arch/powerpc/kvm/book3s_hv.c

   1 /*
   2  * Copyright 2011 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
   3  * Copyright (C) 2009. SUSE Linux Products GmbH. All rights reserved.
   4  *
   5  * Authors:
   6  *    Paul Mackerras <paulus@au1.ibm.com>
   7  *    Alexander Graf <agraf@suse.de>
   8  *    Kevin Wolf <mail@kevin-wolf.de>
   9  *
  10  * Description: KVM functions specific to running on Book 3S
  11  * processors in hypervisor mode (specifically POWER7 and later).
  12  *
  13  * This file is derived from arch/powerpc/kvm/book3s.c,
  14  * by Alexander Graf <agraf@suse.de>.
  15  *
  16  * This program is free software; you can redistribute it and/or modify
  17  * it under the terms of the GNU General Public License, version 2, as
  18  * published by the Free Software Foundation.
  19  */
  20
  21 #include <linux/kvm_host.h>
  22 #include <linux/kernel.h>
  23 #include <linux/err.h>
  24 #include <linux/slab.h>
  25 #include <linux/preempt.h>
  26 #include <linux/sched/signal.h>
  27 #include <linux/sched/stat.h>
  28 #include <linux/delay.h>
  29 #include <linux/export.h>
  30 #include <linux/fs.h>
  31 #include <linux/anon_inodes.h>
  32 #include <linux/cpu.h>
  33 #include <linux/cpumask.h>
  34 #include <linux/spinlock.h>
  35 #include <linux/page-flags.h>
  36 #include <linux/srcu.h>
  37 #include <linux/miscdevice.h>
  38 #include <linux/debugfs.h>
  39 #include <linux/gfp.h>
  40 #include <linux/vmalloc.h>
  41 #include <linux/highmem.h>
  42 #include <linux/hugetlb.h>
  43 #include <linux/kvm_irqfd.h>
  44 #include <linux/irqbypass.h>
  45 #include <linux/module.h>
  46 #include <linux/compiler.h>
  47 #include <linux/of.h>
  48
  49 #include <asm/ftrace.h>
  50 #include <asm/reg.h>
  51 #include <asm/ppc-opcode.h>
  52 #include <asm/asm-prototypes.h>
  53 #include <asm/debug.h>
  54 #include <asm/disassemble.h>
  55 #include <asm/cputable.h>
  56 #include <asm/cacheflush.h>
  57 #include <linux/uaccess.h>
  58 #include <asm/io.h>
  59 #include <asm/kvm_ppc.h>
  60 #include <asm/kvm_book3s.h>
  61 #include <asm/mmu_context.h>
  62 #include <asm/lppaca.h>
  63 #include <asm/processor.h>
  64 #include <asm/cputhreads.h>
  65 #include <asm/page.h>
  66 #include <asm/hvcall.h>
  67 #include <asm/switch_to.h>
  68 #include <asm/smp.h>
  69 #include <asm/dbell.h>
  70 #include <asm/hmi.h>
  71 #include <asm/pnv-pci.h>
  72 #include <asm/mmu.h>
  73 #include <asm/opal.h>
  74 #include <asm/xics.h>
  75 #include <asm/xive.h>
  76
  77 #include "book3s.h"
  78
  79 #define CREATE_TRACE_POINTS
  80 #include "trace_hv.h"
  81
  82 /* #define EXIT_DEBUG */
  83 /* #define EXIT_DEBUG_SIMPLE */
  84 /* #define EXIT_DEBUG_INT */
  85
  86 /* Used to indicate that a guest page fault needs to be handled */
  87 #define RESUME_PAGE_FAULT       (RESUME_GUEST | RESUME_FLAG_ARCH1)
  88 /* Used to indicate that a guest passthrough interrupt needs to be handled */
  89 #define RESUME_PASSTHROUGH      (RESUME_GUEST | RESUME_FLAG_ARCH2)
  90
  91 /* Used as a "null" value for timebase values */
  92 #define TB_NIL  (~(u64)0)
  93
  94 static DECLARE_BITMAP(default_enabled_hcalls, MAX_HCALL_OPCODE/4 + 1);
  95
  96 static int dynamic_mt_modes = 6;
  97 module_param(dynamic_mt_modes, int, 0644);
  98 MODULE_PARM_DESC(dynamic_mt_modes, "Set of allowed dynamic micro-threading modes: 0 (= none), 2, 4, or 6 (= 2 or 4)");
  99 static int target_smt_mode;
 100 module_param(target_smt_mode, int, 0644);
 101 MODULE_PARM_DESC(target_smt_mode, "Target threads per core (0 = max)");
 102
 103 static bool indep_threads_mode = true;
 104 module_param(indep_threads_mode, bool, S_IRUGO | S_IWUSR);
 105 MODULE_PARM_DESC(indep_threads_mode, "Independent-threads mode (only on POWER9)");
 106
 107 #ifdef CONFIG_KVM_XICS
 108 static struct kernel_param_ops module_param_ops = {
 109         .set = param_set_int,
 110         .get = param_get_int,
 111 };
 112
 113 module_param_cb(kvm_irq_bypass, &module_param_ops, &kvm_irq_bypass, 0644);
 114 MODULE_PARM_DESC(kvm_irq_bypass, "Bypass passthrough interrupt optimization");
 115
 116 module_param_cb(h_ipi_redirect, &module_param_ops, &h_ipi_redirect, 0644);
 117 MODULE_PARM_DESC(h_ipi_redirect, "Redirect H_IPI wakeup to a free host core");
 118 #endif
 119
 120 /* If set, the threads on each CPU core have to be in the same MMU mode */
 121 static bool no_mixing_hpt_and_radix;
 122
 123 static void kvmppc_end_cede(struct kvm_vcpu *vcpu);
 124 static int kvmppc_hv_setup_htab_rma(struct kvm_vcpu *vcpu);
 125
 126 /*
 127  * RWMR values for POWER8.  These control the rate at which PURR
 128  * and SPURR count and should be set according to the number of
 129  * online threads in the vcore being run.
 130  */
 131 #define RWMR_RPA_P8_1THREAD     0x164520C62609AECAUL
 132 #define RWMR_RPA_P8_2THREAD     0x7FFF2908450D8DA9UL
 133 #define RWMR_RPA_P8_3THREAD     0x164520C62609AECAUL
 134 #define RWMR_RPA_P8_4THREAD     0x199A421245058DA9UL
 135 #define RWMR_RPA_P8_5THREAD     0x164520C62609AECAUL
 136 #define RWMR_RPA_P8_6THREAD     0x164520C62609AECAUL
 137 #define RWMR_RPA_P8_7THREAD     0x164520C62609AECAUL
 138 #define RWMR_RPA_P8_8THREAD     0x164520C62609AECAUL
 139
 140 static unsigned long p8_rwmr_values[MAX_SMT_THREADS + 1] = {
 141         RWMR_RPA_P8_1THREAD,
 142         RWMR_RPA_P8_1THREAD,
 143         RWMR_RPA_P8_2THREAD,
 144         RWMR_RPA_P8_3THREAD,
 145         RWMR_RPA_P8_4THREAD,
 146         RWMR_RPA_P8_5THREAD,
 147         RWMR_RPA_P8_6THREAD,
 148         RWMR_RPA_P8_7THREAD,
 149         RWMR_RPA_P8_8THREAD,
 150 };
 151
 152 static inline struct kvm_vcpu *next_runnable_thread(struct kvmppc_vcore *vc,
 153                 int *ip)
 154 {
 155         int i = *ip;
 156         struct kvm_vcpu *vcpu;
 157
 158         while (++i < MAX_SMT_THREADS) {
 159                 vcpu = READ_ONCE(vc->runnable_threads[i]);
 160                 if (vcpu) {
 161                         *ip = i;
 162                         return vcpu;
 163                 }
 164         }
 165         return NULL;
 166 }
 167
 168 /* Used to traverse the list of runnable threads for a given vcore */
 169 #define for_each_runnable_thread(i, vcpu, vc) \
 170         for (i = -1; (vcpu = next_runnable_thread(vc, &i)); )
 171
 172 static bool kvmppc_ipi_thread(int cpu)
 173 {
 174         unsigned long msg = PPC_DBELL_TYPE(PPC_DBELL_SERVER);
 175
 176         /* On POWER9 we can use msgsnd to IPI any cpu */
 177         if (cpu_has_feature(CPU_FTR_ARCH_300)) {
 178                 msg |= get_hard_smp_processor_id(cpu);
 179                 smp_mb();
 180                 __asm__ __volatile__ (PPC_MSGSND(%0) : : "r" (msg));
 181                 return true;
 182         }
 183
 184         /* On POWER8 for IPIs to threads in the same core, use msgsnd */
 185         if (cpu_has_feature(CPU_FTR_ARCH_207S)) {
 186                 preempt_disable();
 187                 if (cpu_first_thread_sibling(cpu) ==
 188                     cpu_first_thread_sibling(smp_processor_id())) {
 189                         msg |= cpu_thread_in_core(cpu);
 190                         smp_mb();
 191                         __asm__ __volatile__ (PPC_MSGSND(%0) : : "r" (msg));
 192                         preempt_enable();
 193                         return true;
 194                 }
 195                 preempt_enable();
 196         }
 197
 198 #if defined(CONFIG_PPC_ICP_NATIVE) && defined(CONFIG_SMP)
 199         if (cpu >= 0 && cpu < nr_cpu_ids) {
 200                 if (paca_ptrs[cpu]->kvm_hstate.xics_phys) {
 201                         xics_wake_cpu(cpu);
 202                         return true;
 203                 }
 204                 opal_int_set_mfrr(get_hard_smp_processor_id(cpu), IPI_PRIORITY);
 205                 return true;
 206         }
 207 #endif
 208
 209         return false;
 210 }
 211
 212 static void kvmppc_fast_vcpu_kick_hv(struct kvm_vcpu *vcpu)
 213 {
 214         int cpu;
 215         struct swait_queue_head *wqp;
 216
 217         wqp = kvm_arch_vcpu_wq(vcpu);
 218         if (swq_has_sleeper(wqp)) {
 219                 swake_up_one(wqp);
 220                 ++vcpu->stat.halt_wakeup;
 221         }
 222
 223         cpu = READ_ONCE(vcpu->arch.thread_cpu);
 224         if (cpu >= 0 && kvmppc_ipi_thread(cpu))
 225                 return;
 226
 227         /* CPU points to the first thread of the core */
 228         cpu = vcpu->cpu;
 229         if (cpu >= 0 && cpu < nr_cpu_ids && cpu_online(cpu))
 230                 smp_send_reschedule(cpu);
 231 }
 232
 233 /*
 234  * We use the vcpu_load/put functions to measure stolen time.
 235  * Stolen time is counted as time when either the vcpu is able to
 236  * run as part of a virtual core, but the task running the vcore
 237  * is preempted or sleeping, or when the vcpu needs something done
 238  * in the kernel by the task running the vcpu, but that task is
 239  * preempted or sleeping.  Those two things have to be counted
 240  * separately, since one of the vcpu tasks will take on the job
 241  * of running the core, and the other vcpu tasks in the vcore will
 242  * sleep waiting for it to do that, but that sleep shouldn't count
 243  * as stolen time.
 244  *
 245  * Hence we accumulate stolen time when the vcpu can run as part of
 246  * a vcore using vc->stolen_tb, and the stolen time when the vcpu
 247  * needs its task to do other things in the kernel (for example,
 248  * service a page fault) in busy_stolen.  We don't accumulate
 249  * stolen time for a vcore when it is inactive, or for a vcpu
 250  * when it is in state RUNNING or NOTREADY.  NOTREADY is a bit of
 251  * a misnomer; it means that the vcpu task is not executing in
 252  * the KVM_VCPU_RUN ioctl, i.e. it is in userspace or elsewhere in
 253  * the kernel.  We don't have any way of dividing up that time
 254  * between time that the vcpu is genuinely stopped, time that
 255  * the task is actively working on behalf of the vcpu, and time
 256  * that the task is preempted, so we don't count any of it as
 257  * stolen.
 258  *
 259  * Updates to busy_stolen are protected by arch.tbacct_lock;
 260  * updates to vc->stolen_tb are protected by the vcore->stoltb_lock
 261  * lock.  The stolen times are measured in units of timebase ticks.
 262  * (Note that the != TB_NIL checks below are purely defensive;
 263  * they should never fail.)
 264  */
 265
 266 static void kvmppc_core_start_stolen(struct kvmppc_vcore *vc)
 267 {
 268         unsigned long flags;
 269
 270         spin_lock_irqsave(&vc->stoltb_lock, flags);
 271         vc->preempt_tb = mftb();
 272         spin_unlock_irqrestore(&vc->stoltb_lock, flags);
 273 }
 274
 275 static void kvmppc_core_end_stolen(struct kvmppc_vcore *vc)
 276 {
 277         unsigned long flags;
 278
 279         spin_lock_irqsave(&vc->stoltb_lock, flags);
 280         if (vc->preempt_tb != TB_NIL) {
 281                 vc->stolen_tb += mftb() - vc->preempt_tb;
 282                 vc->preempt_tb = TB_NIL;
 283         }
 284         spin_unlock_irqrestore(&vc->stoltb_lock, flags);
 285 }
 286
 287 static void kvmppc_core_vcpu_load_hv(struct kvm_vcpu *vcpu, int cpu)
 288 {
 289         struct kvmppc_vcore *vc = vcpu->arch.vcore;
 290         unsigned long flags;
 291
 292         /*
 293          * We can test vc->runner without taking the vcore lock,
 294          * because only this task ever sets vc->runner to this
 295          * vcpu, and once it is set to this vcpu, only this task
 296          * ever sets it to NULL.
 297          */
 298         if (vc->runner == vcpu && vc->vcore_state >= VCORE_SLEEPING)
 299                 kvmppc_core_end_stolen(vc);
 300
 301         spin_lock_irqsave(&vcpu->arch.tbacct_lock, flags);
 302         if (vcpu->arch.state == KVMPPC_VCPU_BUSY_IN_HOST &&
 303             vcpu->arch.busy_preempt != TB_NIL) {
 304                 vcpu->arch.busy_stolen += mftb() - vcpu->arch.busy_preempt;
 305                 vcpu->arch.busy_preempt = TB_NIL;
 306         }
 307         spin_unlock_irqrestore(&vcpu->arch.tbacct_lock, flags);
 308 }
 309
 310 static void kvmppc_core_vcpu_put_hv(struct kvm_vcpu *vcpu)
 311 {
 312         struct kvmppc_vcore *vc = vcpu->arch.vcore;
 313         unsigned long flags;
 314
 315         if (vc->runner == vcpu && vc->vcore_state >= VCORE_SLEEPING)
 316                 kvmppc_core_start_stolen(vc);
 317
 318         spin_lock_irqsave(&vcpu->arch.tbacct_lock, flags);
 319         if (vcpu->arch.state == KVMPPC_VCPU_BUSY_IN_HOST)
 320                 vcpu->arch.busy_preempt = mftb();
 321         spin_unlock_irqrestore(&vcpu->arch.tbacct_lock, flags);
 322 }
 323
 324 static void kvmppc_set_msr_hv(struct kvm_vcpu *vcpu, u64 msr)
 325 {
 326         /*
 327          * Check for illegal transactional state bit combination
 328          * and if we find it, force the TS field to a safe state.
 329          */
 330         if ((msr & MSR_TS_MASK) == MSR_TS_MASK)
 331                 msr &= ~MSR_TS_MASK;
 332         vcpu->arch.shregs.msr = msr;
 333         kvmppc_end_cede(vcpu);
 334 }
 335
 336 static void kvmppc_set_pvr_hv(struct kvm_vcpu *vcpu, u32 pvr)
 337 {
 338         vcpu->arch.pvr = pvr;
 339 }
 340
 341 /* Dummy value used in computing PCR value below */
 342 #define PCR_ARCH_300    (PCR_ARCH_207 << 1)
 343
 344 static int kvmppc_set_arch_compat(struct kvm_vcpu *vcpu, u32 arch_compat)
 345 {
 346         unsigned long host_pcr_bit = 0, guest_pcr_bit = 0;
 347         struct kvmppc_vcore *vc = vcpu->arch.vcore;
 348
 349         /* We can (emulate) our own architecture version and anything older */
 350         if (cpu_has_feature(CPU_FTR_ARCH_300))
 351                 host_pcr_bit = PCR_ARCH_300;
 352         else if (cpu_has_feature(CPU_FTR_ARCH_207S))
 353                 host_pcr_bit = PCR_ARCH_207;
 354         else if (cpu_has_feature(CPU_FTR_ARCH_206))
 355                 host_pcr_bit = PCR_ARCH_206;
 356         else
 357                 host_pcr_bit = PCR_ARCH_205;
 358
 359         /* Determine lowest PCR bit needed to run guest in given PVR level */
 360         guest_pcr_bit = host_pcr_bit;
 361         if (arch_compat) {
 362                 switch (arch_compat) {
 363                 case PVR_ARCH_205:
 364                         guest_pcr_bit = PCR_ARCH_205;
 365                         break;
 366                 case PVR_ARCH_206:
 367                 case PVR_ARCH_206p:
 368                         guest_pcr_bit = PCR_ARCH_206;
 369                         break;
 370                 case PVR_ARCH_207:
 371                         guest_pcr_bit = PCR_ARCH_207;
 372                         break;
 373                 case PVR_ARCH_300:
 374                         guest_pcr_bit = PCR_ARCH_300;
 375                         break;
 376                 default:
 377                         return -EINVAL;
 378                 }
 379         }
 380
 381         /* Check requested PCR bits don't exceed our capabilities */
 382         if (guest_pcr_bit > host_pcr_bit)
 383                 return -EINVAL;
 384
 385         spin_lock(&vc->lock);
 386         vc->arch_compat = arch_compat;
 387         /* Set all PCR bits for which guest_pcr_bit <= bit < host_pcr_bit */
 388         vc->pcr = host_pcr_bit - guest_pcr_bit;
 389         spin_unlock(&vc->lock);
 390
 391         return 0;
 392 }
 393
 394 static void kvmppc_dump_regs(struct kvm_vcpu *vcpu)
 395 {
 396         int r;
 397
 398         pr_err("vcpu %p (%d):\n", vcpu, vcpu->vcpu_id);
 399         pr_err("pc  = %.16lx  msr = %.16llx  trap = %x\n",
 400                vcpu->arch.regs.nip, vcpu->arch.shregs.msr, vcpu->arch.trap);
 401         for (r = 0; r < 16; ++r)
 402                 pr_err("r%2d = %.16lx  r%d = %.16lx\n",
 403                        r, kvmppc_get_gpr(vcpu, r),
 404                        r+16, kvmppc_get_gpr(vcpu, r+16));
 405         pr_err("ctr = %.16lx  lr  = %.16lx\n",
 406                vcpu->arch.regs.ctr, vcpu->arch.regs.link);
 407         pr_err("srr0 = %.16llx srr1 = %.16llx\n",
 408                vcpu->arch.shregs.srr0, vcpu->arch.shregs.srr1);
 409         pr_err("sprg0 = %.16llx sprg1 = %.16llx\n",
 410                vcpu->arch.shregs.sprg0, vcpu->arch.shregs.sprg1);
 411         pr_err("sprg2 = %.16llx sprg3 = %.16llx\n",
 412                vcpu->arch.shregs.sprg2, vcpu->arch.shregs.sprg3);
 413         pr_err("cr = %.8x  xer = %.16lx  dsisr = %.8x\n",
 414                vcpu->arch.cr, vcpu->arch.regs.xer, vcpu->arch.shregs.dsisr);
 415         pr_err("dar = %.16llx\n", vcpu->arch.shregs.dar);
 416         pr_err("fault dar = %.16lx dsisr = %.8x\n",
 417                vcpu->arch.fault_dar, vcpu->arch.fault_dsisr);
 418         pr_err("SLB (%d entries):\n", vcpu->arch.slb_max);
 419         for (r = 0; r < vcpu->arch.slb_max; ++r)
 420                 pr_err("  ESID = %.16llx VSID = %.16llx\n",
 421                        vcpu->arch.slb[r].orige, vcpu->arch.slb[r].origv);
 422         pr_err("lpcr = %.16lx sdr1 = %.16lx last_inst = %.8x\n",
 423                vcpu->arch.vcore->lpcr, vcpu->kvm->arch.sdr1,
 424                vcpu->arch.last_inst);
 425 }
 426
 427 static struct kvm_vcpu *kvmppc_find_vcpu(struct kvm *kvm, int id)
 428 {
 429         struct kvm_vcpu *ret;
 430
 431         mutex_lock(&kvm->lock);
 432         ret = kvm_get_vcpu_by_id(kvm, id);
 433         mutex_unlock(&kvm->lock);
 434         return ret;
 435 }
 436
 437 static void init_vpa(struct kvm_vcpu *vcpu, struct lppaca *vpa)
 438 {
 439         vpa->__old_status |= LPPACA_OLD_SHARED_PROC;
 440         vpa->yield_count = cpu_to_be32(1);
 441 }
 442
 443 static int set_vpa(struct kvm_vcpu *vcpu, struct kvmppc_vpa *v,
 444                    unsigned long addr, unsigned long len)
 445 {
 446         /* check address is cacheline aligned */
 447         if (addr & (L1_CACHE_BYTES - 1))
 448                 return -EINVAL;
 449         spin_lock(&vcpu->arch.vpa_update_lock);
 450         if (v->next_gpa != addr || v->len != len) {
 451                 v->next_gpa = addr;
 452                 v->len = addr ? len : 0;
 453                 v->update_pending = 1;
 454         }
 455         spin_unlock(&vcpu->arch.vpa_update_lock);
 456         return 0;
 457 }
 458
 459 /* Length for a per-processor buffer is passed in at offset 4 in the buffer */
 460 struct reg_vpa {
 461         u32 dummy;
 462         union {
 463                 __be16 hword;
 464                 __be32 word;
 465         } length;
 466 };
 467
 468 static int vpa_is_registered(struct kvmppc_vpa *vpap)
 469 {
 470         if (vpap->update_pending)
 471                 return vpap->next_gpa != 0;
 472         return vpap->pinned_addr != NULL;
 473 }
 474
 475 static unsigned long do_h_register_vpa(struct kvm_vcpu *vcpu,
 476                                        unsigned long flags,
 477                                        unsigned long vcpuid, unsigned long vpa)
 478 {
 479         struct kvm *kvm = vcpu->kvm;
 480         unsigned long len, nb;
 481         void *va;
 482         struct kvm_vcpu *tvcpu;
 483         int err;
 484         int subfunc;
 485         struct kvmppc_vpa *vpap;
 486
 487         tvcpu = kvmppc_find_vcpu(kvm, vcpuid);
 488         if (!tvcpu)
 489                 return H_PARAMETER;
 490
 491         subfunc = (flags >> H_VPA_FUNC_SHIFT) & H_VPA_FUNC_MASK;
 492         if (subfunc == H_VPA_REG_VPA || subfunc == H_VPA_REG_DTL ||
 493             subfunc == H_VPA_REG_SLB) {
 494                 /* Registering new area - address must be cache-line aligned */
 495                 if ((vpa & (L1_CACHE_BYTES - 1)) || !vpa)
 496                         return H_PARAMETER;
 497
 498                 /* convert logical addr to kernel addr and read length */
 499                 va = kvmppc_pin_guest_page(kvm, vpa, &nb);
 500                 if (va == NULL)
 501                         return H_PARAMETER;
 502                 if (subfunc == H_VPA_REG_VPA)
 503                         len = be16_to_cpu(((struct reg_vpa *)va)->length.hword);
 504                 else
 505                         len = be32_to_cpu(((struct reg_vpa *)va)->length.word);
 506                 kvmppc_unpin_guest_page(kvm, va, vpa, false);
 507
 508                 /* Check length */
 509                 if (len > nb || len < sizeof(struct reg_vpa))
 510                         return H_PARAMETER;
 511         } else {
 512                 vpa = 0;
 513                 len = 0;
 514         }
 515
 516         err = H_PARAMETER;
 517         vpap = NULL;
 518         spin_lock(&tvcpu->arch.vpa_update_lock);
 519
 520         switch (subfunc) {
 521         case H_VPA_REG_VPA:             /* register VPA */
 522                 /*
 523                  * The size of our lppaca is 1kB because of the way we align
 524                  * it for the guest to avoid crossing a 4kB boundary. We only
 525                  * use 640 bytes of the structure though, so we should accept
 526                  * clients that set a size of 640.
