arch/s390/kvm/vsie.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * kvm nested virtualization support for s390x
   4  *
   5  * Copyright IBM Corp. 2016, 2018
   6  *
   7  *    Author(s): David Hildenbrand <dahi@linux.vnet.ibm.com>
   8  */
   9 #include <linux/vmalloc.h>
  10 #include <linux/kvm_host.h>
  11 #include <linux/bug.h>
  12 #include <linux/list.h>
  13 #include <linux/bitmap.h>
  14 #include <linux/sched/signal.h>
  15
  16 #include <asm/gmap.h>
  17 #include <asm/mmu_context.h>
  18 #include <asm/sclp.h>
  19 #include <asm/nmi.h>
  20 #include <asm/dis.h>
  21 #include <asm/fpu/api.h>
  22 #include "kvm-s390.h"
  23 #include "gaccess.h"
  24
  25 struct vsie_page {
  26         struct kvm_s390_sie_block scb_s;        /* 0x0000 */
  27         /*
  28          * the backup info for machine check. ensure it's at
  29          * the same offset as that in struct sie_page!
  30          */
  31         struct mcck_volatile_info mcck_info;    /* 0x0200 */
  32         /*
  33          * The pinned original scb. Be aware that other VCPUs can modify
  34          * it while we read from it. Values that are used for conditions or
  35          * are reused conditionally, should be accessed via READ_ONCE.
  36          */
  37         struct kvm_s390_sie_block *scb_o;       /* 0x0218 */
  38         /* the shadow gmap in use by the vsie_page */
  39         struct gmap *gmap;                      /* 0x0220 */
  40         /* address of the last reported fault to guest2 */
  41         unsigned long fault_addr;               /* 0x0228 */
  42         /* calculated guest addresses of satellite control blocks */
  43         gpa_t sca_gpa;                          /* 0x0230 */
  44         gpa_t itdba_gpa;                        /* 0x0238 */
  45         gpa_t gvrd_gpa;                         /* 0x0240 */
  46         gpa_t riccbd_gpa;                       /* 0x0248 */
  47         gpa_t sdnx_gpa;                         /* 0x0250 */
  48         __u8 reserved[0x0700 - 0x0258];         /* 0x0258 */
  49         struct kvm_s390_crypto_cb crycb;        /* 0x0700 */
  50         __u8 fac[S390_ARCH_FAC_LIST_SIZE_BYTE]; /* 0x0800 */
  51 };
  52
  53 /* trigger a validity icpt for the given scb */
  54 static int set_validity_icpt(struct kvm_s390_sie_block *scb,
  55                              __u16 reason_code)
  56 {
  57         scb->ipa = 0x1000;
  58         scb->ipb = ((__u32) reason_code) << 16;
  59         scb->icptcode = ICPT_VALIDITY;
  60         return 1;
  61 }
  62
  63 /* mark the prefix as unmapped, this will block the VSIE */
  64 static void prefix_unmapped(struct vsie_page *vsie_page)
  65 {
  66         atomic_or(PROG_REQUEST, &vsie_page->scb_s.prog20);
  67 }
  68
  69 /* mark the prefix as unmapped and wait until the VSIE has been left */
  70 static void prefix_unmapped_sync(struct vsie_page *vsie_page)
  71 {
  72         prefix_unmapped(vsie_page);
  73         if (vsie_page->scb_s.prog0c & PROG_IN_SIE)
  74                 atomic_or(CPUSTAT_STOP_INT, &vsie_page->scb_s.cpuflags);
  75         while (vsie_page->scb_s.prog0c & PROG_IN_SIE)
  76                 cpu_relax();
  77 }
  78
  79 /* mark the prefix as mapped, this will allow the VSIE to run */
