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/reg.h>
  50 #include <asm/ppc-opcode.h>
  51 #include <asm/asm-prototypes.h>
  52 #include <asm/debug.h>
  53 #include <asm/disassemble.h>
  54 #include <asm/cputable.h>
  55 #include <asm/cacheflush.h>
  56 #include <asm/tlbflush.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     0x164520C62609AECA
 132 #define RWMR_RPA_P8_2THREAD     0x7FFF2908450D8DA9
 133 #define RWMR_RPA_P8_3THREAD     0x164520C62609AECA
 134 #define RWMR_RPA_P8_4THREAD     0x199A421245058DA9
 135 #define RWMR_RPA_P8_5THREAD     0x164520C62609AECA
 136 #define RWMR_RPA_P8_6THREAD     0x164520C62609AECA
 137 #define RWMR_RPA_P8_7THREAD     0x164520C62609AECA
 138 #define RWMR_RPA_P8_8THREAD     0x164520C62609AECA
 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(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                 /*
1697                  * POWER9 DD1 has an erratum where writing TBU40 causes
1698                  * the timebase to lose ticks.  So we don't let the
1699                  * timebase offset be changed on P9 DD1.  (It is
1700                  * initialized to zero.)
1701                  */
1702                 if (cpu_has_feature(CPU_FTR_POWER9_DD1))
1703                         break;
1704                 /* round up to multiple of 2^24 */
1705                 vcpu->arch.vcore->tb_offset =
1706                         ALIGN(set_reg_val(id, *val), 1UL << 24);
1707                 break;
1708         case KVM_REG_PPC_LPCR:
1709                 kvmppc_set_lpcr(vcpu, set_reg_val(id, *val), true);
1710                 break;
1711         case KVM_REG_PPC_LPCR_64:
1712                 kvmppc_set_lpcr(vcpu, set_reg_val(id, *val), false);
1713                 break;
1714         case KVM_REG_PPC_PPR:
1715                 vcpu->arch.ppr = set_reg_val(id, *val);
1716                 break;
1717 #ifdef CONFIG_PPC_TRANSACTIONAL_MEM
1718         case KVM_REG_PPC_TFHAR:
1719                 vcpu->arch.tfhar = set_reg_val(id, *val);
1720                 break;
1721         case KVM_REG_PPC_TFIAR:
1722                 vcpu->arch.tfiar = set_reg_val(id, *val);
1723                 break;
1724         case KVM_REG_PPC_TEXASR:
1725                 vcpu->arch.texasr = set_reg_val(id, *val);
1726                 break;
1727         case KVM_REG_PPC_TM_GPR0 ... KVM_REG_PPC_TM_GPR31:
1728                 i = id - KVM_REG_PPC_TM_GPR0;
1729                 vcpu->arch.gpr_tm[i] = set_reg_val(id, *val);
1730                 break;
1731         case KVM_REG_PPC_TM_VSR0 ... KVM_REG_PPC_TM_VSR63:
1732         {
1733                 int j;
1734                 i = id - KVM_REG_PPC_TM_VSR0;
1735                 if (i < 32)
1736                         for (j = 0; j < TS_FPRWIDTH; j++)
1737                                 vcpu->arch.fp_tm.fpr[i][j] = val->vsxval[j];
1738                 else
1739                         if (cpu_has_feature(CPU_FTR_ALTIVEC))
1740                                 vcpu->arch.vr_tm.vr[i-32] = val->vval;
1741                         else
1742                                 r = -ENXIO;
1743                 break;
1744         }
1745         case KVM_REG_PPC_TM_CR:
1746                 vcpu->arch.cr_tm = set_reg_val(id, *val);
1747                 break;
1748         case KVM_REG_PPC_TM_XER:
1749                 vcpu->arch.xer_tm = set_reg_val(id, *val);
1750                 break;
1751         case KVM_REG_PPC_TM_LR:
1752                 vcpu->arch.lr_tm = set_reg_val(id, *val);
1753                 break;
1754         case KVM_REG_PPC_TM_CTR:
1755                 vcpu->arch.ctr_tm = set_reg_val(id, *val);
1756                 break;
1757         case KVM_REG_PPC_TM_FPSCR:
1758                 vcpu->arch.fp_tm.fpscr = set_reg_val(id, *val);
1759                 break;
1760         case KVM_REG_PPC_TM_AMR:
1761                 vcpu->arch.amr_tm = set_reg_val(id, *val);
1762                 break;
1763         case KVM_REG_PPC_TM_PPR:
1764                 vcpu->arch.ppr_tm = set_reg_val(id, *val);
1765                 break;
1766         case KVM_REG_PPC_TM_VRSAVE:
1767                 vcpu->arch.vrsave_tm = set_reg_val(id, *val);
1768                 break;
1769         case KVM_REG_PPC_TM_VSCR:
1770                 if (cpu_has_feature(CPU_FTR_ALTIVEC))
1771                         vcpu->arch.vr.vscr.u[3] = set_reg_val(id, *val);
1772                 else
1773                         r = - ENXIO;
1774                 break;
1775         case KVM_REG_PPC_TM_DSCR:
1776                 vcpu->arch.dscr_tm = set_reg_val(id, *val);
1777                 break;
1778         case KVM_REG_PPC_TM_TAR:
1779                 vcpu->arch.tar_tm = set_reg_val(id, *val);
1780                 break;
1781 #endif
1782         case KVM_REG_PPC_ARCH_COMPAT:
1783                 r = kvmppc_set_arch_compat(vcpu, set_reg_val(id, *val));
1784                 break;
1785         case KVM_REG_PPC_DEC_EXPIRY:
1786                 vcpu->arch.dec_expires = set_reg_val(id, *val) -
1787                         vcpu->arch.vcore->tb_offset;
1788                 break;
1789         case KVM_REG_PPC_ONLINE:
1790                 i = set_reg_val(id, *val);
1791                 if (i && !vcpu->arch.online)
1792                         atomic_inc(&vcpu->arch.vcore->online_count);
1793                 else if (!i && vcpu->arch.online)
1794                         atomic_dec(&vcpu->arch.vcore->online_count);
1795                 vcpu->arch.online = i;
1796                 break;
1797         default:
1798                 r = -EINVAL;
1799                 break;
1800         }
1801
1802         return r;
1803 }
1804
1805 /*
1806  * On POWER9, threads are independent and can be in different partitions.
1807  * Therefore we consider each thread to be a subcore.
1808  * There is a restriction that all threads have to be in the same
1809  * MMU mode (radix or HPT), unfortunately, but since we only support
1810  * HPT guests on a HPT host so far, that isn't an impediment yet.
1811  */
1812 static int threads_per_vcore(struct kvm *kvm)
1813 {
1814         if (kvm->arch.threads_indep)
1815                 return 1;
1816         return threads_per_subcore;
1817 }
1818
1819 static struct kvmppc_vcore *kvmppc_vcore_create(struct kvm *kvm, int core)
1820 {
1821         struct kvmppc_vcore *vcore;
1822
1823         vcore = kzalloc(sizeof(struct kvmppc_vcore), GFP_KERNEL);
1824
1825         if (vcore == NULL)
1826                 return NULL;
1827
1828         spin_lock_init(&vcore->lock);
1829         spin_lock_init(&vcore->stoltb_lock);
1830         init_swait_queue_head(&vcore->wq);
1831         vcore->preempt_tb = TB_NIL;
1832         vcore->lpcr = kvm->arch.lpcr;
1833         vcore->first_vcpuid = core * kvm->arch.smt_mode;
1834         vcore->kvm = kvm;
1835         INIT_LIST_HEAD(&vcore->preempt_list);
1836
1837         return vcore;
1838 }
1839
1840 #ifdef CONFIG_KVM_BOOK3S_HV_EXIT_TIMING
1841 static struct debugfs_timings_element {
1842         const char *name;
1843         size_t offset;
1844 } timings[] = {
1845         {"rm_entry",    offsetof(struct kvm_vcpu, arch.rm_entry)},
1846         {"rm_intr",     offsetof(struct kvm_vcpu, arch.rm_intr)},
1847         {"rm_exit",     offsetof(struct kvm_vcpu, arch.rm_exit)},
1848         {"guest",       offsetof(struct kvm_vcpu, arch.guest_time)},
1849         {"cede",        offsetof(struct kvm_vcpu, arch.cede_time)},
1850 };
1851
1852 #define N_TIMINGS       (ARRAY_SIZE(timings))
1853
1854 struct debugfs_timings_state {
1855         struct kvm_vcpu *vcpu;
1856         unsigned int    buflen;
1857         char            buf[N_TIMINGS * 100];
1858 };
1859
1860 static int debugfs_timings_open(struct inode *inode, struct file *file)
1861 {
1862         struct kvm_vcpu *vcpu = inode->i_private;
1863         struct debugfs_timings_state *p;
1864
1865         p = kzalloc(sizeof(*p), GFP_KERNEL);
1866         if (!p)
1867                 return -ENOMEM;
1868
1869         kvm_get_kvm(vcpu->kvm);
1870         p->vcpu = vcpu;
1871         file->private_data = p;
1872
1873         return nonseekable_open(inode, file);
1874 }
1875
1876 static int debugfs_timings_release(struct inode *inode, struct file *file)
1877 {
1878         struct debugfs_timings_state *p = file->private_data;
1879
1880         kvm_put_kvm(p->vcpu->kvm);
1881         kfree(p);
1882         return 0;
1883 }
1884
1885 static ssize_t debugfs_timings_read(struct file *file, char __user *buf,
1886                                     size_t len, loff_t *ppos)
1887 {
1888         struct debugfs_timings_state *p = file->private_data;
1889         struct kvm_vcpu *vcpu = p->vcpu;
1890         char *s, *buf_end;
1891         struct kvmhv_tb_accumulator tb;
1892         u64 count;
1893         loff_t pos;
1894         ssize_t n;
1895         int i, loops;
1896         bool ok;
1897
1898         if (!p->buflen) {
1899                 s = p->buf;
1900                 buf_end = s + sizeof(p->buf);
1901                 for (i = 0; i < N_TIMINGS; ++i) {
1902                         struct kvmhv_tb_accumulator *acc;
1903
1904                         acc = (struct kvmhv_tb_accumulator *)
1905                                 ((unsigned long)vcpu + timings[i].offset);
1906                         ok = false;
1907                         for (loops = 0; loops < 1000; ++loops) {
1908                                 count = acc->seqcount;
1909                                 if (!(count & 1)) {
1910                                         smp_rmb();
1911                                         tb = *acc;
1912                                         smp_rmb();
1913                                         if (count == acc->seqcount) {
1914                                                 ok = true;
1915                                                 break;
1916                                         }
1917                                 }
1918                                 udelay(1);
1919                         }
1920                         if (!ok)
1921                                 snprintf(s, buf_end - s, "%s: stuck\n",
1922                                         timings[i].name);
1923                         else
1924                                 snprintf(s, buf_end - s,
1925                                         "%s: %llu %llu %llu %llu\n",
1926                                         timings[i].name, count / 2,
1927                                         tb_to_ns(tb.tb_total),
1928                                         tb_to_ns(tb.tb_min),
1929                                         tb_to_ns(tb.tb_max));
1930                         s += strlen(s);
1931                 }
1932                 p->buflen = s - p->buf;
1933         }
1934
1935         pos = *ppos;
1936         if (pos >= p->buflen)
1937                 return 0;
1938         if (len > p->buflen - pos)
1939                 len = p->buflen - pos;
1940         n = copy_to_user(buf, p->buf + pos, len);
1941         if (n) {
1942                 if (n == len)
1943                         return -EFAULT;
1944                 len -= n;
1945         }
1946         *ppos = pos + len;
1947         return len;
1948 }
1949
1950 static ssize_t debugfs_timings_write(struct file *file, const char __user *buf,
1951                                      size_t len, loff_t *ppos)
1952 {
1953         return -EACCES;
1954 }
1955
1956 static const struct file_operations debugfs_timings_ops = {
1957         .owner   = THIS_MODULE,
1958         .open    = debugfs_timings_open,
1959         .release = debugfs_timings_release,
1960         .read    = debugfs_timings_read,
1961         .write   = debugfs_timings_write,
1962         .llseek  = generic_file_llseek,
1963 };
1964
1965 /* Create a debugfs directory for the vcpu */
1966 static void debugfs_vcpu_init(struct kvm_vcpu *vcpu, unsigned int id)
1967 {
1968         char buf[16];
1969         struct kvm *kvm = vcpu->kvm;
1970
1971         snprintf(buf, sizeof(buf), "vcpu%u", id);
1972         if (IS_ERR_OR_NULL(kvm->arch.debugfs_dir))
1973                 return;
1974         vcpu->arch.debugfs_dir = debugfs_create_dir(buf, kvm->arch.debugfs_dir);
1975         if (IS_ERR_OR_NULL(vcpu->arch.debugfs_dir))
1976                 return;
1977         vcpu->arch.debugfs_timings =
1978                 debugfs_create_file("timings", 0444, vcpu->arch.debugfs_dir,
1979                                     vcpu, &debugfs_timings_ops);
1980 }
1981
1982 #else /* CONFIG_KVM_BOOK3S_HV_EXIT_TIMING */
1983 static void debugfs_vcpu_init(struct kvm_vcpu *vcpu, unsigned int id)
1984 {
1985 }
1986 #endif /* CONFIG_KVM_BOOK3S_HV_EXIT_TIMING */
1987
1988 static struct kvm_vcpu *kvmppc_core_vcpu_create_hv(struct kvm *kvm,
1989                                                    unsigned int id)
1990 {
1991         struct kvm_vcpu *vcpu;
1992         int err;
1993         int core;
1994         struct kvmppc_vcore *vcore;
1995
1996         err = -ENOMEM;
1997         vcpu = kmem_cache_zalloc(kvm_vcpu_cache, GFP_KERNEL);
1998         if (!vcpu)
1999                 goto out;
2000
2001         err = kvm_vcpu_init(vcpu, kvm, id);
2002         if (err)
2003                 goto free_vcpu;
2004
2005         vcpu->arch.shared = &vcpu->arch.shregs;
2006 #ifdef CONFIG_KVM_BOOK3S_PR_POSSIBLE
2007         /*
2008          * The shared struct is never shared on HV,
2009          * so we can always use host endianness
2010          */
2011 #ifdef __BIG_ENDIAN__
2012         vcpu->arch.shared_big_endian = true;
2013 #else
2014         vcpu->arch.shared_big_endian = false;
2015 #endif
2016 #endif
2017         vcpu->arch.mmcr[0] = MMCR0_FC;
2018         vcpu->arch.ctrl = CTRL_RUNLATCH;
2019         /* default to host PVR, since we can't spoof it */
2020         kvmppc_set_pvr_hv(vcpu, mfspr(SPRN_PVR));
2021         spin_lock_init(&vcpu->arch.vpa_update_lock);
2022         spin_lock_init(&vcpu->arch.tbacct_lock);
2023         vcpu->arch.busy_preempt = TB_NIL;
2024         vcpu->arch.intr_msr = MSR_SF | MSR_ME;
2025
2026         /*
2027          * Set the default HFSCR for the guest from the host value.
2028          * This value is only used on POWER9.