 527                  */
 528                 BUILD_BUG_ON(sizeof(struct lppaca) != 640);
 529                 if (len < sizeof(struct lppaca))
 530                         break;
 531                 vpap = &tvcpu->arch.vpa;
 532                 err = 0;
 533                 break;
 534
 535         case H_VPA_REG_DTL:             /* register DTL */
 536                 if (len < sizeof(struct dtl_entry))
 537                         break;
 538                 len -= len % sizeof(struct dtl_entry);
 539
 540                 /* Check that they have previously registered a VPA */
 541                 err = H_RESOURCE;
 542                 if (!vpa_is_registered(&tvcpu->arch.vpa))
 543                         break;
 544
 545                 vpap = &tvcpu->arch.dtl;
 546                 err = 0;
 547                 break;
 548
 549         case H_VPA_REG_SLB:             /* register SLB shadow buffer */
 550                 /* Check that they have previously registered a VPA */
 551                 err = H_RESOURCE;
 552                 if (!vpa_is_registered(&tvcpu->arch.vpa))
 553                         break;
 554
 555                 vpap = &tvcpu->arch.slb_shadow;
 556                 err = 0;
 557                 break;
 558
 559         case H_VPA_DEREG_VPA:           /* deregister VPA */
 560                 /* Check they don't still have a DTL or SLB buf registered */
 561                 err = H_RESOURCE;
 562                 if (vpa_is_registered(&tvcpu->arch.dtl) ||
 563                     vpa_is_registered(&tvcpu->arch.slb_shadow))
 564                         break;
 565
 566                 vpap = &tvcpu->arch.vpa;
 567                 err = 0;
 568                 break;
 569
 570         case H_VPA_DEREG_DTL:           /* deregister DTL */
 571                 vpap = &tvcpu->arch.dtl;
 572                 err = 0;
 573                 break;
 574
 575         case H_VPA_DEREG_SLB:           /* deregister SLB shadow buffer */
 576                 vpap = &tvcpu->arch.slb_shadow;
 577                 err = 0;
 578                 break;
 579         }
 580
 581         if (vpap) {
 582                 vpap->next_gpa = vpa;
 583                 vpap->len = len;
 584                 vpap->update_pending = 1;
 585         }
 586
 587         spin_unlock(&tvcpu->arch.vpa_update_lock);
 588
 589         return err;
 590 }
 591
 592 static void kvmppc_update_vpa(struct kvm_vcpu *vcpu, struct kvmppc_vpa *vpap)
 593 {
 594         struct kvm *kvm = vcpu->kvm;
 595         void *va;
 596         unsigned long nb;
 597         unsigned long gpa;
 598
 599         /*
 600          * We need to pin the page pointed to by vpap->next_gpa,
 601          * but we can't call kvmppc_pin_guest_page under the lock
 602          * as it does get_user_pages() and down_read().  So we
 603          * have to drop the lock, pin the page, then get the lock
 604          * again and check that a new area didn't get registered
 605          * in the meantime.
 606          */
 607         for (;;) {
 608                 gpa = vpap->next_gpa;
 609                 spin_unlock(&vcpu->arch.vpa_update_lock);
 610                 va = NULL;
 611                 nb = 0;
 612                 if (gpa)
 613                         va = kvmppc_pin_guest_page(kvm, gpa, &nb);
 614                 spin_lock(&vcpu->arch.vpa_update_lock);
 615                 if (gpa == vpap->next_gpa)
 616                         break;
 617                 /* sigh... unpin that one and try again */
 618                 if (va)
 619                         kvmppc_unpin_guest_page(kvm, va, gpa, false);
 620         }
 621
 622         vpap->update_pending = 0;
 623         if (va && nb < vpap->len) {
 624                 /*
 625                  * If it's now too short, it must be that userspace
 626                  * has changed the mappings underlying guest memory,
 627                  * so unregister the region.
 628                  */
 629                 kvmppc_unpin_guest_page(kvm, va, gpa, false);
 630                 va = NULL;
 631         }
 632         if (vpap->pinned_addr)
 633                 kvmppc_unpin_guest_page(kvm, vpap->pinned_addr, vpap->gpa,
 634                                         vpap->dirty);
 635         vpap->gpa = gpa;
 636         vpap->pinned_addr = va;
 637         vpap->dirty = false;
 638         if (va)
 639                 vpap->pinned_end = va + vpap->len;
 640 }
 641
 642 static void kvmppc_update_vpas(struct kvm_vcpu *vcpu)
 643 {
 644         if (!(vcpu->arch.vpa.update_pending ||
 645               vcpu->arch.slb_shadow.update_pending ||
 646               vcpu->arch.dtl.update_pending))
 647                 return;
 648
 649         spin_lock(&vcpu->arch.vpa_update_lock);
 650         if (vcpu->arch.vpa.update_pending) {
 651                 kvmppc_update_vpa(vcpu, &vcpu->arch.vpa);
 652                 if (vcpu->arch.vpa.pinned_addr)
 653                         init_vpa(vcpu, vcpu->arch.vpa.pinned_addr);
 654         }
 655         if (vcpu->arch.dtl.update_pending) {
 656                 kvmppc_update_vpa(vcpu, &vcpu->arch.dtl);
 657                 vcpu->arch.dtl_ptr = vcpu->arch.dtl.pinned_addr;
 658                 vcpu->arch.dtl_index = 0;
 659         }
 660         if (vcpu->arch.slb_shadow.update_pending)
 661                 kvmppc_update_vpa(vcpu, &vcpu->arch.slb_shadow);
 662         spin_unlock(&vcpu->arch.vpa_update_lock);
 663 }
 664
 665 /*
 666  * Return the accumulated stolen time for the vcore up until `now'.
 667  * The caller should hold the vcore lock.
 668  */
 669 static u64 vcore_stolen_time(struct kvmppc_vcore *vc, u64 now)
 670 {
 671         u64 p;
 672         unsigned long flags;
 673
 674         spin_lock_irqsave(&vc->stoltb_lock, flags);
 675         p = vc->stolen_tb;
 676         if (vc->vcore_state != VCORE_INACTIVE &&
 677             vc->preempt_tb != TB_NIL)
 678                 p += now - vc->preempt_tb;
 679         spin_unlock_irqrestore(&vc->stoltb_lock, flags);
 680         return p;
 681 }
 682
 683 static void kvmppc_create_dtl_entry(struct kvm_vcpu *vcpu,
 684                                     struct kvmppc_vcore *vc)
 685 {
 686         struct dtl_entry *dt;
 687         struct lppaca *vpa;
 688         unsigned long stolen;
 689         unsigned long core_stolen;
 690         u64 now;
 691         unsigned long flags;
 692
 693         dt = vcpu->arch.dtl_ptr;
 694         vpa = vcpu->arch.vpa.pinned_addr;
 695         now = mftb();
 696         core_stolen = vcore_stolen_time(vc, now);
 697         stolen = core_stolen - vcpu->arch.stolen_logged;
 698         vcpu->arch.stolen_logged = core_stolen;
 699         spin_lock_irqsave(&vcpu->arch.tbacct_lock, flags);
 700         stolen += vcpu->arch.busy_stolen;
 701         vcpu->arch.busy_stolen = 0;
 702         spin_unlock_irqrestore(&vcpu->arch.tbacct_lock, flags);
 703         if (!dt || !vpa)
 704                 return;
 705         memset(dt, 0, sizeof(struct dtl_entry));
 706         dt->dispatch_reason = 7;
 707         dt->processor_id = cpu_to_be16(vc->pcpu + vcpu->arch.ptid);
 708         dt->timebase = cpu_to_be64(now + vc->tb_offset);
 709         dt->enqueue_to_dispatch_time = cpu_to_be32(stolen);
 710         dt->srr0 = cpu_to_be64(kvmppc_get_pc(vcpu));
 711         dt->srr1 = cpu_to_be64(vcpu->arch.shregs.msr);
 712         ++dt;
 713         if (dt == vcpu->arch.dtl.pinned_end)
 714                 dt = vcpu->arch.dtl.pinned_addr;
 715         vcpu->arch.dtl_ptr = dt;
 716         /* order writing *dt vs. writing vpa->dtl_idx */
 717         smp_wmb();
 718         vpa->dtl_idx = cpu_to_be64(++vcpu->arch.dtl_index);
 719         vcpu->arch.dtl.dirty = true;
 720 }
 721
 722 /* See if there is a doorbell interrupt pending for a vcpu */
 723 static bool kvmppc_doorbell_pending(struct kvm_vcpu *vcpu)
 724 {
 725         int thr;
 726         struct kvmppc_vcore *vc;
 727
 728         if (vcpu->arch.doorbell_request)
 729                 return true;
 730         /*
 731          * Ensure that the read of vcore->dpdes comes after the read
 732          * of vcpu->doorbell_request.  This barrier matches the
 733          * lwsync in book3s_hv_rmhandlers.S just before the
 734          * fast_guest_return label.
 735          */
 736         smp_rmb();
 737         vc = vcpu->arch.vcore;
 738         thr = vcpu->vcpu_id - vc->first_vcpuid;
 739         return !!(vc->dpdes & (1 << thr));
 740 }
 741
 742 static bool kvmppc_power8_compatible(struct kvm_vcpu *vcpu)
 743 {
 744         if (vcpu->arch.vcore->arch_compat >= PVR_ARCH_207)
 745                 return true;
 746         if ((!vcpu->arch.vcore->arch_compat) &&
 747             cpu_has_feature(CPU_FTR_ARCH_207S))
 748                 return true;
 749         return false;
 750 }
 751
 752 static int kvmppc_h_set_mode(struct kvm_vcpu *vcpu, unsigned long mflags,
 753                              unsigned long resource, unsigned long value1,
 754                              unsigned long value2)
 755 {
 756         switch (resource) {
 757         case H_SET_MODE_RESOURCE_SET_CIABR:
 758                 if (!kvmppc_power8_compatible(vcpu))
 759                         return H_P2;
 760                 if (value2)
 761                         return H_P4;
 762                 if (mflags)
 763                         return H_UNSUPPORTED_FLAG_START;
 764                 /* Guests can't breakpoint the hypervisor */
 765                 if ((value1 & CIABR_PRIV) == CIABR_PRIV_HYPER)
 766                         return H_P3;
 767                 vcpu->arch.ciabr  = value1;
 768                 return H_SUCCESS;
 769         case H_SET_MODE_RESOURCE_SET_DAWR:
 770                 if (!kvmppc_power8_compatible(vcpu))
 771                         return H_P2;
 772                 if (!ppc_breakpoint_available())
 773                         return H_P2;
 774                 if (mflags)
 775                         return H_UNSUPPORTED_FLAG_START;
 776                 if (value2 & DABRX_HYP)
 777                         return H_P4;
 778                 vcpu->arch.dawr  = value1;
 779                 vcpu->arch.dawrx = value2;
 780                 return H_SUCCESS;
 781         default:
 782                 return H_TOO_HARD;
 783         }
 784 }
 785
 786 static int kvm_arch_vcpu_yield_to(struct kvm_vcpu *target)
 787 {
 788         struct kvmppc_vcore *vcore = target->arch.vcore;
 789
 790         /*
 791          * We expect to have been called by the real mode handler
 792          * (kvmppc_rm_h_confer()) which would have directly returned
 793          * H_SUCCESS if the source vcore wasn't idle (e.g. if it may
 794          * have useful work to do and should not confer) so we don't
 795          * recheck that here.
 796          */
 797
 798         spin_lock(&vcore->lock);
 799         if (target->arch.state == KVMPPC_VCPU_RUNNABLE &&
 800             vcore->vcore_state != VCORE_INACTIVE &&
 801             vcore->runner)
 802                 target = vcore->runner;
 803         spin_unlock(&vcore->lock);
 804
 805         return kvm_vcpu_yield_to(target);
 806 }
 807
 808 static int kvmppc_get_yield_count(struct kvm_vcpu *vcpu)
 809 {
 810         int yield_count = 0;
 811         struct lppaca *lppaca;
 812
 813         spin_lock(&vcpu->arch.vpa_update_lock);
 814         lppaca = (struct lppaca *)vcpu->arch.vpa.pinned_addr;
 815         if (lppaca)
 816                 yield_count = be32_to_cpu(lppaca->yield_count);
 817         spin_unlock(&vcpu->arch.vpa_update_lock);
 818         return yield_count;
 819 }
 820
 821 int kvmppc_pseries_do_hcall(struct kvm_vcpu *vcpu)
 822 {
 823         unsigned long req = kvmppc_get_gpr(vcpu, 3);
 824         unsigned long target, ret = H_SUCCESS;
 825         int yield_count;
 826         struct kvm_vcpu *tvcpu;
 827         int idx, rc;
 828
 829         if (req <= MAX_HCALL_OPCODE &&
 830             !test_bit(req/4, vcpu->kvm->arch.enabled_hcalls))
 831                 return RESUME_HOST;
 832
 833         switch (req) {
 834         case H_CEDE:
 835                 break;
 836         case H_PROD:
 837                 target = kvmppc_get_gpr(vcpu, 4);
 838                 tvcpu = kvmppc_find_vcpu(vcpu->kvm, target);
 839                 if (!tvcpu) {
 840                         ret = H_PARAMETER;
 841                         break;
 842                 }
 843                 tvcpu->arch.prodded = 1;
 844                 smp_mb();
 845                 if (tvcpu->arch.ceded)
 846                         kvmppc_fast_vcpu_kick_hv(tvcpu);
 847                 break;
 848         case H_CONFER:
 849                 target = kvmppc_get_gpr(vcpu, 4);
 850                 if (target == -1)
 851                         break;
 852                 tvcpu = kvmppc_find_vcpu(vcpu->kvm, target);
 853                 if (!tvcpu) {
 854                         ret = H_PARAMETER;
 855                         break;
 856                 }
 857                 yield_count = kvmppc_get_gpr(vcpu, 5);
 858                 if (kvmppc_get_yield_count(tvcpu) != yield_count)
 859                         break;
 860                 kvm_arch_vcpu_yield_to(tvcpu);
 861                 break;
 862         case H_REGISTER_VPA:
 863                 ret = do_h_register_vpa(vcpu, kvmppc_get_gpr(vcpu, 4),
 864                                         kvmppc_get_gpr(vcpu, 5),
 865                                         kvmppc_get_gpr(vcpu, 6));
 866                 break;
 867         case H_RTAS:
 868                 if (list_empty(&vcpu->kvm->arch.rtas_tokens))
 869                         return RESUME_HOST;
 870
 871                 idx = srcu_read_lock(&vcpu->kvm->srcu);
 872                 rc = kvmppc_rtas_hcall(vcpu);
 873                 srcu_read_unlock(&vcpu->kvm->srcu, idx);
 874
 875                 if (rc == -ENOENT)
 876                         return RESUME_HOST;
 877                 else if (rc == 0)
 878                         break;
 879
 880                 /* Send the error out to userspace via KVM_RUN */
 881                 return rc;
 882         case H_LOGICAL_CI_LOAD:
 883                 ret = kvmppc_h_logical_ci_load(vcpu);
 884                 if (ret == H_TOO_HARD)
 885                         return RESUME_HOST;
 886                 break;
 887         case H_LOGICAL_CI_STORE:
 888                 ret = kvmppc_h_logical_ci_store(vcpu);
 889                 if (ret == H_TOO_HARD)
 890                         return RESUME_HOST;
 891                 break;
 892         case H_SET_MODE:
 893                 ret = kvmppc_h_set_mode(vcpu, kvmppc_get_gpr(vcpu, 4),
 894                                         kvmppc_get_gpr(vcpu, 5),
 895                                         kvmppc_get_gpr(vcpu, 6),
 896                                         kvmppc_get_gpr(vcpu, 7));
 897                 if (ret == H_TOO_HARD)
 898                         return RESUME_HOST;
 899                 break;
 900         case H_XIRR:
 901         case H_CPPR:
 902         case H_EOI:
 903         case H_IPI:
 904         case H_IPOLL:
 905         case H_XIRR_X:
 906                 if (kvmppc_xics_enabled(vcpu)) {
 907                         if (xive_enabled()) {
 908                                 ret = H_NOT_AVAILABLE;
 909                                 return RESUME_GUEST;
 910                         }
 911                         ret = kvmppc_xics_hcall(vcpu, req);
 912                         break;
 913                 }
 914                 return RESUME_HOST;
 915         case H_PUT_TCE:
 916                 ret = kvmppc_h_put_tce(vcpu, kvmppc_get_gpr(vcpu, 4),
 917                                                 kvmppc_get_gpr(vcpu, 5),
 918                                                 kvmppc_get_gpr(vcpu, 6));
 919                 if (ret == H_TOO_HARD)
 920                         return RESUME_HOST;
 921                 break;
 922         case H_PUT_TCE_INDIRECT:
 923                 ret = kvmppc_h_put_tce_indirect(vcpu, kvmppc_get_gpr(vcpu, 4),
 924                                                 kvmppc_get_gpr(vcpu, 5),
 925                                                 kvmppc_get_gpr(vcpu, 6),
 926                                                 kvmppc_get_gpr(vcpu, 7));
 927                 if (ret == H_TOO_HARD)
 928                         return RESUME_HOST;
 929                 break;
 930         case H_STUFF_TCE:
 931                 ret = kvmppc_h_stuff_tce(vcpu, kvmppc_get_gpr(vcpu, 4),
 932                                                 kvmppc_get_gpr(vcpu, 5),
 933                                                 kvmppc_get_gpr(vcpu, 6),
 934                                                 kvmppc_get_gpr(vcpu, 7));
 935                 if (ret == H_TOO_HARD)
 936                         return RESUME_HOST;
 937                 break;
 938         default:
 939                 return RESUME_HOST;
 940         }
 941         kvmppc_set_gpr(vcpu, 3, ret);
 942         vcpu->arch.hcall_needed = 0;
 943         return RESUME_GUEST;
 944 }
 945
 946 static int kvmppc_hcall_impl_hv(unsigned long cmd)
 947 {
 948         switch (cmd) {
 949         case H_CEDE:
 950         case H_PROD:
 951         case H_CONFER:
 952         case H_REGISTER_VPA:
 953         case H_SET_MODE:
 954         case H_LOGICAL_CI_LOAD:
 955         case H_LOGICAL_CI_STORE:
 956 #ifdef CONFIG_KVM_XICS
 957         case H_XIRR:
 958         case H_CPPR:
 959         case H_EOI:
 960         case H_IPI:
 961         case H_IPOLL:
 962         case H_XIRR_X:
 963 #endif
 964                 return 1;
 965         }
 966
 967         /* See if it's in the real-mode table */
 968         return kvmppc_hcall_impl_hv_realmode(cmd);
 969 }
 970
 971 static int kvmppc_emulate_debug_inst(struct kvm_run *run,
 972                                         struct kvm_vcpu *vcpu)
 973 {
 974         u32 last_inst;
 975
 976         if (kvmppc_get_last_inst(vcpu, INST_GENERIC, &last_inst) !=
 977                                         EMULATE_DONE) {
 978                 /*
 979                  * Fetch failed, so return to guest and
 980                  * try executing it again.
 981                  */
 982                 return RESUME_GUEST;
 983         }
 984
 985         if (last_inst == KVMPPC_INST_SW_BREAKPOINT) {
 986                 run->exit_reason = KVM_EXIT_DEBUG;
 987                 run->debug.arch.address = kvmppc_get_pc(vcpu);
 988                 return RESUME_HOST;
 989         } else {
 990                 kvmppc_core_queue_program(vcpu, SRR1_PROGILL);
 991                 return RESUME_GUEST;
 992         }
 993 }
 994
 995 static void do_nothing(void *x)
 996 {
 997 }
 998
 999 static unsigned long kvmppc_read_dpdes(struct kvm_vcpu *vcpu)
1000 {
1001         int thr, cpu, pcpu, nthreads;
1002         struct kvm_vcpu *v;
1003         unsigned long dpdes;
1004
1005         nthreads = vcpu->kvm->arch.emul_smt_mode;
1006         dpdes = 0;
1007         cpu = vcpu->vcpu_id & ~(nthreads - 1);
1008         for (thr = 0; thr < nthreads; ++thr, ++cpu) {
1009                 v = kvmppc_find_vcpu(vcpu->kvm, cpu);
1010                 if (!v)
1011                         continue;
1012                 /*
1013                  * If the vcpu is currently running on a physical cpu thread,
1014                  * interrupt it in order to pull it out of the guest briefly,
1015                  * which will update its vcore->dpdes value.
1016                  */
1017                 pcpu = READ_ONCE(v->cpu);
1018                 if (pcpu >= 0)
1019                         smp_call_function_single(pcpu, do_nothing, NULL, 1);
1020                 if (kvmppc_doorbell_pending(v))
1021                         dpdes |= 1 << thr;
1022         }
1023         return dpdes;
1024 }
1025
1026 /*
1027  * On POWER9, emulate doorbell-related instructions in order to
1028  * give the guest the illusion of running on a multi-threaded core.
1029  * The instructions emulated are msgsndp, msgclrp, mfspr TIR,
1030  * and mfspr DPDES.
1031  */
1032 static int kvmppc_emulate_doorbell_instr(struct kvm_vcpu *vcpu)
1033 {
1034         u32 inst, rb, thr;
1035         unsigned long arg;
1036         struct kvm *kvm = vcpu->kvm;
1037         struct kvm_vcpu *tvcpu;
1038
1039         if (kvmppc_get_last_inst(vcpu, INST_GENERIC, &inst) != EMULATE_DONE)
1040                 return RESUME_GUEST;
1041         if (get_op(inst) != 31)
1042                 return EMULATE_FAIL;
1043         rb = get_rb(inst);
1044         thr = vcpu->vcpu_id & (kvm->arch.emul_smt_mode - 1);
1045         switch (get_xop(inst)) {
1046         case OP_31_XOP_MSGSNDP:
1047                 arg = kvmppc_get_gpr(vcpu, rb);
1048                 if (((arg >> 27) & 0xf) != PPC_DBELL_SERVER)
1049                         break;
1050                 arg &= 0x3f;
1051                 if (arg >= kvm->arch.emul_smt_mode)
1052                         break;
1053                 tvcpu = kvmppc_find_vcpu(kvm, vcpu->vcpu_id - thr + arg);
1054                 if (!tvcpu)
1055                         break;
1056                 if (!tvcpu->arch.doorbell_request) {
1057                         tvcpu->arch.doorbell_request = 1;
1058                         kvmppc_fast_vcpu_kick_hv(tvcpu);
1059                 }
1060                 break;
1061         case OP_31_XOP_MSGCLRP:
1062                 arg = kvmppc_get_gpr(vcpu, rb);
1063                 if (((arg >> 27) & 0xf) != PPC_DBELL_SERVER)
1064                         break;
1065                 vcpu->arch.vcore->dpdes = 0;
1066                 vcpu->arch.doorbell_request = 0;
1067                 break;
1068         case OP_31_XOP_MFSPR:
1069                 switch (get_sprn(inst)) {
1070                 case SPRN_TIR:
1071                         arg = thr;
1072                         break;
1073                 case SPRN_DPDES:
1074                         arg = kvmppc_read_dpdes(vcpu);
1075                         break;
1076                 default:
1077                         return EMULATE_FAIL;
1078                 }
1079                 kvmppc_set_gpr(vcpu, get_rt(inst), arg);
1080                 break;
1081         default:
1082                 return EMULATE_FAIL;
1083         }
1084         kvmppc_set_pc(vcpu, kvmppc_get_pc(vcpu) + 4);
1085         return RESUME_GUEST;
1086 }
1087
1088 /* Called with vcpu->arch.vcore->lock held */
1089 static int kvmppc_handle_exit_hv(struct kvm_run *run, struct kvm_vcpu *vcpu,
1090                                  struct task_struct *tsk)
1091 {
1092         int r = RESUME_HOST;
1093
1094         vcpu->stat.sum_exits++;
1095
1096         /*
1097          * This can happen if an interrupt occurs in the last stages
1098          * of guest entry or the first stages of guest exit (i.e. after
1099          * setting paca->kvm_hstate.in_guest to KVM_GUEST_MODE_GUEST_HV
1100          * and before setting it to KVM_GUEST_MODE_HOST_HV).
1101          * That can happen due to a bug, or due to a machine check
1102          * occurring at just the wrong time.
