avpnv=%016lx, hash=%016lx, f=%lx, psize: %d ...",
 887                  want_v, slot, flags, psize);
 888
 889         lpar_rc = plpar_pte_protect(flags, slot, want_v);
 890
 891         if (lpar_rc == H_NOT_FOUND) {
 892                 pr_devel("not found !\n");
 893                 return -1;
 894         }
 895
 896         pr_devel("ok\n");
 897
 898         BUG_ON(lpar_rc != H_SUCCESS);
 899
 900         return 0;
 901 }
 902
 903 static long __pSeries_lpar_hpte_find(unsigned long want_v, unsigned long hpte_group)
 904 {
 905         long lpar_rc;
 906         unsigned long i, j;
 907         struct {
 908                 unsigned long pteh;
 909                 unsigned long ptel;
 910         } ptes[4];
 911
 912         for (i = 0; i < HPTES_PER_GROUP; i += 4, hpte_group += 4) {
 913
 914                 lpar_rc = plpar_pte_read_4(0, hpte_group, (void *)ptes);
 915                 if (lpar_rc != H_SUCCESS) {
 916                         pr_info("Failed to read hash page table at %ld err %ld\n",
 917                                 hpte_group, lpar_rc);
 918                         continue;
 919                 }
 920
 921                 for (j = 0; j < 4; j++) {
 922                         if (HPTE_V_COMPARE(ptes[j].pteh, want_v) &&
 923                             (ptes[j].pteh & HPTE_V_VALID))
 924                                 return i + j;
 925                 }
 926         }
 927
 928         return -1;
 929 }
 930
 931 static long pSeries_lpar_hpte_find(unsigned long vpn, int psize, int ssize)
 932 {
 933         long slot;
 934         unsigned long hash;
 935         unsigned long want_v;
 936         unsigned long hpte_group;
 937
 938         hash = hpt_hash(vpn, mmu_psize_defs[psize].shift, ssize);
 939         want_v = hpte_encode_avpn(vpn, psize, ssize);
 940
 941         /* Bolted entries are always in the primary group */
 942         hpte_group = (hash & htab_hash_mask) * HPTES_PER_GROUP;
 943         slot = __pSeries_lpar_hpte_find(want_v, hpte_group);
 944         if (slot < 0)
 945                 return -1;
 946         return hpte_group + slot;
 947 }
 948
 949 static void pSeries_lpar_hpte_updateboltedpp(unsigned long newpp,
 950                                              unsigned long ea,
 951                                              int psize, int ssize)
 952 {
 953         unsigned long vpn;
 954         unsigned long lpar_rc, slot, vsid, flags;
 955
 956         vsid = get_kernel_vsid(ea, ssize);
 957         vpn = hpt_vpn(ea, vsid, ssize);
 958
 959         slot = pSeries_lpar_hpte_find(vpn, psize, ssize);
 960         BUG_ON(slot == -1);
 961
 962         flags = newpp & 7;
 963         if (mmu_has_feature(MMU_FTR_KERNEL_RO))
 964                 /* Move pp0 into bit 8 (IBM 55) */
 965                 flags |= (newpp & HPTE_R_PP0) >> 55;
 966
 967         lpar_rc = plpar_pte_protect(flags, slot, 0);
 968
 969         BUG_ON(lpar_rc != H_SUCCESS);
 970 }
 971
 972 static void pSeries_lpar_hpte_invalidate(unsigned long slot, unsigned long vpn,
 973                                          int psize, int apsize,
 974                                          int ssize, int local)
 975 {
 976         unsigned long want_v;
 977         unsigned long lpar_rc;
 978         unsigned long dummy1, dummy2;
 979
 980         pr_devel("    inval : slot=%lx, vpn=%016lx, psize: %d, local: %d\n",
 981                  slot, vpn, psize, local);
 982
 983         want_v = hpte_encode_avpn(vpn, psize, ssize);
 984         lpar_rc = plpar_pte_remove(H_AVPN, slot, want_v, &dummy1, &dummy2);
 985         if (lpar_rc == H_NOT_FOUND)
 986                 return;
 987
 988         BUG_ON(lpar_rc != H_SUCCESS);
 989 }
 990
 991
 992 /*
 993  * As defined in the PAPR's section 14.5.4.1.8
 994  * The control mask doesn't include the returned reference and change bit from
 995  * the processed PTE.
 996  */
 997 #define HBLKR_AVPN              0x0100000000000000UL
 998 #define HBLKR_CTRL_MASK         0xf800000000000000UL
 999 #define HBLKR_CTRL_SUCCESS      0x8000000000000000UL
1000 #define HBLKR_CTRL_ERRNOTFOUND  0x8800000000000000UL
1001 #define HBLKR_CTRL_ERRBUSY      0xa000000000000000UL
1002
1003 /*
1004  * Returned true if we are supporting this block size for the specified segment
1005  * base page size and actual page size.
1006  *
1007  * Currently, we only support 8 size block.
1008  */
1009 static inline bool is_supported_hlbkrm(int bpsize, int psize)
1010 {
1011         return (hblkrm_size[bpsize][psize] == HBLKRM_SUPPORTED_BLOCK_SIZE);
1012 }
1013
1014 /**
1015  * H_BLOCK_REMOVE caller.
1016  * @idx should point to the latest @param entry set with a PTEX.
1017  * If PTE cannot be processed because another CPUs has already locked that
1018  * group, those entries are put back in @param starting at index 1.
1019  * If entries has to be retried and @retry_busy is set to true, these entries
1020  * are retried until success. If @retry_busy is set to false, the returned
1021  * is the number of entries yet to process.