2029          * On POWER9 DD1, TM doesn't work, so we make sure to
2030          * prevent the guest from using it.
2031          * On POWER9, we want to virtualize the doorbell facility, so we
2032          * turn off the HFSCR bit, which causes those instructions to trap.
2033          */
2034         vcpu->arch.hfscr = mfspr(SPRN_HFSCR);
2035         if (cpu_has_feature(CPU_FTR_P9_TM_HV_ASSIST))
2036                 vcpu->arch.hfscr |= HFSCR_TM;
2037         else if (!cpu_has_feature(CPU_FTR_TM_COMP))
2038                 vcpu->arch.hfscr &= ~HFSCR_TM;
2039         if (cpu_has_feature(CPU_FTR_ARCH_300))
2040                 vcpu->arch.hfscr &= ~HFSCR_MSGP;
2041
2042         kvmppc_mmu_book3s_hv_init(vcpu);
2043
2044         vcpu->arch.state = KVMPPC_VCPU_NOTREADY;
2045
2046         init_waitqueue_head(&vcpu->arch.cpu_run);
2047
2048         mutex_lock(&kvm->lock);
2049         vcore = NULL;
2050         err = -EINVAL;
2051         core = id / kvm->arch.smt_mode;
2052         if (core < KVM_MAX_VCORES) {
2053                 vcore = kvm->arch.vcores[core];
2054                 if (!vcore) {
2055                         err = -ENOMEM;
2056                         vcore = kvmppc_vcore_create(kvm, core);
2057                         kvm->arch.vcores[core] = vcore;
2058                         kvm->arch.online_vcores++;
2059                 }
2060         }
2061         mutex_unlock(&kvm->lock);
2062
2063         if (!vcore)
2064                 goto free_vcpu;
2065
2066         spin_lock(&vcore->lock);
2067         ++vcore->num_threads;
2068         spin_unlock(&vcore->lock);
2069         vcpu->arch.vcore = vcore;
2070         vcpu->arch.ptid = vcpu->vcpu_id - vcore->first_vcpuid;
2071         vcpu->arch.thread_cpu = -1;
2072         vcpu->arch.prev_cpu = -1;
2073
2074         vcpu->arch.cpu_type = KVM_CPU_3S_64;
2075         kvmppc_sanity_check(vcpu);
2076
2077         debugfs_vcpu_init(vcpu, id);
2078
2079         return vcpu;
2080
2081 free_vcpu:
2082         kmem_cache_free(kvm_vcpu_cache, vcpu);
2083 out:
2084         return ERR_PTR(err);
2085 }
2086
2087 static int kvmhv_set_smt_mode(struct kvm *kvm, unsigned long smt_mode,
2088                               unsigned long flags)
2089 {
2090         int err;
2091         int esmt = 0;
2092
2093         if (flags)
2094                 return -EINVAL;
2095         if (smt_mode > MAX_SMT_THREADS || !is_power_of_2(smt_mode))
2096                 return -EINVAL;
2097         if (!cpu_has_feature(CPU_FTR_ARCH_300)) {
2098                 /*
2099                  * On POWER8 (or POWER7), the threading mode is "strict",
2100                  * so we pack smt_mode vcpus per vcore.
2101                  */
2102                 if (smt_mode > threads_per_subcore)
2103                         return -EINVAL;
2104         } else {
2105                 /*
2106                  * On POWER9, the threading mode is "loose",
2107                  * so each vcpu gets its own vcore.
2108                  */
2109                 esmt = smt_mode;
2110                 smt_mode = 1;
2111         }
2112         mutex_lock(&kvm->lock);
2113         err = -EBUSY;
2114         if (!kvm->arch.online_vcores) {
2115                 kvm->arch.smt_mode = smt_mode;
2116                 kvm->arch.emul_smt_mode = esmt;
2117                 err = 0;
2118         }
2119         mutex_unlock(&kvm->lock);
2120
2121         return err;
2122 }
2123
2124 static void unpin_vpa(struct kvm *kvm, struct kvmppc_vpa *vpa)
2125 {
2126         if (vpa->pinned_addr)
2127                 kvmppc_unpin_guest_page(kvm, vpa->pinned_addr, vpa->gpa,
2128                                         vpa->dirty);
2129 }
2130
2131 static void kvmppc_core_vcpu_free_hv(struct kvm_vcpu *vcpu)
2132 {
2133         spin_lock(&vcpu->arch.vpa_update_lock);
2134         unpin_vpa(vcpu->kvm, &vcpu->arch.dtl);
2135         unpin_vpa(vcpu->kvm, &vcpu->arch.slb_shadow);
2136         unpin_vpa(vcpu->kvm, &vcpu->arch.vpa);
2137         spin_unlock(&vcpu->arch.vpa_update_lock);
2138         kvm_vcpu_uninit(vcpu);
2139         kmem_cache_free(kvm_vcpu_cache, vcpu);
2140 }
2141
2142 static int kvmppc_core_check_requests_hv(struct kvm_vcpu *vcpu)
2143 {
2144         /* Indicate we want to get back into the guest */
2145         return 1;
2146 }
2147
2148 static void kvmppc_set_timer(struct kvm_vcpu *vcpu)
2149 {
2150         unsigned long dec_nsec, now;
2151
2152         now = get_tb();
2153         if (now > vcpu->arch.dec_expires) {
2154                 /* decrementer has already gone negative */
2155                 kvmppc_core_queue_dec(vcpu);
2156                 kvmppc_core_prepare_to_enter(vcpu);
2157                 return;
2158         }
2159         dec_nsec = (vcpu->arch.dec_expires - now) * NSEC_PER_SEC
2160                    / tb_ticks_per_sec;
2161         hrtimer_start(&vcpu->arch.dec_timer, dec_nsec, HRTIMER_MODE_REL);
2162         vcpu->arch.timer_running = 1;
2163 }
2164
2165 static void kvmppc_end_cede(struct kvm_vcpu *vcpu)
2166 {
2167         vcpu->arch.ceded = 0;
2168         if (vcpu->arch.timer_running) {
2169                 hrtimer_try_to_cancel(&vcpu->arch.dec_timer);
2170                 vcpu->arch.timer_running = 0;
2171         }
2172 }
2173
2174 extern int __kvmppc_vcore_entry(void);
2175
2176 static void kvmppc_remove_runnable(struct kvmppc_vcore *vc,
2177                                    struct kvm_vcpu *vcpu)
2178 {
2179         u64 now;
2180
2181         if (vcpu->arch.state != KVMPPC_VCPU_RUNNABLE)
2182                 return;
2183         spin_lock_irq(&vcpu->arch.tbacct_lock);
2184         now = mftb();
2185         vcpu->arch.busy_stolen += vcore_stolen_time(vc, now) -
2186                 vcpu->arch.stolen_logged;
2187         vcpu->arch.busy_preempt = now;
2188         vcpu->arch.state = KVMPPC_VCPU_BUSY_IN_HOST;
2189         spin_unlock_irq(&vcpu->arch.tbacct_lock);
2190         --vc->n_runnable;
2191         WRITE_ONCE(vc->runnable_threads[vcpu->arch.ptid], NULL);
2192 }
2193
2194 static int kvmppc_grab_hwthread(int cpu)
2195 {
2196         struct paca_struct *tpaca;
2197         long timeout = 10000;
2198
2199         tpaca = paca_ptrs[cpu];
2200
2201         /* Ensure the thread won't go into the kernel if it wakes */
2202         tpaca->kvm_hstate.kvm_vcpu = NULL;
2203         tpaca->kvm_hstate.kvm_vcore = NULL;
2204         tpaca->kvm_hstate.napping = 0;
2205         smp_wmb();
2206         tpaca->kvm_hstate.hwthread_req = 1;
2207
2208         /*
2209          * If the thread is already executing in the kernel (e.g. handling
2210          * a stray interrupt), wait for it to get back to nap mode.
2211          * The smp_mb() is to ensure that our setting of hwthread_req
2212          * is visible before we look at hwthread_state, so if this
2213          * races with the code at system_reset_pSeries and the thread
2214          * misses our setting of hwthread_req, we are sure to see its
2215          * setting of hwthread_state, and vice versa.
2216          */
2217         smp_mb();
2218         while (tpaca->kvm_hstate.hwthread_state == KVM_HWTHREAD_IN_KERNEL) {
2219                 if (--timeout <= 0) {
2220                         pr_err("KVM: couldn't grab cpu %d\n", cpu);
2221                         return -EBUSY;
2222                 }
2223                 udelay(1);
2224         }
2225         return 0;
2226 }
2227
2228 static void kvmppc_release_hwthread(int cpu)
2229 {
2230         struct paca_struct *tpaca;
2231
2232         tpaca = paca_ptrs[cpu];
2233         tpaca->kvm_hstate.hwthread_req = 0;
2234         tpaca->kvm_hstate.kvm_vcpu = NULL;
2235         tpaca->kvm_hstate.kvm_vcore = NULL;
2236         tpaca->kvm_hstate.kvm_split_mode = NULL;
2237 }
2238
2239 static void radix_flush_cpu(struct kvm *kvm, int cpu, struct kvm_vcpu *vcpu)
2240 {
2241         int i;
2242
2243         cpu = cpu_first_thread_sibling(cpu);
2244         cpumask_set_cpu(cpu, &kvm->arch.need_tlb_flush);
2245         /*
2246          * Make sure setting of bit in need_tlb_flush precedes
2247          * testing of cpu_in_guest bits.  The matching barrier on
2248          * the other side is the first smp_mb() in kvmppc_run_core().
2249          */
2250         smp_mb();
2251         for (i = 0; i < threads_per_core; ++i)
2252                 if (cpumask_test_cpu(cpu + i, &kvm->arch.cpu_in_guest))
2253                         smp_call_function_single(cpu + i, do_nothing, NULL, 1);
2254 }
2255
2256 static void kvmppc_prepare_radix_vcpu(struct kvm_vcpu *vcpu, int pcpu)
2257 {
2258         struct kvm *kvm = vcpu->kvm;
2259
2260         /*
2261          * With radix, the guest can do TLB invalidations itself,
2262          * and it could choose to use the local form (tlbiel) if
2263          * it is invalidating a translation that has only ever been
2264          * used on one vcpu.  However, that doesn't mean it has
2265          * only ever been used on one physical cpu, since vcpus
2266          * can move around between pcpus.  To cope with this, when
2267          * a vcpu moves from one pcpu to another, we need to tell
2268          * any vcpus running on the same core as this vcpu previously
2269          * ran to flush the TLB.  The TLB is shared between threads,
2270          * so we use a single bit in .need_tlb_flush for all 4 threads.
2271          */
2272         if (vcpu->arch.prev_cpu != pcpu) {
2273                 if (vcpu->arch.prev_cpu >= 0 &&
2274                     cpu_first_thread_sibling(vcpu->arch.prev_cpu) !=
2275                     cpu_first_thread_sibling(pcpu))
2276                         radix_flush_cpu(kvm, vcpu->arch.prev_cpu, vcpu);
2277                 vcpu->arch.prev_cpu = pcpu;
2278         }
2279 }
2280
2281 static void kvmppc_start_thread(struct kvm_vcpu *vcpu, struct kvmppc_vcore *vc)
2282 {
2283         int cpu;
2284         struct paca_struct *tpaca;
2285         struct kvm *kvm = vc->kvm;
2286
2287         cpu = vc->pcpu;
2288         if (vcpu) {
2289                 if (vcpu->arch.timer_running) {
2290                         hrtimer_try_to_cancel(&vcpu->arch.dec_timer);
2291                         vcpu->arch.timer_running = 0;
2292                 }
2293                 cpu += vcpu->arch.ptid;
2294                 vcpu->cpu = vc->pcpu;
2295                 vcpu->arch.thread_cpu = cpu;
2296                 cpumask_set_cpu(cpu, &kvm->arch.cpu_in_guest);
2297         }
2298         tpaca = paca_ptrs[cpu];
2299         tpaca->kvm_hstate.kvm_vcpu = vcpu;
2300         tpaca->kvm_hstate.ptid = cpu - vc->pcpu;
2301         tpaca->kvm_hstate.fake_suspend = 0;
2302         /* Order stores to hstate.kvm_vcpu etc. before store to kvm_vcore */
2303         smp_wmb();
2304         tpaca->kvm_hstate.kvm_vcore = vc;
2305         if (cpu != smp_processor_id())
2306                 kvmppc_ipi_thread(cpu);
2307 }
2308
2309 static void kvmppc_wait_for_nap(int n_threads)
2310 {
2311         int cpu = smp_processor_id();
2312         int i, loops;
2313
2314         if (n_threads <= 1)
2315                 return;
2316         for (loops = 0; loops < 1000000; ++loops) {
2317                 /*
2318                  * Check if all threads are finished.
2319                  * We set the vcore pointer when starting a thread
2320                  * and the thread clears it when finished, so we look
2321                  * for any threads that still have a non-NULL vcore ptr.
2322                  */
2323                 for (i = 1; i < n_threads; ++i)
2324                         if (paca_ptrs[cpu + i]->kvm_hstate.kvm_vcore)
2325                                 break;
2326                 if (i == n_threads) {
2327                         HMT_medium();
2328                         return;
2329                 }
2330                 HMT_low();
2331         }
2332         HMT_medium();
2333         for (i = 1; i < n_threads; ++i)
2334                 if (paca_ptrs[cpu + i]->kvm_hstate.kvm_vcore)
2335                         pr_err("KVM: CPU %d seems to be stuck\n", cpu + i);
2336 }
2337
2338 /*
2339  * Check that we are on thread 0 and that any other threads in
2340  * this core are off-line.  Then grab the threads so they can't
2341  * enter the kernel.
2342  */
2343 static int on_primary_thread(void)
2344 {
2345         int cpu = smp_processor_id();
2346         int thr;
2347
2348         /* Are we on a primary subcore? */
2349         if (cpu_thread_in_subcore(cpu))
2350                 return 0;
2351
2352         thr = 0;
2353         while (++thr < threads_per_subcore)
2354                 if (cpu_online(cpu + thr))
2355                         return 0;
2356
2357         /* Grab all hw threads so they can't go into the kernel */
2358         for (thr = 1; thr < threads_per_subcore; ++thr) {
2359                 if (kvmppc_grab_hwthread(cpu + thr)) {
2360                         /* Couldn't grab one; let the others go */
2361                         do {
2362                                 kvmppc_release_hwthread(cpu + thr);
2363                         } while (--thr > 0);
2364                         return 0;
2365                 }
2366         }
2367         return 1;
2368 }
2369
2370 /*
2371  * A list of virtual cores for each physical CPU.
2372  * These are vcores that could run but their runner VCPU tasks are
2373  * (or may be) preempted.