1103          */
1104         if (vcpu->arch.shregs.msr & MSR_HV) {
1105                 printk(KERN_EMERG "KVM trap in HV mode!\n");
1106                 printk(KERN_EMERG "trap=0x%x | pc=0x%lx | msr=0x%llx\n",
1107                         vcpu->arch.trap, kvmppc_get_pc(vcpu),
1108                         vcpu->arch.shregs.msr);
1109                 kvmppc_dump_regs(vcpu);
1110                 run->exit_reason = KVM_EXIT_INTERNAL_ERROR;
1111                 run->hw.hardware_exit_reason = vcpu->arch.trap;
1112                 return RESUME_HOST;
1113         }
1114         run->exit_reason = KVM_EXIT_UNKNOWN;
1115         run->ready_for_interrupt_injection = 1;
1116         switch (vcpu->arch.trap) {
1117         /* We're good on these - the host merely wanted to get our attention */
1118         case BOOK3S_INTERRUPT_HV_DECREMENTER:
1119                 vcpu->stat.dec_exits++;
1120                 r = RESUME_GUEST;
1121                 break;
1122         case BOOK3S_INTERRUPT_EXTERNAL:
1123         case BOOK3S_INTERRUPT_H_DOORBELL:
1124         case BOOK3S_INTERRUPT_H_VIRT:
1125                 vcpu->stat.ext_intr_exits++;
1126                 r = RESUME_GUEST;
1127                 break;
1128         /* SR/HMI/PMI are HV interrupts that host has handled. Resume guest.*/
1129         case BOOK3S_INTERRUPT_HMI:
1130         case BOOK3S_INTERRUPT_PERFMON:
1131         case BOOK3S_INTERRUPT_SYSTEM_RESET:
1132                 r = RESUME_GUEST;
1133                 break;
1134         case BOOK3S_INTERRUPT_MACHINE_CHECK:
1135                 /* Exit to guest with KVM_EXIT_NMI as exit reason */
1136                 run->exit_reason = KVM_EXIT_NMI;
1137                 run->hw.hardware_exit_reason = vcpu->arch.trap;
1138                 /* Clear out the old NMI status from run->flags */
1139                 run->flags &= ~KVM_RUN_PPC_NMI_DISP_MASK;
1140                 /* Now set the NMI status */
1141                 if (vcpu->arch.mce_evt.disposition == MCE_DISPOSITION_RECOVERED)
1142                         run->flags |= KVM_RUN_PPC_NMI_DISP_FULLY_RECOV;
1143                 else
1144                         run->flags |= KVM_RUN_PPC_NMI_DISP_NOT_RECOV;
1145
1146                 r = RESUME_HOST;
1147                 /* Print the MCE event to host console. */
1148                 machine_check_print_event_info(&vcpu->arch.mce_evt, false);
1149                 break;
1150         case BOOK3S_INTERRUPT_PROGRAM:
1151         {
1152                 ulong flags;
1153                 /*
1154                  * Normally program interrupts are delivered directly
1155                  * to the guest by the hardware, but we can get here
1156                  * as a result of a hypervisor emulation interrupt
1157                  * (e40) getting turned into a 700 by BML RTAS.
1158                  */
1159                 flags = vcpu->arch.shregs.msr & 0x1f0000ull;
1160                 kvmppc_core_queue_program(vcpu, flags);
1161                 r = RESUME_GUEST;
1162                 break;
1163         }
1164         case BOOK3S_INTERRUPT_SYSCALL:
1165         {
1166                 /* hcall - punt to userspace */
1167                 int i;
1168
1169                 /* hypercall with MSR_PR has already been handled in rmode,
1170                  * and never reaches here.
1171                  */
1172
1173                 run->papr_hcall.nr = kvmppc_get_gpr(vcpu, 3);
1174                 for (i = 0; i < 9; ++i)
1175                         run->papr_hcall.args[i] = kvmppc_get_gpr(vcpu, 4 + i);
1176                 run->exit_reason = KVM_EXIT_PAPR_HCALL;
1177                 vcpu->arch.hcall_needed = 1;
1178                 r = RESUME_HOST;
1179                 break;
1180         }
1181         /*
1182          * We get these next two if the guest accesses a page which it thinks
1183          * it has mapped but which is not actually present, either because
1184          * it is for an emulated I/O device or because the corresonding
1185          * host page has been paged out.  Any other HDSI/HISI interrupts
1186          * have been handled already.
1187          */
1188         case BOOK3S_INTERRUPT_H_DATA_STORAGE:
1189                 r = RESUME_PAGE_FAULT;
1190                 break;
1191         case BOOK3S_INTERRUPT_H_INST_STORAGE:
1192                 vcpu->arch.fault_dar = kvmppc_get_pc(vcpu);
1193                 vcpu->arch.fault_dsisr = 0;
1194                 r = RESUME_PAGE_FAULT;
1195                 break;
1196         /*
1197          * This occurs if the guest executes an illegal instruction.
1198          * If the guest debug is disabled, generate a program interrupt
1199          * to the guest. If guest debug is enabled, we need to check
1200          * whether the instruction is a software breakpoint instruction.
1201          * Accordingly return to Guest or Host.
1202          */
1203         case BOOK3S_INTERRUPT_H_EMUL_ASSIST:
1204                 if (vcpu->arch.emul_inst != KVM_INST_FETCH_FAILED)
1205                         vcpu->arch.last_inst = kvmppc_need_byteswap(vcpu) ?
1206                                 swab32(vcpu->arch.emul_inst) :
1207                                 vcpu->arch.emul_inst;
1208                 if (vcpu->guest_debug & KVM_GUESTDBG_USE_SW_BP) {
1209                         /* Need vcore unlocked to call kvmppc_get_last_inst */
1210                         spin_unlock(&vcpu->arch.vcore->lock);
1211                         r = kvmppc_emulate_debug_inst(run, vcpu);
1212                         spin_lock(&vcpu->arch.vcore->lock);
1213                 } else {
1214                         kvmppc_core_queue_program(vcpu, SRR1_PROGILL);
1215                         r = RESUME_GUEST;
1216                 }
1217                 break;
1218         /*
1219          * This occurs if the guest (kernel or userspace), does something that
1220          * is prohibited by HFSCR.
1221          * On POWER9, this could be a doorbell instruction that we need
1222          * to emulate.
1223          * Otherwise, we just generate a program interrupt to the guest.
1224          */
1225         case BOOK3S_INTERRUPT_H_FAC_UNAVAIL:
1226                 r = EMULATE_FAIL;
1227                 if (((vcpu->arch.hfscr >> 56) == FSCR_MSGP_LG) &&
1228                     cpu_has_feature(CPU_FTR_ARCH_300)) {
1229                         /* Need vcore unlocked to call kvmppc_get_last_inst */
1230                         spin_unlock(&vcpu->arch.vcore->lock);
1231                         r = kvmppc_emulate_doorbell_instr(vcpu);
1232                         spin_lock(&vcpu->arch.vcore->lock);
1233                 }
1234                 if (r == EMULATE_FAIL) {
1235                         kvmppc_core_queue_program(vcpu, SRR1_PROGILL);
1236                         r = RESUME_GUEST;
1237                 }
1238                 break;
1239
1240 #ifdef CONFIG_PPC_TRANSACTIONAL_MEM
1241         case BOOK3S_INTERRUPT_HV_SOFTPATCH:
1242                 /*
1243                  * This occurs for various TM-related instructions that
1244                  * we need to emulate on POWER9 DD2.2.  We have already
1245                  * handled the cases where the guest was in real-suspend
1246                  * mode and was transitioning to transactional state.
1247                  */
1248                 r = kvmhv_p9_tm_emulation(vcpu);
1249                 break;
1250 #endif
1251
1252         case BOOK3S_INTERRUPT_HV_RM_HARD:
1253                 r = RESUME_PASSTHROUGH;
1254                 break;
1255         default:
1256                 kvmppc_dump_regs(vcpu);
1257                 printk(KERN_EMERG "trap=0x%x | pc=0x%lx | msr=0x%llx\n",
1258                         vcpu->arch.trap, kvmppc_get_pc(vcpu),
1259                         vcpu->arch.shregs.msr);
1260                 run->hw.hardware_exit_reason = vcpu->arch.trap;
1261                 r = RESUME_HOST;
1262                 break;
1263         }
1264
1265         return r;
1266 }
1267
1268 static int kvm_arch_vcpu_ioctl_get_sregs_hv(struct kvm_vcpu *vcpu,
1269                                             struct kvm_sregs *sregs)
1270 {
1271         int i;
1272
1273         memset(sregs, 0, sizeof(struct kvm_sregs));
1274         sregs->pvr = vcpu->arch.pvr;
1275         for (i = 0; i < vcpu->arch.slb_max; i++) {
1276                 sregs->u.s.ppc64.slb[i].slbe = vcpu->arch.slb[i].orige;
1277                 sregs->u.s.ppc64.slb[i].slbv = vcpu->arch.slb[i].origv;
1278         }
1279
1280         return 0;
1281 }
1282
1283 static int kvm_arch_vcpu_ioctl_set_sregs_hv(struct kvm_vcpu *vcpu,
1284                                             struct kvm_sregs *sregs)
1285 {
1286         int i, j;
1287
1288         /* Only accept the same PVR as the host's, since we can't spoof it */
1289         if (sregs->pvr != vcpu->arch.pvr)
1290                 return -EINVAL;
1291
1292         j = 0;
1293         for (i = 0; i < vcpu->arch.slb_nr; i++) {
1294                 if (sregs->u.s.ppc64.slb[i].slbe & SLB_ESID_V) {
1295                         vcpu->arch.slb[j].orige = sregs->u.s.ppc64.slb[i].slbe;
1296                         vcpu->arch.slb[j].origv = sregs->u.s.ppc64.slb[i].slbv;
1297                         ++j;
1298                 }
1299         }
1300         vcpu->arch.slb_max = j;
1301
1302         return 0;
1303 }
1304
1305 static void kvmppc_set_lpcr(struct kvm_vcpu *vcpu, u64 new_lpcr,
1306                 bool preserve_top32)
1307 {
1308         struct kvm *kvm = vcpu->kvm;
1309         struct kvmppc_vcore *vc = vcpu->arch.vcore;
1310         u64 mask;
1311
1312         mutex_lock(&kvm->lock);
1313         spin_lock(&vc->lock);
1314         /*
1315          * If ILE (interrupt little-endian) has changed, update the
1316          * MSR_LE bit in the intr_msr for each vcpu in this vcore.
1317          */
1318         if ((new_lpcr & LPCR_ILE) != (vc->lpcr & LPCR_ILE)) {
1319                 struct kvm_vcpu *vcpu;
1320                 int i;
1321
1322                 kvm_for_each_vcpu(i, vcpu, kvm) {
1323                         if (vcpu->arch.vcore != vc)
1324                                 continue;
1325                         if (new_lpcr & LPCR_ILE)
1326                                 vcpu->arch.intr_msr |= MSR_LE;
1327                         else
1328                                 vcpu->arch.intr_msr &= ~MSR_LE;
1329                 }
1330         }
1331
1332         /*
1333          * Userspace can only modify DPFD (default prefetch depth),
1334          * ILE (interrupt little-endian) and TC (translation control).
1335          * On POWER8 and POWER9 userspace can also modify AIL (alt. interrupt loc.).
1336          */
1337         mask = LPCR_DPFD | LPCR_ILE | LPCR_TC;
1338         if (cpu_has_feature(CPU_FTR_ARCH_207S))
1339                 mask |= LPCR_AIL;
1340         /*
1341          * On POWER9, allow userspace to enable large decrementer for the
1342          * guest, whether or not the host has it enabled.
1343          */
1344         if (cpu_has_feature(CPU_FTR_ARCH_300))
1345                 mask |= LPCR_LD;
1346
1347         /* Broken 32-bit version of LPCR must not clear top bits */
1348         if (preserve_top32)
1349                 mask &= 0xFFFFFFFF;
1350         vc->lpcr = (vc->lpcr & ~mask) | (new_lpcr & mask);
1351         spin_unlock(&vc->lock);
1352         mutex_unlock(&kvm->lock);
1353 }
1354
1355 static int kvmppc_get_one_reg_hv(struct kvm_vcpu *vcpu, u64 id,
1356                                  union kvmppc_one_reg *val)
1357 {
1358         int r = 0;
1359         long int i;
1360
1361         switch (id) {
1362         case KVM_REG_PPC_DEBUG_INST:
1363                 *val = get_reg_val(id, KVMPPC_INST_SW_BREAKPOINT);
1364                 break;
1365         case KVM_REG_PPC_HIOR:
1366                 *val = get_reg_val(id, 0);
1367                 break;
1368         case KVM_REG_PPC_DABR:
1369                 *val = get_reg_val(id, vcpu->arch.dabr);
1370                 break;
1371         case KVM_REG_PPC_DABRX:
1372                 *val = get_reg_val(id, vcpu->arch.dabrx);
1373                 break;
1374         case KVM_REG_PPC_DSCR:
1375                 *val = get_reg_val(id, vcpu->arch.dscr);
1376                 break;
1377         case KVM_REG_PPC_PURR:
1378                 *val = get_reg_val(id, vcpu->arch.purr);
1379                 break;
1380         case KVM_REG_PPC_SPURR:
1381                 *val = get_reg_val(id, vcpu->arch.spurr);
1382                 break;
1383         case KVM_REG_PPC_AMR:
1384                 *val = get_reg_val(id, vcpu->arch.amr);
1385                 break;
1386         case KVM_REG_PPC_UAMOR:
1387                 *val = get_reg_val(id, vcpu->arch.uamor);
1388                 break;
1389         case KVM_REG_PPC_MMCR0 ... KVM_REG_PPC_MMCRS:
1390                 i = id - KVM_REG_PPC_MMCR0;
1391                 *val = get_reg_val(id, vcpu->arch.mmcr[i]);
1392                 break;
1393         case KVM_REG_PPC_PMC1 ... KVM_REG_PPC_PMC8:
1394                 i = id - KVM_REG_PPC_PMC1;
1395                 *val = get_reg_val(id, vcpu->arch.pmc[i]);
1396                 break;
1397         case KVM_REG_PPC_SPMC1 ... KVM_REG_PPC_SPMC2:
1398                 i = id - KVM_REG_PPC_SPMC1;
1399                 *val = get_reg_val(id, vcpu->arch.spmc[i]);
1400                 break;
1401         case KVM_REG_PPC_SIAR:
1402                 *val = get_reg_val(id, vcpu->arch.siar);
1403                 break;
1404         case KVM_REG_PPC_SDAR:
1405                 *val = get_reg_val(id, vcpu->arch.sdar);
1406                 break;
1407         case KVM_REG_PPC_SIER:
1408                 *val = get_reg_val(id, vcpu->arch.sier);
1409                 break;
1410         case KVM_REG_PPC_IAMR:
1411                 *val = get_reg_val(id, vcpu->arch.iamr);
1412                 break;
1413         case KVM_REG_PPC_PSPB:
1414                 *val = get_reg_val(id, vcpu->arch.pspb);
1415                 break;
1416         case KVM_REG_PPC_DPDES:
1417                 *val = get_reg_val(id, vcpu->arch.vcore->dpdes);
1418                 break;
1419         case KVM_REG_PPC_VTB:
1420                 *val = get_reg_val(id, vcpu->arch.vcore->vtb);
1421                 break;
1422         case KVM_REG_PPC_DAWR:
1423                 *val = get_reg_val(id, vcpu->arch.dawr);
1424                 break;
1425         case KVM_REG_PPC_DAWRX:
1426                 *val = get_reg_val(id, vcpu->arch.dawrx);
1427                 break;
1428         case KVM_REG_PPC_CIABR:
1429                 *val = get_reg_val(id, vcpu->arch.ciabr);
1430                 break;
1431         case KVM_REG_PPC_CSIGR:
1432                 *val = get_reg_val(id, vcpu->arch.csigr);
1433                 break;
1434         case KVM_REG_PPC_TACR:
1435                 *val = get_reg_val(id, vcpu->arch.tacr);
1436                 break;
1437         case KVM_REG_PPC_TCSCR:
1438                 *val = get_reg_val(id, vcpu->arch.tcscr);
1439                 break;
1440         case KVM_REG_PPC_PID:
1441                 *val = get_reg_val(id, vcpu->arch.pid);
1442                 break;
1443         case KVM_REG_PPC_ACOP:
1444                 *val = get_reg_val(id, vcpu->arch.acop);
1445                 break;
1446         case KVM_REG_PPC_WORT:
1447                 *val = get_reg_val(id, vcpu->arch.wort);
1448                 break;
1449         case KVM_REG_PPC_TIDR:
1450                 *val = get_reg_val(id, vcpu->arch.tid);
1451                 break;
1452         case KVM_REG_PPC_PSSCR:
1453                 *val = get_reg_val(id, vcpu->arch.psscr);
1454                 break;
1455         case KVM_REG_PPC_VPA_ADDR:
1456                 spin_lock(&vcpu->arch.vpa_update_lock);
1457                 *val = get_reg_val(id, vcpu->arch.vpa.next_gpa);
1458                 spin_unlock(&vcpu->arch.vpa_update_lock);
1459                 break;
1460         case KVM_REG_PPC_VPA_SLB:
1461                 spin_lock(&vcpu->arch.vpa_update_lock);
1462                 val->vpaval.addr = vcpu->arch.slb_shadow.next_gpa;
1463                 val->vpaval.length = vcpu->arch.slb_shadow.len;
1464                 spin_unlock(&vcpu->arch.vpa_update_lock);
1465                 break;
1466         case KVM_REG_PPC_VPA_DTL:
1467                 spin_lock(&vcpu->arch.vpa_update_lock);
1468                 val->vpaval.addr = vcpu->arch.dtl.next_gpa;
1469                 val->vpaval.length = vcpu->arch.dtl.len;
1470                 spin_unlock(&vcpu->arch.vpa_update_lock);
1471                 break;
1472         case KVM_REG_PPC_TB_OFFSET:
1473                 *val = get_reg_val(id, vcpu->arch.vcore->tb_offset);
1474                 break;
1475         case KVM_REG_PPC_LPCR:
1476         case KVM_REG_PPC_LPCR_64:
1477                 *val = get_reg_val(id, vcpu->arch.vcore->lpcr);
1478                 break;
1479         case KVM_REG_PPC_PPR:
1480                 *val = get_reg_val(id, vcpu->arch.ppr);
1481                 break;
1482 #ifdef CONFIG_PPC_TRANSACTIONAL_MEM
1483         case KVM_REG_PPC_TFHAR:
1484                 *val = get_reg_val(id, vcpu->arch.tfhar);
1485                 break;
1486         case KVM_REG_PPC_TFIAR:
1487                 *val = get_reg_val(id, vcpu->arch.tfiar);
1488                 break;
1489         case KVM_REG_PPC_TEXASR:
1490                 *val = get_reg_val(id, vcpu->arch.texasr);
1491                 break;
1492         case KVM_REG_PPC_TM_GPR0 ... KVM_REG_PPC_TM_GPR31:
1493                 i = id - KVM_REG_PPC_TM_GPR0;
1494                 *val = get_reg_val(id, vcpu->arch.gpr_tm[i]);
1495                 break;
1496         case KVM_REG_PPC_TM_VSR0 ... KVM_REG_PPC_TM_VSR63:
1497         {
1498                 int j;
1499                 i = id - KVM_REG_PPC_TM_VSR0;
1500                 if (i < 32)
1501                         for (j = 0; j < TS_FPRWIDTH; j++)
1502                                 val->vsxval[j] = vcpu->arch.fp_tm.fpr[i][j];
1503                 else {
1504                         if (cpu_has_feature(CPU_FTR_ALTIVEC))
1505                                 val->vval = vcpu->arch.vr_tm.vr[i-32];
1506                         else
1507                                 r = -ENXIO;
1508                 }
1509                 break;
1510         }
1511         case KVM_REG_PPC_TM_CR:
1512                 *val = get_reg_val(id, vcpu->arch.cr_tm);
1513                 break;
1514         case KVM_REG_PPC_TM_XER:
1515                 *val = get_reg_val(id, vcpu->arch.xer_tm);
1516                 break;
1517         case KVM_REG_PPC_TM_LR:
1518                 *val = get_reg_val(id, vcpu->arch.lr_tm);
1519                 break;
1520         case KVM_REG_PPC_TM_CTR:
1521                 *val = get_reg_val(id, vcpu->arch.ctr_tm);
1522                 break;
1523         case KVM_REG_PPC_TM_FPSCR:
1524                 *val = get_reg_val(id, vcpu->arch.fp_tm.fpscr);
1525                 break;
1526         case KVM_REG_PPC_TM_AMR:
1527                 *val = get_reg_val(id, vcpu->arch.amr_tm);
1528                 break;
1529         case KVM_REG_PPC_TM_PPR:
1530                 *val = get_reg_val(id, vcpu->arch.ppr_tm);
1531                 break;
1532         case KVM_REG_PPC_TM_VRSAVE:
1533                 *val = get_reg_val(id, vcpu->arch.vrsave_tm);
1534                 break;
1535         case KVM_REG_PPC_TM_VSCR:
1536                 if (cpu_has_feature(CPU_FTR_ALTIVEC))
1537                         *val = get_reg_val(id, vcpu->arch.vr_tm.vscr.u[3]);
1538                 else
1539                         r = -ENXIO;
1540                 break;
1541         case KVM_REG_PPC_TM_DSCR:
1542                 *val = get_reg_val(id, vcpu->arch.dscr_tm);
1543                 break;
1544         case KVM_REG_PPC_TM_TAR:
1545                 *val = get_reg_val(id, vcpu->arch.tar_tm);
1546                 break;
1547 #endif
1548         case KVM_REG_PPC_ARCH_COMPAT:
1549                 *val = get_reg_val(id, vcpu->arch.vcore->arch_compat);
1550                 break;
1551         case KVM_REG_PPC_DEC_EXPIRY:
1552                 *val = get_reg_val(id, vcpu->arch.dec_expires +
1553                                    vcpu->arch.vcore->tb_offset);
1554                 break;
1555         case KVM_REG_PPC_ONLINE:
1556                 *val = get_reg_val(id, vcpu->arch.online);
1557                 break;
1558         default:
1559                 r = -EINVAL;
1560                 break;
1561         }
1562
1563         return r;
1564 }
1565
1566 static int kvmppc_set_one_reg_hv(struct kvm_vcpu *vcpu, u64 id,
1567                                  union kvmppc_one_reg *val)
1568 {
1569         int r = 0;
1570         long int i;
1571         unsigned long addr, len;
1572
1573         switch (id) {
1574         case KVM_REG_PPC_HIOR:
1575                 /* Only allow this to be set to zero */
1576                 if (set_reg_val(id, *val))
1577                         r = -EINVAL;
1578                 break;
1579         case KVM_REG_PPC_DABR:
1580                 vcpu->arch.dabr = set_reg_val(id, *val);
1581                 break;
1582         case KVM_REG_PPC_DABRX:
1583                 vcpu->arch.dabrx = set_reg_val(id, *val) & ~DABRX_HYP;
1584                 break;
1585         case KVM_REG_PPC_DSCR:
1586                 vcpu->arch.dscr = set_reg_val(id, *val);
1587                 break;
1588         case KVM_REG_PPC_PURR:
1589                 vcpu->arch.purr = set_reg_val(id, *val);
1590                 break;
1591         case KVM_REG_PPC_SPURR:
1592                 vcpu->arch.spurr = set_reg_val(id, *val);
1593                 break;
1594         case KVM_REG_PPC_AMR:
1595                 vcpu->arch.amr = set_reg_val(id, *val);
1596                 break;
1597         case KVM_REG_PPC_UAMOR:
1598                 vcpu->arch.uamor = set_reg_val(id, *val);
1599                 break;
1600         case KVM_REG_PPC_MMCR0 ... KVM_REG_PPC_MMCRS:
1601                 i = id - KVM_REG_PPC_MMCR0;
1602                 vcpu->arch.mmcr[i] = set_reg_val(id, *val);
1603                 break;
1604         case KVM_REG_PPC_PMC1 ... KVM_REG_PPC_PMC8:
1605                 i = id - KVM_REG_PPC_PMC1;
1606                 vcpu->arch.pmc[i] = set_reg_val(id, *val);
1607                 break;
1608         case KVM_REG_PPC_SPMC1 ... KVM_REG_PPC_SPMC2:
1609                 i = id - KVM_REG_PPC_SPMC1;
1610                 vcpu->arch.spmc[i] = set_reg_val(id, *val);
1611                 break;
1612         case KVM_REG_PPC_SIAR:
1613                 vcpu->arch.siar = set_reg_val(id, *val);
1614                 break;
1615         case KVM_REG_PPC_SDAR:
1616                 vcpu->arch.sdar = set_reg_val(id, *val);
1617                 break;
1618         case KVM_REG_PPC_SIER:
1619                 vcpu->arch.sier = set_reg_val(id, *val);
1620                 break;
1621         case KVM_REG_PPC_IAMR:
1622                 vcpu->arch.iamr = set_reg_val(id, *val);
1623                 break;
1624         case KVM_REG_PPC_PSPB:
1625                 vcpu->arch.pspb = set_reg_val(id, *val);
1626                 break;
1627         case KVM_REG_PPC_DPDES:
1628                 vcpu->arch.vcore->dpdes = set_reg_val(id, *val);
1629                 break;