1022  */
1023 static unsigned long call_block_remove(unsigned long idx, unsigned long *param,
1024                                        bool retry_busy)
1025 {
1026         unsigned long i, rc, new_idx;
1027         unsigned long retbuf[PLPAR_HCALL9_BUFSIZE];
1028
1029         if (idx < 2) {
1030                 pr_warn("Unexpected empty call to H_BLOCK_REMOVE");
1031                 return 0;
1032         }
1033 again:
1034         new_idx = 0;
1035         if (idx > PLPAR_HCALL9_BUFSIZE) {
1036                 pr_err("Too many PTEs (%lu) for H_BLOCK_REMOVE", idx);
1037                 idx = PLPAR_HCALL9_BUFSIZE;
1038         } else if (idx < PLPAR_HCALL9_BUFSIZE)
1039                 param[idx] = HBR_END;
1040
1041         rc = plpar_hcall9(H_BLOCK_REMOVE, retbuf,
1042                           param[0], /* AVA */
1043                           param[1],  param[2],  param[3],  param[4], /* TS0-7 */
1044                           param[5],  param[6],  param[7],  param[8]);
1045         if (rc == H_SUCCESS)
1046                 return 0;
1047
1048         BUG_ON(rc != H_PARTIAL);
1049
1050         /* Check that the unprocessed entries were 'not found' or 'busy' */
1051         for (i = 0; i < idx-1; i++) {
1052                 unsigned long ctrl = retbuf[i] & HBLKR_CTRL_MASK;
1053
1054                 if (ctrl == HBLKR_CTRL_ERRBUSY) {
1055                         param[++new_idx] = param[i+1];
1056                         continue;
1057                 }
1058
1059                 BUG_ON(ctrl != HBLKR_CTRL_SUCCESS
1060                        && ctrl != HBLKR_CTRL_ERRNOTFOUND);
1061         }
1062
1063         /*
1064          * If there were entries found busy, retry these entries if requested,
1065          * of if all the entries have to be retried.
1066          */
1067         if (new_idx && (retry_busy || new_idx == (PLPAR_HCALL9_BUFSIZE-1))) {
1068                 idx = new_idx + 1;
1069                 goto again;
1070         }
1071
1072         return new_idx;
1073 }
1074
1075 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1076 /*
1077  * Limit iterations holding pSeries_lpar_tlbie_lock to 3. We also need
1078  * to make sure that we avoid bouncing the hypervisor tlbie lock.
1079  */
1080 #define PPC64_HUGE_HPTE_BATCH 12
1081
1082 static void hugepage_block_invalidate(unsigned long *slot, unsigned long *vpn,
1083                                       int count, int psize, int ssize)
1084 {
1085         unsigned long param[PLPAR_HCALL9_BUFSIZE];
1086         unsigned long shift, current_vpgb, vpgb;
1087         int i, pix = 0;
1088
1089         shift = mmu_psize_defs[psize].shift;
1090
1091         for (i = 0; i < count; i++) {
1092                 /*
1093                  * Shifting 3 bits more on the right to get a
1094                  * 8 pages aligned virtual addresse.
1095                  */
1096                 vpgb = (vpn[i] >> (shift - VPN_SHIFT + 3));
1097                 if (!pix || vpgb != current_vpgb) {
1098                         /*
1099                          * Need to start a new 8 pages block, flush
1100                          * the current one if needed.
1101                          */
1102                         if (pix)
1103                                 (void)call_block_remove(pix, param, true);
1104                         current_vpgb = vpgb;
1105                         param[0] = hpte_encode_avpn(vpn[i], psize, ssize);
1106                         pix = 1;
1107                 }
1108
1109                 param[pix++] = HBR_REQUEST | HBLKR_AVPN | slot[i];
1110                 if (pix == PLPAR_HCALL9_BUFSIZE) {
1111                         pix = call_block_remove(pix, param, false);
1112                         /*
1113                          * pix = 0 means that all the entries were
1114                          * removed, we can start a new block.
1115                          * Otherwise, this means that there are entries
1116                          * to retry, and pix points to latest one, so
1117                          * we should increment it and try to continue
1118                          * the same block.