2374  */
2375 struct preempted_vcore_list {
2376         struct list_head        list;
2377         spinlock_t              lock;
2378 };
2379
2380 static DEFINE_PER_CPU(struct preempted_vcore_list, preempted_vcores);
2381
2382 static void init_vcore_lists(void)
2383 {
2384         int cpu;
2385
2386         for_each_possible_cpu(cpu) {
2387                 struct preempted_vcore_list *lp = &per_cpu(preempted_vcores, cpu);
2388                 spin_lock_init(&lp->lock);
2389                 INIT_LIST_HEAD(&lp->list);
2390         }
2391 }
2392
2393 static void kvmppc_vcore_preempt(struct kvmppc_vcore *vc)
2394 {
2395         struct preempted_vcore_list *lp = this_cpu_ptr(&preempted_vcores);
2396
2397         vc->vcore_state = VCORE_PREEMPT;
2398         vc->pcpu = smp_processor_id();
2399         if (vc->num_threads < threads_per_vcore(vc->kvm)) {
2400                 spin_lock(&lp->lock);
2401                 list_add_tail(&vc->preempt_list, &lp->list);
2402                 spin_unlock(&lp->lock);
2403         }
2404
2405         /* Start accumulating stolen time */
2406         kvmppc_core_start_stolen(vc);
2407 }
2408
2409 static void kvmppc_vcore_end_preempt(struct kvmppc_vcore *vc)
2410 {
2411         struct preempted_vcore_list *lp;
2412
2413         kvmppc_core_end_stolen(vc);
2414         if (!list_empty(&vc->preempt_list)) {
2415                 lp = &per_cpu(preempted_vcores, vc->pcpu);
2416                 spin_lock(&lp->lock);
2417                 list_del_init(&vc->preempt_list);
2418                 spin_unlock(&lp->lock);
2419         }
2420         vc->vcore_state = VCORE_INACTIVE;
2421 }
2422
2423 /*
2424  * This stores information about the virtual cores currently
2425  * assigned to a physical core.
2426  */
2427 struct core_info {
2428         int             n_subcores;
2429         int             max_subcore_threads;
2430         int             total_threads;
2431         int             subcore_threads[MAX_SUBCORES];
2432         struct kvmppc_vcore *vc[MAX_SUBCORES];
2433 };
2434
2435 /*
2436  * This mapping means subcores 0 and 1 can use threads 0-3 and 4-7
2437  * respectively in 2-way micro-threading (split-core) mode on POWER8.
2438  */
2439 static int subcore_thread_map[MAX_SUBCORES] = { 0, 4, 2, 6 };
2440
2441 static void init_core_info(struct core_info *cip, struct kvmppc_vcore *vc)
2442 {
2443         memset(cip, 0, sizeof(*cip));
2444         cip->n_subcores = 1;
2445         cip->max_subcore_threads = vc->num_threads;
2446         cip->total_threads = vc->num_threads;
2447         cip->subcore_threads[0] = vc->num_threads;
2448         cip->vc[0] = vc;
2449 }
2450
2451 static bool subcore_config_ok(int n_subcores, int n_threads)
2452 {
2453         /*
2454          * POWER9 "SMT4" cores are permanently in what is effectively a 4-way
2455          * split-core mode, with one thread per subcore.
2456          */
2457         if (cpu_has_feature(CPU_FTR_ARCH_300))
2458                 return n_subcores <= 4 && n_threads == 1;
2459
2460         /* On POWER8, can only dynamically split if unsplit to begin with */
2461         if (n_subcores > 1 && threads_per_subcore < MAX_SMT_THREADS)
2462                 return false;
2463         if (n_subcores > MAX_SUBCORES)
2464                 return false;
2465         if (n_subcores > 1) {
2466                 if (!(dynamic_mt_modes & 2))
2467                         n_subcores = 4;
2468                 if (n_subcores > 2 && !(dynamic_mt_modes & 4))
2469                         return false;
2470         }
2471
2472         return n_subcores * roundup_pow_of_two(n_threads) <= MAX_SMT_THREADS;
2473 }
2474
2475 static void init_vcore_to_run(struct kvmppc_vcore *vc)
2476 {
2477         vc->entry_exit_map = 0;
2478         vc->in_guest = 0;
2479         vc->napping_threads = 0;
2480         vc->conferring_threads = 0;
2481         vc->tb_offset_applied = 0;
2482 }
2483
2484 static bool can_dynamic_split(struct kvmppc_vcore *vc, struct core_info *cip)
2485 {
2486         int n_threads = vc->num_threads;
2487         int sub;
2488
2489         if (!cpu_has_feature(CPU_FTR_ARCH_207S))
2490                 return false;
2491
2492         /* Some POWER9 chips require all threads to be in the same MMU mode */
2493         if (no_mixing_hpt_and_radix &&
2494             kvm_is_radix(vc->kvm) != kvm_is_radix(cip->vc[0]->kvm))
2495                 return false;
2496
2497         if (n_threads < cip->max_subcore_threads)
2498                 n_threads = cip->max_subcore_threads;
2499         if (!subcore_config_ok(cip->n_subcores + 1, n_threads))
2500                 return false;
2501         cip->max_subcore_threads = n_threads;
2502
2503         sub = cip->n_subcores;
2504         ++cip->n_subcores;
2505         cip->total_threads += vc->num_threads;
2506         cip->subcore_threads[sub] = vc->num_threads;
2507         cip->vc[sub] = vc;
2508         init_vcore_to_run(vc);
2509         list_del_init(&vc->preempt_list);
2510
2511         return true;
2512 }
2513
2514 /*
2515  * Work out whether it is possible to piggyback the execution of
2516  * vcore *pvc onto the execution of the other vcores described in *cip.
2517  */
2518 static bool can_piggyback(struct kvmppc_vcore *pvc, struct core_info *cip,
2519                           int target_threads)
2520 {
2521         if (cip->total_threads + pvc->num_threads > target_threads)
2522                 return false;
2523
2524         return can_dynamic_split(pvc, cip);
2525 }
2526
2527 static void prepare_threads(struct kvmppc_vcore *vc)
2528 {
2529         int i;
2530         struct kvm_vcpu *vcpu;
2531
2532         for_each_runnable_thread(i, vcpu, vc) {
2533                 if (signal_pending(vcpu->arch.run_task))
2534                         vcpu->arch.ret = -EINTR;
2535                 else if (vcpu->arch.vpa.update_pending ||
2536                          vcpu->arch.slb_shadow.update_pending ||
2537                          vcpu->arch.dtl.update_pending)
2538                         vcpu->arch.ret = RESUME_GUEST;
2539                 else
2540                         continue;
2541                 kvmppc_remove_runnable(vc, vcpu);
2542                 wake_up(&vcpu->arch.cpu_run);
2543         }
2544 }
2545
2546 static void collect_piggybacks(struct core_info *cip, int target_threads)
2547 {
2548         struct preempted_vcore_list *lp = this_cpu_ptr(&preempted_vcores);
2549         struct kvmppc_vcore *pvc, *vcnext;
2550
2551         spin_lock(&lp->lock);
2552         list_for_each_entry_safe(pvc, vcnext, &lp->list, preempt_list) {
2553                 if (!spin_trylock(&pvc->lock))
2554                         continue;
2555                 prepare_threads(pvc);
2556                 if (!pvc->n_runnable) {
2557                         list_del_init(&pvc->preempt_list);
2558                         if (pvc->runner == NULL) {
2559                                 pvc->vcore_state = VCORE_INACTIVE;
2560                                 kvmppc_core_end_stolen(pvc);
2561                         }
2562                         spin_unlock(&pvc->lock);
2563                         continue;
2564                 }
2565                 if (!can_piggyback(pvc, cip, target_threads)) {
2566                         spin_unlock(&pvc->lock);
2567                         continue;
2568                 }
2569                 kvmppc_core_end_stolen(pvc);
2570                 pvc->vcore_state = VCORE_PIGGYBACK;
2571                 if (cip->total_threads >= target_threads)
2572                         break;
2573         }
2574         spin_unlock(&lp->lock);
2575 }
2576
2577 static bool recheck_signals(struct core_info *cip)
2578 {
2579         int sub, i;
2580         struct kvm_vcpu *vcpu;
2581
2582         for (sub = 0; sub < cip->n_subcores; ++sub)
2583                 for_each_runnable_thread(i, vcpu, cip->vc[sub])
2584                         if (signal_pending(vcpu->arch.run_task))
2585                                 return true;
2586         return false;
2587 }
2588
2589 static void post_guest_process(struct kvmppc_vcore *vc, bool is_master)
2590 {
2591         int still_running = 0, i;
2592         u64 now;
2593         long ret;
2594         struct kvm_vcpu *vcpu;
2595
2596         spin_lock(&vc->lock);
2597         now = get_tb();
2598         for_each_runnable_thread(i, vcpu, vc) {
2599                 /* cancel pending dec exception if dec is positive */
2600                 if (now < vcpu->arch.dec_expires &&
2601                     kvmppc_core_pending_dec(vcpu))
2602                         kvmppc_core_dequeue_dec(vcpu);
2603
2604                 trace_kvm_guest_exit(vcpu);
2605
2606                 ret = RESUME_GUEST;
2607                 if (vcpu->arch.trap)
2608                         ret = kvmppc_handle_exit_hv(vcpu->arch.kvm_run, vcpu,
2609                                                     vcpu->arch.run_task);
2610
2611                 vcpu->arch.ret = ret;
2612                 vcpu->arch.trap = 0;
2613
2614                 if (is_kvmppc_resume_guest(vcpu->arch.ret)) {
2615                         if (vcpu->arch.pending_exceptions)
2616                                 kvmppc_core_prepare_to_enter(vcpu);
2617                         if (vcpu->arch.ceded)
2618                                 kvmppc_set_timer(vcpu);
2619                         else
2620                                 ++still_running;
2621                 } else {
2622                         kvmppc_remove_runnable(vc, vcpu);
2623                         wake_up(&vcpu->arch.cpu_run);
2624                 }
2625         }
2626         if (!is_master) {
2627                 if (still_running > 0) {
2628                         kvmppc_vcore_preempt(vc);
2629                 } else if (vc->runner) {
2630                         vc->vcore_state = VCORE_PREEMPT;
2631                         kvmppc_core_start_stolen(vc);
2632                 } else {
2633                         vc->vcore_state = VCORE_INACTIVE;
2634                 }
2635                 if (vc->n_runnable > 0 && vc->runner == NULL) {
2636                         /* make sure there's a candidate runner awake */
2637                         i = -1;
2638                         vcpu = next_runnable_thread(vc, &i);
2639                         wake_up(&vcpu->arch.cpu_run);
2640                 }
2641         }
2642         spin_unlock(&vc->lock);
2643 }
2644
2645 /*
2646  * Clear core from the list of active host cores as we are about to
2647  * enter the guest. Only do this if it is the primary thread of the
2648  * core (not if a subcore) that is entering the guest.
2649  */
2650 static inline int kvmppc_clear_host_core(unsigned int cpu)
2651 {
2652         int core;
2653
2654         if (!kvmppc_host_rm_ops_hv || cpu_thread_in_core(cpu))
2655                 return 0;
2656         /*
2657          * Memory barrier can be omitted here as we will do a smp_wmb()
2658          * later in kvmppc_start_thread and we need ensure that state is
2659          * visible to other CPUs only after we enter guest.
2660          */
2661         core = cpu >> threads_shift;
2662         kvmppc_host_rm_ops_hv->rm_core[core].rm_state.in_host = 0;
2663         return 0;
2664 }
2665
2666 /*
2667  * Advertise this core as an active host core since we exited the guest
2668  * Only need to do this if it is the primary thread of the core that is
2669  * exiting.
2670  */
2671 static inline int kvmppc_set_host_core(unsigned int cpu)
2672 {
2673         int core;
2674
2675         if (!kvmppc_host_rm_ops_hv || cpu_thread_in_core(cpu))
2676                 return 0;
2677
2678         /*
2679          * Memory barrier can be omitted here because we do a spin_unlock
2680          * immediately after this which provides the memory barrier.
2681          */
2682         core = cpu >> threads_shift;
2683         kvmppc_host_rm_ops_hv->rm_core[core].rm_state.in_host = 1;
2684         return 0;
2685 }
2686
2687 static void set_irq_happened(int trap)
2688 {
2689         switch (trap) {
2690         case BOOK3S_INTERRUPT_EXTERNAL:
2691                 local_paca->irq_happened |= PACA_IRQ_EE;
2692                 break;
2693         case BOOK3S_INTERRUPT_H_DOORBELL:
2694                 local_paca->irq_happened |= PACA_IRQ_DBELL;
2695                 break;
2696         case BOOK3S_INTERRUPT_HMI:
2697                 local_paca->irq_happened |= PACA_IRQ_HMI;
2698                 break;
2699         case BOOK3S_INTERRUPT_SYSTEM_RESET:
2700                 replay_system_reset();
2701                 break;
2702         }
2703 }
2704
2705 /*
2706  * Run a set of guest threads on a physical core.
2707  * Called with vc->lock held.
2708  */
2709 static noinline void kvmppc_run_core(struct kvmppc_vcore *vc)
2710 {
2711         struct kvm_vcpu *vcpu;
2712         int i;
2713         int srcu_idx;
2714         struct core_info core_info;
2715         struct kvmppc_vcore *pvc;
2716         struct kvm_split_mode split_info, *sip;
2717         int split, subcore_size, active;
2718         int sub;
2719         bool thr0_done;
2720         unsigned long cmd_bit, stat_bit;
2721         int pcpu, thr;
2722         int target_threads;
2723         int controlled_threads;
2724         int trap;
2725         bool is_power8;
2726         bool hpt_on_radix;
2727
2728         /*
2729          * Remove from the list any threads that have a signal pending
2730          * or need a VPA update done
2731          */
2732         prepare_threads(vc);
2733
2734         /* if the runner is no longer runnable, let the caller pick a new one */
2735         if (vc->runner->arch.state != KVMPPC_VCPU_RUNNABLE)
2736                 return;
2737
2738         /*
2739          * Initialize *vc.
2740          */
2741         init_vcore_to_run(vc);
2742         vc->preempt_tb = TB_NIL;
2743
2744         /*
2745          * Number of threads that we will be controlling: the same as
2746          * the number of threads per subcore, except on POWER9,
2747          * where it's 1 because the threads are (mostly) independent.
2748          */
2749         controlled_threads = threads_per_vcore(vc->kvm);
2750
2751         /*
2752          * Make sure we are running on primary threads, and that secondary
2753          * threads are offline.  Also check if the number of threads in this
2754          * guest are greater than the current system threads per guest.
2755          * On POWER9, we need to be not in independent-threads mode if
2756          * this is a HPT guest on a radix host machine where the
2757          * CPU threads may not be in different MMU modes.
2758          */
2759         hpt_on_radix = no_mixing_hpt_and_radix && radix_enabled() &&
2760                 !kvm_is_radix(vc->kvm);
2761         if (((controlled_threads > 1) &&
2762              ((vc->num_threads > threads_per_subcore) || !on_primary_thread())) ||
2763             (hpt_on_radix && vc->kvm->arch.threads_indep)) {
2764                 for_each_runnable_thread(i, vcpu, vc) {
2765                         vcpu->arch.ret = -EBUSY;
2766                         kvmppc_remove_runnable(vc, vcpu);
2767                         wake_up(&vcpu->arch.cpu_run);
2768                 }
2769                 goto out;
2770         }
2771
2772         /*
2773          * See if we could run any other vcores on the physical core
2774          * along with this one.