1630         case KVM_REG_PPC_VTB:
1631                 vcpu->arch.vcore->vtb = set_reg_val(id, *val);
1632                 break;
1633         case KVM_REG_PPC_DAWR:
1634                 vcpu->arch.dawr = set_reg_val(id, *val);
1635                 break;
1636         case KVM_REG_PPC_DAWRX:
1637                 vcpu->arch.dawrx = set_reg_val(id, *val) & ~DAWRX_HYP;
1638                 break;
1639         case KVM_REG_PPC_CIABR:
1640                 vcpu->arch.ciabr = set_reg_val(id, *val);
1641                 /* Don't allow setting breakpoints in hypervisor code */
1642                 if ((vcpu->arch.ciabr & CIABR_PRIV) == CIABR_PRIV_HYPER)
1643                         vcpu->arch.ciabr &= ~CIABR_PRIV;        /* disable */
1644                 break;
1645         case KVM_REG_PPC_CSIGR:
1646                 vcpu->arch.csigr = set_reg_val(id, *val);
1647                 break;
1648         case KVM_REG_PPC_TACR:
1649                 vcpu->arch.tacr = set_reg_val(id, *val);
1650                 break;
1651         case KVM_REG_PPC_TCSCR:
1652                 vcpu->arch.tcscr = set_reg_val(id, *val);
1653                 break;
1654         case KVM_REG_PPC_PID:
1655                 vcpu->arch.pid = set_reg_val(id, *val);
1656                 break;
1657         case KVM_REG_PPC_ACOP:
1658                 vcpu->arch.acop = set_reg_val(id, *val);
1659                 break;
1660         case KVM_REG_PPC_WORT:
1661                 vcpu->arch.wort = set_reg_val(id, *val);
1662                 break;
1663         case KVM_REG_PPC_TIDR:
1664                 vcpu->arch.tid = set_reg_val(id, *val);
1665                 break;
1666         case KVM_REG_PPC_PSSCR:
1667                 vcpu->arch.psscr = set_reg_val(id, *val) & PSSCR_GUEST_VIS;
1668                 break;
1669         case KVM_REG_PPC_VPA_ADDR:
1670                 addr = set_reg_val(id, *val);
1671                 r = -EINVAL;
1672                 if (!addr && (vcpu->arch.slb_shadow.next_gpa ||
1673                               vcpu->arch.dtl.next_gpa))
1674                         break;
1675                 r = set_vpa(vcpu, &vcpu->arch.vpa, addr, sizeof(struct lppaca));
1676                 break;
1677         case KVM_REG_PPC_VPA_SLB:
1678                 addr = val->vpaval.addr;
1679                 len = val->vpaval.length;
1680                 r = -EINVAL;
1681                 if (addr && !vcpu->arch.vpa.next_gpa)
1682                         break;
1683                 r = set_vpa(vcpu, &vcpu->arch.slb_shadow, addr, len);
1684                 break;
1685         case KVM_REG_PPC_VPA_DTL:
1686                 addr = val->vpaval.addr;
1687                 len = val->vpaval.length;
1688                 r = -EINVAL;
1689                 if (addr && (len < sizeof(struct dtl_entry) ||
1690                              !vcpu->arch.vpa.next_gpa))
1691                         break;
1692                 len -= len % sizeof(struct dtl_entry);
1693                 r = set_vpa(vcpu, &vcpu->arch.dtl, addr, len);
1694                 break;
1695         case KVM_REG_PPC_TB_OFFSET:
1696                 /* round up to multiple of 2^24 */
1697                 vcpu->arch.vcore->tb_offset =
1698                         ALIGN(set_reg_val(id, *val), 1UL << 24);
1699                 break;
1700         case KVM_REG_PPC_LPCR:
1701                 kvmppc_set_lpcr(vcpu, set_reg_val(id, *val), true);
1702                 break;
1703         case KVM_REG_PPC_LPCR_64:
1704                 kvmppc_set_lpcr(vcpu, set_reg_val(id, *val), false);
1705                 break;
1706         case KVM_REG_PPC_PPR:
1707                 vcpu->arch.ppr = set_reg_val(id, *val);
1708                 break;
1709 #ifdef CONFIG_PPC_TRANSACTIONAL_MEM
1710         case KVM_REG_PPC_TFHAR:
1711                 vcpu->arch.tfhar = set_reg_val(id, *val);
1712                 break;
1713         case KVM_REG_PPC_TFIAR:
1714                 vcpu->arch.tfiar = set_reg_val(id, *val);
1715                 break;
1716         case KVM_REG_PPC_TEXASR:
1717                 vcpu->arch.texasr = set_reg_val(id, *val);
1718                 break;
1719         case KVM_REG_PPC_TM_GPR0 ... KVM_REG_PPC_TM_GPR31:
1720                 i = id - KVM_REG_PPC_TM_GPR0;
1721                 vcpu->arch.gpr_tm[i] = set_reg_val(id, *val);
1722                 break;
1723         case KVM_REG_PPC_TM_VSR0 ... KVM_REG_PPC_TM_VSR63:
1724         {
1725                 int j;
1726                 i = id - KVM_REG_PPC_TM_VSR0;
1727                 if (i < 32)
1728                         for (j = 0; j < TS_FPRWIDTH; j++)
1729                                 vcpu->arch.fp_tm.fpr[i][j] = val->vsxval[j];
1730                 else
1731                         if (cpu_has_feature(CPU_FTR_ALTIVEC))
1732                                 vcpu->arch.vr_tm.vr[i-32] = val->vval;
1733                         else
1734                                 r = -ENXIO;
1735                 break;
1736         }
1737         case KVM_REG_PPC_TM_CR:
1738                 vcpu->arch.cr_tm = set_reg_val(id, *val);
1739                 break;
1740         case KVM_REG_PPC_TM_XER:
1741                 vcpu->arch.xer_tm = set_reg_val(id, *val);
1742                 break;
1743         case KVM_REG_PPC_TM_LR:
1744                 vcpu->arch.lr_tm = set_reg_val(id, *val);
1745                 break;
1746         case KVM_REG_PPC_TM_CTR:
1747                 vcpu->arch.ctr_tm = set_reg_val(id, *val);
1748                 break;
1749         case KVM_REG_PPC_TM_FPSCR:
1750                 vcpu->arch.fp_tm.fpscr = set_reg_val(id, *val);
1751                 break;
1752         case KVM_REG_PPC_TM_AMR:
1753                 vcpu->arch.amr_tm = set_reg_val(id, *val);
1754                 break;
1755         case KVM_REG_PPC_TM_PPR:
1756                 vcpu->arch.ppr_tm = set_reg_val(id, *val);
1757                 break;
1758         case KVM_REG_PPC_TM_VRSAVE:
1759                 vcpu->arch.vrsave_tm = set_reg_val(id, *val);
1760                 break;
1761         case KVM_REG_PPC_TM_VSCR:
1762                 if (cpu_has_feature(CPU_FTR_ALTIVEC))
1763                         vcpu->arch.vr.vscr.u[3] = set_reg_val(id, *val);
1764                 else
1765                         r = - ENXIO;
1766                 break;
1767         case KVM_REG_PPC_TM_DSCR:
1768                 vcpu->arch.dscr_tm = set_reg_val(id, *val);
1769                 break;
1770         case KVM_REG_PPC_TM_TAR:
1771                 vcpu->arch.tar_tm = set_reg_val(id, *val);
1772                 break;
1773 #endif
1774         case KVM_REG_PPC_ARCH_COMPAT:
1775                 r = kvmppc_set_arch_compat(vcpu, set_reg_val(id, *val));
1776                 break;
1777         case KVM_REG_PPC_DEC_EXPIRY:
1778                 vcpu->arch.dec_expires = set_reg_val(id, *val) -
1779                         vcpu->arch.vcore->tb_offset;
1780                 break;
1781         case KVM_REG_PPC_ONLINE:
1782                 i = set_reg_val(id, *val);
1783                 if (i && !vcpu->arch.online)
1784                         atomic_inc(&vcpu->arch.vcore->online_count);
1785                 else if (!i && vcpu->arch.online)
1786                         atomic_dec(&vcpu->arch.vcore->online_count);
1787                 vcpu->arch.online = i;
1788                 break;
1789         default:
1790                 r = -EINVAL;
1791                 break;
1792         }
1793
1794         return r;
1795 }
1796
1797 /*
1798  * On POWER9, threads are independent and can be in different partitions.
1799  * Therefore we consider each thread to be a subcore.
1800  * There is a restriction that all threads have to be in the same
1801  * MMU mode (radix or HPT), unfortunately, but since we only support
1802  * HPT guests on a HPT host so far, that isn't an impediment yet.
1803  */
1804 static int threads_per_vcore(struct kvm *kvm)
1805 {
1806         if (kvm->arch.threads_indep)
1807                 return 1;
1808         return threads_per_subcore;
1809 }
1810
1811 static struct kvmppc_vcore *kvmppc_vcore_create(struct kvm *kvm, int id)
1812 {
1813         struct kvmppc_vcore *vcore;
1814
1815         vcore = kzalloc(sizeof(struct kvmppc_vcore), GFP_KERNEL);
1816
1817         if (vcore == NULL)
1818                 return NULL;
1819
1820         spin_lock_init(&vcore->lock);
1821         spin_lock_init(&vcore->stoltb_lock);
1822         init_swait_queue_head(&vcore->wq);
1823         vcore->preempt_tb = TB_NIL;
1824         vcore->lpcr = kvm->arch.lpcr;
1825         vcore->first_vcpuid = id;
1826         vcore->kvm = kvm;
1827         INIT_LIST_HEAD(&vcore->preempt_list);
1828
1829         return vcore;
1830 }
1831
1832 #ifdef CONFIG_KVM_BOOK3S_HV_EXIT_TIMING
1833 static struct debugfs_timings_element {
1834         const char *name;
1835         size_t offset;
1836 } timings[] = {
1837         {"rm_entry",    offsetof(struct kvm_vcpu, arch.rm_entry)},
1838         {"rm_intr",     offsetof(struct kvm_vcpu, arch.rm_intr)},
1839         {"rm_exit",     offsetof(struct kvm_vcpu, arch.rm_exit)},
1840         {"guest",       offsetof(struct kvm_vcpu, arch.guest_time)},
1841         {"cede",        offsetof(struct kvm_vcpu, arch.cede_time)},
1842 };
1843
1844 #define N_TIMINGS       (ARRAY_SIZE(timings))
1845
1846 struct debugfs_timings_state {
1847         struct kvm_vcpu *vcpu;
1848         unsigned int    buflen;
1849         char            buf[N_TIMINGS * 100];
1850 };
1851
1852 static int debugfs_timings_open(struct inode *inode, struct file *file)
1853 {
1854         struct kvm_vcpu *vcpu = inode->i_private;
1855         struct debugfs_timings_state *p;
1856
1857         p = kzalloc(sizeof(*p), GFP_KERNEL);
1858         if (!p)
1859                 return -ENOMEM;
1860
1861         kvm_get_kvm(vcpu->kvm);
1862         p->vcpu = vcpu;
1863         file->private_data = p;
1864
1865         return nonseekable_open(inode, file);
1866 }
1867
1868 static int debugfs_timings_release(struct inode *inode, struct file *file)
1869 {
1870         struct debugfs_timings_state *p = file->private_data;
1871
1872         kvm_put_kvm(p->vcpu->kvm);
1873         kfree(p);
1874         return 0;
1875 }
1876
1877 static ssize_t debugfs_timings_read(struct file *file, char __user *buf,
1878                                     size_t len, loff_t *ppos)
1879 {
1880         struct debugfs_timings_state *p = file->private_data;
1881         struct kvm_vcpu *vcpu = p->vcpu;
1882         char *s, *buf_end;
1883         struct kvmhv_tb_accumulator tb;
1884         u64 count;
1885         loff_t pos;
1886         ssize_t n;
1887         int i, loops;
1888         bool ok;
1889
1890         if (!p->buflen) {
1891                 s = p->buf;
1892                 buf_end = s + sizeof(p->buf);
1893                 for (i = 0; i < N_TIMINGS; ++i) {
1894                         struct kvmhv_tb_accumulator *acc;
1895
1896                         acc = (struct kvmhv_tb_accumulator *)
1897                                 ((unsigned long)vcpu + timings[i].offset);
1898                         ok = false;
1899                         for (loops = 0; loops < 1000; ++loops) {
1900                                 count = acc->seqcount;
1901                                 if (!(count & 1)) {
1902                                         smp_rmb();
1903                                         tb = *acc;
1904                                         smp_rmb();
1905                                         if (count == acc->seqcount) {
1906                                                 ok = true;
1907                                                 break;
1908                                         }
1909                                 }
1910                                 udelay(1);
1911                         }
1912                         if (!ok)
1913                                 snprintf(s, buf_end - s, "%s: stuck\n",
1914                                         timings[i].name);
1915                         else
1916                                 snprintf(s, buf_end - s,
1917                                         "%s: %llu %llu %llu %llu\n",
1918                                         timings[i].name, count / 2,
1919                                         tb_to_ns(tb.tb_total),
1920                                         tb_to_ns(tb.tb_min),
1921                                         tb_to_ns(tb.tb_max));
1922                         s += strlen(s);
1923                 }
1924                 p->buflen = s - p->buf;
1925         }
1926
1927         pos = *ppos;
1928         if (pos >= p->buflen)
1929                 return 0;
1930         if (len > p->buflen - pos)
1931                 len = p->buflen - pos;
1932         n = copy_to_user(buf, p->buf + pos, len);
1933         if (n) {
1934                 if (n == len)
1935                         return -EFAULT;
1936                 len -= n;
1937         }
1938         *ppos = pos + len;
1939         return len;
1940 }
1941
1942 static ssize_t debugfs_timings_write(struct file *file, const char __user *buf,
1943                                      size_t len, loff_t *ppos)
1944 {
1945         return -EACCES;
1946 }
1947
1948 static const struct file_operations debugfs_timings_ops = {
1949         .owner   = THIS_MODULE,
1950         .open    = debugfs_timings_open,
1951         .release = debugfs_timings_release,
1952         .read    = debugfs_timings_read,
1953         .write   = debugfs_timings_write,
1954         .llseek  = generic_file_llseek,
1955 };
1956
1957 /* Create a debugfs directory for the vcpu */
1958 static void debugfs_vcpu_init(struct kvm_vcpu *vcpu, unsigned int id)
1959 {
1960         char buf[16];
1961         struct kvm *kvm = vcpu->kvm;
1962
1963         snprintf(buf, sizeof(buf), "vcpu%u", id);
1964         if (IS_ERR_OR_NULL(kvm->arch.debugfs_dir))
1965                 return;
1966         vcpu->arch.debugfs_dir = debugfs_create_dir(buf, kvm->arch.debugfs_dir);
1967         if (IS_ERR_OR_NULL(vcpu->arch.debugfs_dir))
1968                 return;
1969         vcpu->arch.debugfs_timings =
1970                 debugfs_create_file("timings", 0444, vcpu->arch.debugfs_dir,
1971                                     vcpu, &debugfs_timings_ops);
1972 }
1973
1974 #else /* CONFIG_KVM_BOOK3S_HV_EXIT_TIMING */
1975 static void debugfs_vcpu_init(struct kvm_vcpu *vcpu, unsigned int id)
1976 {
1977 }
1978 #endif /* CONFIG_KVM_BOOK3S_HV_EXIT_TIMING */
1979
1980 static struct kvm_vcpu *kvmppc_core_vcpu_create_hv(struct kvm *kvm,
1981                                                    unsigned int id)
1982 {
1983         struct kvm_vcpu *vcpu;
1984         int err;
1985         int core;
1986         struct kvmppc_vcore *vcore;
1987
1988         err = -ENOMEM;
1989         vcpu = kmem_cache_zalloc(kvm_vcpu_cache, GFP_KERNEL);
1990         if (!vcpu)
1991                 goto out;
1992
1993         err = kvm_vcpu_init(vcpu, kvm, id);
1994         if (err)
1995                 goto free_vcpu;
1996
1997         vcpu->arch.shared = &vcpu->arch.shregs;
1998 #ifdef CONFIG_KVM_BOOK3S_PR_POSSIBLE
1999         /*
2000          * The shared struct is never shared on HV,
2001          * so we can always use host endianness
2002          */
2003 #ifdef __BIG_ENDIAN__
2004         vcpu->arch.shared_big_endian = true;
2005 #else
2006         vcpu->arch.shared_big_endian = false;
2007 #endif
2008 #endif
2009         vcpu->arch.mmcr[0] = MMCR0_FC;
2010         vcpu->arch.ctrl = CTRL_RUNLATCH;
2011         /* default to host PVR, since we can't spoof it */
2012         kvmppc_set_pvr_hv(vcpu, mfspr(SPRN_PVR));
2013         spin_lock_init(&vcpu->arch.vpa_update_lock);
2014         spin_lock_init(&vcpu->arch.tbacct_lock);
2015         vcpu->arch.busy_preempt = TB_NIL;
2016         vcpu->arch.intr_msr = MSR_SF | MSR_ME;
2017
2018         /*
2019          * Set the default HFSCR for the guest from the host value.
2020          * This value is only used on POWER9.
2021          * On POWER9, we want to virtualize the doorbell facility, so we
2022          * turn off the HFSCR bit, which causes those instructions to trap.
2023          */
2024         vcpu->arch.hfscr = mfspr(SPRN_HFSCR);
2025         if (cpu_has_feature(CPU_FTR_P9_TM_HV_ASSIST))
2026                 vcpu->arch.hfscr |= HFSCR_TM;
2027         else if (!cpu_has_feature(CPU_FTR_TM_COMP))
2028                 vcpu->arch.hfscr &= ~HFSCR_TM;
2029         if (cpu_has_feature(CPU_FTR_ARCH_300))
2030                 vcpu->arch.hfscr &= ~HFSCR_MSGP;
2031
2032         kvmppc_mmu_book3s_hv_init(vcpu);
2033
2034         vcpu->arch.state = KVMPPC_VCPU_NOTREADY;
2035
2036         init_waitqueue_head(&vcpu->arch.cpu_run);
2037
2038         mutex_lock(&kvm->lock);
2039         vcore = NULL;
2040         err = -EINVAL;
2041         if (cpu_has_feature(CPU_FTR_ARCH_300)) {
2042                 if (id >= (KVM_MAX_VCPUS * kvm->arch.emul_smt_mode)) {
2043                         pr_devel("KVM: VCPU ID too high\n");
2044                         core = KVM_MAX_VCORES;
2045                 } else {
2046                         BUG_ON(kvm->arch.smt_mode != 1);
2047                         core = kvmppc_pack_vcpu_id(kvm, id);
2048                 }
2049         } else {
2050                 core = id / kvm->arch.smt_mode;
2051         }
2052         if (core < KVM_MAX_VCORES) {
2053                 vcore = kvm->arch.vcores[core];
2054                 if (vcore && cpu_has_feature(CPU_FTR_ARCH_300)) {
2055                         pr_devel("KVM: collision on id %u", id);
2056                         vcore = NULL;
2057                 } else if (!vcore) {
2058                         err = -ENOMEM;
2059                         vcore = kvmppc_vcore_create(kvm,
2060                                         id & ~(kvm->arch.smt_mode - 1));
2061                         kvm->arch.vcores[core] = vcore;
2062                         kvm->arch.online_vcores++;
2063                 }
2064         }
2065         mutex_unlock(&kvm->lock);
2066
2067         if (!vcore)
2068                 goto free_vcpu;
2069
2070         spin_lock(&vcore->lock);
2071         ++vcore->num_threads;
2072         spin_unlock(&vcore->lock);
2073         vcpu->arch.vcore = vcore;
2074         vcpu->arch.ptid = vcpu->vcpu_id - vcore->first_vcpuid;
2075         vcpu->arch.thread_cpu = -1;
2076         vcpu->arch.prev_cpu = -1;
2077
2078         vcpu->arch.cpu_type = KVM_CPU_3S_64;
2079         kvmppc_sanity_check(vcpu);
2080
2081         debugfs_vcpu_init(vcpu, id);
2082
2083         return vcpu;
2084
2085 free_vcpu:
2086         kmem_cache_free(kvm_vcpu_cache, vcpu);
2087 out:
2088         return ERR_PTR(err);
2089 }
2090
2091 static int kvmhv_set_smt_mode(struct kvm *kvm, unsigned long smt_mode,
2092                               unsigned long flags)
2093 {
2094         int err;
2095         int esmt = 0;
2096
2097         if (flags)
2098                 return -EINVAL;
2099         if (smt_mode > MAX_SMT_THREADS || !is_power_of_2(smt_mode))
2100                 return -EINVAL;
2101         if (!cpu_has_feature(CPU_FTR_ARCH_300)) {
2102                 /*
2103                  * On POWER8 (or POWER7), the threading mode is "strict",
2104                  * so we pack smt_mode vcpus per vcore.
2105                  */
2106                 if (smt_mode > threads_per_subcore)
2107                         return -EINVAL;
2108         } else {
2109                 /*
2110                  * On POWER9, the threading mode is "loose",
2111                  * so each vcpu gets its own vcore.
2112                  */
2113                 esmt = smt_mode;
2114                 smt_mode = 1;
2115         }
2116         mutex_lock(&kvm->lock);
2117         err = -EBUSY;
2118         if (!kvm->arch.online_vcores) {
2119                 kvm->arch.smt_mode = smt_mode;
2120                 kvm->arch.emul_smt_mode = esmt;
2121                 err = 0;
2122         }
2123         mutex_unlock(&kvm->lock);
2124
2125         return err;
2126 }
2127
2128 static void unpin_vpa(struct kvm *kvm, struct kvmppc_vpa *vpa)
2129 {
2130         if (vpa->pinned_addr)
2131                 kvmppc_unpin_guest_page(kvm, vpa->pinned_addr, vpa->gpa,
2132                                         vpa->dirty);
2133 }
2134
2135 static void kvmppc_core_vcpu_free_hv(struct kvm_vcpu *vcpu)
2136 {
2137         spin_lock(&vcpu->arch.vpa_update_lock);
2138         unpin_vpa(vcpu->kvm, &vcpu->arch.dtl);
2139         unpin_vpa(vcpu->kvm, &vcpu->arch.slb_shadow);
2140         unpin_vpa(vcpu->kvm, &vcpu->arch.vpa);
2141         spin_unlock(&vcpu->arch.vpa_update_lock);
2142         kvm_vcpu_uninit(vcpu);
2143         kmem_cache_free(kvm_vcpu_cache, vcpu);
2144 }
2145
2146 static int kvmppc_core_check_requests_hv(struct kvm_vcpu *vcpu)
2147 {
2148         /* Indicate we want to get back into the guest */
2149         return 1;
2150 }
2151
2152 static void kvmppc_set_timer(struct kvm_vcpu *vcpu)
2153 {
2154         unsigned long dec_nsec, now;
2155
2156         now = get_tb();
2157         if (now > vcpu->arch.dec_expires) {
2158                 /* decrementer has already gone negative */
2159                 kvmppc_core_queue_dec(vcpu);
2160                 kvmppc_core_prepare_to_enter(vcpu);
2161                 return;
2162         }
2163         dec_nsec = (vcpu->arch.dec_expires - now) * NSEC_PER_SEC
2164                    / tb_ticks_per_sec;
2165         hrtimer_start(&vcpu->arch.dec_timer, dec_nsec, HRTIMER_MODE_REL);
2166         vcpu->arch.timer_running = 1;
2167 }
2168
2169 static void kvmppc_end_cede(struct kvm_vcpu *vcpu)
2170 {
2171         vcpu->arch.ceded = 0;
2172         if (vcpu->arch.timer_running) {
2173                 hrtimer_try_to_cancel(&vcpu->arch.dec_timer);
2174                 vcpu->arch.timer_running = 0;
2175         }
2176 }
2177
2178 extern int __kvmppc_vcore_entry(void);
2179
2180 static void kvmppc_remove_runnable(struct kvmppc_vcore *vc,
2181                                    struct kvm_vcpu *vcpu)
2182 {
2183         u64 now;
2184
2185         if (vcpu->arch.state != KVMPPC_VCPU_RUNNABLE)
2186                 return;
2187         spin_lock_irq(&vcpu->arch.tbacct_lock);
2188         now = mftb();
2189         vcpu->arch.busy_stolen += vcore_stolen_time(vc, now) -
2190                 vcpu->arch.stolen_logged;
2191         vcpu->arch.busy_preempt = now;
2192         vcpu->arch.state = KVMPPC_VCPU_BUSY_IN_HOST;
2193         spin_unlock_irq(&vcpu->arch.tbacct_lock);
2194         --vc->n_runnable;
2195         WRITE_ONCE(vc->runnable_threads[vcpu->arch.ptid], NULL);
2196 }
2197
2198 static int kvmppc_grab_hwthread(int cpu)
2199 {
2200         struct paca_struct *tpaca;
2201         long timeout = 10000;
2202
2203         tpaca = paca_ptrs[cpu];
2204
2205         /* Ensure the thread won't go into the kernel if it wakes */
2206         tpaca->kvm_hstate.kvm_vcpu = NULL;
2207         tpaca->kvm_hstate.kvm_vcore = NULL;
2208         tpaca->kvm_hstate.napping = 0;
2209         smp_wmb();
2210         tpaca->kvm_hstate.hwthread_req = 1;
2211
2212         /*
2213          * If the thread is already executing in the kernel (e.g. handling
2214          * a stray interrupt), wait for it to get back to nap mode.