1119                          */
1120                         if (pix)
1121                                 pix++;
1122                 }
1123         }
1124         if (pix)
1125                 (void)call_block_remove(pix, param, true);
1126 }
1127
1128 static void hugepage_bulk_invalidate(unsigned long *slot, unsigned long *vpn,
1129                                      int count, int psize, int ssize)
1130 {
1131         unsigned long param[PLPAR_HCALL9_BUFSIZE];
1132         int i = 0, pix = 0, rc;
1133
1134         for (i = 0; i < count; i++) {
1135
1136                 if (!firmware_has_feature(FW_FEATURE_BULK_REMOVE)) {
1137                         pSeries_lpar_hpte_invalidate(slot[i], vpn[i], psize, 0,
1138                                                      ssize, 0);
1139                 } else {
1140                         param[pix] = HBR_REQUEST | HBR_AVPN | slot[i];
1141                         param[pix+1] = hpte_encode_avpn(vpn[i], psize, ssize);
1142                         pix += 2;
1143                         if (pix == 8) {
1144                                 rc = plpar_hcall9(H_BULK_REMOVE, param,
1145                                                   param[0], param[1], param[2],
1146                                                   param[3], param[4], param[5],
1147                                                   param[6], param[7]);
1148                                 BUG_ON(rc != H_SUCCESS);
1149                                 pix = 0;
1150                         }
1151                 }
1152         }
1153         if (pix) {
1154                 param[pix] = HBR_END;
1155                 rc = plpar_hcall9(H_BULK_REMOVE, param, param[0], param[1],
1156                                   param[2], param[3], param[4], param[5],
1157                                   param[6], param[7]);
1158                 BUG_ON(rc != H_SUCCESS);
1159         }
1160 }
1161
1162 static inline void __pSeries_lpar_hugepage_invalidate(unsigned long *slot,
1163                                                       unsigned long *vpn,
1164                                                       int count, int psize,
1165                                                       int ssize)
1166 {
1167         unsigned long flags = 0;
1168         int lock_tlbie = !mmu_has_feature(MMU_FTR_LOCKLESS_TLBIE);
1169
1170         if (lock_tlbie)
1171                 spin_lock_irqsave(&pSeries_lpar_tlbie_lock, flags);
1172
1173         /* Assuming THP size is 16M */
1174         if (is_supported_hlbkrm(psize, MMU_PAGE_16M))
1175                 hugepage_block_invalidate(slot, vpn, count, psize, ssize);
1176         else
1177                 hugepage_bulk_invalidate(slot, vpn, count, psize, ssize);
1178
1179         if (lock_tlbie)
1180                 spin_unlock_irqrestore(&pSeries_lpar_tlbie_lock, flags);
1181 }
1182
1183 static void pSeries_lpar_hugepage_invalidate(unsigned long vsid,
1184                                              unsigned long addr,
1185                                              unsigned char *hpte_slot_array,
1186                                              int psize, int ssize, int local)
1187 {
1188         int i, index = 0;
1189         unsigned long s_addr = addr;
1190         unsigned int max_hpte_count, valid;
1191         unsigned long vpn_array[PPC64_HUGE_HPTE_BATCH];
1192         unsigned long slot_array[PPC64_HUGE_HPTE_BATCH];
1193         unsigned long shift, hidx, vpn = 0, hash, slot;
1194
1195         shift = mmu_psize_defs[psize].shift;
1196         max_hpte_count = 1U << (PMD_SHIFT - shift);
1197
1198         for (i = 0; i < max_hpte_count; i++) {
1199                 valid = hpte_valid(hpte_slot_array, i);
1200                 if (!valid)
1201                         continue;
1202                 hidx =  hpte_hash_index(hpte_slot_array, i);
1203
1204                 /* get the vpn */
1205                 addr = s_addr + (i * (1ul << shift));
1206                 vpn = hpt_vpn(addr, vsid, ssize);
1207                 hash = hpt_hash(vpn, shift, ssize);
1208                 if (hidx & _PTEIDX_SECONDARY)
1209                         hash = ~hash;
1210
1211                 slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
1212                 slot += hidx & _PTEIDX_GROUP_IX;
1213
1214                 slot_array[index] = slot;
1215                 vpn_array[index] = vpn;
1216                 if (index == PPC64_HUGE_HPTE_BATCH - 1) {
1217                         /*
1218                          * Now do a bluk invalidate
1219                          */
1220                         __pSeries_lpar_hugepage_invalidate(slot_array,
1221                                                            vpn_array,
1222                                                            PPC64_HUGE_HPTE_BATCH,
1223                                                            psize, ssize);
1224                         index = 0;
1225                 } else
1226                         index++;
1227         }
1228         if (index)
1229                 __pSeries_lpar_hugepage_invalidate(slot_array, vpn_array,
1230                                                    index, psize, ssize);
1231 }
1232 #else
1233 static void pSeries_lpar_hugepage_invalidate(unsigned long vsid,
1234                                              unsigned long addr,
1235                                              unsigned char *hpte_slot_array,
1236                                              int psize, int ssize, int local)
1237 {
1238         WARN(1, "%s called without THP support\n", __func__);
1239 }
1240 #endif
1241
1242 static int pSeries_lpar_hpte_removebolted(unsigned long ea,
1243                                           int psize, int ssize)
1244 {
1245         unsigned long vpn;
1246         unsigned long slot, vsid;
1247
1248         vsid = get_kernel_vsid(ea, ssize);
1249         vpn = hpt_vpn(ea, vsid, ssize);
1250
1251         slot = pSeries_lpar_hpte_find(vpn, psize, ssize);
1252         if (slot == -1)
1253                 return -ENOENT;
1254
1255         /*
1256          * lpar doesn't use the passed actual page size
1257          */
1258         pSeries_lpar_hpte_invalidate(slot, vpn, psize, 0, ssize, 0);
1259         return 0;
1260 }
1261
1262
1263 static inline unsigned long compute_slot(real_pte_t pte,
1264                                          unsigned long vpn,
1265                                          unsigned long index,
1266                                          unsigned long shift,
1267                                          int ssize)
1268 {
1269         unsigned long slot, hash, hidx;
1270
1271         hash = hpt_hash(vpn, shift, ssize);
1272         hidx = __rpte_to_hidx(pte, index);
1273         if (hidx & _PTEIDX_SECONDARY)
1274                 hash = ~hash;
1275         slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
1276         slot += hidx & _PTEIDX_GROUP_IX;
1277         return slot;
1278 }
1279
1280 /**
1281  * The hcall H_BLOCK_REMOVE implies that the virtual pages to processed are
1282  * "all within the same naturally aligned 8 page virtual address block".
1283  */
1284 static void do_block_remove(unsigned long number, struct ppc64_tlb_batch *batch,
1285                             unsigned long *param)
1286 {
1287         unsigned long vpn;
1288         unsigned long i, pix = 0;
1289         unsigned long index, shift, slot, current_vpgb, vpgb;
1290         real_pte_t pte;
1291         int psize, ssize;
1292
1293         psize = batch->psize;
1294         ssize = batch->ssize;
1295
1296         for (i = 0; i < number; i++) {
1297                 vpn = batch->vpn[i];
1298                 pte = batch->pte[i];
1299                 pte_iterate_hashed_subpages(pte, psize, vpn, index, shift) {
1300                         /*
1301                          * Shifting 3 bits more on the right to get a
1302                          * 8 pages aligned virtual addresse.