2775          */
2776         init_core_info(&core_info, vc);
2777         pcpu = smp_processor_id();
2778         target_threads = controlled_threads;
2779         if (target_smt_mode && target_smt_mode < target_threads)
2780                 target_threads = target_smt_mode;
2781         if (vc->num_threads < target_threads)
2782                 collect_piggybacks(&core_info, target_threads);
2783
2784         /*
2785          * On radix, arrange for TLB flushing if necessary.
2786          * This has to be done before disabling interrupts since
2787          * it uses smp_call_function().
2788          */
2789         pcpu = smp_processor_id();
2790         if (kvm_is_radix(vc->kvm)) {
2791                 for (sub = 0; sub < core_info.n_subcores; ++sub)
2792                         for_each_runnable_thread(i, vcpu, core_info.vc[sub])
2793                                 kvmppc_prepare_radix_vcpu(vcpu, pcpu);
2794         }
2795
2796         /*
2797          * Hard-disable interrupts, and check resched flag and signals.
2798          * If we need to reschedule or deliver a signal, clean up
2799          * and return without going into the guest(s).
2800          * If the mmu_ready flag has been cleared, don't go into the
2801          * guest because that means a HPT resize operation is in progress.
2802          */
2803         local_irq_disable();
2804         hard_irq_disable();
2805         if (lazy_irq_pending() || need_resched() ||
2806             recheck_signals(&core_info) || !vc->kvm->arch.mmu_ready) {
2807                 local_irq_enable();
2808                 vc->vcore_state = VCORE_INACTIVE;
2809                 /* Unlock all except the primary vcore */
2810                 for (sub = 1; sub < core_info.n_subcores; ++sub) {
2811                         pvc = core_info.vc[sub];
2812                         /* Put back on to the preempted vcores list */
2813                         kvmppc_vcore_preempt(pvc);
2814                         spin_unlock(&pvc->lock);
2815                 }
2816                 for (i = 0; i < controlled_threads; ++i)
2817                         kvmppc_release_hwthread(pcpu + i);
2818                 return;
2819         }
2820
2821         kvmppc_clear_host_core(pcpu);
2822
2823         /* Decide on micro-threading (split-core) mode */
2824         subcore_size = threads_per_subcore;
2825         cmd_bit = stat_bit = 0;
2826         split = core_info.n_subcores;
2827         sip = NULL;
2828         is_power8 = cpu_has_feature(CPU_FTR_ARCH_207S)
2829                 && !cpu_has_feature(CPU_FTR_ARCH_300);
2830
2831         if (split > 1 || hpt_on_radix) {
2832                 sip = &split_info;
2833                 memset(&split_info, 0, sizeof(split_info));
2834                 for (sub = 0; sub < core_info.n_subcores; ++sub)
2835                         split_info.vc[sub] = core_info.vc[sub];
2836
2837                 if (is_power8) {
2838                         if (split == 2 && (dynamic_mt_modes & 2)) {
2839                                 cmd_bit = HID0_POWER8_1TO2LPAR;
2840                                 stat_bit = HID0_POWER8_2LPARMODE;
2841                         } else {
2842                                 split = 4;
2843                                 cmd_bit = HID0_POWER8_1TO4LPAR;
2844                                 stat_bit = HID0_POWER8_4LPARMODE;
2845                         }
2846                         subcore_size = MAX_SMT_THREADS / split;
2847                         split_info.rpr = mfspr(SPRN_RPR);
2848                         split_info.pmmar = mfspr(SPRN_PMMAR);
2849                         split_info.ldbar = mfspr(SPRN_LDBAR);
2850                         split_info.subcore_size = subcore_size;
2851                 } else {
2852                         split_info.subcore_size = 1;
2853                         if (hpt_on_radix) {
2854                                 /* Use the split_info for LPCR/LPIDR changes */
2855                                 split_info.lpcr_req = vc->lpcr;
2856                                 split_info.lpidr_req = vc->kvm->arch.lpid;
2857                                 split_info.host_lpcr = vc->kvm->arch.host_lpcr;
2858                                 split_info.do_set = 1;
2859                         }
2860                 }
2861
2862                 /* order writes to split_info before kvm_split_mode pointer */
2863                 smp_wmb();
2864         }
2865
2866         for (thr = 0; thr < controlled_threads; ++thr) {
2867                 struct paca_struct *paca = paca_ptrs[pcpu + thr];
2868
2869                 paca->kvm_hstate.tid = thr;
2870                 paca->kvm_hstate.napping = 0;
2871                 paca->kvm_hstate.kvm_split_mode = sip;
2872         }
2873
2874         /* Initiate micro-threading (split-core) on POWER8 if required */
2875         if (cmd_bit) {
2876                 unsigned long hid0 = mfspr(SPRN_HID0);
2877
2878                 hid0 |= cmd_bit | HID0_POWER8_DYNLPARDIS;
2879                 mb();
2880                 mtspr(SPRN_HID0, hid0);
2881                 isync();
2882                 for (;;) {
2883                         hid0 = mfspr(SPRN_HID0);
2884                         if (hid0 & stat_bit)
2885                                 break;
2886                         cpu_relax();
2887                 }
2888         }
2889
2890         /*
2891          * On POWER8, set RWMR register.
2892          * Since it only affects PURR and SPURR, it doesn't affect
2893          * the host, so we don't save/restore the host value.
2894          */
2895         if (is_power8) {
2896                 unsigned long rwmr_val = RWMR_RPA_P8_8THREAD;
2897                 int n_online = atomic_read(&vc->online_count);
2898
2899                 /*
2900                  * Use the 8-thread value if we're doing split-core
2901                  * or if the vcore's online count looks bogus.
2902                  */
2903                 if (split == 1 && threads_per_subcore == MAX_SMT_THREADS &&
2904                     n_online >= 1 && n_online <= MAX_SMT_THREADS)
2905                         rwmr_val = p8_rwmr_values[n_online];
2906                 mtspr(SPRN_RWMR, rwmr_val);
2907         }
2908
2909         /* Start all the threads */
2910         active = 0;
2911         for (sub = 0; sub < core_info.n_subcores; ++sub) {
2912                 thr = is_power8 ? subcore_thread_map[sub] : sub;
2913                 thr0_done = false;
2914                 active |= 1 << thr;
2915                 pvc = core_info.vc[sub];
2916                 pvc->pcpu = pcpu + thr;
2917                 for_each_runnable_thread(i, vcpu, pvc) {
2918                         kvmppc_start_thread(vcpu, pvc);
2919                         kvmppc_create_dtl_entry(vcpu, pvc);
2920                         trace_kvm_guest_enter(vcpu);
2921                         if (!vcpu->arch.ptid)
2922                                 thr0_done = true;
2923                         active |= 1 << (thr + vcpu->arch.ptid);
2924                 }
2925                 /*
2926                  * We need to start the first thread of each subcore
2927                  * even if it doesn't have a vcpu.
2928                  */
2929                 if (!thr0_done)
2930                         kvmppc_start_thread(NULL, pvc);
2931         }
2932
2933         /*
2934          * Ensure that split_info.do_nap is set after setting
2935          * the vcore pointer in the PACA of the secondaries.
2936          */
2937         smp_mb();
2938
2939         /*
2940          * When doing micro-threading, poke the inactive threads as well.
2941          * This gets them to the nap instruction after kvm_do_nap,
2942          * which reduces the time taken to unsplit later.
2943          * For POWER9 HPT guest on radix host, we need all the secondary
2944          * threads woken up so they can do the LPCR/LPIDR change.
2945          */
2946         if (cmd_bit || hpt_on_radix) {
2947                 split_info.do_nap = 1;  /* ask secondaries to nap when done */
2948                 for (thr = 1; thr < threads_per_subcore; ++thr)
2949                         if (!(active & (1 << thr)))
2950                                 kvmppc_ipi_thread(pcpu + thr);
2951         }
2952
2953         vc->vcore_state = VCORE_RUNNING;
2954         preempt_disable();
2955
2956         trace_kvmppc_run_core(vc, 0);
2957
2958         for (sub = 0; sub < core_info.n_subcores; ++sub)
2959                 spin_unlock(&core_info.vc[sub]->lock);
2960
2961         if (kvm_is_radix(vc->kvm)) {
2962                 int tmp = pcpu;
2963
2964                 /*
2965                  * Do we need to flush the process scoped TLB for the LPAR?
2966                  *
2967                  * On POWER9, individual threads can come in here, but the
2968                  * TLB is shared between the 4 threads in a core, hence
2969                  * invalidating on one thread invalidates for all.
2970                  * Thus we make all 4 threads use the same bit here.
2971                  *
2972                  * Hash must be flushed in realmode in order to use tlbiel.
2973                  */
2974                 mtspr(SPRN_LPID, vc->kvm->arch.lpid);
2975                 isync();
2976
2977                 if (cpu_has_feature(CPU_FTR_ARCH_300))
2978                         tmp &= ~0x3UL;
2979
2980                 if (cpumask_test_cpu(tmp, &vc->kvm->arch.need_tlb_flush)) {
2981                         radix__local_flush_tlb_lpid_guest(vc->kvm->arch.lpid);
2982                         /* Clear the bit after the TLB flush */
2983                         cpumask_clear_cpu(tmp, &vc->kvm->arch.need_tlb_flush);
2984                 }
2985         }
2986
2987         /*
2988          * Interrupts will be enabled once we get into the guest,
2989          * so tell lockdep that we're about to enable interrupts.
2990          */
2991         trace_hardirqs_on();
2992
2993         guest_enter_irqoff();
2994
2995         srcu_idx = srcu_read_lock(&vc->kvm->srcu);
2996
2997         this_cpu_disable_ftrace();
2998
2999         trap = __kvmppc_vcore_entry();
3000
3001         this_cpu_enable_ftrace();
3002
3003         srcu_read_unlock(&vc->kvm->srcu, srcu_idx);
3004
3005         trace_hardirqs_off();
3006         set_irq_happened(trap);
3007
3008         spin_lock(&vc->lock);
3009         /* prevent other vcpu threads from doing kvmppc_start_thread() now */
3010         vc->vcore_state = VCORE_EXITING;
3011
3012         /* wait for secondary threads to finish writing their state to memory */
3013         kvmppc_wait_for_nap(controlled_threads);
3014
3015         /* Return to whole-core mode if we split the core earlier */
3016         if (cmd_bit) {
3017                 unsigned long hid0 = mfspr(SPRN_HID0);
3018                 unsigned long loops = 0;
3019
3020                 hid0 &= ~HID0_POWER8_DYNLPARDIS;
3021                 stat_bit = HID0_POWER8_2LPARMODE | HID0_POWER8_4LPARMODE;
3022                 mb();
3023                 mtspr(SPRN_HID0, hid0);
3024                 isync();
3025                 for (;;) {
3026                         hid0 = mfspr(SPRN_HID0);
3027                         if (!(hid0 & stat_bit))
3028                                 break;
3029                         cpu_relax();
3030                         ++loops;
3031                 }
3032         } else if (hpt_on_radix) {
3033                 /* Wait for all threads to have seen final sync */
3034                 for (thr = 1; thr < controlled_threads; ++thr) {
3035                         struct paca_struct *paca = paca_ptrs[pcpu + thr];
3036
3037                         while (paca->kvm_hstate.kvm_split_mode) {
3038                                 HMT_low();
3039                                 barrier();
3040                         }
3041                         HMT_medium();
3042                 }
3043         }
3044         split_info.do_nap = 0;
3045
3046         kvmppc_set_host_core(pcpu);
3047
3048         local_irq_enable();
3049         guest_exit();
3050
3051         /* Let secondaries go back to the offline loop */
3052         for (i = 0; i < controlled_threads; ++i) {
3053                 kvmppc_release_hwthread(pcpu + i);
3054                 if (sip && sip->napped[i])
3055                         kvmppc_ipi_thread(pcpu + i);
3056                 cpumask_clear_cpu(pcpu + i, &vc->kvm->arch.cpu_in_guest);
3057         }
3058
3059         spin_unlock(&vc->lock);
3060
3061         /* make sure updates to secondary vcpu structs are visible now */
3062         smp_mb();
3063
3064         preempt_enable();
3065
3066         for (sub = 0; sub < core_info.n_subcores; ++sub) {
3067                 pvc = core_info.vc[sub];
3068                 post_guest_process(pvc, pvc == vc);
3069         }
3070
3071         spin_lock(&vc->lock);
3072
3073  out:
3074         vc->vcore_state = VCORE_INACTIVE;
3075         trace_kvmppc_run_core(vc, 1);
3076 }
3077
3078 /*
3079  * Wait for some other vcpu thread to execute us, and
3080  * wake us up when we need to handle something in the host.
3081  */
3082 static void kvmppc_wait_for_exec(struct kvmppc_vcore *vc,
3083                                  struct kvm_vcpu *vcpu, int wait_state)
3084 {
3085         DEFINE_WAIT(wait);
3086
3087         prepare_to_wait(&vcpu->arch.cpu_run, &wait, wait_state);
3088         if (vcpu->arch.state == KVMPPC_VCPU_RUNNABLE) {
3089                 spin_unlock(&vc->lock);
3090                 schedule();
3091                 spin_lock(&vc->lock);
3092         }
3093         finish_wait(&vcpu->arch.cpu_run, &wait);
3094 }
3095
3096 static void grow_halt_poll_ns(struct kvmppc_vcore *vc)
3097 {
3098         /* 10us base */
3099         if (vc->halt_poll_ns == 0 && halt_poll_ns_grow)
3100                 vc->halt_poll_ns = 10000;
3101         else
3102                 vc->halt_poll_ns *= halt_poll_ns_grow;
3103 }
3104
3105 static void shrink_halt_poll_ns(struct kvmppc_vcore *vc)
3106 {
3107         if (halt_poll_ns_shrink == 0)
3108                 vc->halt_poll_ns = 0;
3109         else
3110                 vc->halt_poll_ns /= halt_poll_ns_shrink;
3111 }
3112
3113 #ifdef CONFIG_KVM_XICS
3114 static inline bool xive_interrupt_pending(struct kvm_vcpu *vcpu)
3115 {
3116         if (!xive_enabled())
3117                 return false;
3118         return vcpu->arch.irq_pending || vcpu->arch.xive_saved_state.pipr <
3119                 vcpu->arch.xive_saved_state.cppr;
3120 }
3121 #else
3122 static inline bool xive_interrupt_pending(struct kvm_vcpu *vcpu)
3123 {
3124         return false;
3125 }
3126 #endif /* CONFIG_KVM_XICS */
3127
3128 static bool kvmppc_vcpu_woken(struct kvm_vcpu *vcpu)
3129 {
3130         if (vcpu->arch.pending_exceptions || vcpu->arch.prodded ||
3131             kvmppc_doorbell_pending(vcpu) || xive_interrupt_pending(vcpu))
3132                 return true;
3133
3134         return false;
3135 }
3136
3137 /*
3138  * Check to see if any of the runnable vcpus on the vcore have pending
3139  * exceptions or are no longer ceded
3140  */
3141 static int kvmppc_vcore_check_block(struct kvmppc_vcore *vc)
3142 {
3143         struct kvm_vcpu *vcpu;
3144         int i;
3145
3146         for_each_runnable_thread(i, vcpu, vc) {
3147                 if (!vcpu->arch.ceded || kvmppc_vcpu_woken(vcpu))
3148                         return 1;
3149         }
3150
3151         return 0;
3152 }
3153
3154 /*
3155  * All the vcpus in this vcore are idle, so wait for a decrementer
3156  * or external interrupt to one of the vcpus.  vc->lock is held.