2215          * The smp_mb() is to ensure that our setting of hwthread_req
2216          * is visible before we look at hwthread_state, so if this
2217          * races with the code at system_reset_pSeries and the thread
2218          * misses our setting of hwthread_req, we are sure to see its
2219          * setting of hwthread_state, and vice versa.
2220          */
2221         smp_mb();
2222         while (tpaca->kvm_hstate.hwthread_state == KVM_HWTHREAD_IN_KERNEL) {
2223                 if (--timeout <= 0) {
2224                         pr_err("KVM: couldn't grab cpu %d\n", cpu);
2225                         return -EBUSY;
2226                 }
2227                 udelay(1);
2228         }
2229         return 0;
2230 }
2231
2232 static void kvmppc_release_hwthread(int cpu)
2233 {
2234         struct paca_struct *tpaca;
2235
2236         tpaca = paca_ptrs[cpu];
2237         tpaca->kvm_hstate.hwthread_req = 0;
2238         tpaca->kvm_hstate.kvm_vcpu = NULL;
2239         tpaca->kvm_hstate.kvm_vcore = NULL;
2240         tpaca->kvm_hstate.kvm_split_mode = NULL;
2241 }
2242
2243 static void radix_flush_cpu(struct kvm *kvm, int cpu, struct kvm_vcpu *vcpu)
2244 {
2245         int i;
2246
2247         cpu = cpu_first_thread_sibling(cpu);
2248         cpumask_set_cpu(cpu, &kvm->arch.need_tlb_flush);
2249         /*
2250          * Make sure setting of bit in need_tlb_flush precedes
2251          * testing of cpu_in_guest bits.  The matching barrier on
2252          * the other side is the first smp_mb() in kvmppc_run_core().
2253          */
2254         smp_mb();
2255         for (i = 0; i < threads_per_core; ++i)
2256                 if (cpumask_test_cpu(cpu + i, &kvm->arch.cpu_in_guest))
2257                         smp_call_function_single(cpu + i, do_nothing, NULL, 1);
2258 }
2259
2260 static void kvmppc_prepare_radix_vcpu(struct kvm_vcpu *vcpu, int pcpu)
2261 {
2262         struct kvm *kvm = vcpu->kvm;
2263
2264         /*
2265          * With radix, the guest can do TLB invalidations itself,
2266          * and it could choose to use the local form (tlbiel) if
2267          * it is invalidating a translation that has only ever been
2268          * used on one vcpu.  However, that doesn't mean it has
2269          * only ever been used on one physical cpu, since vcpus
2270          * can move around between pcpus.  To cope with this, when
2271          * a vcpu moves from one pcpu to another, we need to tell
2272          * any vcpus running on the same core as this vcpu previously
2273          * ran to flush the TLB.  The TLB is shared between threads,
2274          * so we use a single bit in .need_tlb_flush for all 4 threads.
2275          */
2276         if (vcpu->arch.prev_cpu != pcpu) {
2277                 if (vcpu->arch.prev_cpu >= 0 &&
2278                     cpu_first_thread_sibling(vcpu->arch.prev_cpu) !=
2279                     cpu_first_thread_sibling(pcpu))
2280                         radix_flush_cpu(kvm, vcpu->arch.prev_cpu, vcpu);
2281                 vcpu->arch.prev_cpu = pcpu;
2282         }
2283 }
2284
2285 static void kvmppc_start_thread(struct kvm_vcpu *vcpu, struct kvmppc_vcore *vc)
2286 {
2287         int cpu;
2288         struct paca_struct *tpaca;
2289         struct kvm *kvm = vc->kvm;
2290
2291         cpu = vc->pcpu;
2292         if (vcpu) {
2293                 if (vcpu->arch.timer_running) {
2294                         hrtimer_try_to_cancel(&vcpu->arch.dec_timer);
2295                         vcpu->arch.timer_running = 0;
2296                 }
2297                 cpu += vcpu->arch.ptid;
2298                 vcpu->cpu = vc->pcpu;
2299                 vcpu->arch.thread_cpu = cpu;
2300                 cpumask_set_cpu(cpu, &kvm->arch.cpu_in_guest);
2301         }
2302         tpaca = paca_ptrs[cpu];
2303         tpaca->kvm_hstate.kvm_vcpu = vcpu;
2304         tpaca->kvm_hstate.ptid = cpu - vc->pcpu;
2305         tpaca->kvm_hstate.fake_suspend = 0;
2306         /* Order stores to hstate.kvm_vcpu etc. before store to kvm_vcore */
2307         smp_wmb();
2308         tpaca->kvm_hstate.kvm_vcore = vc;
2309         if (cpu != smp_processor_id())
2310                 kvmppc_ipi_thread(cpu);
2311 }
2312
2313 static void kvmppc_wait_for_nap(int n_threads)
2314 {
2315         int cpu = smp_processor_id();
2316         int i, loops;
2317
2318         if (n_threads <= 1)
2319                 return;
2320         for (loops = 0; loops < 1000000; ++loops) {
2321                 /*
2322                  * Check if all threads are finished.
2323                  * We set the vcore pointer when starting a thread
2324                  * and the thread clears it when finished, so we look
2325                  * for any threads that still have a non-NULL vcore ptr.
2326                  */
2327                 for (i = 1; i < n_threads; ++i)
2328                         if (paca_ptrs[cpu + i]->kvm_hstate.kvm_vcore)
2329                                 break;
2330                 if (i == n_threads) {
2331                         HMT_medium();
2332                         return;
2333                 }
2334                 HMT_low();
2335         }
2336         HMT_medium();
2337         for (i = 1; i < n_threads; ++i)
2338                 if (paca_ptrs[cpu + i]->kvm_hstate.kvm_vcore)
2339                         pr_err("KVM: CPU %d seems to be stuck\n", cpu + i);
2340 }
2341
2342 /*
2343  * Check that we are on thread 0 and that any other threads in
2344  * this core are off-line.  Then grab the threads so they can't
2345  * enter the kernel.
2346  */
2347 static int on_primary_thread(void)
2348 {
2349         int cpu = smp_processor_id();
2350         int thr;
2351
2352         /* Are we on a primary subcore? */
2353         if (cpu_thread_in_subcore(cpu))
2354                 return 0;
2355
2356         thr = 0;
2357         while (++thr < threads_per_subcore)
2358                 if (cpu_online(cpu + thr))
2359                         return 0;
2360
2361         /* Grab all hw threads so they can't go into the kernel */
2362         for (thr = 1; thr < threads_per_subcore; ++thr) {
2363                 if (kvmppc_grab_hwthread(cpu + thr)) {
2364                         /* Couldn't grab one; let the others go */
2365                         do {
2366                                 kvmppc_release_hwthread(cpu + thr);
2367                         } while (--thr > 0);
2368                         return 0;
2369                 }
2370         }
2371         return 1;
2372 }
2373
2374 /*
2375  * A list of virtual cores for each physical CPU.
2376  * These are vcores that could run but their runner VCPU tasks are
2377  * (or may be) preempted.
2378  */
2379 struct preempted_vcore_list {
2380         struct list_head        list;
2381         spinlock_t              lock;
2382 };
2383
2384 static DEFINE_PER_CPU(struct preempted_vcore_list, preempted_vcores);
2385
2386 static void init_vcore_lists(void)
2387 {
2388         int cpu;
2389
2390         for_each_possible_cpu(cpu) {
2391                 struct preempted_vcore_list *lp = &per_cpu(preempted_vcores, cpu);
2392                 spin_lock_init(&lp->lock);
2393                 INIT_LIST_HEAD(&lp->list);
2394         }
2395 }
2396
2397 static void kvmppc_vcore_preempt(struct kvmppc_vcore *vc)
2398 {
2399         struct preempted_vcore_list *lp = this_cpu_ptr(&preempted_vcores);
2400
2401         vc->vcore_state = VCORE_PREEMPT;
2402         vc->pcpu = smp_processor_id();
2403         if (vc->num_threads < threads_per_vcore(vc->kvm)) {
2404                 spin_lock(&lp->lock);
2405                 list_add_tail(&vc->preempt_list, &lp->list);
2406                 spin_unlock(&lp->lock);
2407         }
2408
2409         /* Start accumulating stolen time */
2410         kvmppc_core_start_stolen(vc);
2411 }
2412
2413 static void kvmppc_vcore_end_preempt(struct kvmppc_vcore *vc)
2414 {
2415         struct preempted_vcore_list *lp;
2416
2417         kvmppc_core_end_stolen(vc);
2418         if (!list_empty(&vc->preempt_list)) {
2419                 lp = &per_cpu(preempted_vcores, vc->pcpu);
2420                 spin_lock(&lp->lock);
2421                 list_del_init(&vc->preempt_list);
2422                 spin_unlock(&lp->lock);
2423         }
2424         vc->vcore_state = VCORE_INACTIVE;
2425 }
2426
2427 /*
2428  * This stores information about the virtual cores currently
2429  * assigned to a physical core.
2430  */
2431 struct core_info {
2432         int             n_subcores;
2433         int             max_subcore_threads;
2434         int             total_threads;
2435         int             subcore_threads[MAX_SUBCORES];
2436         struct kvmppc_vcore *vc[MAX_SUBCORES];
2437 };
2438
2439 /*
2440  * This mapping means subcores 0 and 1 can use threads 0-3 and 4-7
2441  * respectively in 2-way micro-threading (split-core) mode on POWER8.
2442  */
2443 static int subcore_thread_map[MAX_SUBCORES] = { 0, 4, 2, 6 };
2444
2445 static void init_core_info(struct core_info *cip, struct kvmppc_vcore *vc)
2446 {
2447         memset(cip, 0, sizeof(*cip));
2448         cip->n_subcores = 1;
2449         cip->max_subcore_threads = vc->num_threads;
2450         cip->total_threads = vc->num_threads;
2451         cip->subcore_threads[0] = vc->num_threads;
2452         cip->vc[0] = vc;
2453 }
2454
2455 static bool subcore_config_ok(int n_subcores, int n_threads)
2456 {
2457         /*
2458          * POWER9 "SMT4" cores are permanently in what is effectively a 4-way
2459          * split-core mode, with one thread per subcore.
2460          */
2461         if (cpu_has_feature(CPU_FTR_ARCH_300))
2462                 return n_subcores <= 4 && n_threads == 1;
2463
2464         /* On POWER8, can only dynamically split if unsplit to begin with */
2465         if (n_subcores > 1 && threads_per_subcore < MAX_SMT_THREADS)
2466                 return false;
2467         if (n_subcores > MAX_SUBCORES)
2468                 return false;
2469         if (n_subcores > 1) {
2470                 if (!(dynamic_mt_modes & 2))
2471                         n_subcores = 4;
2472                 if (n_subcores > 2 && !(dynamic_mt_modes & 4))
2473                         return false;
2474         }
2475
2476         return n_subcores * roundup_pow_of_two(n_threads) <= MAX_SMT_THREADS;
2477 }
2478
2479 static void init_vcore_to_run(struct kvmppc_vcore *vc)
2480 {
2481         vc->entry_exit_map = 0;
2482         vc->in_guest = 0;
2483         vc->napping_threads = 0;
2484         vc->conferring_threads = 0;
2485         vc->tb_offset_applied = 0;
2486 }
2487
2488 static bool can_dynamic_split(struct kvmppc_vcore *vc, struct core_info *cip)
2489 {
2490         int n_threads = vc->num_threads;
2491         int sub;
2492
2493         if (!cpu_has_feature(CPU_FTR_ARCH_207S))
2494                 return false;
2495
2496         /* Some POWER9 chips require all threads to be in the same MMU mode */
2497         if (no_mixing_hpt_and_radix &&
2498             kvm_is_radix(vc->kvm) != kvm_is_radix(cip->vc[0]->kvm))
2499                 return false;
2500
2501         if (n_threads < cip->max_subcore_threads)
2502                 n_threads = cip->max_subcore_threads;
2503         if (!subcore_config_ok(cip->n_subcores + 1, n_threads))
2504                 return false;
2505         cip->max_subcore_threads = n_threads;
2506
2507         sub = cip->n_subcores;
2508         ++cip->n_subcores;
2509         cip->total_threads += vc->num_threads;
2510         cip->subcore_threads[sub] = vc->num_threads;
2511         cip->vc[sub] = vc;
2512         init_vcore_to_run(vc);
2513         list_del_init(&vc->preempt_list);
2514
2515         return true;
2516 }
2517
2518 /*
2519  * Work out whether it is possible to piggyback the execution of
2520  * vcore *pvc onto the execution of the other vcores described in *cip.
2521  */
2522 static bool can_piggyback(struct kvmppc_vcore *pvc, struct core_info *cip,
2523                           int target_threads)
2524 {
2525         if (cip->total_threads + pvc->num_threads > target_threads)
2526                 return false;
2527
2528         return can_dynamic_split(pvc, cip);
2529 }
2530
2531 static void prepare_threads(struct kvmppc_vcore *vc)
2532 {
2533         int i;
2534         struct kvm_vcpu *vcpu;
2535
2536         for_each_runnable_thread(i, vcpu, vc) {
2537                 if (signal_pending(vcpu->arch.run_task))
2538                         vcpu->arch.ret = -EINTR;
2539                 else if (vcpu->arch.vpa.update_pending ||
2540                          vcpu->arch.slb_shadow.update_pending ||
2541                          vcpu->arch.dtl.update_pending)
2542                         vcpu->arch.ret = RESUME_GUEST;
2543                 else
2544                         continue;
2545                 kvmppc_remove_runnable(vc, vcpu);
2546                 wake_up(&vcpu->arch.cpu_run);
2547         }
2548 }
2549
2550 static void collect_piggybacks(struct core_info *cip, int target_threads)
2551 {
2552         struct preempted_vcore_list *lp = this_cpu_ptr(&preempted_vcores);
2553         struct kvmppc_vcore *pvc, *vcnext;
2554
2555         spin_lock(&lp->lock);
2556         list_for_each_entry_safe(pvc, vcnext, &lp->list, preempt_list) {
2557                 if (!spin_trylock(&pvc->lock))
2558                         continue;
2559                 prepare_threads(pvc);
2560                 if (!pvc->n_runnable) {
2561                         list_del_init(&pvc->preempt_list);
2562                         if (pvc->runner == NULL) {
2563                                 pvc->vcore_state = VCORE_INACTIVE;
2564                                 kvmppc_core_end_stolen(pvc);
2565                         }
2566                         spin_unlock(&pvc->lock);
2567                         continue;
2568                 }
2569                 if (!can_piggyback(pvc, cip, target_threads)) {
2570                         spin_unlock(&pvc->lock);
2571                         continue;
2572                 }
2573                 kvmppc_core_end_stolen(pvc);
2574                 pvc->vcore_state = VCORE_PIGGYBACK;
2575                 if (cip->total_threads >= target_threads)
2576                         break;
2577         }
2578         spin_unlock(&lp->lock);
2579 }
2580
2581 static bool recheck_signals(struct core_info *cip)
2582 {
2583         int sub, i;
2584         struct kvm_vcpu *vcpu;
2585
2586         for (sub = 0; sub < cip->n_subcores; ++sub)
2587                 for_each_runnable_thread(i, vcpu, cip->vc[sub])
2588                         if (signal_pending(vcpu->arch.run_task))
2589                                 return true;
2590         return false;
2591 }
2592
2593 static void post_guest_process(struct kvmppc_vcore *vc, bool is_master)
2594 {
2595         int still_running = 0, i;
2596         u64 now;
2597         long ret;
2598         struct kvm_vcpu *vcpu;
2599
2600         spin_lock(&vc->lock);
2601         now = get_tb();
2602         for_each_runnable_thread(i, vcpu, vc) {
2603                 /* cancel pending dec exception if dec is positive */
2604                 if (now < vcpu->arch.dec_expires &&
2605                     kvmppc_core_pending_dec(vcpu))
2606                         kvmppc_core_dequeue_dec(vcpu);
2607
2608                 trace_kvm_guest_exit(vcpu);
2609
2610                 ret = RESUME_GUEST;
2611                 if (vcpu->arch.trap)
2612                         ret = kvmppc_handle_exit_hv(vcpu->arch.kvm_run, vcpu,
2613                                                     vcpu->arch.run_task);
2614
2615                 vcpu->arch.ret = ret;
2616                 vcpu->arch.trap = 0;
2617
2618                 if (is_kvmppc_resume_guest(vcpu->arch.ret)) {
2619                         if (vcpu->arch.pending_exceptions)
2620                                 kvmppc_core_prepare_to_enter(vcpu);
2621                         if (vcpu->arch.ceded)
2622                                 kvmppc_set_timer(vcpu);
2623                         else
2624                                 ++still_running;
2625                 } else {
2626                         kvmppc_remove_runnable(vc, vcpu);
2627                         wake_up(&vcpu->arch.cpu_run);
2628                 }
2629         }
2630         if (!is_master) {
2631                 if (still_running > 0) {
2632                         kvmppc_vcore_preempt(vc);
2633                 } else if (vc->runner) {
2634                         vc->vcore_state = VCORE_PREEMPT;
2635                         kvmppc_core_start_stolen(vc);
2636                 } else {
2637                         vc->vcore_state = VCORE_INACTIVE;
2638                 }
2639                 if (vc->n_runnable > 0 && vc->runner == NULL) {
2640                         /* make sure there's a candidate runner awake */
2641                         i = -1;
2642                         vcpu = next_runnable_thread(vc, &i);
2643                         wake_up(&vcpu->arch.cpu_run);
2644                 }
2645         }
2646         spin_unlock(&vc->lock);
2647 }
2648
2649 /*
2650  * Clear core from the list of active host cores as we are about to
2651  * enter the guest. Only do this if it is the primary thread of the
2652  * core (not if a subcore) that is entering the guest.
2653  */
2654 static inline int kvmppc_clear_host_core(unsigned int cpu)
2655 {
2656         int core;
2657
2658         if (!kvmppc_host_rm_ops_hv || cpu_thread_in_core(cpu))
2659                 return 0;
2660         /*
2661          * Memory barrier can be omitted here as we will do a smp_wmb()
2662          * later in kvmppc_start_thread and we need ensure that state is
2663          * visible to other CPUs only after we enter guest.
2664          */
2665         core = cpu >> threads_shift;
2666         kvmppc_host_rm_ops_hv->rm_core[core].rm_state.in_host = 0;
2667         return 0;
2668 }
2669
2670 /*
2671  * Advertise this core as an active host core since we exited the guest
2672  * Only need to do this if it is the primary thread of the core that is
2673  * exiting.
2674  */
2675 static inline int kvmppc_set_host_core(unsigned int cpu)
2676 {
2677         int core;
2678
2679         if (!kvmppc_host_rm_ops_hv || cpu_thread_in_core(cpu))
2680                 return 0;
2681
2682         /*
2683          * Memory barrier can be omitted here because we do a spin_unlock
2684          * immediately after this which provides the memory barrier.
2685          */
2686         core = cpu >> threads_shift;
2687         kvmppc_host_rm_ops_hv->rm_core[core].rm_state.in_host = 1;
2688         return 0;
2689 }
2690
2691 static void set_irq_happened(int trap)
2692 {
2693         switch (trap) {
2694         case BOOK3S_INTERRUPT_EXTERNAL:
2695                 local_paca->irq_happened |= PACA_IRQ_EE;
2696                 break;
2697         case BOOK3S_INTERRUPT_H_DOORBELL:
2698                 local_paca->irq_happened |= PACA_IRQ_DBELL;
2699                 break;
2700         case BOOK3S_INTERRUPT_HMI:
2701                 local_paca->irq_happened |= PACA_IRQ_HMI;
2702                 break;
2703         case BOOK3S_INTERRUPT_SYSTEM_RESET:
2704                 replay_system_reset();
2705                 break;
2706         }
2707 }
2708
2709 /*
2710  * Run a set of guest threads on a physical core.
2711  * Called with vc->lock held.
2712  */
2713 static noinline void kvmppc_run_core(struct kvmppc_vcore *vc)
2714 {
2715         struct kvm_vcpu *vcpu;
2716         int i;
2717         int srcu_idx;
2718         struct core_info core_info;
2719         struct kvmppc_vcore *pvc;
2720         struct kvm_split_mode split_info, *sip;
2721         int split, subcore_size, active;
2722         int sub;
2723         bool thr0_done;
2724         unsigned long cmd_bit, stat_bit;
2725         int pcpu, thr;
2726         int target_threads;
2727         int controlled_threads;
2728         int trap;
2729         bool is_power8;
2730         bool hpt_on_radix;
2731
2732         /*
2733          * Remove from the list any threads that have a signal pending
2734          * or need a VPA update done
2735          */
2736         prepare_threads(vc);
2737
2738         /* if the runner is no longer runnable, let the caller pick a new one */
2739         if (vc->runner->arch.state != KVMPPC_VCPU_RUNNABLE)
2740                 return;
2741
2742         /*
2743          * Initialize *vc.
2744          */
2745         init_vcore_to_run(vc);
2746         vc->preempt_tb = TB_NIL;
2747
2748         /*
2749          * Number of threads that we will be controlling: the same as
2750          * the number of threads per subcore, except on POWER9,
2751          * where it's 1 because the threads are (mostly) independent.
2752          */
2753         controlled_threads = threads_per_vcore(vc->kvm);
2754
2755         /*
2756          * Make sure we are running on primary threads, and that secondary
2757          * threads are offline.  Also check if the number of threads in this
2758          * guest are greater than the current system threads per guest.
2759          * On POWER9, we need to be not in independent-threads mode if
2760          * this is a HPT guest on a radix host machine where the
2761          * CPU threads may not be in different MMU modes.
2762          */
2763         hpt_on_radix = no_mixing_hpt_and_radix && radix_enabled() &&
2764                 !kvm_is_radix(vc->kvm);
2765         if (((controlled_threads > 1) &&
2766              ((vc->num_threads > threads_per_subcore) || !on_primary_thread())) ||
2767             (hpt_on_radix && vc->kvm->arch.threads_indep)) {
2768                 for_each_runnable_thread(i, vcpu, vc) {
2769                         vcpu->arch.ret = -EBUSY;
2770                         kvmppc_remove_runnable(vc, vcpu);
2771                         wake_up(&vcpu->arch.cpu_run);
2772                 }
2773                 goto out;
2774         }
2775
2776         /*
2777          * See if we could run any other vcores on the physical core
2778          * along with this one.
2779          */
2780         init_core_info(&core_info, vc);
2781         pcpu = smp_processor_id();
2782         target_threads = controlled_threads;
2783         if (target_smt_mode && target_smt_mode < target_threads)
2784                 target_threads = target_smt_mode;
2785         if (vc->num_threads < target_threads)
2786                 collect_piggybacks(&core_info, target_threads);
2787
2788         /*
2789          * On radix, arrange for TLB flushing if necessary.
2790          * This has to be done before disabling interrupts since
2791          * it uses smp_call_function().
2792          */
2793         pcpu = smp_processor_id();
2794         if (kvm_is_radix(vc->kvm)) {
2795                 for (sub = 0; sub < core_info.n_subcores; ++sub)
2796                         for_each_runnable_thread(i, vcpu, core_info.vc[sub])
2797                                 kvmppc_prepare_radix_vcpu(vcpu, pcpu);
2798         }
2799
2800         /*
2801          * Hard-disable interrupts, and check resched flag and signals.
2802          * If we need to reschedule or deliver a signal, clean up
2803          * and return without going into the guest(s).
2804          * If the mmu_ready flag has been cleared, don't go into the
2805          * guest because that means a HPT resize operation is in progress.