1303                          */
1304                         vpgb = (vpn >> (shift - VPN_SHIFT + 3));
1305                         if (!pix || vpgb != current_vpgb) {
1306                                 /*
1307                                  * Need to start a new 8 pages block, flush
1308                                  * the current one if needed.
1309                                  */
1310                                 if (pix)
1311                                         (void)call_block_remove(pix, param,
1312                                                                 true);
1313                                 current_vpgb = vpgb;
1314                                 param[0] = hpte_encode_avpn(vpn, psize,
1315                                                             ssize);
1316                                 pix = 1;
1317                         }
1318
1319                         slot = compute_slot(pte, vpn, index, shift, ssize);
1320                         param[pix++] = HBR_REQUEST | HBLKR_AVPN | slot;
1321
1322                         if (pix == PLPAR_HCALL9_BUFSIZE) {
1323                                 pix = call_block_remove(pix, param, false);
1324                                 /*
1325                                  * pix = 0 means that all the entries were
1326                                  * removed, we can start a new block.
1327                                  * Otherwise, this means that there are entries
1328                                  * to retry, and pix points to latest one, so
1329                                  * we should increment it and try to continue
1330                                  * the same block.
1331                                  */
1332                                 if (pix)
1333                                         pix++;
1334                         }
1335                 } pte_iterate_hashed_end();
1336         }
1337
1338         if (pix)
1339                 (void)call_block_remove(pix, param, true);
1340 }
1341
1342 /*
1343  * TLB Block Invalidate Characteristics
1344  *
1345  * These characteristics define the size of the block the hcall H_BLOCK_REMOVE
1346  * is able to process for each couple segment base page size, actual page size.
1347  *
1348  * The ibm,get-system-parameter properties is returning a buffer with the
1349  * following layout:
1350  *
1351  * [ 2 bytes size of the RTAS buffer (excluding these 2 bytes) ]
1352  * -----------------
1353  * TLB Block Invalidate Specifiers:
1354  * [ 1 byte LOG base 2 of the TLB invalidate block size being specified ]
1355  * [ 1 byte Number of page sizes (N) that are supported for the specified
1356  *          TLB invalidate block size ]
1357  * [ 1 byte Encoded segment base page size and actual page size
1358  *          MSB=0 means 4k segment base page size and actual page size
1359  *          MSB=1 the penc value in mmu_psize_def ]
1360  * ...
1361  * -----------------
1362  * Next TLB Block Invalidate Specifiers...
1363  * -----------------
1364  * [ 0 ]
1365  */
1366 static inline void set_hblkrm_bloc_size(int bpsize, int psize,
1367                                         unsigned int block_size)
1368 {
1369         if (block_size > hblkrm_size[bpsize][psize])
1370                 hblkrm_size[bpsize][psize] = block_size;
1371 }
1372
1373 /*
1374  * Decode the Encoded segment base page size and actual page size.
1375  * PAPR specifies:
1376  *   - bit 7 is the L bit
1377  *   - bits 0-5 are the penc value
1378  * If the L bit is 0, this means 4K segment base page size and actual page size
1379  * otherwise the penc value should be read.
1380  */
1381 #define HBLKRM_L_MASK           0x80
1382 #define HBLKRM_PENC_MASK        0x3f
1383 static inline void __init check_lp_set_hblkrm(unsigned int lp,
1384                                               unsigned int block_size)
1385 {
1386         unsigned int bpsize, psize;
1387
1388         /* First, check the L bit, if not set, this means 4K */
1389         if ((lp & HBLKRM_L_MASK) == 0) {
1390                 set_hblkrm_bloc_size(MMU_PAGE_4K, MMU_PAGE_4K, block_size);
1391                 return;
1392         }
1393
1394         lp &= HBLKRM_PENC_MASK;
1395         for (bpsize = 0; bpsize < MMU_PAGE_COUNT; bpsize++) {
1396                 struct mmu_psize_def *def = &mmu_psize_defs[bpsize];
1397
1398                 for (psize = 0; psize < MMU_PAGE_COUNT; psize++) {
1399                         if (def->penc[psize] == lp) {
1400                                 set_hblkrm_bloc_size(bpsize, psize, block_size);
1401                                 return;
1402                         }
1403                 }
1404         }
1405 }
1406
1407 #define SPLPAR_TLB_BIC_TOKEN            50
1408
1409 /*
1410  * The size of the TLB Block Invalidate Characteristics is variable. But at the
1411  * maximum it will be the number of possible page sizes *2 + 10 bytes.
1412  * Currently MMU_PAGE_COUNT is 16, which means 42 bytes. Use a cache line size
1413  * (128 bytes) for the buffer to get plenty of space.
1414  */
1415 #define SPLPAR_TLB_BIC_MAXLENGTH        128
1416
1417 void __init pseries_lpar_read_hblkrm_characteristics(void)
1418 {
1419         unsigned char local_buffer[SPLPAR_TLB_BIC_MAXLENGTH];
1420         int call_status, len, idx, bpsize;
1421
1422         if (!firmware_has_feature(FW_FEATURE_BLOCK_REMOVE))
1423                 return;
1424
1425         spin_lock(&rtas_data_buf_lock);
1426         memset(rtas_data_buf, 0, RTAS_DATA_BUF_SIZE);
1427         call_status = rtas_call(rtas_token("ibm,get-system-parameter"), 3, 1,
1428                                 NULL,
1429                                 SPLPAR_TLB_BIC_TOKEN,
1430                                 __pa(rtas_data_buf),
1431                                 RTAS_DATA_BUF_SIZE);
1432         memcpy(local_buffer, rtas_data_buf, SPLPAR_TLB_BIC_MAXLENGTH);
1433         local_buffer[SPLPAR_TLB_BIC_MAXLENGTH - 1] = '\0';
1434         spin_unlock(&rtas_data_buf_lock);
1435
1436         if (call_status != 0) {
1437                 pr_warn("%s %s Error calling get-system-parameter (0x%x)\n",
1438                         __FILE__, __func__, call_status);
1439                 return;
1440         }
1441
1442         /*
1443          * The first two (2) bytes of the data in the buffer are the length of
1444          * the returned data, not counting these first two (2) bytes.