3157  */
3158 static void kvmppc_vcore_blocked(struct kvmppc_vcore *vc)
3159 {
3160         ktime_t cur, start_poll, start_wait;
3161         int do_sleep = 1;
3162         u64 block_ns;
3163         DECLARE_SWAITQUEUE(wait);
3164
3165         /* Poll for pending exceptions and ceded state */
3166         cur = start_poll = ktime_get();
3167         if (vc->halt_poll_ns) {
3168                 ktime_t stop = ktime_add_ns(start_poll, vc->halt_poll_ns);
3169                 ++vc->runner->stat.halt_attempted_poll;
3170
3171                 vc->vcore_state = VCORE_POLLING;
3172                 spin_unlock(&vc->lock);
3173
3174                 do {
3175                         if (kvmppc_vcore_check_block(vc)) {
3176                                 do_sleep = 0;
3177                                 break;
3178                         }
3179                         cur = ktime_get();
3180                 } while (single_task_running() && ktime_before(cur, stop));
3181
3182                 spin_lock(&vc->lock);
3183                 vc->vcore_state = VCORE_INACTIVE;
3184
3185                 if (!do_sleep) {
3186                         ++vc->runner->stat.halt_successful_poll;
3187                         goto out;
3188                 }
3189         }
3190
3191         prepare_to_swait(&vc->wq, &wait, TASK_INTERRUPTIBLE);
3192
3193         if (kvmppc_vcore_check_block(vc)) {
3194                 finish_swait(&vc->wq, &wait);
3195                 do_sleep = 0;
3196                 /* If we polled, count this as a successful poll */
3197                 if (vc->halt_poll_ns)
3198                         ++vc->runner->stat.halt_successful_poll;
3199                 goto out;
3200         }
3201
3202         start_wait = ktime_get();
3203
3204         vc->vcore_state = VCORE_SLEEPING;
3205         trace_kvmppc_vcore_blocked(vc, 0);
3206         spin_unlock(&vc->lock);
3207         schedule();
3208         finish_swait(&vc->wq, &wait);
3209         spin_lock(&vc->lock);
3210         vc->vcore_state = VCORE_INACTIVE;
3211         trace_kvmppc_vcore_blocked(vc, 1);
3212         ++vc->runner->stat.halt_successful_wait;
3213
3214         cur = ktime_get();
3215
3216 out:
3217         block_ns = ktime_to_ns(cur) - ktime_to_ns(start_poll);
3218
3219         /* Attribute wait time */
3220         if (do_sleep) {
3221                 vc->runner->stat.halt_wait_ns +=
3222                         ktime_to_ns(cur) - ktime_to_ns(start_wait);
3223                 /* Attribute failed poll time */
3224                 if (vc->halt_poll_ns)
3225                         vc->runner->stat.halt_poll_fail_ns +=
3226                                 ktime_to_ns(start_wait) -
3227                                 ktime_to_ns(start_poll);
3228         } else {
3229                 /* Attribute successful poll time */
3230                 if (vc->halt_poll_ns)
3231                         vc->runner->stat.halt_poll_success_ns +=
3232                                 ktime_to_ns(cur) -
3233                                 ktime_to_ns(start_poll);
3234         }
3235
3236         /* Adjust poll time */
3237         if (halt_poll_ns) {
3238                 if (block_ns <= vc->halt_poll_ns)
3239                         ;
3240                 /* We slept and blocked for longer than the max halt time */
3241                 else if (vc->halt_poll_ns && block_ns > halt_poll_ns)
3242                         shrink_halt_poll_ns(vc);
3243                 /* We slept and our poll time is too small */
3244                 else if (vc->halt_poll_ns < halt_poll_ns &&
3245                                 block_ns < halt_poll_ns)
3246                         grow_halt_poll_ns(vc);
3247                 if (vc->halt_poll_ns > halt_poll_ns)
3248                         vc->halt_poll_ns = halt_poll_ns;
3249         } else
3250                 vc->halt_poll_ns = 0;
3251
3252         trace_kvmppc_vcore_wakeup(do_sleep, block_ns);
3253 }
3254
3255 static int kvmhv_setup_mmu(struct kvm_vcpu *vcpu)
3256 {
3257         int r = 0;
3258         struct kvm *kvm = vcpu->kvm;
3259
3260         mutex_lock(&kvm->lock);
3261         if (!kvm->arch.mmu_ready) {
3262                 if (!kvm_is_radix(kvm))
3263                         r = kvmppc_hv_setup_htab_rma(vcpu);
3264                 if (!r) {
3265                         if (cpu_has_feature(CPU_FTR_ARCH_300))
3266                                 kvmppc_setup_partition_table(kvm);
3267                         kvm->arch.mmu_ready = 1;
3268                 }
3269         }
3270         mutex_unlock(&kvm->lock);
3271         return r;
3272 }
3273
3274 static int kvmppc_run_vcpu(struct kvm_run *kvm_run, struct kvm_vcpu *vcpu)
3275 {
3276         int n_ceded, i, r;
3277         struct kvmppc_vcore *vc;
3278         struct kvm_vcpu *v;
3279
3280         trace_kvmppc_run_vcpu_enter(vcpu);
3281
3282         kvm_run->exit_reason = 0;
3283         vcpu->arch.ret = RESUME_GUEST;
3284         vcpu->arch.trap = 0;
3285         kvmppc_update_vpas(vcpu);
3286
3287         /*
3288          * Synchronize with other threads in this virtual core
3289          */
3290         vc = vcpu->arch.vcore;
3291         spin_lock(&vc->lock);
3292         vcpu->arch.ceded = 0;
3293         vcpu->arch.run_task = current;
3294         vcpu->arch.kvm_run = kvm_run;
3295         vcpu->arch.stolen_logged = vcore_stolen_time(vc, mftb());
3296         vcpu->arch.state = KVMPPC_VCPU_RUNNABLE;
3297         vcpu->arch.busy_preempt = TB_NIL;
3298         WRITE_ONCE(vc->runnable_threads[vcpu->arch.ptid], vcpu);
3299         ++vc->n_runnable;
3300
3301         /*
3302          * This happens the first time this is called for a vcpu.
3303          * If the vcore is already running, we may be able to start
3304          * this thread straight away and have it join in.
3305          */
3306         if (!signal_pending(current)) {
3307                 if ((vc->vcore_state == VCORE_PIGGYBACK ||
3308                      vc->vcore_state == VCORE_RUNNING) &&
3309                            !VCORE_IS_EXITING(vc)) {
3310                         kvmppc_create_dtl_entry(vcpu, vc);
3311                         kvmppc_start_thread(vcpu, vc);
3312                         trace_kvm_guest_enter(vcpu);
3313                 } else if (vc->vcore_state == VCORE_SLEEPING) {
3314                         swake_up(&vc->wq);
3315                 }
3316
3317         }
3318
3319         while (vcpu->arch.state == KVMPPC_VCPU_RUNNABLE &&
3320                !signal_pending(current)) {
3321                 /* See if the MMU is ready to go */
3322                 if (!vcpu->kvm->arch.mmu_ready) {
3323                         spin_unlock(&vc->lock);
3324                         r = kvmhv_setup_mmu(vcpu);
3325                         spin_lock(&vc->lock);
3326                         if (r) {
3327                                 kvm_run->exit_reason = KVM_EXIT_FAIL_ENTRY;
3328                                 kvm_run->fail_entry.
3329                                         hardware_entry_failure_reason = 0;
3330                                 vcpu->arch.ret = r;
3331                                 break;
3332                         }
3333                 }
3334
3335                 if (vc->vcore_state == VCORE_PREEMPT && vc->runner == NULL)
3336                         kvmppc_vcore_end_preempt(vc);
3337
3338                 if (vc->vcore_state != VCORE_INACTIVE) {
3339                         kvmppc_wait_for_exec(vc, vcpu, TASK_INTERRUPTIBLE);
3340                         continue;
3341                 }
3342                 for_each_runnable_thread(i, v, vc) {
3343                         kvmppc_core_prepare_to_enter(v);
3344                         if (signal_pending(v->arch.run_task)) {
3345                                 kvmppc_remove_runnable(vc, v);
3346                                 v->stat.signal_exits++;
3347                                 v->arch.kvm_run->exit_reason = KVM_EXIT_INTR;
3348                                 v->arch.ret = -EINTR;
3349                                 wake_up(&v->arch.cpu_run);
3350                         }
3351                 }
3352                 if (!vc->n_runnable || vcpu->arch.state != KVMPPC_VCPU_RUNNABLE)
3353                         break;
3354                 n_ceded = 0;
3355                 for_each_runnable_thread(i, v, vc) {
3356                         if (!kvmppc_vcpu_woken(v))
3357                                 n_ceded += v->arch.ceded;
3358                         else
3359                                 v->arch.ceded = 0;
3360                 }
3361                 vc->runner = vcpu;
3362                 if (n_ceded == vc->n_runnable) {
3363                         kvmppc_vcore_blocked(vc);
3364                 } else if (need_resched()) {
3365                         kvmppc_vcore_preempt(vc);
3366                         /* Let something else run */
3367                         cond_resched_lock(&vc->lock);
3368                         if (vc->vcore_state == VCORE_PREEMPT)
3369                                 kvmppc_vcore_end_preempt(vc);
3370                 } else {
3371                         kvmppc_run_core(vc);
3372                 }
3373                 vc->runner = NULL;
3374         }
3375
3376         while (vcpu->arch.state == KVMPPC_VCPU_RUNNABLE &&
3377                (vc->vcore_state == VCORE_RUNNING ||
3378                 vc->vcore_state == VCORE_EXITING ||
3379                 vc->vcore_state == VCORE_PIGGYBACK))
3380                 kvmppc_wait_for_exec(vc, vcpu, TASK_UNINTERRUPTIBLE);
3381
3382         if (vc->vcore_state == VCORE_PREEMPT && vc->runner == NULL)
3383                 kvmppc_vcore_end_preempt(vc);
3384
3385         if (vcpu->arch.state == KVMPPC_VCPU_RUNNABLE) {
3386                 kvmppc_remove_runnable(vc, vcpu);
3387                 vcpu->stat.signal_exits++;
3388                 kvm_run->exit_reason = KVM_EXIT_INTR;
3389                 vcpu->arch.ret = -EINTR;
3390         }
3391
3392         if (vc->n_runnable && vc->vcore_state == VCORE_INACTIVE) {
3393                 /* Wake up some vcpu to run the core */
3394                 i = -1;
3395                 v = next_runnable_thread(vc, &i);
3396                 wake_up(&v->arch.cpu_run);
3397         }
3398
3399         trace_kvmppc_run_vcpu_exit(vcpu, kvm_run);
3400         spin_unlock(&vc->lock);
3401         return vcpu->arch.ret;
3402 }
3403
3404 static int kvmppc_vcpu_run_hv(struct kvm_run *run, struct kvm_vcpu *vcpu)
3405 {
3406         int r;
3407         int srcu_idx;
3408         unsigned long ebb_regs[3] = {}; /* shut up GCC */
3409         unsigned long user_tar = 0;
3410         unsigned int user_vrsave;
3411         struct kvm *kvm;
3412
3413         if (!vcpu->arch.sane) {
3414                 run->exit_reason = KVM_EXIT_INTERNAL_ERROR;
3415                 return -EINVAL;
3416         }
3417
3418         /*
3419          * Don't allow entry with a suspended transaction, because
3420          * the guest entry/exit code will lose it.
3421          * If the guest has TM enabled, save away their TM-related SPRs
3422          * (they will get restored by the TM unavailable interrupt).
3423          */
3424 #ifdef CONFIG_PPC_TRANSACTIONAL_MEM
3425         if (cpu_has_feature(CPU_FTR_TM) && current->thread.regs &&
3426             (current->thread.regs->msr & MSR_TM)) {
3427                 if (MSR_TM_ACTIVE(current->thread.regs->msr)) {
3428                         run->exit_reason = KVM_EXIT_FAIL_ENTRY;
3429                         run->fail_entry.hardware_entry_failure_reason = 0;
3430                         return -EINVAL;
3431                 }
3432                 /* Enable TM so we can read the TM SPRs */
3433                 mtmsr(mfmsr() | MSR_TM);
3434                 current->thread.tm_tfhar = mfspr(SPRN_TFHAR);
3435                 current->thread.tm_tfiar = mfspr(SPRN_TFIAR);
3436                 current->thread.tm_texasr = mfspr(SPRN_TEXASR);
3437                 current->thread.regs->msr &= ~MSR_TM;
3438         }
3439 #endif
3440
3441         /*
3442          * Force online to 1 for the sake of old userspace which doesn't
3443          * set it.
3444          */
3445         if (!vcpu->arch.online) {
3446                 atomic_inc(&vcpu->arch.vcore->online_count);
3447                 vcpu->arch.online = 1;
3448         }
3449
3450         kvmppc_core_prepare_to_enter(vcpu);
3451
3452         /* No need to go into the guest when all we'll do is come back out */
3453         if (signal_pending(current)) {
3454                 run->exit_reason = KVM_EXIT_INTR;
3455                 return -EINTR;
3456         }
3457
3458         kvm = vcpu->kvm;
3459         atomic_inc(&kvm->arch.vcpus_running);
3460         /* Order vcpus_running vs. mmu_ready, see kvmppc_alloc_reset_hpt */
3461         smp_mb();
3462
3463         flush_all_to_thread(current);
3464
3465         /* Save userspace EBB and other register values */
3466         if (cpu_has_feature(CPU_FTR_ARCH_207S)) {
3467                 ebb_regs[0] = mfspr(SPRN_EBBHR);
3468                 ebb_regs[1] = mfspr(SPRN_EBBRR);
3469                 ebb_regs[2] = mfspr(SPRN_BESCR);
3470                 user_tar = mfspr(SPRN_TAR);
3471         }
3472         user_vrsave = mfspr(SPRN_VRSAVE);
3473
3474         vcpu->arch.wqp = &vcpu->arch.vcore->wq;
3475         vcpu->arch.pgdir = current->mm->pgd;
3476         vcpu->arch.state = KVMPPC_VCPU_BUSY_IN_HOST;
3477
3478         do {
3479                 r = kvmppc_run_vcpu(run, vcpu);
3480
3481                 if (run->exit_reason == KVM_EXIT_PAPR_HCALL &&
3482                     !(vcpu->arch.shregs.msr & MSR_PR)) {
3483                         trace_kvm_hcall_enter(vcpu);
3484                         r = kvmppc_pseries_do_hcall(vcpu);
3485                         trace_kvm_hcall_exit(vcpu, r);
3486                         kvmppc_core_prepare_to_enter(vcpu);
3487                 } else if (r == RESUME_PAGE_FAULT) {
3488                         srcu_idx = srcu_read_lock(&kvm->srcu);
3489                         r = kvmppc_book3s_hv_page_fault(run, vcpu,
3490                                 vcpu->arch.fault_dar, vcpu->arch.fault_dsisr);
3491                         srcu_read_unlock(&kvm->srcu, srcu_idx);
3492                 } else if (r == RESUME_PASSTHROUGH) {
3493                         if (WARN_ON(xive_enabled()))
3494                                 r = H_SUCCESS;
3495                         else
3496                                 r = kvmppc_xics_rm_complete(vcpu, 0);
3497                 }
3498         } while (is_kvmppc_resume_guest(r));
3499
3500         /* Restore userspace EBB and other register values */
3501         if (cpu_has_feature(CPU_FTR_ARCH_207S)) {
3502                 mtspr(SPRN_EBBHR, ebb_regs[0]);
3503                 mtspr(SPRN_EBBRR, ebb_regs[1]);
3504                 mtspr(SPRN_BESCR, ebb_regs[2]);
3505                 mtspr(SPRN_TAR, user_tar);
3506                 mtspr(SPRN_FSCR, current->thread.fscr);
3507         }
3508         mtspr(SPRN_VRSAVE, user_vrsave);
3509
3510         vcpu->arch.state = KVMPPC_VCPU_NOTREADY;
3511         atomic_dec(&kvm->arch.vcpus_running);
3512         return r;
3513 }
3514
3515 static void kvmppc_add_seg_page_size(struct kvm_ppc_one_seg_page_size **sps,
3516                                      int shift, int sllp)
3517 {
3518         (*sps)->page_shift = shift;
3519         (*sps)->slb_enc = sllp;
3520         (*sps)->enc[0].page_shift = shift;
3521         (*sps)->enc[0].pte_enc = kvmppc_pgsize_lp_encoding(shift, shift);
3522         /*
3523          * Add 16MB MPSS support (may get filtered out by userspace)
3524          */
3525         if (shift != 24) {
3526                 int penc = kvmppc_pgsize_lp_encoding(shift, 24);
3527                 if (penc != -1) {
3528                         (*sps)->enc[1].page_shift = 24;
3529                         (*sps)->enc[1].pte_enc = penc;
3530                 }
3531         }
3532         (*sps)++;
3533 }
3534
3535 static int kvm_vm_ioctl_get_smmu_info_hv(struct kvm *kvm,
3536                                          struct kvm_ppc_smmu_info *info)
3537 {
3538         struct kvm_ppc_one_seg_page_size *sps;
3539
3540         /*
3541          * POWER7, POWER8 and POWER9 all support 32 storage keys for data.