2806          */
2807         local_irq_disable();
2808         hard_irq_disable();
2809         if (lazy_irq_pending() || need_resched() ||
2810             recheck_signals(&core_info) || !vc->kvm->arch.mmu_ready) {
2811                 local_irq_enable();
2812                 vc->vcore_state = VCORE_INACTIVE;
2813                 /* Unlock all except the primary vcore */
2814                 for (sub = 1; sub < core_info.n_subcores; ++sub) {
2815                         pvc = core_info.vc[sub];
2816                         /* Put back on to the preempted vcores list */
2817                         kvmppc_vcore_preempt(pvc);
2818                         spin_unlock(&pvc->lock);
2819                 }
2820                 for (i = 0; i < controlled_threads; ++i)
2821                         kvmppc_release_hwthread(pcpu + i);
2822                 return;
2823         }
2824
2825         kvmppc_clear_host_core(pcpu);
2826
2827         /* Decide on micro-threading (split-core) mode */
2828         subcore_size = threads_per_subcore;
2829         cmd_bit = stat_bit = 0;
2830         split = core_info.n_subcores;
2831         sip = NULL;
2832         is_power8 = cpu_has_feature(CPU_FTR_ARCH_207S)
2833                 && !cpu_has_feature(CPU_FTR_ARCH_300);
2834
2835         if (split > 1 || hpt_on_radix) {
2836                 sip = &split_info;
2837                 memset(&split_info, 0, sizeof(split_info));
2838                 for (sub = 0; sub < core_info.n_subcores; ++sub)
2839                         split_info.vc[sub] = core_info.vc[sub];
2840
2841                 if (is_power8) {
2842                         if (split == 2 && (dynamic_mt_modes & 2)) {
2843                                 cmd_bit = HID0_POWER8_1TO2LPAR;
2844                                 stat_bit = HID0_POWER8_2LPARMODE;
2845                         } else {
2846                                 split = 4;
2847                                 cmd_bit = HID0_POWER8_1TO4LPAR;
2848                                 stat_bit = HID0_POWER8_4LPARMODE;
2849                         }
2850                         subcore_size = MAX_SMT_THREADS / split;
2851                         split_info.rpr = mfspr(SPRN_RPR);
2852                         split_info.pmmar = mfspr(SPRN_PMMAR);
2853                         split_info.ldbar = mfspr(SPRN_LDBAR);
2854                         split_info.subcore_size = subcore_size;
2855                 } else {
2856                         split_info.subcore_size = 1;
2857                         if (hpt_on_radix) {
2858                                 /* Use the split_info for LPCR/LPIDR changes */
2859                                 split_info.lpcr_req = vc->lpcr;
2860                                 split_info.lpidr_req = vc->kvm->arch.lpid;
2861                                 split_info.host_lpcr = vc->kvm->arch.host_lpcr;
2862                                 split_info.do_set = 1;
2863                         }
2864                 }
2865
2866                 /* order writes to split_info before kvm_split_mode pointer */
2867                 smp_wmb();
2868         }
2869
2870         for (thr = 0; thr < controlled_threads; ++thr) {
2871                 struct paca_struct *paca = paca_ptrs[pcpu + thr];
2872
2873                 paca->kvm_hstate.tid = thr;
2874                 paca->kvm_hstate.napping = 0;
2875                 paca->kvm_hstate.kvm_split_mode = sip;
2876         }
2877
2878         /* Initiate micro-threading (split-core) on POWER8 if required */
2879         if (cmd_bit) {
2880                 unsigned long hid0 = mfspr(SPRN_HID0);
2881
2882                 hid0 |= cmd_bit | HID0_POWER8_DYNLPARDIS;
2883                 mb();
2884                 mtspr(SPRN_HID0, hid0);
2885                 isync();
2886                 for (;;) {
2887                         hid0 = mfspr(SPRN_HID0);
2888                         if (hid0 & stat_bit)
2889                                 break;
2890                         cpu_relax();
2891                 }
2892         }
2893
2894         /*
2895          * On POWER8, set RWMR register.
2896          * Since it only affects PURR and SPURR, it doesn't affect
2897          * the host, so we don't save/restore the host value.
2898          */
2899         if (is_power8) {
2900                 unsigned long rwmr_val = RWMR_RPA_P8_8THREAD;
2901                 int n_online = atomic_read(&vc->online_count);
2902
2903                 /*
2904                  * Use the 8-thread value if we're doing split-core
2905                  * or if the vcore's online count looks bogus.
2906                  */
2907                 if (split == 1 && threads_per_subcore == MAX_SMT_THREADS &&
2908                     n_online >= 1 && n_online <= MAX_SMT_THREADS)
2909                         rwmr_val = p8_rwmr_values[n_online];
2910                 mtspr(SPRN_RWMR, rwmr_val);
2911         }
2912
2913         /* Start all the threads */
2914         active = 0;
2915         for (sub = 0; sub < core_info.n_subcores; ++sub) {
2916                 thr = is_power8 ? subcore_thread_map[sub] : sub;
2917                 thr0_done = false;
2918                 active |= 1 << thr;
2919                 pvc = core_info.vc[sub];
2920                 pvc->pcpu = pcpu + thr;
2921                 for_each_runnable_thread(i, vcpu, pvc) {
2922                         kvmppc_start_thread(vcpu, pvc);
2923                         kvmppc_create_dtl_entry(vcpu, pvc);
2924                         trace_kvm_guest_enter(vcpu);
2925                         if (!vcpu->arch.ptid)
2926                                 thr0_done = true;
2927                         active |= 1 << (thr + vcpu->arch.ptid);
2928                 }
2929                 /*
2930                  * We need to start the first thread of each subcore
2931                  * even if it doesn't have a vcpu.
2932                  */
2933                 if (!thr0_done)
2934                         kvmppc_start_thread(NULL, pvc);
2935         }
2936
2937         /*
2938          * Ensure that split_info.do_nap is set after setting
2939          * the vcore pointer in the PACA of the secondaries.
2940          */
2941         smp_mb();
2942
2943         /*
2944          * When doing micro-threading, poke the inactive threads as well.
2945          * This gets them to the nap instruction after kvm_do_nap,
2946          * which reduces the time taken to unsplit later.
2947          * For POWER9 HPT guest on radix host, we need all the secondary
2948          * threads woken up so they can do the LPCR/LPIDR change.
2949          */
2950         if (cmd_bit || hpt_on_radix) {
2951                 split_info.do_nap = 1;  /* ask secondaries to nap when done */
2952                 for (thr = 1; thr < threads_per_subcore; ++thr)
2953                         if (!(active & (1 << thr)))
2954                                 kvmppc_ipi_thread(pcpu + thr);
2955         }
2956
2957         vc->vcore_state = VCORE_RUNNING;
2958         preempt_disable();
2959
2960         trace_kvmppc_run_core(vc, 0);
2961
2962         for (sub = 0; sub < core_info.n_subcores; ++sub)
2963                 spin_unlock(&core_info.vc[sub]->lock);
2964
2965         if (kvm_is_radix(vc->kvm)) {
2966                 int tmp = pcpu;
2967
2968                 /*
2969                  * Do we need to flush the process scoped TLB for the LPAR?
2970                  *
2971                  * On POWER9, individual threads can come in here, but the
2972                  * TLB is shared between the 4 threads in a core, hence
2973                  * invalidating on one thread invalidates for all.
2974                  * Thus we make all 4 threads use the same bit here.
2975                  *
2976                  * Hash must be flushed in realmode in order to use tlbiel.
2977                  */
2978                 mtspr(SPRN_LPID, vc->kvm->arch.lpid);
2979                 isync();
2980
2981                 if (cpu_has_feature(CPU_FTR_ARCH_300))
2982                         tmp &= ~0x3UL;
2983
2984                 if (cpumask_test_cpu(tmp, &vc->kvm->arch.need_tlb_flush)) {
2985                         radix__local_flush_tlb_lpid_guest(vc->kvm->arch.lpid);
2986                         /* Clear the bit after the TLB flush */
2987                         cpumask_clear_cpu(tmp, &vc->kvm->arch.need_tlb_flush);
2988                 }
2989         }
2990
2991         /*
2992          * Interrupts will be enabled once we get into the guest,
2993          * so tell lockdep that we're about to enable interrupts.
2994          */
2995         trace_hardirqs_on();
2996
2997         guest_enter_irqoff();
2998
2999         srcu_idx = srcu_read_lock(&vc->kvm->srcu);
3000
3001         this_cpu_disable_ftrace();
3002
3003         trap = __kvmppc_vcore_entry();
3004
3005         this_cpu_enable_ftrace();
3006
3007         srcu_read_unlock(&vc->kvm->srcu, srcu_idx);
3008
3009         trace_hardirqs_off();
3010         set_irq_happened(trap);
3011
3012         spin_lock(&vc->lock);
3013         /* prevent other vcpu threads from doing kvmppc_start_thread() now */
3014         vc->vcore_state = VCORE_EXITING;
3015
3016         /* wait for secondary threads to finish writing their state to memory */
3017         kvmppc_wait_for_nap(controlled_threads);
3018
3019         /* Return to whole-core mode if we split the core earlier */
3020         if (cmd_bit) {
3021                 unsigned long hid0 = mfspr(SPRN_HID0);
3022                 unsigned long loops = 0;
3023
3024                 hid0 &= ~HID0_POWER8_DYNLPARDIS;
3025                 stat_bit = HID0_POWER8_2LPARMODE | HID0_POWER8_4LPARMODE;
3026                 mb();
3027                 mtspr(SPRN_HID0, hid0);
3028                 isync();
3029                 for (;;) {
3030                         hid0 = mfspr(SPRN_HID0);
3031                         if (!(hid0 & stat_bit))
3032                                 break;
3033                         cpu_relax();
3034                         ++loops;
3035                 }
3036         } else if (hpt_on_radix) {
3037                 /* Wait for all threads to have seen final sync */
3038                 for (thr = 1; thr < controlled_threads; ++thr) {
3039                         struct paca_struct *paca = paca_ptrs[pcpu + thr];
3040
3041                         while (paca->kvm_hstate.kvm_split_mode) {
3042                                 HMT_low();
3043                                 barrier();
3044                         }
3045                         HMT_medium();
3046                 }
3047         }
3048         split_info.do_nap = 0;
3049
3050         kvmppc_set_host_core(pcpu);
3051
3052         local_irq_enable();
3053         guest_exit();
3054
3055         /* Let secondaries go back to the offline loop */
3056         for (i = 0; i < controlled_threads; ++i) {
3057                 kvmppc_release_hwthread(pcpu + i);
3058                 if (sip && sip->napped[i])
3059                         kvmppc_ipi_thread(pcpu + i);
3060                 cpumask_clear_cpu(pcpu + i, &vc->kvm->arch.cpu_in_guest);
3061         }
3062
3063         spin_unlock(&vc->lock);
3064
3065         /* make sure updates to secondary vcpu structs are visible now */
3066         smp_mb();
3067
3068         preempt_enable();
3069
3070         for (sub = 0; sub < core_info.n_subcores; ++sub) {
3071                 pvc = core_info.vc[sub];
3072                 post_guest_process(pvc, pvc == vc);
3073         }
3074
3075         spin_lock(&vc->lock);
3076
3077  out:
3078         vc->vcore_state = VCORE_INACTIVE;
3079         trace_kvmppc_run_core(vc, 1);
3080 }
3081
3082 /*
3083  * Wait for some other vcpu thread to execute us, and
3084  * wake us up when we need to handle something in the host.
3085  */
3086 static void kvmppc_wait_for_exec(struct kvmppc_vcore *vc,
3087                                  struct kvm_vcpu *vcpu, int wait_state)
3088 {
3089         DEFINE_WAIT(wait);
3090
3091         prepare_to_wait(&vcpu->arch.cpu_run, &wait, wait_state);
3092         if (vcpu->arch.state == KVMPPC_VCPU_RUNNABLE) {
3093                 spin_unlock(&vc->lock);
3094                 schedule();
3095                 spin_lock(&vc->lock);
3096         }
3097         finish_wait(&vcpu->arch.cpu_run, &wait);
3098 }
3099
3100 static void grow_halt_poll_ns(struct kvmppc_vcore *vc)
3101 {
3102         /* 10us base */
3103         if (vc->halt_poll_ns == 0 && halt_poll_ns_grow)
3104                 vc->halt_poll_ns = 10000;
3105         else
3106                 vc->halt_poll_ns *= halt_poll_ns_grow;
3107 }
3108
3109 static void shrink_halt_poll_ns(struct kvmppc_vcore *vc)
3110 {
3111         if (halt_poll_ns_shrink == 0)
3112                 vc->halt_poll_ns = 0;
3113         else
3114                 vc->halt_poll_ns /= halt_poll_ns_shrink;
3115 }
3116
3117 #ifdef CONFIG_KVM_XICS
3118 static inline bool xive_interrupt_pending(struct kvm_vcpu *vcpu)
3119 {
3120         if (!xive_enabled())
3121                 return false;
3122         return vcpu->arch.irq_pending || vcpu->arch.xive_saved_state.pipr <
3123                 vcpu->arch.xive_saved_state.cppr;
3124 }
3125 #else
3126 static inline bool xive_interrupt_pending(struct kvm_vcpu *vcpu)
3127 {
3128         return false;
3129 }
3130 #endif /* CONFIG_KVM_XICS */
3131
3132 static bool kvmppc_vcpu_woken(struct kvm_vcpu *vcpu)
3133 {
3134         if (vcpu->arch.pending_exceptions || vcpu->arch.prodded ||
3135             kvmppc_doorbell_pending(vcpu) || xive_interrupt_pending(vcpu))
3136                 return true;
3137
3138         return false;
3139 }
3140
3141 /*
3142  * Check to see if any of the runnable vcpus on the vcore have pending
3143  * exceptions or are no longer ceded
3144  */
3145 static int kvmppc_vcore_check_block(struct kvmppc_vcore *vc)
3146 {
3147         struct kvm_vcpu *vcpu;
3148         int i;
3149
3150         for_each_runnable_thread(i, vcpu, vc) {
3151                 if (!vcpu->arch.ceded || kvmppc_vcpu_woken(vcpu))
3152                         return 1;
3153         }
3154
3155         return 0;
3156 }
3157
3158 /*
3159  * All the vcpus in this vcore are idle, so wait for a decrementer
3160  * or external interrupt to one of the vcpus.  vc->lock is held.
3161  */
3162 static void kvmppc_vcore_blocked(struct kvmppc_vcore *vc)
3163 {
3164         ktime_t cur, start_poll, start_wait;
3165         int do_sleep = 1;
3166         u64 block_ns;
3167         DECLARE_SWAITQUEUE(wait);
3168
3169         /* Poll for pending exceptions and ceded state */
3170         cur = start_poll = ktime_get();
3171         if (vc->halt_poll_ns) {
3172                 ktime_t stop = ktime_add_ns(start_poll, vc->halt_poll_ns);
3173                 ++vc->runner->stat.halt_attempted_poll;
3174
3175                 vc->vcore_state = VCORE_POLLING;
3176                 spin_unlock(&vc->lock);
3177
3178                 do {
3179                         if (kvmppc_vcore_check_block(vc)) {
3180                                 do_sleep = 0;
3181                                 break;
3182                         }
3183                         cur = ktime_get();
3184                 } while (single_task_running() && ktime_before(cur, stop));
3185
3186                 spin_lock(&vc->lock);
3187                 vc->vcore_state = VCORE_INACTIVE;
3188
3189                 if (!do_sleep) {
3190                         ++vc->runner->stat.halt_successful_poll;
3191                         goto out;
3192                 }
3193         }
3194
3195         prepare_to_swait_exclusive(&vc->wq, &wait, TASK_INTERRUPTIBLE);
3196
3197         if (kvmppc_vcore_check_block(vc)) {
3198                 finish_swait(&vc->wq, &wait);
3199                 do_sleep = 0;
3200                 /* If we polled, count this as a successful poll */
3201                 if (vc->halt_poll_ns)
3202                         ++vc->runner->stat.halt_successful_poll;
3203                 goto out;
3204         }
3205
3206         start_wait = ktime_get();
3207
3208         vc->vcore_state = VCORE_SLEEPING;
3209         trace_kvmppc_vcore_blocked(vc, 0);
3210         spin_unlock(&vc->lock);
3211         schedule();
3212         finish_swait(&vc->wq, &wait);
3213         spin_lock(&vc->lock);
3214         vc->vcore_state = VCORE_INACTIVE;
3215         trace_kvmppc_vcore_blocked(vc, 1);
3216         ++vc->runner->stat.halt_successful_wait;
3217
3218         cur = ktime_get();
3219
3220 out:
3221         block_ns = ktime_to_ns(cur) - ktime_to_ns(start_poll);
3222
3223         /* Attribute wait time */
3224         if (do_sleep) {
3225                 vc->runner->stat.halt_wait_ns +=
3226                         ktime_to_ns(cur) - ktime_to_ns(start_wait);
3227                 /* Attribute failed poll time */
3228                 if (vc->halt_poll_ns)
3229                         vc->runner->stat.halt_poll_fail_ns +=
3230                                 ktime_to_ns(start_wait) -
3231                                 ktime_to_ns(start_poll);
3232         } else {
3233                 /* Attribute successful poll time */
3234                 if (vc->halt_poll_ns)
3235                         vc->runner->stat.halt_poll_success_ns +=
3236                                 ktime_to_ns(cur) -
3237                                 ktime_to_ns(start_poll);
3238         }
3239
3240         /* Adjust poll time */
3241         if (halt_poll_ns) {
3242                 if (block_ns <= vc->halt_poll_ns)
3243                         ;
3244                 /* We slept and blocked for longer than the max halt time */
3245                 else if (vc->halt_poll_ns && block_ns > halt_poll_ns)
3246                         shrink_halt_poll_ns(vc);
3247                 /* We slept and our poll time is too small */
3248                 else if (vc->halt_poll_ns < halt_poll_ns &&
3249                                 block_ns < halt_poll_ns)
3250                         grow_halt_poll_ns(vc);
3251                 if (vc->halt_poll_ns > halt_poll_ns)
3252                         vc->halt_poll_ns = halt_poll_ns;
3253         } else
3254                 vc->halt_poll_ns = 0;
3255
3256         trace_kvmppc_vcore_wakeup(do_sleep, block_ns);
3257 }
3258
3259 static int kvmhv_setup_mmu(struct kvm_vcpu *vcpu)
3260 {
3261         int r = 0;
3262         struct kvm *kvm = vcpu->kvm;
3263
3264         mutex_lock(&kvm->lock);
3265         if (!kvm->arch.mmu_ready) {
3266                 if (!kvm_is_radix(kvm))
3267                         r = kvmppc_hv_setup_htab_rma(vcpu);
3268                 if (!r) {
3269                         if (cpu_has_feature(CPU_FTR_ARCH_300))
3270                                 kvmppc_setup_partition_table(kvm);
3271                         kvm->arch.mmu_ready = 1;
3272                 }
3273         }
3274         mutex_unlock(&kvm->lock);
3275         return r;
3276 }
3277
3278 static int kvmppc_run_vcpu(struct kvm_run *kvm_run, struct kvm_vcpu *vcpu)
3279 {
3280         int n_ceded, i, r;
3281         struct kvmppc_vcore *vc;
3282         struct kvm_vcpu *v;
3283
3284         trace_kvmppc_run_vcpu_enter(vcpu);
3285
3286         kvm_run->exit_reason = 0;
3287         vcpu->arch.ret = RESUME_GUEST;
3288         vcpu->arch.trap = 0;
3289         kvmppc_update_vpas(vcpu);
3290
3291         /*
3292          * Synchronize with other threads in this virtual core
3293          */
3294         vc = vcpu->arch.vcore;
3295         spin_lock(&vc->lock);
3296         vcpu->arch.ceded = 0;
3297         vcpu->arch.run_task = current;
3298         vcpu->arch.kvm_run = kvm_run;
3299         vcpu->arch.stolen_logged = vcore_stolen_time(vc, mftb());
3300         vcpu->arch.state = KVMPPC_VCPU_RUNNABLE;
3301         vcpu->arch.busy_preempt = TB_NIL;
3302         WRITE_ONCE(vc->runnable_threads[vcpu->arch.ptid], vcpu);
3303         ++vc->n_runnable;
3304
3305         /*
3306          * This happens the first time this is called for a vcpu.
3307          * If the vcore is already running, we may be able to start
3308          * this thread straight away and have it join in.
3309          */
3310         if (!signal_pending(current)) {
3311                 if ((vc->vcore_state == VCORE_PIGGYBACK ||
3312                      vc->vcore_state == VCORE_RUNNING) &&
3313                            !VCORE_IS_EXITING(vc)) {
3314                         kvmppc_create_dtl_entry(vcpu, vc);
3315                         kvmppc_start_thread(vcpu, vc);
3316                         trace_kvm_guest_enter(vcpu);
3317                 } else if (vc->vcore_state == VCORE_SLEEPING) {
3318                         swake_up_one(&vc->wq);
3319                 }
3320
3321         }
3322
3323         while (vcpu->arch.state == KVMPPC_VCPU_RUNNABLE &&
3324                !signal_pending(current)) {
3325                 /* See if the MMU is ready to go */
3326                 if (!vcpu->kvm->arch.mmu_ready) {
3327                         spin_unlock(&vc->lock);
3328                         r = kvmhv_setup_mmu(vcpu);
3329                         spin_lock(&vc->lock);
3330                         if (r) {
3331                                 kvm_run->exit_reason = KVM_EXIT_FAIL_ENTRY;
3332                                 kvm_run->fail_entry.
3333                                         hardware_entry_failure_reason = 0;
3334                                 vcpu->arch.ret = r;
3335                                 break;
3336                         }
3337                 }
3338
3339                 if (vc->vcore_state == VCORE_PREEMPT && vc->runner == NULL)
3340                         kvmppc_vcore_end_preempt(vc);
3341
3342                 if (vc->vcore_state != VCORE_INACTIVE) {
3343                         kvmppc_wait_for_exec(vc, vcpu, TASK_INTERRUPTIBLE);
3344                         continue;
3345                 }
3346                 for_each_runnable_thread(i, v, vc) {
3347                         kvmppc_core_prepare_to_enter(v);
3348                         if (signal_pending(v->arch.run_task)) {
3349                                 kvmppc_remove_runnable(vc, v);
3350                                 v->stat.signal_exits++;
3351                                 v->arch.kvm_run->exit_reason = KVM_EXIT_INTR;
3352                                 v->arch.ret = -EINTR;
3353                                 wake_up(&v->arch.cpu_run);
3354                         }
3355                 }
3356                 if (!vc->n_runnable || vcpu->arch.state != KVMPPC_VCPU_RUNNABLE)
3357                         break;
3358                 n_ceded = 0;
3359                 for_each_runnable_thread(i, v, vc) {
3360                         if (!kvmppc_vcpu_woken(v))
3361                                 n_ceded += v->arch.ceded;
3362                         else
3363                                 v->arch.ceded = 0;
3364                 }
3365                 vc->runner = vcpu;
3366                 if (n_ceded == vc->n_runnable) {
3367                         kvmppc_vcore_blocked(vc);
3368                 } else if (need_resched()) {
3369                         kvmppc_vcore_preempt(vc);
3370                         /* Let something else run */
3371                         cond_resched_lock(&vc->lock);
3372                         if (vc->vcore_state == VCORE_PREEMPT)
3373                                 kvmppc_vcore_end_preempt(vc);
3374                 } else {
3375                         kvmppc_run_core(vc);
3376                 }
3377                 vc->runner = NULL;
3378         }
3379
3380         while (vcpu->arch.state == KVMPPC_VCPU_RUNNABLE &&
3381                (vc->vcore_state == VCORE_RUNNING ||
3382                 vc->vcore_state == VCORE_EXITING ||
3383                 vc->vcore_state == VCORE_PIGGYBACK))
3384                 kvmppc_wait_for_exec(vc, vcpu, TASK_UNINTERRUPTIBLE);
3385
3386         if (vc->vcore_state == VCORE_PREEMPT && vc->runner == NULL)
3387                 kvmppc_vcore_end_preempt(vc);
3388
3389         if (vcpu->arch.state == KVMPPC_VCPU_RUNNABLE) {
3390                 kvmppc_remove_runnable(vc, vcpu);
3391                 vcpu->stat.signal_exits++;
3392                 kvm_run->exit_reason = KVM_EXIT_INTR;
3393                 vcpu->arch.ret = -EINTR;
3394         }
3395
3396         if (vc->n_runnable && vc->vcore_state == VCORE_INACTIVE) {
3397                 /* Wake up some vcpu to run the core */
3398                 i = -1;
3399                 v = next_runnable_thread(vc, &i);
3400                 wake_up(&v->arch.cpu_run);
3401         }
3402
3403         trace_kvmppc_run_vcpu_exit(vcpu, kvm_run);
3404         spin_unlock(&vc->lock);
3405         return vcpu->arch.ret;
3406 }
3407
3408 static int kvmppc_vcpu_run_hv(struct kvm_run *run, struct kvm_vcpu *vcpu)
3409 {
3410         int r;
3411         int srcu_idx;
3412         unsigned long ebb_regs[3] = {}; /* shut up GCC */
3413         unsigned long user_tar = 0;
3414         unsigned int user_vrsave;
3415         struct kvm *kvm;
3416
3417         if (!vcpu->arch.sane) {
3418                 run->exit_reason = KVM_EXIT_INTERNAL_ERROR;
3419                 return -EINVAL;
3420         }
3421
3422         /*
3423          * Don't allow entry with a suspended transaction, because
3424          * the guest entry/exit code will lose it.