1445          */
1446         len = be16_to_cpu(*((u16 *)local_buffer)) + 2;
1447         if (len > SPLPAR_TLB_BIC_MAXLENGTH) {
1448                 pr_warn("%s too large returned buffer %d", __func__, len);
1449                 return;
1450         }
1451
1452         idx = 2;
1453         while (idx < len) {
1454                 u8 block_shift = local_buffer[idx++];
1455                 u32 block_size;
1456                 unsigned int npsize;
1457
1458                 if (!block_shift)
1459                         break;
1460
1461                 block_size = 1 << block_shift;
1462
1463                 for (npsize = local_buffer[idx++];
1464                      npsize > 0 && idx < len; npsize--)
1465                         check_lp_set_hblkrm((unsigned int) local_buffer[idx++],
1466                                             block_size);
1467         }
1468
1469         for (bpsize = 0; bpsize < MMU_PAGE_COUNT; bpsize++)
1470                 for (idx = 0; idx < MMU_PAGE_COUNT; idx++)
1471                         if (hblkrm_size[bpsize][idx])
1472                                 pr_info("H_BLOCK_REMOVE supports base psize:%d psize:%d block size:%d",
1473                                         bpsize, idx, hblkrm_size[bpsize][idx]);
1474 }
1475
1476 /*
1477  * Take a spinlock around flushes to avoid bouncing the hypervisor tlbie
1478  * lock.
1479  */
1480 static void pSeries_lpar_flush_hash_range(unsigned long number, int local)
1481 {
1482         unsigned long vpn;
1483         unsigned long i, pix, rc;
1484         unsigned long flags = 0;
1485         struct ppc64_tlb_batch *batch = this_cpu_ptr(&ppc64_tlb_batch);
1486         int lock_tlbie = !mmu_has_feature(MMU_FTR_LOCKLESS_TLBIE);
1487         unsigned long param[PLPAR_HCALL9_BUFSIZE];
1488         unsigned long index, shift, slot;
1489         real_pte_t pte;
1490         int psize, ssize;
1491
1492         if (lock_tlbie)
1493                 spin_lock_irqsave(&pSeries_lpar_tlbie_lock, flags);
1494
1495         if (is_supported_hlbkrm(batch->psize, batch->psize)) {
1496                 do_block_remove(number, batch, param);
1497                 goto out;
1498         }
1499
1500         psize = batch->psize;
1501         ssize = batch->ssize;
1502         pix = 0;
1503         for (i = 0; i < number; i++) {
1504                 vpn = batch->vpn[i];
1505                 pte = batch->pte[i];
1506                 pte_iterate_hashed_subpages(pte, psize, vpn, index, shift) {
1507                         slot = compute_slot(pte, vpn, index, shift, ssize);
1508                         if (!firmware_has_feature(FW_FEATURE_BULK_REMOVE)) {
1509                                 /*
1510                                  * lpar doesn't use the passed actual page size
1511                                  */
1512                                 pSeries_lpar_hpte_invalidate(slot, vpn, psize,
1513                                                              0, ssize, local);
1514                         } else {
1515                                 param[pix] = HBR_REQUEST | HBR_AVPN | slot;
1516                                 param[pix+1] = hpte_encode_avpn(vpn, psize,
1517                                                                 ssize);
1518                                 pix += 2;
1519                                 if (pix == 8) {
1520                                         rc = plpar_hcall9(H_BULK_REMOVE, param,
1521                                                 param[0], param[1], param[2],
1522                                                 param[3], param[4], param[5],
1523                                                 param[6], param[7]);
1524                                         BUG_ON(rc != H_SUCCESS);
1525                                         pix = 0;
1526                                 }
1527                         }
1528                 } pte_iterate_hashed_end();
1529         }
1530         if (pix) {
1531                 param[pix] = HBR_END;
1532                 rc = plpar_hcall9(H_BULK_REMOVE, param, param[0], param[1],
1533                                   param[2], param[3], param[4], param[5],
1534                                   param[6], param[7]);
1535                 BUG_ON(rc != H_SUCCESS);
1536         }
1537
1538 out:
1539         if (lock_tlbie)
1540                 spin_unlock_irqrestore(&pSeries_lpar_tlbie_lock, flags);
1541 }
1542
1543 static int __init disable_bulk_remove(char *str)
1544 {
1545         if (strcmp(str, "off") == 0 &&
1546             firmware_has_feature(FW_FEATURE_BULK_REMOVE)) {
1547                 pr_info("Disabling BULK_REMOVE firmware feature");
1548                 powerpc_firmware_features &= ~FW_FEATURE_BULK_REMOVE;
1549         }
1550         return 1;
1551 }
1552
1553 __setup("bulk_remove=", disable_bulk_remove);
1554
1555 #define HPT_RESIZE_TIMEOUT      10000 /* ms */
1556
1557 struct hpt_resize_state {
1558         unsigned long shift;
1559         int commit_rc;
1560 };
1561
1562 static int pseries_lpar_resize_hpt_commit(void *data)
1563 {
1564         struct hpt_resize_state *state = data;
1565
1566         state->commit_rc = plpar_resize_hpt_commit(0, state->shift);
1567         if (state->commit_rc != H_SUCCESS)
1568                 return -EIO;
1569
1570         /* Hypervisor has transitioned the HTAB, update our globals */
1571         ppc64_pft_size = state->shift;
1572         htab_size_bytes = 1UL << ppc64_pft_size;
1573         htab_hash_mask = (htab_size_bytes >> 7) - 1;
1574
1575         return 0;
1576 }
1577
1578 /*
1579  * Must be called in process context. The caller must hold the
1580  * cpus_lock.