3542          * POWER7 doesn't support keys for instruction accesses,
3543          * POWER8 and POWER9 do.
3544          */
3545         info->data_keys = 32;
3546         info->instr_keys = cpu_has_feature(CPU_FTR_ARCH_207S) ? 32 : 0;
3547
3548         /* POWER7, 8 and 9 all have 1T segments and 32-entry SLB */
3549         info->flags = KVM_PPC_PAGE_SIZES_REAL | KVM_PPC_1T_SEGMENTS;
3550         info->slb_size = 32;
3551
3552         /* We only support these sizes for now, and no muti-size segments */
3553         sps = &info->sps[0];
3554         kvmppc_add_seg_page_size(&sps, 12, 0);
3555         kvmppc_add_seg_page_size(&sps, 16, SLB_VSID_L | SLB_VSID_LP_01);
3556         kvmppc_add_seg_page_size(&sps, 24, SLB_VSID_L);
3557
3558         return 0;
3559 }
3560
3561 /*
3562  * Get (and clear) the dirty memory log for a memory slot.
3563  */
3564 static int kvm_vm_ioctl_get_dirty_log_hv(struct kvm *kvm,
3565                                          struct kvm_dirty_log *log)
3566 {
3567         struct kvm_memslots *slots;
3568         struct kvm_memory_slot *memslot;
3569         int i, r;
3570         unsigned long n;
3571         unsigned long *buf, *p;
3572         struct kvm_vcpu *vcpu;
3573
3574         mutex_lock(&kvm->slots_lock);
3575
3576         r = -EINVAL;
3577         if (log->slot >= KVM_USER_MEM_SLOTS)
3578                 goto out;
3579
3580         slots = kvm_memslots(kvm);
3581         memslot = id_to_memslot(slots, log->slot);
3582         r = -ENOENT;
3583         if (!memslot->dirty_bitmap)
3584                 goto out;
3585
3586         /*
3587          * Use second half of bitmap area because both HPT and radix
3588          * accumulate bits in the first half.
3589          */
3590         n = kvm_dirty_bitmap_bytes(memslot);
3591         buf = memslot->dirty_bitmap + n / sizeof(long);
3592         memset(buf, 0, n);
3593
3594         if (kvm_is_radix(kvm))
3595                 r = kvmppc_hv_get_dirty_log_radix(kvm, memslot, buf);
3596         else
3597                 r = kvmppc_hv_get_dirty_log_hpt(kvm, memslot, buf);
3598         if (r)
3599                 goto out;
3600
3601         /*
3602          * We accumulate dirty bits in the first half of the
3603          * memslot's dirty_bitmap area, for when pages are paged
3604          * out or modified by the host directly.  Pick up these
3605          * bits and add them to the map.
3606          */
3607         p = memslot->dirty_bitmap;
3608         for (i = 0; i < n / sizeof(long); ++i)
3609                 buf[i] |= xchg(&p[i], 0);
3610
3611         /* Harvest dirty bits from VPA and DTL updates */
3612         /* Note: we never modify the SLB shadow buffer areas */
3613         kvm_for_each_vcpu(i, vcpu, kvm) {
3614                 spin_lock(&vcpu->arch.vpa_update_lock);
3615                 kvmppc_harvest_vpa_dirty(&vcpu->arch.vpa, memslot, buf);
3616                 kvmppc_harvest_vpa_dirty(&vcpu->arch.dtl, memslot, buf);
3617                 spin_unlock(&vcpu->arch.vpa_update_lock);
3618         }
3619
3620         r = -EFAULT;
3621         if (copy_to_user(log->dirty_bitmap, buf, n))
3622                 goto out;
3623
3624         r = 0;
3625 out:
3626         mutex_unlock(&kvm->slots_lock);
3627         return r;
3628 }
3629
3630 static void kvmppc_core_free_memslot_hv(struct kvm_memory_slot *free,
3631                                         struct kvm_memory_slot *dont)
3632 {
3633         if (!dont || free->arch.rmap != dont->arch.rmap) {
3634                 vfree(free->arch.rmap);
3635                 free->arch.rmap = NULL;
3636         }
3637 }
3638
3639 static int kvmppc_core_create_memslot_hv(struct kvm_memory_slot *slot,
3640                                          unsigned long npages)
3641 {
3642         slot->arch.rmap = vzalloc(array_size(npages, sizeof(*slot->arch.rmap)));
3643         if (!slot->arch.rmap)
3644                 return -ENOMEM;
3645
3646         return 0;
3647 }
3648
3649 static int kvmppc_core_prepare_memory_region_hv(struct kvm *kvm,
3650                                         struct kvm_memory_slot *memslot,
3651                                         const struct kvm_userspace_memory_region *mem)
3652 {
3653         return 0;
3654 }
3655
3656 static void kvmppc_core_commit_memory_region_hv(struct kvm *kvm,
3657                                 const struct kvm_userspace_memory_region *mem,
3658                                 const struct kvm_memory_slot *old,
3659                                 const struct kvm_memory_slot *new)
3660 {
3661         unsigned long npages = mem->memory_size >> PAGE_SHIFT;
3662
3663         /*
3664          * If we are making a new memslot, it might make
3665          * some address that was previously cached as emulated
3666          * MMIO be no longer emulated MMIO, so invalidate
3667          * all the caches of emulated MMIO translations.
3668          */
3669         if (npages)
3670                 atomic64_inc(&kvm->arch.mmio_update);
3671 }
3672
3673 /*
3674  * Update LPCR values in kvm->arch and in vcores.
3675  * Caller must hold kvm->lock.
3676  */
3677 void kvmppc_update_lpcr(struct kvm *kvm, unsigned long lpcr, unsigned long mask)
3678 {
3679         long int i;
3680         u32 cores_done = 0;
3681
3682         if ((kvm->arch.lpcr & mask) == lpcr)
3683                 return;
3684
3685         kvm->arch.lpcr = (kvm->arch.lpcr & ~mask) | lpcr;
3686
3687         for (i = 0; i < KVM_MAX_VCORES; ++i) {
3688                 struct kvmppc_vcore *vc = kvm->arch.vcores[i];
3689                 if (!vc)
3690                         continue;
3691                 spin_lock(&vc->lock);
3692                 vc->lpcr = (vc->lpcr & ~mask) | lpcr;
3693                 spin_unlock(&vc->lock);
3694                 if (++cores_done >= kvm->arch.online_vcores)
3695                         break;
3696         }
3697 }
3698
3699 static void kvmppc_mmu_destroy_hv(struct kvm_vcpu *vcpu)
3700 {
3701         return;
3702 }
3703
3704 void kvmppc_setup_partition_table(struct kvm *kvm)
3705 {
3706         unsigned long dw0, dw1;
3707
3708         if (!kvm_is_radix(kvm)) {
3709                 /* PS field - page size for VRMA */
3710                 dw0 = ((kvm->arch.vrma_slb_v & SLB_VSID_L) >> 1) |
3711                         ((kvm->arch.vrma_slb_v & SLB_VSID_LP) << 1);
3712                 /* HTABSIZE and HTABORG fields */
3713                 dw0 |= kvm->arch.sdr1;
3714
3715                 /* Second dword as set by userspace */
3716                 dw1 = kvm->arch.process_table;
3717         } else {
3718                 dw0 = PATB_HR | radix__get_tree_size() |
3719                         __pa(kvm->arch.pgtable) | RADIX_PGD_INDEX_SIZE;
3720                 dw1 = PATB_GR | kvm->arch.process_table;
3721         }
3722
3723         mmu_partition_table_set_entry(kvm->arch.lpid, dw0, dw1);
3724 }
3725
3726 /*
3727  * Set up HPT (hashed page table) and RMA (real-mode area).
3728  * Must be called with kvm->lock held.
3729  */
3730 static int kvmppc_hv_setup_htab_rma(struct kvm_vcpu *vcpu)
3731 {
3732         int err = 0;
3733         struct kvm *kvm = vcpu->kvm;
3734         unsigned long hva;
3735         struct kvm_memory_slot *memslot;
3736         struct vm_area_struct *vma;
3737         unsigned long lpcr = 0, senc;
3738         unsigned long psize, porder;
3739         int srcu_idx;
3740
3741         /* Allocate hashed page table (if not done already) and reset it */
3742         if (!kvm->arch.hpt.virt) {
3743                 int order = KVM_DEFAULT_HPT_ORDER;
3744                 struct kvm_hpt_info info;
3745
3746                 err = kvmppc_allocate_hpt(&info, order);
3747                 /* If we get here, it means userspace didn't specify a
3748                  * size explicitly.  So, try successively smaller
3749                  * sizes if the default failed. */
3750                 while ((err == -ENOMEM) && --order >= PPC_MIN_HPT_ORDER)
3751                         err  = kvmppc_allocate_hpt(&info, order);
3752
3753                 if (err < 0) {
3754                         pr_err("KVM: Couldn't alloc HPT\n");
3755                         goto out;
3756                 }
3757
3758                 kvmppc_set_hpt(kvm, &info);
3759         }
3760
3761         /* Look up the memslot for guest physical address 0 */
3762         srcu_idx = srcu_read_lock(&kvm->srcu);
3763         memslot = gfn_to_memslot(kvm, 0);
3764
3765         /* We must have some memory at 0 by now */
3766         err = -EINVAL;
3767         if (!memslot || (memslot->flags & KVM_MEMSLOT_INVALID))
3768                 goto out_srcu;
3769
3770         /* Look up the VMA for the start of this memory slot */
3771         hva = memslot->userspace_addr;
3772         down_read(&current->mm->mmap_sem);
3773         vma = find_vma(current->mm, hva);
3774         if (!vma || vma->vm_start > hva || (vma->vm_flags & VM_IO))
3775                 goto up_out;
3776
3777         psize = vma_kernel_pagesize(vma);
3778
3779         up_read(&current->mm->mmap_sem);
3780
3781         /* We can handle 4k, 64k or 16M pages in the VRMA */
3782         if (psize >= 0x1000000)
3783                 psize = 0x1000000;
3784         else if (psize >= 0x10000)
3785                 psize = 0x10000;
3786         else
3787                 psize = 0x1000;
3788         porder = __ilog2(psize);
3789
3790         senc = slb_pgsize_encoding(psize);
3791         kvm->arch.vrma_slb_v = senc | SLB_VSID_B_1T |
3792                 (VRMA_VSID << SLB_VSID_SHIFT_1T);
3793         /* Create HPTEs in the hash page table for the VRMA */
3794         kvmppc_map_vrma(vcpu, memslot, porder);
3795
3796         /* Update VRMASD field in the LPCR */
3797         if (!cpu_has_feature(CPU_FTR_ARCH_300)) {
3798                 /* the -4 is to account for senc values starting at 0x10 */
3799                 lpcr = senc << (LPCR_VRMASD_SH - 4);
3800                 kvmppc_update_lpcr(kvm, lpcr, LPCR_VRMASD);
3801         }
3802
3803         /* Order updates to kvm->arch.lpcr etc. vs. mmu_ready */
3804         smp_wmb();
3805         err = 0;
3806  out_srcu:
3807         srcu_read_unlock(&kvm->srcu, srcu_idx);
3808  out:
3809         return err;
3810
3811  up_out:
3812         up_read(&current->mm->mmap_sem);
3813         goto out_srcu;
3814 }
3815
3816 /* Must be called with kvm->lock held and mmu_ready = 0 and no vcpus running */
3817 int kvmppc_switch_mmu_to_hpt(struct kvm *kvm)
3818 {
3819         kvmppc_free_radix(kvm);
3820         kvmppc_update_lpcr(kvm, LPCR_VPM1,
3821                            LPCR_VPM1 | LPCR_UPRT | LPCR_GTSE | LPCR_HR);
3822         kvmppc_rmap_reset(kvm);
3823         kvm->arch.radix = 0;
3824         kvm->arch.process_table = 0;
3825         return 0;
3826 }
3827
3828 /* Must be called with kvm->lock held and mmu_ready = 0 and no vcpus running */
3829 int kvmppc_switch_mmu_to_radix(struct kvm *kvm)
3830 {
3831         int err;
3832
3833         err = kvmppc_init_vm_radix(kvm);
3834         if (err)
3835                 return err;
3836
3837         kvmppc_free_hpt(&kvm->arch.hpt);
3838         kvmppc_update_lpcr(kvm, LPCR_UPRT | LPCR_GTSE | LPCR_HR,
3839                            LPCR_VPM1 | LPCR_UPRT | LPCR_GTSE | LPCR_HR);
3840         kvm->arch.radix = 1;
3841         return 0;
3842 }
3843
3844 #ifdef CONFIG_KVM_XICS
3845 /*
3846  * Allocate a per-core structure for managing state about which cores are
3847  * running in the host versus the guest and for exchanging data between
3848  * real mode KVM and CPU running in the host.
3849  * This is only done for the first VM.
3850  * The allocated structure stays even if all VMs have stopped.
3851  * It is only freed when the kvm-hv module is unloaded.