3425          * If the guest has TM enabled, save away their TM-related SPRs
3426          * (they will get restored by the TM unavailable interrupt).
3427          */
3428 #ifdef CONFIG_PPC_TRANSACTIONAL_MEM
3429         if (cpu_has_feature(CPU_FTR_TM) && current->thread.regs &&
3430             (current->thread.regs->msr & MSR_TM)) {
3431                 if (MSR_TM_ACTIVE(current->thread.regs->msr)) {
3432                         run->exit_reason = KVM_EXIT_FAIL_ENTRY;
3433                         run->fail_entry.hardware_entry_failure_reason = 0;
3434                         return -EINVAL;
3435                 }
3436                 /* Enable TM so we can read the TM SPRs */
3437                 mtmsr(mfmsr() | MSR_TM);
3438                 current->thread.tm_tfhar = mfspr(SPRN_TFHAR);
3439                 current->thread.tm_tfiar = mfspr(SPRN_TFIAR);
3440                 current->thread.tm_texasr = mfspr(SPRN_TEXASR);
3441                 current->thread.regs->msr &= ~MSR_TM;
3442         }
3443 #endif
3444
3445         /*
3446          * Force online to 1 for the sake of old userspace which doesn't
3447          * set it.
3448          */
3449         if (!vcpu->arch.online) {
3450                 atomic_inc(&vcpu->arch.vcore->online_count);
3451                 vcpu->arch.online = 1;
3452         }
3453
3454         kvmppc_core_prepare_to_enter(vcpu);
3455
3456         /* No need to go into the guest when all we'll do is come back out */
3457         if (signal_pending(current)) {
3458                 run->exit_reason = KVM_EXIT_INTR;
3459                 return -EINTR;
3460         }
3461
3462         kvm = vcpu->kvm;
3463         atomic_inc(&kvm->arch.vcpus_running);
3464         /* Order vcpus_running vs. mmu_ready, see kvmppc_alloc_reset_hpt */
3465         smp_mb();
3466
3467         flush_all_to_thread(current);
3468
3469         /* Save userspace EBB and other register values */
3470         if (cpu_has_feature(CPU_FTR_ARCH_207S)) {
3471                 ebb_regs[0] = mfspr(SPRN_EBBHR);
3472                 ebb_regs[1] = mfspr(SPRN_EBBRR);
3473                 ebb_regs[2] = mfspr(SPRN_BESCR);
3474                 user_tar = mfspr(SPRN_TAR);
3475         }
3476         user_vrsave = mfspr(SPRN_VRSAVE);
3477
3478         vcpu->arch.wqp = &vcpu->arch.vcore->wq;
3479         vcpu->arch.pgdir = current->mm->pgd;
3480         vcpu->arch.state = KVMPPC_VCPU_BUSY_IN_HOST;
3481
3482         do {
3483                 r = kvmppc_run_vcpu(run, vcpu);
3484
3485                 if (run->exit_reason == KVM_EXIT_PAPR_HCALL &&
3486                     !(vcpu->arch.shregs.msr & MSR_PR)) {
3487                         trace_kvm_hcall_enter(vcpu);
3488                         r = kvmppc_pseries_do_hcall(vcpu);
3489                         trace_kvm_hcall_exit(vcpu, r);
3490                         kvmppc_core_prepare_to_enter(vcpu);
3491                 } else if (r == RESUME_PAGE_FAULT) {
3492                         srcu_idx = srcu_read_lock(&kvm->srcu);
3493                         r = kvmppc_book3s_hv_page_fault(run, vcpu,
3494                                 vcpu->arch.fault_dar, vcpu->arch.fault_dsisr);
3495                         srcu_read_unlock(&kvm->srcu, srcu_idx);
3496                 } else if (r == RESUME_PASSTHROUGH) {
3497                         if (WARN_ON(xive_enabled()))
3498                                 r = H_SUCCESS;
3499                         else
3500                                 r = kvmppc_xics_rm_complete(vcpu, 0);
3501                 }
3502         } while (is_kvmppc_resume_guest(r));
3503
3504         /* Restore userspace EBB and other register values */
3505         if (cpu_has_feature(CPU_FTR_ARCH_207S)) {
3506                 mtspr(SPRN_EBBHR, ebb_regs[0]);
3507                 mtspr(SPRN_EBBRR, ebb_regs[1]);
3508                 mtspr(SPRN_BESCR, ebb_regs[2]);
3509                 mtspr(SPRN_TAR, user_tar);
3510                 mtspr(SPRN_FSCR, current->thread.fscr);
3511         }
3512         mtspr(SPRN_VRSAVE, user_vrsave);
3513
3514         vcpu->arch.state = KVMPPC_VCPU_NOTREADY;
3515         atomic_dec(&kvm->arch.vcpus_running);
3516         return r;
3517 }
3518
3519 static void kvmppc_add_seg_page_size(struct kvm_ppc_one_seg_page_size **sps,
3520                                      int shift, int sllp)
3521 {
3522         (*sps)->page_shift = shift;
3523         (*sps)->slb_enc = sllp;
3524         (*sps)->enc[0].page_shift = shift;
3525         (*sps)->enc[0].pte_enc = kvmppc_pgsize_lp_encoding(shift, shift);
3526         /*
3527          * Add 16MB MPSS support (may get filtered out by userspace)
3528          */
3529         if (shift != 24) {
3530                 int penc = kvmppc_pgsize_lp_encoding(shift, 24);
3531                 if (penc != -1) {
3532                         (*sps)->enc[1].page_shift = 24;
3533                         (*sps)->enc[1].pte_enc = penc;
3534                 }
3535         }
3536         (*sps)++;
3537 }
3538
3539 static int kvm_vm_ioctl_get_smmu_info_hv(struct kvm *kvm,
3540                                          struct kvm_ppc_smmu_info *info)
3541 {
3542         struct kvm_ppc_one_seg_page_size *sps;
3543
3544         /*
3545          * POWER7, POWER8 and POWER9 all support 32 storage keys for data.
3546          * POWER7 doesn't support keys for instruction accesses,
3547          * POWER8 and POWER9 do.
3548          */
3549         info->data_keys = 32;
3550         info->instr_keys = cpu_has_feature(CPU_FTR_ARCH_207S) ? 32 : 0;
3551
3552         /* POWER7, 8 and 9 all have 1T segments and 32-entry SLB */
3553         info->flags = KVM_PPC_PAGE_SIZES_REAL | KVM_PPC_1T_SEGMENTS;
3554         info->slb_size = 32;
3555
3556         /* We only support these sizes for now, and no muti-size segments */
3557         sps = &info->sps[0];
3558         kvmppc_add_seg_page_size(&sps, 12, 0);
3559         kvmppc_add_seg_page_size(&sps, 16, SLB_VSID_L | SLB_VSID_LP_01);
3560         kvmppc_add_seg_page_size(&sps, 24, SLB_VSID_L);
3561
3562         return 0;
3563 }
3564
3565 /*
3566  * Get (and clear) the dirty memory log for a memory slot.
3567  */
3568 static int kvm_vm_ioctl_get_dirty_log_hv(struct kvm *kvm,
3569                                          struct kvm_dirty_log *log)
3570 {
3571         struct kvm_memslots *slots;
3572         struct kvm_memory_slot *memslot;
3573         int i, r;
3574         unsigned long n;
3575         unsigned long *buf, *p;
3576         struct kvm_vcpu *vcpu;
3577
3578         mutex_lock(&kvm->slots_lock);
3579
3580         r = -EINVAL;
3581         if (log->slot >= KVM_USER_MEM_SLOTS)
3582                 goto out;
3583
3584         slots = kvm_memslots(kvm);
3585         memslot = id_to_memslot(slots, log->slot);
3586         r = -ENOENT;
3587         if (!memslot->dirty_bitmap)
3588                 goto out;
3589
3590         /*
3591          * Use second half of bitmap area because both HPT and radix
3592          * accumulate bits in the first half.
3593          */
3594         n = kvm_dirty_bitmap_bytes(memslot);
3595         buf = memslot->dirty_bitmap + n / sizeof(long);
3596         memset(buf, 0, n);
3597
3598         if (kvm_is_radix(kvm))
3599                 r = kvmppc_hv_get_dirty_log_radix(kvm, memslot, buf);
3600         else
3601                 r = kvmppc_hv_get_dirty_log_hpt(kvm, memslot, buf);
3602         if (r)
3603                 goto out;
3604
3605         /*
3606          * We accumulate dirty bits in the first half of the
3607          * memslot's dirty_bitmap area, for when pages are paged
3608          * out or modified by the host directly.  Pick up these
3609          * bits and add them to the map.
3610          */
3611         p = memslot->dirty_bitmap;
3612         for (i = 0; i < n / sizeof(long); ++i)
3613                 buf[i] |= xchg(&p[i], 0);
3614
3615         /* Harvest dirty bits from VPA and DTL updates */
3616         /* Note: we never modify the SLB shadow buffer areas */
3617         kvm_for_each_vcpu(i, vcpu, kvm) {
3618                 spin_lock(&vcpu->arch.vpa_update_lock);
3619                 kvmppc_harvest_vpa_dirty(&vcpu->arch.vpa, memslot, buf);
3620                 kvmppc_harvest_vpa_dirty(&vcpu->arch.dtl, memslot, buf);
3621                 spin_unlock(&vcpu->arch.vpa_update_lock);
3622         }
3623
3624         r = -EFAULT;
3625         if (copy_to_user(log->dirty_bitmap, buf, n))
3626                 goto out;
3627
3628         r = 0;
3629 out:
3630         mutex_unlock(&kvm->slots_lock);
3631         return r;
3632 }
3633
3634 static void kvmppc_core_free_memslot_hv(struct kvm_memory_slot *free,
3635                                         struct kvm_memory_slot *dont)
3636 {
3637         if (!dont || free->arch.rmap != dont->arch.rmap) {
3638                 vfree(free->arch.rmap);
3639                 free->arch.rmap = NULL;
3640         }
3641 }
3642
3643 static int kvmppc_core_create_memslot_hv(struct kvm_memory_slot *slot,
3644                                          unsigned long npages)
3645 {
3646         slot->arch.rmap = vzalloc(array_size(npages, sizeof(*slot->arch.rmap)));
3647         if (!slot->arch.rmap)
3648                 return -ENOMEM;
3649
3650         return 0;
3651 }
3652
3653 static int kvmppc_core_prepare_memory_region_hv(struct kvm *kvm,
3654                                         struct kvm_memory_slot *memslot,
3655                                         const struct kvm_userspace_memory_region *mem)
3656 {
3657         return 0;
3658 }
3659
3660 static void kvmppc_core_commit_memory_region_hv(struct kvm *kvm,
3661                                 const struct kvm_userspace_memory_region *mem,
3662                                 const struct kvm_memory_slot *old,
3663                                 const struct kvm_memory_slot *new)
3664 {
3665         unsigned long npages = mem->memory_size >> PAGE_SHIFT;
3666
3667         /*
3668          * If we are making a new memslot, it might make
3669          * some address that was previously cached as emulated
3670          * MMIO be no longer emulated MMIO, so invalidate
3671          * all the caches of emulated MMIO translations.
3672          */
3673         if (npages)
3674                 atomic64_inc(&kvm->arch.mmio_update);
3675 }
3676
3677 /*
3678  * Update LPCR values in kvm->arch and in vcores.
3679  * Caller must hold kvm->lock.
3680  */
3681 void kvmppc_update_lpcr(struct kvm *kvm, unsigned long lpcr, unsigned long mask)
3682 {
3683         long int i;
3684         u32 cores_done = 0;
3685
3686         if ((kvm->arch.lpcr & mask) == lpcr)
3687                 return;
3688
3689         kvm->arch.lpcr = (kvm->arch.lpcr & ~mask) | lpcr;
3690
3691         for (i = 0; i < KVM_MAX_VCORES; ++i) {
3692                 struct kvmppc_vcore *vc = kvm->arch.vcores[i];
3693                 if (!vc)
3694                         continue;
3695                 spin_lock(&vc->lock);
3696                 vc->lpcr = (vc->lpcr & ~mask) | lpcr;
3697                 spin_unlock(&vc->lock);
3698                 if (++cores_done >= kvm->arch.online_vcores)
3699                         break;
3700         }
3701 }
3702
3703 static void kvmppc_mmu_destroy_hv(struct kvm_vcpu *vcpu)
3704 {
3705         return;
3706 }
3707
3708 void kvmppc_setup_partition_table(struct kvm *kvm)
3709 {
3710         unsigned long dw0, dw1;
3711
3712         if (!kvm_is_radix(kvm)) {
3713                 /* PS field - page size for VRMA */
3714                 dw0 = ((kvm->arch.vrma_slb_v & SLB_VSID_L) >> 1) |
3715                         ((kvm->arch.vrma_slb_v & SLB_VSID_LP) << 1);
3716                 /* HTABSIZE and HTABORG fields */
3717                 dw0 |= kvm->arch.sdr1;
3718
3719                 /* Second dword as set by userspace */
3720                 dw1 = kvm->arch.process_table;
3721         } else {
3722                 dw0 = PATB_HR | radix__get_tree_size() |
3723                         __pa(kvm->arch.pgtable) | RADIX_PGD_INDEX_SIZE;
3724                 dw1 = PATB_GR | kvm->arch.process_table;
3725         }
3726
3727         mmu_partition_table_set_entry(kvm->arch.lpid, dw0, dw1);
3728 }
3729
3730 /*
3731  * Set up HPT (hashed page table) and RMA (real-mode area).
3732  * Must be called with kvm->lock held.
3733  */
3734 static int kvmppc_hv_setup_htab_rma(struct kvm_vcpu *vcpu)
3735 {
3736         int err = 0;
3737         struct kvm *kvm = vcpu->kvm;
3738         unsigned long hva;
3739         struct kvm_memory_slot *memslot;
3740         struct vm_area_struct *vma;
3741         unsigned long lpcr = 0, senc;
3742         unsigned long psize, porder;
3743         int srcu_idx;
3744
3745         /* Allocate hashed page table (if not done already) and reset it */
3746         if (!kvm->arch.hpt.virt) {
3747                 int order = KVM_DEFAULT_HPT_ORDER;
3748                 struct kvm_hpt_info info;
3749
3750                 err = kvmppc_allocate_hpt(&info, order);
3751                 /* If we get here, it means userspace didn't specify a
3752                  * size explicitly.  So, try successively smaller
3753                  * sizes if the default failed. */
3754                 while ((err == -ENOMEM) && --order >= PPC_MIN_HPT_ORDER)
3755                         err  = kvmppc_allocate_hpt(&info, order);
3756
3757                 if (err < 0) {
3758                         pr_err("KVM: Couldn't alloc HPT\n");
3759                         goto out;
3760                 }
3761
3762                 kvmppc_set_hpt(kvm, &info);
3763         }
3764
3765         /* Look up the memslot for guest physical address 0 */
3766         srcu_idx = srcu_read_lock(&kvm->srcu);
3767         memslot = gfn_to_memslot(kvm, 0);
3768
3769         /* We must have some memory at 0 by now */
3770         err = -EINVAL;
3771         if (!memslot || (memslot->flags & KVM_MEMSLOT_INVALID))
3772                 goto out_srcu;
3773
3774         /* Look up the VMA for the start of this memory slot */
3775         hva = memslot->userspace_addr;
3776         down_read(&current->mm->mmap_sem);
3777         vma = find_vma(current->mm, hva);
3778         if (!vma || vma->vm_start > hva || (vma->vm_flags & VM_IO))
3779                 goto up_out;
3780
3781         psize = vma_kernel_pagesize(vma);
3782
3783         up_read(&current->mm->mmap_sem);
3784
3785         /* We can handle 4k, 64k or 16M pages in the VRMA */
3786         if (psize >= 0x1000000)
3787                 psize = 0x1000000;
3788         else if (psize >= 0x10000)
3789                 psize = 0x10000;
3790         else
3791                 psize = 0x1000;
3792         porder = __ilog2(psize);
3793
3794         senc = slb_pgsize_encoding(psize);
3795         kvm->arch.vrma_slb_v = senc | SLB_VSID_B_1T |
3796                 (VRMA_VSID << SLB_VSID_SHIFT_1T);
3797         /* Create HPTEs in the hash page table for the VRMA */
3798         kvmppc_map_vrma(vcpu, memslot, porder);
3799
3800         /* Update VRMASD field in the LPCR */
3801         if (!cpu_has_feature(CPU_FTR_ARCH_300)) {
3802                 /* the -4 is to account for senc values starting at 0x10 */
3803                 lpcr = senc << (LPCR_VRMASD_SH - 4);
3804                 kvmppc_update_lpcr(kvm, lpcr, LPCR_VRMASD);
3805         }
3806
3807         /* Order updates to kvm->arch.lpcr etc. vs. mmu_ready */
3808         smp_wmb();
3809         err = 0;
3810  out_srcu:
3811         srcu_read_unlock(&kvm->srcu, srcu_idx);
3812  out:
3813         return err;
3814
3815  up_out:
3816         up_read(&current->mm->mmap_sem);
3817         goto out_srcu;
3818 }
3819
3820 /* Must be called with kvm->lock held and mmu_ready = 0 and no vcpus running */
3821 int kvmppc_switch_mmu_to_hpt(struct kvm *kvm)
3822 {
3823         kvmppc_free_radix(kvm);
3824         kvmppc_update_lpcr(kvm, LPCR_VPM1,
3825                            LPCR_VPM1 | LPCR_UPRT | LPCR_GTSE | LPCR_HR);
3826         kvmppc_rmap_reset(kvm);
3827         kvm->arch.radix = 0;
3828         kvm->arch.process_table = 0;
3829         return 0;
3830 }
3831
3832 /* Must be called with kvm->lock held and mmu_ready = 0 and no vcpus running */
3833 int kvmppc_switch_mmu_to_radix(struct kvm *kvm)
3834 {
3835         int err;
3836
3837         err = kvmppc_init_vm_radix(kvm);
3838         if (err)
3839                 return err;
3840
3841         kvmppc_free_hpt(&kvm->arch.hpt);
3842         kvmppc_update_lpcr(kvm, LPCR_UPRT | LPCR_GTSE | LPCR_HR,
3843                            LPCR_VPM1 | LPCR_UPRT | LPCR_GTSE | LPCR_HR);
3844         kvm->arch.radix = 1;
3845         return 0;
3846 }
3847
3848 #ifdef CONFIG_KVM_XICS
3849 /*
3850  * Allocate a per-core structure for managing state about which cores are
3851  * running in the host versus the guest and for exchanging data between
3852  * real mode KVM and CPU running in the host.
3853  * This is only done for the first VM.
3854  * The allocated structure stays even if all VMs have stopped.
3855  * It is only freed when the kvm-hv module is unloaded.
3856  * It's OK for this routine to fail, we just don't support host
3857  * core operations like redirecting H_IPI wakeups.
3858  */
3859 void kvmppc_alloc_host_rm_ops(void)
3860 {
3861         struct kvmppc_host_rm_ops *ops;
3862         unsigned long l_ops;
3863         int cpu, core;
3864         int size;
3865
3866         /* Not the first time here ? */
3867         if (kvmppc_host_rm_ops_hv != NULL)
3868                 return;
3869
3870         ops = kzalloc(sizeof(struct kvmppc_host_rm_ops), GFP_KERNEL);
3871         if (!ops)
3872                 return;
3873
3874         size = cpu_nr_cores() * sizeof(struct kvmppc_host_rm_core);
3875         ops->rm_core = kzalloc(size, GFP_KERNEL);
3876
3877         if (!ops->rm_core) {
3878                 kfree(ops);
3879                 return;
3880         }
3881
3882         cpus_read_lock();
3883
3884         for (cpu = 0; cpu < nr_cpu_ids; cpu += threads_per_core) {
3885                 if (!cpu_online(cpu))
3886                         continue;
3887
3888                 core = cpu >> threads_shift;
3889                 ops->rm_core[core].rm_state.in_host = 1;
3890         }
3891
3892         ops->vcpu_kick = kvmppc_fast_vcpu_kick_hv;
3893
3894         /*
3895          * Make the contents of the kvmppc_host_rm_ops structure visible
3896          * to other CPUs before we assign it to the global variable.
3897          * Do an atomic assignment (no locks used here), but if someone
3898          * beats us to it, just free our copy and return.
3899          */
3900         smp_wmb();
3901         l_ops = (unsigned long) ops;
3902
3903         if (cmpxchg64((unsigned long *)&kvmppc_host_rm_ops_hv, 0, l_ops)) {
3904                 cpus_read_unlock();
3905                 kfree(ops->rm_core);
3906                 kfree(ops);
3907                 return;
3908         }
3909
3910         cpuhp_setup_state_nocalls_cpuslocked(CPUHP_KVM_PPC_BOOK3S_PREPARE,
3911                                              "ppc/kvm_book3s:prepare",
3912                                              kvmppc_set_host_core,
3913                                              kvmppc_clear_host_core);
3914         cpus_read_unlock();
3915 }
3916
3917 void kvmppc_free_host_rm_ops(void)
3918 {
3919         if (kvmppc_host_rm_ops_hv) {
3920                 cpuhp_remove_state_nocalls(CPUHP_KVM_PPC_BOOK3S_PREPARE);
3921                 kfree(kvmppc_host_rm_ops_hv->rm_core);
3922                 kfree(kvmppc_host_rm_ops_hv);
3923                 kvmppc_host_rm_ops_hv = NULL;
3924         }
3925 }
3926 #endif
3927
3928 static int kvmppc_core_init_vm_hv(struct kvm *kvm)
3929 {
3930         unsigned long lpcr, lpid;
3931         char buf[32];
3932         int ret;
3933
3934         /* Allocate the guest's logical partition ID */
3935
3936         lpid = kvmppc_alloc_lpid();
3937         if ((long)lpid < 0)
3938                 return -ENOMEM;
3939         kvm->arch.lpid = lpid;
3940
3941         kvmppc_alloc_host_rm_ops();
3942
3943         /*
3944          * Since we don't flush the TLB when tearing down a VM,
3945          * and this lpid might have previously been used,
3946          * make sure we flush on each core before running the new VM.
3947          * On POWER9, the tlbie in mmu_partition_table_set_entry()
3948          * does this flush for us.
3949          */
3950         if (!cpu_has_feature(CPU_FTR_ARCH_300))
3951                 cpumask_setall(&kvm->arch.need_tlb_flush);
3952
3953         /* Start out with the default set of hcalls enabled */
3954         memcpy(kvm->arch.enabled_hcalls, default_enabled_hcalls,
3955                sizeof(kvm->arch.enabled_hcalls));
3956
3957         if (!cpu_has_feature(CPU_FTR_ARCH_300))
3958                 kvm->arch.host_sdr1 = mfspr(SPRN_SDR1);
3959
3960         /* Init LPCR for virtual RMA mode */
3961         kvm->arch.host_lpid = mfspr(SPRN_LPID);
3962         kvm->arch.host_lpcr = lpcr = mfspr(SPRN_LPCR);
3963         lpcr &= LPCR_PECE | LPCR_LPES;
3964         lpcr |= (4UL << LPCR_DPFD_SH) | LPCR_HDICE |
3965                 LPCR_VPM0 | LPCR_VPM1;
3966         kvm->arch.vrma_slb_v = SLB_VSID_B_1T |
3967                 (VRMA_VSID << SLB_VSID_SHIFT_1T);
3968         /* On POWER8 turn on online bit to enable PURR/SPURR */
3969         if (cpu_has_feature(CPU_FTR_ARCH_207S))
3970                 lpcr |= LPCR_ONL;
3971         /*
3972          * On POWER9, VPM0 bit is reserved (VPM0=1 behaviour is assumed)
3973          * Set HVICE bit to enable hypervisor virtualization interrupts.
3974          * Set HEIC to prevent OS interrupts to go to hypervisor (should
3975          * be unnecessary but better safe than sorry in case we re-enable
3976          * EE in HV mode with this LPCR still set)
3977          */
3978         if (cpu_has_feature(CPU_FTR_ARCH_300)) {
3979                 lpcr &= ~LPCR_VPM0;
3980                 lpcr |= LPCR_HVICE | LPCR_HEIC;
3981
3982                 /*
3983                  * If xive is enabled, we route 0x500 interrupts directly
3984                  * to the guest.