1581  */
1582 static int pseries_lpar_resize_hpt(unsigned long shift)
1583 {
1584         struct hpt_resize_state state = {
1585                 .shift = shift,
1586                 .commit_rc = H_FUNCTION,
1587         };
1588         unsigned int delay, total_delay = 0;
1589         int rc;
1590         ktime_t t0, t1, t2;
1591
1592         might_sleep();
1593
1594         if (!firmware_has_feature(FW_FEATURE_HPT_RESIZE))
1595                 return -ENODEV;
1596
1597         pr_info("Attempting to resize HPT to shift %lu\n", shift);
1598
1599         t0 = ktime_get();
1600
1601         rc = plpar_resize_hpt_prepare(0, shift);
1602         while (H_IS_LONG_BUSY(rc)) {
1603                 delay = get_longbusy_msecs(rc);
1604                 total_delay += delay;
1605                 if (total_delay > HPT_RESIZE_TIMEOUT) {
1606                         /* prepare with shift==0 cancels an in-progress resize */
1607                         rc = plpar_resize_hpt_prepare(0, 0);
1608                         if (rc != H_SUCCESS)
1609                                 pr_warn("Unexpected error %d cancelling timed out HPT resize\n",
1610                                        rc);
1611                         return -ETIMEDOUT;
1612                 }
1613                 msleep(delay);
1614                 rc = plpar_resize_hpt_prepare(0, shift);
1615         };
1616
1617         switch (rc) {
1618         case H_SUCCESS:
1619                 /* Continue on */
1620                 break;
1621
1622         case H_PARAMETER:
1623                 pr_warn("Invalid argument from H_RESIZE_HPT_PREPARE\n");
1624                 return -EINVAL;
1625         case H_RESOURCE:
1626                 pr_warn("Operation not permitted from H_RESIZE_HPT_PREPARE\n");
1627                 return -EPERM;
1628         default:
1629                 pr_warn("Unexpected error %d from H_RESIZE_HPT_PREPARE\n", rc);
1630                 return -EIO;
1631         }
1632
1633         t1 = ktime_get();
1634
1635         rc = stop_machine_cpuslocked(pseries_lpar_resize_hpt_commit,
1636                                      &state, NULL);
1637
1638         t2 = ktime_get();
1639
1640         if (rc != 0) {
1641                 switch (state.commit_rc) {
1642                 case H_PTEG_FULL:
1643                         return -ENOSPC;
1644
1645                 default:
1646                         pr_warn("Unexpected error %d from H_RESIZE_HPT_COMMIT\n",
1647                                 state.commit_rc);
1648                         return -EIO;
1649                 };
1650         }
1651
1652         pr_info("HPT resize to shift %lu complete (%lld ms / %lld ms)\n",
1653                 shift, (long long) ktime_ms_delta(t1, t0),
1654                 (long long) ktime_ms_delta(t2, t1));
1655
1656         return 0;
1657 }
1658
1659 static int pseries_lpar_register_process_table(unsigned long base,
1660                         unsigned long page_size, unsigned long table_size)
1661 {
1662         long rc;
1663         unsigned long flags = 0;
1664
1665         if (table_size)
1666                 flags |= PROC_TABLE_NEW;
1667         if (radix_enabled())
1668                 flags |= PROC_TABLE_RADIX | PROC_TABLE_GTSE;
1669         else
1670                 flags |= PROC_TABLE_HPT_SLB;
1671         for (;;) {
1672                 rc = plpar_hcall_norets(H_REGISTER_PROC_TBL, flags, base,
1673                                         page_size, table_size);
1674                 if (!H_IS_LONG_BUSY(rc))
1675                         break;
1676                 mdelay(get_longbusy_msecs(rc));
1677         }
1678         if (rc != H_SUCCESS) {
1679                 pr_err("Failed to register process table (rc=%ld)\n", rc);
1680                 BUG();
1681         }
1682         return rc;
1683 }
1684
1685 void __init hpte_init_pseries(void)
1686 {
1687         mmu_hash_ops.hpte_invalidate     = pSeries_lpar_hpte_invalidate;
1688         mmu_hash_ops.hpte_updatepp       = pSeries_lpar_hpte_updatepp;
1689         mmu_hash_ops.hpte_updateboltedpp = pSeries_lpar_hpte_updateboltedpp;
1690         mmu_hash_ops.hpte_insert         = pSeries_lpar_hpte_insert;
1691         mmu_hash_ops.hpte_remove         = pSeries_lpar_hpte_remove;
1692         mmu_hash_ops.hpte_removebolted   = pSeries_lpar_hpte_removebolted;
1693         mmu_hash_ops.flush_hash_range    = pSeries_lpar_flush_hash_range;
1694         mmu_hash_ops.hpte_clear_all      = pseries_hpte_clear_all;
1695         mmu_hash_ops.hugepage_invalidate = pSeries_lpar_hugepage_invalidate;
1696
1697         if (firmware_has_feature(FW_FEATURE_HPT_RESIZE))
1698                 mmu_hash_ops.resize_hpt = pseries_lpar_resize_hpt;
1699
1700         /*
1701          * On POWER9, we need to do a H_REGISTER_PROC_TBL hcall
1702          * to inform the hypervisor that we wish to use the HPT.