3852  * It's OK for this routine to fail, we just don't support host
3853  * core operations like redirecting H_IPI wakeups.
3854  */
3855 void kvmppc_alloc_host_rm_ops(void)
3856 {
3857         struct kvmppc_host_rm_ops *ops;
3858         unsigned long l_ops;
3859         int cpu, core;
3860         int size;
3861
3862         /* Not the first time here ? */
3863         if (kvmppc_host_rm_ops_hv != NULL)
3864                 return;
3865
3866         ops = kzalloc(sizeof(struct kvmppc_host_rm_ops), GFP_KERNEL);
3867         if (!ops)
3868                 return;
3869
3870         size = cpu_nr_cores() * sizeof(struct kvmppc_host_rm_core);
3871         ops->rm_core = kzalloc(size, GFP_KERNEL);
3872
3873         if (!ops->rm_core) {
3874                 kfree(ops);
3875                 return;
3876         }
3877
3878         cpus_read_lock();
3879
3880         for (cpu = 0; cpu < nr_cpu_ids; cpu += threads_per_core) {
3881                 if (!cpu_online(cpu))
3882                         continue;
3883
3884                 core = cpu >> threads_shift;
3885                 ops->rm_core[core].rm_state.in_host = 1;
3886         }
3887
3888         ops->vcpu_kick = kvmppc_fast_vcpu_kick_hv;
3889
3890         /*
3891          * Make the contents of the kvmppc_host_rm_ops structure visible
3892          * to other CPUs before we assign it to the global variable.
3893          * Do an atomic assignment (no locks used here), but if someone
3894          * beats us to it, just free our copy and return.
3895          */
3896         smp_wmb();
3897         l_ops = (unsigned long) ops;
3898
3899         if (cmpxchg64((unsigned long *)&kvmppc_host_rm_ops_hv, 0, l_ops)) {
3900                 cpus_read_unlock();
3901                 kfree(ops->rm_core);
3902                 kfree(ops);
3903                 return;
3904         }
3905
3906         cpuhp_setup_state_nocalls_cpuslocked(CPUHP_KVM_PPC_BOOK3S_PREPARE,
3907                                              "ppc/kvm_book3s:prepare",
3908                                              kvmppc_set_host_core,
3909                                              kvmppc_clear_host_core);
3910         cpus_read_unlock();
3911 }
3912
3913 void kvmppc_free_host_rm_ops(void)
3914 {
3915         if (kvmppc_host_rm_ops_hv) {
3916                 cpuhp_remove_state_nocalls(CPUHP_KVM_PPC_BOOK3S_PREPARE);
3917                 kfree(kvmppc_host_rm_ops_hv->rm_core);
3918                 kfree(kvmppc_host_rm_ops_hv);
3919                 kvmppc_host_rm_ops_hv = NULL;
3920         }
3921 }
3922 #endif
3923
3924 static int kvmppc_core_init_vm_hv(struct kvm *kvm)
3925 {
3926         unsigned long lpcr, lpid;
3927         char buf[32];
3928         int ret;
3929
3930         /* Allocate the guest's logical partition ID */
3931
3932         lpid = kvmppc_alloc_lpid();
3933         if ((long)lpid < 0)
3934                 return -ENOMEM;
3935         kvm->arch.lpid = lpid;
3936
3937         kvmppc_alloc_host_rm_ops();
3938
3939         /*
3940          * Since we don't flush the TLB when tearing down a VM,
3941          * and this lpid might have previously been used,
3942          * make sure we flush on each core before running the new VM.
3943          * On POWER9, the tlbie in mmu_partition_table_set_entry()
3944          * does this flush for us.
3945          */
3946         if (!cpu_has_feature(CPU_FTR_ARCH_300))
3947                 cpumask_setall(&kvm->arch.need_tlb_flush);
3948
3949         /* Start out with the default set of hcalls enabled */
3950         memcpy(kvm->arch.enabled_hcalls, default_enabled_hcalls,
3951                sizeof(kvm->arch.enabled_hcalls));
3952
3953         if (!cpu_has_feature(CPU_FTR_ARCH_300))
3954                 kvm->arch.host_sdr1 = mfspr(SPRN_SDR1);
3955
3956         /* Init LPCR for virtual RMA mode */
3957         kvm->arch.host_lpid = mfspr(SPRN_LPID);
3958         kvm->arch.host_lpcr = lpcr = mfspr(SPRN_LPCR);
3959         lpcr &= LPCR_PECE | LPCR_LPES;
3960         lpcr |= (4UL << LPCR_DPFD_SH) | LPCR_HDICE |
3961                 LPCR_VPM0 | LPCR_VPM1;
3962         kvm->arch.vrma_slb_v = SLB_VSID_B_1T |
3963                 (VRMA_VSID << SLB_VSID_SHIFT_1T);
3964         /* On POWER8 turn on online bit to enable PURR/SPURR */
3965         if (cpu_has_feature(CPU_FTR_ARCH_207S))
3966                 lpcr |= LPCR_ONL;
3967         /*
3968          * On POWER9, VPM0 bit is reserved (VPM0=1 behaviour is assumed)
3969          * Set HVICE bit to enable hypervisor virtualization interrupts.
3970          * Set HEIC to prevent OS interrupts to go to hypervisor (should
3971          * be unnecessary but better safe than sorry in case we re-enable
3972          * EE in HV mode with this LPCR still set)
3973          */
3974         if (cpu_has_feature(CPU_FTR_ARCH_300)) {
3975                 lpcr &= ~LPCR_VPM0;
3976                 lpcr |= LPCR_HVICE | LPCR_HEIC;
3977
3978                 /*
3979                  * If xive is enabled, we route 0x500 interrupts directly
3980                  * to the guest.
3981                  */
3982                 if (xive_enabled())
3983                         lpcr |= LPCR_LPES;
3984         }
3985
3986         /*
3987          * If the host uses radix, the guest starts out as radix.
3988          */
3989         if (radix_enabled()) {
3990                 kvm->arch.radix = 1;
3991                 kvm->arch.mmu_ready = 1;
3992                 lpcr &= ~LPCR_VPM1;
3993                 lpcr |= LPCR_UPRT | LPCR_GTSE | LPCR_HR;
3994                 ret = kvmppc_init_vm_radix(kvm);
3995                 if (ret) {
3996                         kvmppc_free_lpid(kvm->arch.lpid);
3997                         return ret;
3998                 }
3999                 kvmppc_setup_partition_table(kvm);
4000         }
4001
4002         kvm->arch.lpcr = lpcr;
4003
4004         /* Initialization for future HPT resizes */
4005         kvm->arch.resize_hpt = NULL;
4006
4007         /*
4008          * Work out how many sets the TLB has, for the use of
4009          * the TLB invalidation loop in book3s_hv_rmhandlers.S.
4010          */
4011         if (radix_enabled())
4012                 kvm->arch.tlb_sets = POWER9_TLB_SETS_RADIX;     /* 128 */
4013         else if (cpu_has_feature(CPU_FTR_ARCH_300))
4014                 kvm->arch.tlb_sets = POWER9_TLB_SETS_HASH;      /* 256 */
4015         else if (cpu_has_feature(CPU_FTR_ARCH_207S))
4016                 kvm->arch.tlb_sets = POWER8_TLB_SETS;           /* 512 */
4017         else
4018                 kvm->arch.tlb_sets = POWER7_TLB_SETS;           /* 128 */
4019
4020         /*
4021          * Track that we now have a HV mode VM active. This blocks secondary
4022          * CPU threads from coming online.
4023          * On POWER9, we only need to do this if the "indep_threads_mode"
4024          * module parameter has been set to N.
4025          */
4026         if (cpu_has_feature(CPU_FTR_ARCH_300))
4027                 kvm->arch.threads_indep = indep_threads_mode;
4028         if (!kvm->arch.threads_indep)
4029                 kvm_hv_vm_activated();
4030
4031         /*
4032          * Initialize smt_mode depending on processor.
4033          * POWER8 and earlier have to use "strict" threading, where
4034          * all vCPUs in a vcore have to run on the same (sub)core,
4035          * whereas on POWER9 the threads can each run a different
4036          * guest.
4037          */
4038         if (!cpu_has_feature(CPU_FTR_ARCH_300))
4039                 kvm->arch.smt_mode = threads_per_subcore;
4040         else
4041                 kvm->arch.smt_mode = 1;
4042         kvm->arch.emul_smt_mode = 1;
4043
4044         /*
4045          * Create a debugfs directory for the VM
4046          */
4047         snprintf(buf, sizeof(buf), "vm%d", current->pid);
4048         kvm->arch.debugfs_dir = debugfs_create_dir(buf, kvm_debugfs_dir);
4049         kvmppc_mmu_debugfs_init(kvm);
4050
4051         return 0;
4052 }
4053
4054 static void kvmppc_free_vcores(struct kvm *kvm)
4055 {
4056         long int i;
4057
4058         for (i = 0; i < KVM_MAX_VCORES; ++i)
4059                 kfree(kvm->arch.vcores[i]);
4060         kvm->arch.online_vcores = 0;
4061 }
4062
4063 static void kvmppc_core_destroy_vm_hv(struct kvm *kvm)
4064 {
4065         debugfs_remove_recursive(kvm->arch.debugfs_dir);
4066
4067         if (!kvm->arch.threads_indep)
4068                 kvm_hv_vm_deactivated();
4069
4070         kvmppc_free_vcores(kvm);
4071
4072         kvmppc_free_lpid(kvm->arch.lpid);
4073
4074         if (kvm_is_radix(kvm))
4075                 kvmppc_free_radix(kvm);
4076         else
4077                 kvmppc_free_hpt(&kvm->arch.hpt);
4078
4079         kvmppc_free_pimap(kvm);
4080 }
4081
4082 /* We don't need to emulate any privileged instructions or dcbz */
4083 static int kvmppc_core_emulate_op_hv(struct kvm_run *run, struct kvm_vcpu *vcpu,
4084                                      unsigned int inst, int *advance)
4085 {
4086         return EMULATE_FAIL;
4087 }
4088
4089 static int kvmppc_core_emulate_mtspr_hv(struct kvm_vcpu *vcpu, int sprn,
4090                                         ulong spr_val)
4091 {
4092         return EMULATE_FAIL;
4093 }
4094
4095 static int kvmppc_core_emulate_mfspr_hv(struct kvm_vcpu *vcpu, int sprn,
4096                                         ulong *spr_val)
4097 {
4098         return EMULATE_FAIL;
4099 }
4100
4101 static int kvmppc_core_check_processor_compat_hv(void)
4102 {
4103         if (!cpu_has_feature(CPU_FTR_HVMODE) ||
4104             !cpu_has_feature(CPU_FTR_ARCH_206))
4105                 return -EIO;
4106
4107         return 0;
4108 }
4109
4110 #ifdef CONFIG_KVM_XICS
4111
4112 void kvmppc_free_pimap(struct kvm *kvm)
4113 {
4114         kfree(kvm->arch.pimap);
4115 }
4116
4117 static struct kvmppc_passthru_irqmap *kvmppc_alloc_pimap(void)
4118 {
4119         return kzalloc(sizeof(struct kvmppc_passthru_irqmap), GFP_KERNEL);
4120 }
4121
4122 static int kvmppc_set_passthru_irq(struct kvm *kvm, int host_irq, int guest_gsi)
4123 {
4124         struct irq_desc *desc;
4125         struct kvmppc_irq_map *irq_map;
4126         struct kvmppc_passthru_irqmap *pimap;
4127         struct irq_chip *chip;
4128         int i, rc = 0;
4129
4130         if (!kvm_irq_bypass)
4131                 return 1;
4132
4133         desc = irq_to_desc(host_irq);
4134         if (!desc)
4135                 return -EIO;
4136
4137         mutex_lock(&kvm->lock);
4138
4139         pimap = kvm->arch.pimap;
4140         if (pimap == NULL) {
4141                 /* First call, allocate structure to hold IRQ map */
4142                 pimap = kvmppc_alloc_pimap();
4143                 if (pimap == NULL) {
4144                         mutex_unlock(&kvm->lock);
4145                         return -ENOMEM;
4146                 }
4147                 kvm->arch.pimap = pimap;
4148         }
4149
4150         /*
4151          * For now, we only support interrupts for which the EOI operation
4152          * is an OPAL call followed by a write to XIRR, since that's
4153          * what our real-mode EOI code does, or a XIVE interrupt
4154          */
4155         chip = irq_data_get_irq_chip(&desc->irq_data);
4156         if (!chip || !(is_pnv_opal_msi(chip) || is_xive_irq(chip))) {
4157                 pr_warn("kvmppc_set_passthru_irq_hv: Could not assign IRQ map for (%d,%d)\n",
4158                         host_irq, guest_gsi);
4159                 mutex_unlock(&kvm->lock);
4160                 return -ENOENT;
4161         }
4162
4163         /*
4164          * See if we already have an entry for this guest IRQ number.
4165          * If it's mapped to a hardware IRQ number, that's an error,
4166          * otherwise re-use this entry.
4167          */
4168         for (i = 0; i < pimap->n_mapped; i++) {
4169                 if (guest_gsi == pimap->mapped[i].v_hwirq) {
4170                         if (pimap->mapped[i].r_hwirq) {
4171                                 mutex_unlock(&kvm->lock);
4172                                 return -EINVAL;
4173                         }
4174                         break;
4175                 }
4176         }
4177
4178         if (i == KVMPPC_PIRQ_MAPPED) {
4179                 mutex_unlock(&kvm->lock);
4180                 return -EAGAIN;         /* table is full */
4181         }
4182
4183         irq_map = &pimap->mapped[i];
4184
4185         irq_map->v_hwirq = guest_gsi;
4186         irq_map->desc = desc;
4187
4188         /*
4189          * Order the above two stores before the next to serialize with
4190          * the KVM real mode handler.
4191          */
4192         smp_wmb();
4193         irq_map->r_hwirq = desc->irq_data.hwirq;
4194
4195         if (i == pimap->n_mapped)
4196                 pimap->n_mapped++;
4197
4198         if (xive_enabled())
4199                 rc = kvmppc_xive_set_mapped(kvm, guest_gsi, desc);
4200         else
4201                 kvmppc_xics_set_mapped(kvm, guest_gsi, desc->irq_data.hwirq);
4202         if (rc)
4203                 irq_map->r_hwirq = 0;
4204
4205         mutex_unlock(&kvm->lock);
4206
4207         return 0;
4208 }
4209
4210 static int kvmppc_clr_passthru_irq(struct kvm *kvm, int host_irq, int guest_gsi)
4211 {
4212         struct irq_desc *desc;
4213         struct kvmppc_passthru_irqmap *pimap;
4214         int i, rc = 0;
4215
4216         if (!kvm_irq_bypass)
4217                 return 0;
4218
4219         desc = irq_to_desc(host_irq);
4220         if (!desc)
4221                 return -EIO;
4222
4223         mutex_lock(&kvm->lock);
4224         if (!kvm->arch.pimap)
4225                 goto unlock;
4226
4227         pimap = kvm->arch.pimap;
4228
4229         for (i = 0; i < pimap->n_mapped; i++) {
4230                 if (guest_gsi == pimap->mapped[i].v_hwirq)
4231                         break;
4232         }
4233
4234         if (i == pimap->n_mapped) {
4235                 mutex_unlock(&kvm->lock);
4236                 return -ENODEV;
4237         }
4238
4239         if (xive_enabled())
4240                 rc = kvmppc_xive_clr_mapped(kvm, guest_gsi, pimap->mapped[i].desc);
4241         else
4242                 kvmppc_xics_clr_mapped(kvm, guest_gsi, pimap->mapped[i].r_hwirq);
4243
4244         /* invalidate the entry (what do do on error from the above ?) */
4245         pimap->mapped[i].r_hwirq = 0;
4246
4247         /*
4248          * We don't free this structure even when the count goes to
4249          * zero. The structure is freed when we destroy the VM.