3985                  */
3986                 if (xive_enabled())
3987                         lpcr |= LPCR_LPES;
3988         }
3989
3990         /*
3991          * If the host uses radix, the guest starts out as radix.
3992          */
3993         if (radix_enabled()) {
3994                 kvm->arch.radix = 1;
3995                 kvm->arch.mmu_ready = 1;
3996                 lpcr &= ~LPCR_VPM1;
3997                 lpcr |= LPCR_UPRT | LPCR_GTSE | LPCR_HR;
3998                 ret = kvmppc_init_vm_radix(kvm);
3999                 if (ret) {
4000                         kvmppc_free_lpid(kvm->arch.lpid);
4001                         return ret;
4002                 }
4003                 kvmppc_setup_partition_table(kvm);
4004         }
4005
4006         kvm->arch.lpcr = lpcr;
4007
4008         /* Initialization for future HPT resizes */
4009         kvm->arch.resize_hpt = NULL;
4010
4011         /*
4012          * Work out how many sets the TLB has, for the use of
4013          * the TLB invalidation loop in book3s_hv_rmhandlers.S.
4014          */
4015         if (radix_enabled())
4016                 kvm->arch.tlb_sets = POWER9_TLB_SETS_RADIX;     /* 128 */
4017         else if (cpu_has_feature(CPU_FTR_ARCH_300))
4018                 kvm->arch.tlb_sets = POWER9_TLB_SETS_HASH;      /* 256 */
4019         else if (cpu_has_feature(CPU_FTR_ARCH_207S))
4020                 kvm->arch.tlb_sets = POWER8_TLB_SETS;           /* 512 */
4021         else
4022                 kvm->arch.tlb_sets = POWER7_TLB_SETS;           /* 128 */
4023
4024         /*
4025          * Track that we now have a HV mode VM active. This blocks secondary
4026          * CPU threads from coming online.
4027          * On POWER9, we only need to do this if the "indep_threads_mode"
4028          * module parameter has been set to N.
4029          */
4030         if (cpu_has_feature(CPU_FTR_ARCH_300))
4031                 kvm->arch.threads_indep = indep_threads_mode;
4032         if (!kvm->arch.threads_indep)
4033                 kvm_hv_vm_activated();
4034
4035         /*
4036          * Initialize smt_mode depending on processor.
4037          * POWER8 and earlier have to use "strict" threading, where
4038          * all vCPUs in a vcore have to run on the same (sub)core,
4039          * whereas on POWER9 the threads can each run a different
4040          * guest.
4041          */
4042         if (!cpu_has_feature(CPU_FTR_ARCH_300))
4043                 kvm->arch.smt_mode = threads_per_subcore;
4044         else
4045                 kvm->arch.smt_mode = 1;
4046         kvm->arch.emul_smt_mode = 1;
4047
4048         /*
4049          * Create a debugfs directory for the VM
4050          */
4051         snprintf(buf, sizeof(buf), "vm%d", current->pid);
4052         kvm->arch.debugfs_dir = debugfs_create_dir(buf, kvm_debugfs_dir);
4053         kvmppc_mmu_debugfs_init(kvm);
4054
4055         return 0;
4056 }
4057
4058 static void kvmppc_free_vcores(struct kvm *kvm)
4059 {
4060         long int i;
4061
4062         for (i = 0; i < KVM_MAX_VCORES; ++i)
4063                 kfree(kvm->arch.vcores[i]);
4064         kvm->arch.online_vcores = 0;
4065 }
4066
4067 static void kvmppc_core_destroy_vm_hv(struct kvm *kvm)
4068 {
4069         debugfs_remove_recursive(kvm->arch.debugfs_dir);
4070
4071         if (!kvm->arch.threads_indep)
4072                 kvm_hv_vm_deactivated();
4073
4074         kvmppc_free_vcores(kvm);
4075
4076         kvmppc_free_lpid(kvm->arch.lpid);
4077
4078         if (kvm_is_radix(kvm))
4079                 kvmppc_free_radix(kvm);
4080         else
4081                 kvmppc_free_hpt(&kvm->arch.hpt);
4082
4083         kvmppc_free_pimap(kvm);
4084 }
4085
4086 /* We don't need to emulate any privileged instructions or dcbz */
4087 static int kvmppc_core_emulate_op_hv(struct kvm_run *run, struct kvm_vcpu *vcpu,
4088                                      unsigned int inst, int *advance)
4089 {
4090         return EMULATE_FAIL;
4091 }
4092
4093 static int kvmppc_core_emulate_mtspr_hv(struct kvm_vcpu *vcpu, int sprn,
4094                                         ulong spr_val)
4095 {
4096         return EMULATE_FAIL;
4097 }
4098
4099 static int kvmppc_core_emulate_mfspr_hv(struct kvm_vcpu *vcpu, int sprn,
4100                                         ulong *spr_val)
4101 {
4102         return EMULATE_FAIL;
4103 }
4104
4105 static int kvmppc_core_check_processor_compat_hv(void)
4106 {
4107         if (!cpu_has_feature(CPU_FTR_HVMODE) ||
4108             !cpu_has_feature(CPU_FTR_ARCH_206))
4109                 return -EIO;
4110
4111         return 0;
4112 }
4113
4114 #ifdef CONFIG_KVM_XICS
4115
4116 void kvmppc_free_pimap(struct kvm *kvm)
4117 {
4118         kfree(kvm->arch.pimap);
4119 }
4120
4121 static struct kvmppc_passthru_irqmap *kvmppc_alloc_pimap(void)
4122 {
4123         return kzalloc(sizeof(struct kvmppc_passthru_irqmap), GFP_KERNEL);
4124 }
4125
4126 static int kvmppc_set_passthru_irq(struct kvm *kvm, int host_irq, int guest_gsi)
4127 {
4128         struct irq_desc *desc;
4129         struct kvmppc_irq_map *irq_map;
4130         struct kvmppc_passthru_irqmap *pimap;
4131         struct irq_chip *chip;
4132         int i, rc = 0;
4133
4134         if (!kvm_irq_bypass)
4135                 return 1;
4136
4137         desc = irq_to_desc(host_irq);
4138         if (!desc)
4139                 return -EIO;
4140
4141         mutex_lock(&kvm->lock);
4142
4143         pimap = kvm->arch.pimap;
4144         if (pimap == NULL) {
4145                 /* First call, allocate structure to hold IRQ map */
4146                 pimap = kvmppc_alloc_pimap();
4147                 if (pimap == NULL) {
4148                         mutex_unlock(&kvm->lock);
4149                         return -ENOMEM;
4150                 }
4151                 kvm->arch.pimap = pimap;
4152         }
4153
4154         /*
4155          * For now, we only support interrupts for which the EOI operation
4156          * is an OPAL call followed by a write to XIRR, since that's
4157          * what our real-mode EOI code does, or a XIVE interrupt
4158          */
4159         chip = irq_data_get_irq_chip(&desc->irq_data);
4160         if (!chip || !(is_pnv_opal_msi(chip) || is_xive_irq(chip))) {
4161                 pr_warn("kvmppc_set_passthru_irq_hv: Could not assign IRQ map for (%d,%d)\n",
4162                         host_irq, guest_gsi);
4163                 mutex_unlock(&kvm->lock);
4164                 return -ENOENT;
4165         }
4166
4167         /*
4168          * See if we already have an entry for this guest IRQ number.
4169          * If it's mapped to a hardware IRQ number, that's an error,
4170          * otherwise re-use this entry.
4171          */
4172         for (i = 0; i < pimap->n_mapped; i++) {
4173                 if (guest_gsi == pimap->mapped[i].v_hwirq) {
4174                         if (pimap->mapped[i].r_hwirq) {
4175                                 mutex_unlock(&kvm->lock);
4176                                 return -EINVAL;
4177                         }
4178                         break;
4179                 }
4180         }
4181
4182         if (i == KVMPPC_PIRQ_MAPPED) {
4183                 mutex_unlock(&kvm->lock);
4184                 return -EAGAIN;         /* table is full */
4185         }
4186
4187         irq_map = &pimap->mapped[i];
4188
4189         irq_map->v_hwirq = guest_gsi;
4190         irq_map->desc = desc;
4191
4192         /*
4193          * Order the above two stores before the next to serialize with
4194          * the KVM real mode handler.
4195          */
4196         smp_wmb();
4197         irq_map->r_hwirq = desc->irq_data.hwirq;
4198
4199         if (i == pimap->n_mapped)
4200                 pimap->n_mapped++;
4201
4202         if (xive_enabled())
4203                 rc = kvmppc_xive_set_mapped(kvm, guest_gsi, desc);
4204         else
4205                 kvmppc_xics_set_mapped(kvm, guest_gsi, desc->irq_data.hwirq);
4206         if (rc)
4207                 irq_map->r_hwirq = 0;
4208
4209         mutex_unlock(&kvm->lock);
4210
4211         return 0;
4212 }
4213
4214 static int kvmppc_clr_passthru_irq(struct kvm *kvm, int host_irq, int guest_gsi)
4215 {
4216         struct irq_desc *desc;
4217         struct kvmppc_passthru_irqmap *pimap;
4218         int i, rc = 0;
4219
4220         if (!kvm_irq_bypass)
4221                 return 0;
4222
4223         desc = irq_to_desc(host_irq);
4224         if (!desc)
4225                 return -EIO;
4226
4227         mutex_lock(&kvm->lock);
4228         if (!kvm->arch.pimap)
4229                 goto unlock;
4230
4231         pimap = kvm->arch.pimap;
4232
4233         for (i = 0; i < pimap->n_mapped; i++) {
4234                 if (guest_gsi == pimap->mapped[i].v_hwirq)
4235                         break;
4236         }
4237
4238         if (i == pimap->n_mapped) {
4239                 mutex_unlock(&kvm->lock);
4240                 return -ENODEV;
4241         }
4242
4243         if (xive_enabled())
4244                 rc = kvmppc_xive_clr_mapped(kvm, guest_gsi, pimap->mapped[i].desc);
4245         else
4246                 kvmppc_xics_clr_mapped(kvm, guest_gsi, pimap->mapped[i].r_hwirq);
4247
4248         /* invalidate the entry (what do do on error from the above ?) */
4249         pimap->mapped[i].r_hwirq = 0;
4250
4251         /*
4252          * We don't free this structure even when the count goes to
4253          * zero. The structure is freed when we destroy the VM.
4254          */
4255  unlock:
4256         mutex_unlock(&kvm->lock);
4257         return rc;
4258 }
4259
4260 static int kvmppc_irq_bypass_add_producer_hv(struct irq_bypass_consumer *cons,
4261                                              struct irq_bypass_producer *prod)
4262 {
4263         int ret = 0;
4264         struct kvm_kernel_irqfd *irqfd =
4265                 container_of(cons, struct kvm_kernel_irqfd, consumer);
4266
4267         irqfd->producer = prod;
4268
4269         ret = kvmppc_set_passthru_irq(irqfd->kvm, prod->irq, irqfd->gsi);
4270         if (ret)
4271                 pr_info("kvmppc_set_passthru_irq (irq %d, gsi %d) fails: %d\n",
4272                         prod->irq, irqfd->gsi, ret);
4273
4274         return ret;
4275 }
4276
4277 static void kvmppc_irq_bypass_del_producer_hv(struct irq_bypass_consumer *cons,
4278                                               struct irq_bypass_producer *prod)
4279 {
4280         int ret;
4281         struct kvm_kernel_irqfd *irqfd =
4282                 container_of(cons, struct kvm_kernel_irqfd, consumer);
4283
4284         irqfd->producer = NULL;
4285
4286         /*
4287          * When producer of consumer is unregistered, we change back to
4288          * default external interrupt handling mode - KVM real mode
4289          * will switch back to host.
4290          */
4291         ret = kvmppc_clr_passthru_irq(irqfd->kvm, prod->irq, irqfd->gsi);
4292         if (ret)
4293                 pr_warn("kvmppc_clr_passthru_irq (irq %d, gsi %d) fails: %d\n",
4294                         prod->irq, irqfd->gsi, ret);
4295 }
4296 #endif
4297
4298 static long kvm_arch_vm_ioctl_hv(struct file *filp,
4299                                  unsigned int ioctl, unsigned long arg)
4300 {
4301         struct kvm *kvm __maybe_unused = filp->private_data;
4302         void __user *argp = (void __user *)arg;
4303         long r;
4304
4305         switch (ioctl) {
4306
4307         case KVM_PPC_ALLOCATE_HTAB: {
4308                 u32 htab_order;
4309
4310                 r = -EFAULT;
4311                 if (get_user(htab_order, (u32 __user *)argp))
4312                         break;
4313                 r = kvmppc_alloc_reset_hpt(kvm, htab_order);
4314                 if (r)
4315                         break;
4316                 r = 0;
4317                 break;
4318         }
4319
4320         case KVM_PPC_GET_HTAB_FD: {
4321                 struct kvm_get_htab_fd ghf;
4322
4323                 r = -EFAULT;
4324                 if (copy_from_user(&ghf, argp, sizeof(ghf)))
4325                         break;
4326                 r = kvm_vm_ioctl_get_htab_fd(kvm, &ghf);
4327                 break;
4328         }
4329
4330         case KVM_PPC_RESIZE_HPT_PREPARE: {
4331                 struct kvm_ppc_resize_hpt rhpt;
4332
4333                 r = -EFAULT;
4334                 if (copy_from_user(&rhpt, argp, sizeof(rhpt)))
4335                         break;
4336
4337                 r = kvm_vm_ioctl_resize_hpt_prepare(kvm, &rhpt);
4338                 break;
4339         }
4340
4341         case KVM_PPC_RESIZE_HPT_COMMIT: {
4342                 struct kvm_ppc_resize_hpt rhpt;
4343
4344                 r = -EFAULT;
4345                 if (copy_from_user(&rhpt, argp, sizeof(rhpt)))
4346                         break;
4347
4348                 r = kvm_vm_ioctl_resize_hpt_commit(kvm, &rhpt);
4349                 break;
4350         }
4351
4352         default:
4353                 r = -ENOTTY;
4354         }
4355
4356         return r;
4357 }
4358
4359 /*
4360  * List of hcall numbers to enable by default.
4361  * For compatibility with old userspace, we enable by default
4362  * all hcalls that were implemented before the hcall-enabling
4363  * facility was added.  Note this list should not include H_RTAS.
4364  */
4365 static unsigned int default_hcall_list[] = {
4366         H_REMOVE,
4367         H_ENTER,
4368         H_READ,
4369         H_PROTECT,
4370         H_BULK_REMOVE,
4371         H_GET_TCE,
4372         H_PUT_TCE,
4373         H_SET_DABR,
4374         H_SET_XDABR,
4375         H_CEDE,
4376         H_PROD,
4377         H_CONFER,
4378         H_REGISTER_VPA,
4379 #ifdef CONFIG_KVM_XICS
4380         H_EOI,
4381         H_CPPR,
4382         H_IPI,
4383         H_IPOLL,
4384         H_XIRR,
4385         H_XIRR_X,
4386 #endif
4387         0
4388 };
4389
4390 static void init_default_hcalls(void)
4391 {
4392         int i;
4393         unsigned int hcall;
4394
4395         for (i = 0; default_hcall_list[i]; ++i) {
4396                 hcall = default_hcall_list[i];
4397                 WARN_ON(!kvmppc_hcall_impl_hv(hcall));
4398                 __set_bit(hcall / 4, default_enabled_hcalls);
4399         }
4400 }
4401
4402 static int kvmhv_configure_mmu(struct kvm *kvm, struct kvm_ppc_mmuv3_cfg *cfg)
4403 {
4404         unsigned long lpcr;
4405         int radix;
4406         int err;
4407
4408         /* If not on a POWER9, reject it */
4409         if (!cpu_has_feature(CPU_FTR_ARCH_300))
4410                 return -ENODEV;
4411
4412         /* If any unknown flags set, reject it */
4413         if (cfg->flags & ~(KVM_PPC_MMUV3_RADIX | KVM_PPC_MMUV3_GTSE))
4414                 return -EINVAL;
4415
4416         /* GR (guest radix) bit in process_table field must match */
4417         radix = !!(cfg->flags & KVM_PPC_MMUV3_RADIX);
4418         if (!!(cfg->process_table & PATB_GR) != radix)
4419                 return -EINVAL;
4420
4421         /* Process table size field must be reasonable, i.e. <= 24 */
4422         if ((cfg->process_table & PRTS_MASK) > 24)
4423                 return -EINVAL;
4424
4425         /* We can change a guest to/from radix now, if the host is radix */
4426         if (radix && !radix_enabled())
4427                 return -EINVAL;
4428
4429         mutex_lock(&kvm->lock);
4430         if (radix != kvm_is_radix(kvm)) {
4431                 if (kvm->arch.mmu_ready) {
4432                         kvm->arch.mmu_ready = 0;
4433                         /* order mmu_ready vs. vcpus_running */
4434                         smp_mb();
4435                         if (atomic_read(&kvm->arch.vcpus_running)) {
4436                                 kvm->arch.mmu_ready = 1;
4437                                 err = -EBUSY;
4438                                 goto out_unlock;
4439                         }
4440                 }
4441                 if (radix)
4442                         err = kvmppc_switch_mmu_to_radix(kvm);
4443                 else
4444                         err = kvmppc_switch_mmu_to_hpt(kvm);
4445                 if (err)
4446                         goto out_unlock;
4447         }
4448
4449         kvm->arch.process_table = cfg->process_table;
4450         kvmppc_setup_partition_table(kvm);
4451
4452         lpcr = (cfg->flags & KVM_PPC_MMUV3_GTSE) ? LPCR_GTSE : 0;
4453         kvmppc_update_lpcr(kvm, lpcr, LPCR_GTSE);
4454         err = 0;
4455
4456  out_unlock:
4457         mutex_unlock(&kvm->lock);
4458         return err;
4459 }
4460
4461 static struct kvmppc_ops kvm_ops_hv = {
4462         .get_sregs = kvm_arch_vcpu_ioctl_get_sregs_hv,
4463         .set_sregs = kvm_arch_vcpu_ioctl_set_sregs_hv,
4464         .get_one_reg = kvmppc_get_one_reg_hv,
4465         .set_one_reg = kvmppc_set_one_reg_hv,
4466         .vcpu_load   = kvmppc_core_vcpu_load_hv,
4467         .vcpu_put    = kvmppc_core_vcpu_put_hv,
4468         .set_msr     = kvmppc_set_msr_hv,
4469         .vcpu_run    = kvmppc_vcpu_run_hv,
4470         .vcpu_create = kvmppc_core_vcpu_create_hv,
4471         .vcpu_free   = kvmppc_core_vcpu_free_hv,
4472         .check_requests = kvmppc_core_check_requests_hv,
4473         .get_dirty_log  = kvm_vm_ioctl_get_dirty_log_hv,
4474         .flush_memslot  = kvmppc_core_flush_memslot_hv,
4475         .prepare_memory_region = kvmppc_core_prepare_memory_region_hv,
4476         .commit_memory_region  = kvmppc_core_commit_memory_region_hv,
4477         .unmap_hva_range = kvm_unmap_hva_range_hv,
4478         .age_hva  = kvm_age_hva_hv,
4479         .test_age_hva = kvm_test_age_hva_hv,
4480         .set_spte_hva = kvm_set_spte_hva_hv,
4481         .mmu_destroy  = kvmppc_mmu_destroy_hv,
4482         .free_memslot = kvmppc_core_free_memslot_hv,
4483         .create_memslot = kvmppc_core_create_memslot_hv,
4484         .init_vm =  kvmppc_core_init_vm_hv,
4485         .destroy_vm = kvmppc_core_destroy_vm_hv,
4486         .get_smmu_info = kvm_vm_ioctl_get_smmu_info_hv,
4487         .emulate_op = kvmppc_core_emulate_op_hv,
4488         .emulate_mtspr = kvmppc_core_emulate_mtspr_hv,
4489         .emulate_mfspr = kvmppc_core_emulate_mfspr_hv,
4490         .fast_vcpu_kick = kvmppc_fast_vcpu_kick_hv,
4491         .arch_vm_ioctl  = kvm_arch_vm_ioctl_hv,
4492         .hcall_implemented = kvmppc_hcall_impl_hv,
4493 #ifdef CONFIG_KVM_XICS
4494         .irq_bypass_add_producer = kvmppc_irq_bypass_add_producer_hv,
4495         .irq_bypass_del_producer = kvmppc_irq_bypass_del_producer_hv,
4496 #endif
4497         .configure_mmu = kvmhv_configure_mmu,
4498         .get_rmmu_info = kvmhv_get_rmmu_info,
4499         .set_smt_mode = kvmhv_set_smt_mode,
4500 };
4501
4502 static int kvm_init_subcore_bitmap(void)
4503 {
4504         int i, j;
4505         int nr_cores = cpu_nr_cores();
4506         struct sibling_subcore_state *sibling_subcore_state;
4507
4508         for (i = 0; i < nr_cores; i++) {
4509                 int first_cpu = i * threads_per_core;
4510                 int node = cpu_to_node(first_cpu);
4511
4512                 /* Ignore if it is already allocated. */
4513                 if (paca_ptrs[first_cpu]->sibling_subcore_state)
4514                         continue;
4515
4516                 sibling_subcore_state =
4517                         kmalloc_node(sizeof(struct sibling_subcore_state),
4518                                                         GFP_KERNEL, node);
4519                 if (!sibling_subcore_state)
4520                         return -ENOMEM;
4521
4522                 memset(sibling_subcore_state, 0,
4523                                 sizeof(struct sibling_subcore_state));
4524
4525                 for (j = 0; j < threads_per_core; j++) {
4526                         int cpu = first_cpu + j;
4527
4528                         paca_ptrs[cpu]->sibling_subcore_state =
4529                                                 sibling_subcore_state;
4530                 }
4531         }
4532         return 0;
4533 }
4534
4535 static int kvmppc_radix_possible(void)
4536 {
4537         return cpu_has_feature(CPU_FTR_ARCH_300) && radix_enabled();
4538 }
4539
4540 static int kvmppc_book3s_init_hv(void)
4541 {
4542         int r;
4543         /*
4544          * FIXME!! Do we need to check on all cpus ?
4545          */
4546         r = kvmppc_core_check_processor_compat_hv();
4547         if (r < 0)
4548                 return -ENODEV;
4549
4550         r = kvm_init_subcore_bitmap();
4551         if (r)
4552                 return r;
4553
4554         /*
4555          * We need a way of accessing the XICS interrupt controller,
4556          * either directly, via paca_ptrs[cpu]->kvm_hstate.xics_phys, or
4557          * indirectly, via OPAL.
4558          */
4559 #ifdef CONFIG_SMP
4560         if (!xive_enabled() && !local_paca->kvm_hstate.xics_phys) {
4561                 struct device_node *np;
4562
4563                 np = of_find_compatible_node(NULL, NULL, "ibm,opal-intc");
4564                 if (!np) {
4565                         pr_err("KVM-HV: Cannot determine method for accessing XICS\n");
4566                         return -ENODEV;
4567                 }
4568                 /* presence of intc confirmed - node can be dropped again */
4569                 of_node_put(np);
4570         }
4571 #endif
4572
4573         kvm_ops_hv.owner = THIS_MODULE;
4574         kvmppc_hv_ops = &kvm_ops_hv;
4575
4576         init_default_hcalls();
4577
4578         init_vcore_lists();
4579
4580         r = kvmppc_mmu_hv_init();
4581         if (r)
4582                 return r;
4583
4584         if (kvmppc_radix_possible())
4585                 r = kvmppc_radix_init();
4586
4587         /*
4588          * POWER9 chips before version 2.02 can't have some threads in
4589          * HPT mode and some in radix mode on the same core.
4590          */
4591         if (cpu_has_feature(CPU_FTR_ARCH_300)) {
4592                 unsigned int pvr = mfspr(SPRN_PVR);
4593                 if ((pvr >> 16) == PVR_POWER9 &&
4594                     (((pvr & 0xe000) == 0 && (pvr & 0xfff) < 0x202) ||
4595                      ((pvr & 0xe000) == 0x2000 && (pvr & 0xfff) < 0x101)))
4596                         no_mixing_hpt_and_radix = true;
4597         }
4598
4599         return r;
4600 }
4601
4602 static void kvmppc_book3s_exit_hv(void)
4603 {
4604         kvmppc_free_host_rm_ops();
4605         if (kvmppc_radix_possible())
4606                 kvmppc_radix_exit();
4607         kvmppc_hv_ops = NULL;
4608 }
4609
4610 module_init(kvmppc_book3s_init_hv);
4611 module_exit(kvmppc_book3s_exit_hv);
4612 MODULE_LICENSE("GPL");
4613 MODULE_ALIAS_MISCDEV(KVM_MINOR);
4614 MODULE_ALIAS("devname:kvm");