1703          */
1704         if (cpu_has_feature(CPU_FTR_ARCH_300))
1705                 pseries_lpar_register_process_table(0, 0, 0);
1706 }
1707
1708 void radix_init_pseries(void)
1709 {
1710         pr_info("Using radix MMU under hypervisor\n");
1711
1712         pseries_lpar_register_process_table(__pa(process_tb),
1713                                                 0, PRTB_SIZE_SHIFT - 12);
1714 }
1715
1716 #ifdef CONFIG_PPC_SMLPAR
1717 #define CMO_FREE_HINT_DEFAULT 1
1718 static int cmo_free_hint_flag = CMO_FREE_HINT_DEFAULT;
1719
1720 static int __init cmo_free_hint(char *str)
1721 {
1722         char *parm;
1723         parm = strstrip(str);
1724
1725         if (strcasecmp(parm, "no") == 0 || strcasecmp(parm, "off") == 0) {
1726                 pr_info("%s: CMO free page hinting is not active.\n", __func__);
1727                 cmo_free_hint_flag = 0;
1728                 return 1;
1729         }
1730
1731         cmo_free_hint_flag = 1;
1732         pr_info("%s: CMO free page hinting is active.\n", __func__);
1733
1734         if (strcasecmp(parm, "yes") == 0 || strcasecmp(parm, "on") == 0)
1735                 return 1;
1736
1737         return 0;
1738 }
1739
1740 __setup("cmo_free_hint=", cmo_free_hint);
1741
1742 static void pSeries_set_page_state(struct page *page, int order,
1743                                    unsigned long state)
1744 {
1745         int i, j;
1746         unsigned long cmo_page_sz, addr;
1747
1748         cmo_page_sz = cmo_get_page_size();
1749         addr = __pa((unsigned long)page_address(page));
1750
1751         for (i = 0; i < (1 << order); i++, addr += PAGE_SIZE) {
1752                 for (j = 0; j < PAGE_SIZE; j += cmo_page_sz)
1753                         plpar_hcall_norets(H_PAGE_INIT, state, addr + j, 0);
1754         }
1755 }
1756
1757 void arch_free_page(struct page *page, int order)
1758 {
1759         if (radix_enabled())
1760                 return;
1761         if (!cmo_free_hint_flag || !firmware_has_feature(FW_FEATURE_CMO))
1762                 return;
1763
1764         pSeries_set_page_state(page, order, H_PAGE_SET_UNUSED);
1765 }
1766 EXPORT_SYMBOL(arch_free_page);
1767
1768 #endif /* CONFIG_PPC_SMLPAR */
1769 #endif /* CONFIG_PPC_BOOK3S_64 */
1770
1771 #ifdef CONFIG_TRACEPOINTS
1772 #ifdef CONFIG_JUMP_LABEL
1773 struct static_key hcall_tracepoint_key = STATIC_KEY_INIT;
1774
1775 int hcall_tracepoint_regfunc(void)
1776 {
1777         static_key_slow_inc(&hcall_tracepoint_key);
1778         return 0;
1779 }
1780
1781 void hcall_tracepoint_unregfunc(void)
1782 {
1783         static_key_slow_dec(&hcall_tracepoint_key);
1784 }
1785 #else
1786 /*
1787  * We optimise our hcall path by placing hcall_tracepoint_refcount
1788  * directly in the TOC so we can check if the hcall tracepoints are
1789  * enabled via a single load.
1790  */
1791
1792 /* NB: reg/unreg are called while guarded with the tracepoints_mutex */
1793 extern long hcall_tracepoint_refcount;
1794
1795 int hcall_tracepoint_regfunc(void)
1796 {
1797         hcall_tracepoint_refcount++;
1798         return 0;
1799 }
1800
1801 void hcall_tracepoint_unregfunc(void)
1802 {
1803         hcall_tracepoint_refcount--;
1804 }
1805 #endif
1806
1807 /*
1808  * Since the tracing code might execute hcalls we need to guard against
1809  * recursion. One example of this are spinlocks calling H_YIELD on
1810  * shared processor partitions.
1811  */
1812 static DEFINE_PER_CPU(unsigned int, hcall_trace_depth);
1813
1814
1815 void __trace_hcall_entry(unsigned long opcode, unsigned long *args)
1816 {
1817         unsigned long flags;
1818         unsigned int *depth;
1819
1820         /*
1821          * We cannot call tracepoints inside RCU idle regions which
1822          * means we must not trace H_CEDE.