4250          */
4251  unlock:
4252         mutex_unlock(&kvm->lock);
4253         return rc;
4254 }
4255
4256 static int kvmppc_irq_bypass_add_producer_hv(struct irq_bypass_consumer *cons,
4257                                              struct irq_bypass_producer *prod)
4258 {
4259         int ret = 0;
4260         struct kvm_kernel_irqfd *irqfd =
4261                 container_of(cons, struct kvm_kernel_irqfd, consumer);
4262
4263         irqfd->producer = prod;
4264
4265         ret = kvmppc_set_passthru_irq(irqfd->kvm, prod->irq, irqfd->gsi);
4266         if (ret)
4267                 pr_info("kvmppc_set_passthru_irq (irq %d, gsi %d) fails: %d\n",
4268                         prod->irq, irqfd->gsi, ret);
4269
4270         return ret;
4271 }
4272
4273 static void kvmppc_irq_bypass_del_producer_hv(struct irq_bypass_consumer *cons,
4274                                               struct irq_bypass_producer *prod)
4275 {
4276         int ret;
4277         struct kvm_kernel_irqfd *irqfd =
4278                 container_of(cons, struct kvm_kernel_irqfd, consumer);
4279
4280         irqfd->producer = NULL;
4281
4282         /*
4283          * When producer of consumer is unregistered, we change back to
4284          * default external interrupt handling mode - KVM real mode
4285          * will switch back to host.
4286          */
4287         ret = kvmppc_clr_passthru_irq(irqfd->kvm, prod->irq, irqfd->gsi);
4288         if (ret)
4289                 pr_warn("kvmppc_clr_passthru_irq (irq %d, gsi %d) fails: %d\n",
4290                         prod->irq, irqfd->gsi, ret);
4291 }
4292 #endif
4293
4294 static long kvm_arch_vm_ioctl_hv(struct file *filp,
4295                                  unsigned int ioctl, unsigned long arg)
4296 {
4297         struct kvm *kvm __maybe_unused = filp->private_data;
4298         void __user *argp = (void __user *)arg;
4299         long r;
4300
4301         switch (ioctl) {
4302
4303         case KVM_PPC_ALLOCATE_HTAB: {
4304                 u32 htab_order;
4305
4306                 r = -EFAULT;
4307                 if (get_user(htab_order, (u32 __user *)argp))
4308                         break;
4309                 r = kvmppc_alloc_reset_hpt(kvm, htab_order);
4310                 if (r)
4311                         break;
4312                 r = 0;
4313                 break;
4314         }
4315
4316         case KVM_PPC_GET_HTAB_FD: {
4317                 struct kvm_get_htab_fd ghf;
4318
4319                 r = -EFAULT;
4320                 if (copy_from_user(&ghf, argp, sizeof(ghf)))
4321                         break;
4322                 r = kvm_vm_ioctl_get_htab_fd(kvm, &ghf);
4323                 break;
4324         }
4325
4326         case KVM_PPC_RESIZE_HPT_PREPARE: {
4327                 struct kvm_ppc_resize_hpt rhpt;
4328
4329                 r = -EFAULT;
4330                 if (copy_from_user(&rhpt, argp, sizeof(rhpt)))
4331                         break;
4332
4333                 r = kvm_vm_ioctl_resize_hpt_prepare(kvm, &rhpt);
4334                 break;
4335         }
4336
4337         case KVM_PPC_RESIZE_HPT_COMMIT: {
4338                 struct kvm_ppc_resize_hpt rhpt;
4339
4340                 r = -EFAULT;
4341                 if (copy_from_user(&rhpt, argp, sizeof(rhpt)))
4342                         break;
4343
4344                 r = kvm_vm_ioctl_resize_hpt_commit(kvm, &rhpt);
4345                 break;
4346         }
4347
4348         default:
4349                 r = -ENOTTY;
4350         }
4351
4352         return r;
4353 }
4354
4355 /*
4356  * List of hcall numbers to enable by default.
4357  * For compatibility with old userspace, we enable by default
4358  * all hcalls that were implemented before the hcall-enabling
4359  * facility was added.  Note this list should not include H_RTAS.
4360  */
4361 static unsigned int default_hcall_list[] = {
4362         H_REMOVE,
4363         H_ENTER,
4364         H_READ,
4365         H_PROTECT,
4366         H_BULK_REMOVE,
4367         H_GET_TCE,
4368         H_PUT_TCE,
4369         H_SET_DABR,
4370         H_SET_XDABR,
4371         H_CEDE,
4372         H_PROD,
4373         H_CONFER,
4374         H_REGISTER_VPA,
4375 #ifdef CONFIG_KVM_XICS
4376         H_EOI,
4377         H_CPPR,
4378         H_IPI,
4379         H_IPOLL,
4380         H_XIRR,
4381         H_XIRR_X,
4382 #endif
4383         0
4384 };
4385
4386 static void init_default_hcalls(void)
4387 {
4388         int i;
4389         unsigned int hcall;
4390
4391         for (i = 0; default_hcall_list[i]; ++i) {
4392                 hcall = default_hcall_list[i];
4393                 WARN_ON(!kvmppc_hcall_impl_hv(hcall));
4394                 __set_bit(hcall / 4, default_enabled_hcalls);
4395         }
4396 }
4397
4398 static int kvmhv_configure_mmu(struct kvm *kvm, struct kvm_ppc_mmuv3_cfg *cfg)
4399 {
4400         unsigned long lpcr;
4401         int radix;
4402         int err;
4403
4404         /* If not on a POWER9, reject it */
4405         if (!cpu_has_feature(CPU_FTR_ARCH_300))
4406                 return -ENODEV;
4407
4408         /* If any unknown flags set, reject it */
4409         if (cfg->flags & ~(KVM_PPC_MMUV3_RADIX | KVM_PPC_MMUV3_GTSE))
4410                 return -EINVAL;
4411
4412         /* GR (guest radix) bit in process_table field must match */
4413         radix = !!(cfg->flags & KVM_PPC_MMUV3_RADIX);
4414         if (!!(cfg->process_table & PATB_GR) != radix)
4415                 return -EINVAL;
4416
4417         /* Process table size field must be reasonable, i.e. <= 24 */
4418         if ((cfg->process_table & PRTS_MASK) > 24)
4419                 return -EINVAL;
4420
4421         /* We can change a guest to/from radix now, if the host is radix */
4422         if (radix && !radix_enabled())
4423                 return -EINVAL;
4424
4425         mutex_lock(&kvm->lock);
4426         if (radix != kvm_is_radix(kvm)) {
4427                 if (kvm->arch.mmu_ready) {
4428                         kvm->arch.mmu_ready = 0;
4429                         /* order mmu_ready vs. vcpus_running */
4430                         smp_mb();
4431                         if (atomic_read(&kvm->arch.vcpus_running)) {
4432                                 kvm->arch.mmu_ready = 1;
4433                                 err = -EBUSY;
4434                                 goto out_unlock;
4435                         }
4436                 }
4437                 if (radix)
4438                         err = kvmppc_switch_mmu_to_radix(kvm);
4439                 else
4440                         err = kvmppc_switch_mmu_to_hpt(kvm);
4441                 if (err)
4442                         goto out_unlock;
4443         }
4444
4445         kvm->arch.process_table = cfg->process_table;
4446         kvmppc_setup_partition_table(kvm);
4447
4448         lpcr = (cfg->flags & KVM_PPC_MMUV3_GTSE) ? LPCR_GTSE : 0;
4449         kvmppc_update_lpcr(kvm, lpcr, LPCR_GTSE);
4450         err = 0;
4451
4452  out_unlock:
4453         mutex_unlock(&kvm->lock);
4454         return err;
4455 }
4456
4457 static struct kvmppc_ops kvm_ops_hv = {
4458         .get_sregs = kvm_arch_vcpu_ioctl_get_sregs_hv,
4459         .set_sregs = kvm_arch_vcpu_ioctl_set_sregs_hv,
4460         .get_one_reg = kvmppc_get_one_reg_hv,
4461         .set_one_reg = kvmppc_set_one_reg_hv,
4462         .vcpu_load   = kvmppc_core_vcpu_load_hv,
4463         .vcpu_put    = kvmppc_core_vcpu_put_hv,
4464         .set_msr     = kvmppc_set_msr_hv,
4465         .vcpu_run    = kvmppc_vcpu_run_hv,
4466         .vcpu_create = kvmppc_core_vcpu_create_hv,
4467         .vcpu_free   = kvmppc_core_vcpu_free_hv,
4468         .check_requests = kvmppc_core_check_requests_hv,
4469         .get_dirty_log  = kvm_vm_ioctl_get_dirty_log_hv,
4470         .flush_memslot  = kvmppc_core_flush_memslot_hv,
4471         .prepare_memory_region = kvmppc_core_prepare_memory_region_hv,
4472         .commit_memory_region  = kvmppc_core_commit_memory_region_hv,
4473         .unmap_hva_range = kvm_unmap_hva_range_hv,
4474         .age_hva  = kvm_age_hva_hv,
4475         .test_age_hva = kvm_test_age_hva_hv,
4476         .set_spte_hva = kvm_set_spte_hva_hv,
4477         .mmu_destroy  = kvmppc_mmu_destroy_hv,
4478         .free_memslot = kvmppc_core_free_memslot_hv,
4479         .create_memslot = kvmppc_core_create_memslot_hv,
4480         .init_vm =  kvmppc_core_init_vm_hv,
4481         .destroy_vm = kvmppc_core_destroy_vm_hv,
4482         .get_smmu_info = kvm_vm_ioctl_get_smmu_info_hv,
4483         .emulate_op = kvmppc_core_emulate_op_hv,
4484         .emulate_mtspr = kvmppc_core_emulate_mtspr_hv,
4485         .emulate_mfspr = kvmppc_core_emulate_mfspr_hv,
4486         .fast_vcpu_kick = kvmppc_fast_vcpu_kick_hv,
4487         .arch_vm_ioctl  = kvm_arch_vm_ioctl_hv,
4488         .hcall_implemented = kvmppc_hcall_impl_hv,
4489 #ifdef CONFIG_KVM_XICS
4490         .irq_bypass_add_producer = kvmppc_irq_bypass_add_producer_hv,
4491         .irq_bypass_del_producer = kvmppc_irq_bypass_del_producer_hv,
4492 #endif
4493         .configure_mmu = kvmhv_configure_mmu,
4494         .get_rmmu_info = kvmhv_get_rmmu_info,
4495         .set_smt_mode = kvmhv_set_smt_mode,
4496 };
4497
4498 static int kvm_init_subcore_bitmap(void)
4499 {
4500         int i, j;
4501         int nr_cores = cpu_nr_cores();
4502         struct sibling_subcore_state *sibling_subcore_state;
4503
4504         for (i = 0; i < nr_cores; i++) {
4505                 int first_cpu = i * threads_per_core;
4506                 int node = cpu_to_node(first_cpu);
4507
4508                 /* Ignore if it is already allocated. */
4509                 if (paca_ptrs[first_cpu]->sibling_subcore_state)
4510                         continue;
4511
4512                 sibling_subcore_state =
4513                         kmalloc_node(sizeof(struct sibling_subcore_state),
4514                                                         GFP_KERNEL, node);
4515                 if (!sibling_subcore_state)
4516                         return -ENOMEM;
4517
4518                 memset(sibling_subcore_state, 0,
4519                                 sizeof(struct sibling_subcore_state));
4520
4521                 for (j = 0; j < threads_per_core; j++) {
4522                         int cpu = first_cpu + j;
4523
4524                         paca_ptrs[cpu]->sibling_subcore_state =
4525                                                 sibling_subcore_state;
4526                 }
4527         }
4528         return 0;
4529 }
4530
4531 static int kvmppc_radix_possible(void)
4532 {
4533         return cpu_has_feature(CPU_FTR_ARCH_300) && radix_enabled();
4534 }
4535
4536 static int kvmppc_book3s_init_hv(void)
4537 {
4538         int r;
4539         /*
4540          * FIXME!! Do we need to check on all cpus ?
4541          */
4542         r = kvmppc_core_check_processor_compat_hv();
4543         if (r < 0)
4544                 return -ENODEV;
4545
4546         r = kvm_init_subcore_bitmap();
4547         if (r)
4548                 return r;
4549
4550         /*
4551          * We need a way of accessing the XICS interrupt controller,
4552          * either directly, via paca_ptrs[cpu]->kvm_hstate.xics_phys, or
4553          * indirectly, via OPAL.
4554          */
4555 #ifdef CONFIG_SMP
4556         if (!xive_enabled() && !local_paca->kvm_hstate.xics_phys) {
4557                 struct device_node *np;
4558
4559                 np = of_find_compatible_node(NULL, NULL, "ibm,opal-intc");
4560                 if (!np) {
4561                         pr_err("KVM-HV: Cannot determine method for accessing XICS\n");
4562                         return -ENODEV;
4563                 }
4564         }
4565 #endif
4566
4567         kvm_ops_hv.owner = THIS_MODULE;
4568         kvmppc_hv_ops = &kvm_ops_hv;
4569
4570         init_default_hcalls();
4571
4572         init_vcore_lists();
4573
4574         r = kvmppc_mmu_hv_init();
4575         if (r)
4576                 return r;
4577
4578         if (kvmppc_radix_possible())
4579                 r = kvmppc_radix_init();
4580
4581         /*
4582          * POWER9 chips before version 2.02 can't have some threads in
4583          * HPT mode and some in radix mode on the same core.
4584          */
4585         if (cpu_has_feature(CPU_FTR_ARCH_300)) {
4586                 unsigned int pvr = mfspr(SPRN_PVR);
4587                 if ((pvr >> 16) == PVR_POWER9 &&
4588                     (((pvr & 0xe000) == 0 && (pvr & 0xfff) < 0x202) ||
4589                      ((pvr & 0xe000) == 0x2000 && (pvr & 0xfff) < 0x101)))
4590                         no_mixing_hpt_and_radix = true;
4591         }
4592
4593         return r;
4594 }
4595
4596 static void kvmppc_book3s_exit_hv(void)
4597 {
4598         kvmppc_free_host_rm_ops();
4599         if (kvmppc_radix_possible())
4600                 kvmppc_radix_exit();
4601         kvmppc_hv_ops = NULL;
4602 }
4603
4604 module_init(kvmppc_book3s_init_hv);
4605 module_exit(kvmppc_book3s_exit_hv);
4606 MODULE_LICENSE("GPL");
4607 MODULE_ALIAS_MISCDEV(KVM_MINOR);
4608 MODULE_ALIAS("devname:kvm");