1823          */
1824         if (opcode == H_CEDE)
1825                 return;
1826
1827         local_irq_save(flags);
1828
1829         depth = this_cpu_ptr(&hcall_trace_depth);
1830
1831         if (*depth)
1832                 goto out;
1833
1834         (*depth)++;
1835         preempt_disable();
1836         trace_hcall_entry(opcode, args);
1837         (*depth)--;
1838
1839 out:
1840         local_irq_restore(flags);
1841 }
1842
1843 void __trace_hcall_exit(long opcode, long retval, unsigned long *retbuf)
1844 {
1845         unsigned long flags;
1846         unsigned int *depth;
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         trace_hcall_exit(opcode, retval, retbuf);
1860         preempt_enable();
1861         (*depth)--;
1862
1863 out:
1864         local_irq_restore(flags);
1865 }
1866 #endif
1867
1868 /**
1869  * h_get_mpp
1870  * H_GET_MPP hcall returns info in 7 parms
1871  */
1872 int h_get_mpp(struct hvcall_mpp_data *mpp_data)
1873 {
1874         int rc;
1875         unsigned long retbuf[PLPAR_HCALL9_BUFSIZE];
1876
1877         rc = plpar_hcall9(H_GET_MPP, retbuf);
1878
1879         mpp_data->entitled_mem = retbuf[0];
1880         mpp_data->mapped_mem = retbuf[1];
1881
1882         mpp_data->group_num = (retbuf[2] >> 2 * 8) & 0xffff;
1883         mpp_data->pool_num = retbuf[2] & 0xffff;
1884
1885         mpp_data->mem_weight = (retbuf[3] >> 7 * 8) & 0xff;
1886         mpp_data->unallocated_mem_weight = (retbuf[3] >> 6 * 8) & 0xff;
1887         mpp_data->unallocated_entitlement = retbuf[3] & 0xffffffffffffUL;
1888
1889         mpp_data->pool_size = retbuf[4];
1890         mpp_data->loan_request = retbuf[5];
1891         mpp_data->backing_mem = retbuf[6];
1892
1893         return rc;
1894 }
1895 EXPORT_SYMBOL(h_get_mpp);
1896
1897 int h_get_mpp_x(struct hvcall_mpp_x_data *mpp_x_data)
1898 {
1899         int rc;
1900         unsigned long retbuf[PLPAR_HCALL9_BUFSIZE] = { 0 };
1901
1902         rc = plpar_hcall9(H_GET_MPP_X, retbuf);
1903
1904         mpp_x_data->coalesced_bytes = retbuf[0];
1905         mpp_x_data->pool_coalesced_bytes = retbuf[1];
1906         mpp_x_data->pool_purr_cycles = retbuf[2];
1907         mpp_x_data->pool_spurr_cycles = retbuf[3];
1908
1909         return rc;
1910 }
1911
1912 static unsigned long vsid_unscramble(unsigned long vsid, int ssize)
1913 {
1914         unsigned long protovsid;
1915         unsigned long va_bits = VA_BITS;
1916         unsigned long modinv, vsid_modulus;
1917         unsigned long max_mod_inv, tmp_modinv;
1918
1919         if (!mmu_has_feature(MMU_FTR_68_BIT_VA))
1920                 va_bits = 65;
1921
1922         if (ssize == MMU_SEGSIZE_256M) {
1923                 modinv = VSID_MULINV_256M;
1924                 vsid_modulus = ((1UL << (va_bits - SID_SHIFT)) - 1);
1925         } else {
1926                 modinv = VSID_MULINV_1T;
1927                 vsid_modulus = ((1UL << (va_bits - SID_SHIFT_1T)) - 1);
1928         }
1929
1930         /*
1931          * vsid outside our range.
1932          */
1933         if (vsid >= vsid_modulus)
1934                 return 0;
1935
1936         /*
1937          * If modinv is the modular multiplicate inverse of (x % vsid_modulus)
1938          * and vsid = (protovsid * x) % vsid_modulus, then we say:
1939          *   protovsid = (vsid * modinv) % vsid_modulus
1940          */
1941
1942         /* Check if (vsid * modinv) overflow (63 bits) */
1943         max_mod_inv = 0x7fffffffffffffffull / vsid;
1944         if (modinv < max_mod_inv)
1945                 return (vsid * modinv) % vsid_modulus;
1946
1947         tmp_modinv = modinv/max_mod_inv;
1948         modinv %= max_mod_inv;
1949
1950         protovsid = (((vsid * max_mod_inv) % vsid_modulus) * tmp_modinv) % vsid_modulus;
1951         protovsid = (protovsid + vsid * modinv) % vsid_modulus;
1952
1953         return protovsid;
1954 }
1955
1956 static int __init reserve_vrma_context_id(void)
1957 {
1958         unsigned long protovsid;
1959
1960         /*
1961          * Reserve context ids which map to reserved virtual addresses. For now
1962          * we only reserve the context id which maps to the VRMA VSID. We ignore
1963          * the addresses in "ibm,adjunct-virtual-addresses" because we don't
1964          * enable adjunct support via the "ibm,client-architecture-support"
1965          * interface.
1966          */
1967         protovsid = vsid_unscramble(VRMA_VSID, MMU_SEGSIZE_1T);
1968         hash__reserve_context_id(protovsid >> ESID_BITS_1T);
1969         return 0;
1970 }
1971 machine_device_initcall(pseries, reserve_vrma_context_id);
1972
1973 #ifdef CONFIG_DEBUG_FS
1974 /* debugfs file interface for vpa data */
1975 static ssize_t vpa_file_read(struct file *filp, char __user *buf, size_t len,
1976                               loff_t *pos)
1977 {
1978         int cpu = (long)filp->private_data;
1979         struct lppaca *lppaca = &lppaca_of(cpu);
1980
1981         return simple_read_from_buffer(buf, len, pos, lppaca,
1982                                 sizeof(struct lppaca));
1983 }
1984
1985 static const struct file_operations vpa_fops = {
1986         .open           = simple_open,
1987         .read           = vpa_file_read,
1988         .llseek         = default_llseek,
1989 };
1990
1991 static int __init vpa_debugfs_init(void)
1992 {
1993         char name[16];
1994         long i;
1995         static struct dentry *vpa_dir;
1996
1997         if (!firmware_has_feature(FW_FEATURE_SPLPAR))
1998                 return 0;
1999
2000         vpa_dir = debugfs_create_dir("vpa", powerpc_debugfs_root);
2001
2002         /* set up the per-cpu vpa file*/
2003         for_each_possible_cpu(i) {
2004                 sprintf(name, "cpu-%ld", i);
2005                 debugfs_create_file(name, 0400, vpa_dir, (void *)i, &vpa_fops);
2006         }
2007
2008         return 0;
2009 }
2010 machine_arch_initcall(pseries, vpa_debugfs_init);
2011 #endif /* CONFIG_DEBUG_FS */