  80 static void prefix_mapped(struct vsie_page *vsie_page)
  81 {
  82         atomic_andnot(PROG_REQUEST, &vsie_page->scb_s.prog20);
  83 }
  84
  85 /* test if the prefix is mapped into the gmap shadow */
  86 static int prefix_is_mapped(struct vsie_page *vsie_page)
  87 {
  88         return !(atomic_read(&vsie_page->scb_s.prog20) & PROG_REQUEST);
  89 }
  90
  91 /* copy the updated intervention request bits into the shadow scb */
  92 static void update_intervention_requests(struct vsie_page *vsie_page)
  93 {
  94         const int bits = CPUSTAT_STOP_INT | CPUSTAT_IO_INT | CPUSTAT_EXT_INT;
  95         int cpuflags;
  96
  97         cpuflags = atomic_read(&vsie_page->scb_o->cpuflags);
  98         atomic_andnot(bits, &vsie_page->scb_s.cpuflags);
  99         atomic_or(cpuflags & bits, &vsie_page->scb_s.cpuflags);
 100 }
 101
 102 /* shadow (filter and validate) the cpuflags  */
 103 static int prepare_cpuflags(struct kvm_vcpu *vcpu, struct vsie_page *vsie_page)
 104 {
 105         struct kvm_s390_sie_block *scb_s = &vsie_page->scb_s;
 106         struct kvm_s390_sie_block *scb_o = vsie_page->scb_o;
 107         int newflags, cpuflags = atomic_read(&scb_o->cpuflags);
 108
 109         /* we don't allow ESA/390 guests */
 110         if (!(cpuflags & CPUSTAT_ZARCH))
 111                 return set_validity_icpt(scb_s, 0x0001U);
 112
 113         if (cpuflags & (CPUSTAT_RRF | CPUSTAT_MCDS))
 114                 return set_validity_icpt(scb_s, 0x0001U);
 115         else if (cpuflags & (CPUSTAT_SLSV | CPUSTAT_SLSR))
 116                 return set_validity_icpt(scb_s, 0x0007U);
 117
 118         /* intervention requests will be set later */
 119         newflags = CPUSTAT_ZARCH;
 120         if (cpuflags & CPUSTAT_GED && test_kvm_facility(vcpu->kvm, 8))
 121                 newflags |= CPUSTAT_GED;
 122         if (cpuflags & CPUSTAT_GED2 && test_kvm_facility(vcpu->kvm, 78)) {
 123                 if (cpuflags & CPUSTAT_GED)
 124                         return set_validity_icpt(scb_s, 0x0001U);
 125                 newflags |= CPUSTAT_GED2;
 126         }
 127         if (test_kvm_cpu_feat(vcpu->kvm, KVM_S390_VM_CPU_FEAT_GPERE))
 128                 newflags |= cpuflags & CPUSTAT_P;
 129         if (test_kvm_cpu_feat(vcpu->kvm, KVM_S390_VM_CPU_FEAT_GSLS))
 130                 newflags |= cpuflags & CPUSTAT_SM;
 131         if (test_kvm_cpu_feat(vcpu->kvm, KVM_S390_VM_CPU_FEAT_IBS))
 132                 newflags |= cpuflags & CPUSTAT_IBS;
 133         if (test_kvm_cpu_feat(vcpu->kvm, KVM_S390_VM_CPU_FEAT_KSS))
 134                 newflags |= cpuflags & CPUSTAT_KSS;
 135
 136         atomic_set(&scb_s->cpuflags, newflags);
 137         return 0;
 138 }
 139 /* Copy to APCB FORMAT1 from APCB FORMAT0 */
 140 static int setup_apcb10(struct kvm_vcpu *vcpu, struct kvm_s390_apcb1 *apcb_s,
 141                         unsigned long apcb_o, struct kvm_s390_apcb1 *apcb_h)
 142 {
 143         struct kvm_s390_apcb0 tmp;
 144
 145         if (read_guest_real(vcpu, apcb_o, &tmp, sizeof(struct kvm_s390_apcb0)))
 146                 return -EFAULT;
 147
 148         apcb_s->apm[0] = apcb_h->apm[0] & tmp.apm[0];
 149         apcb_s->aqm[0] = apcb_h->aqm[0] & tmp.aqm[0] & 0xffff000000000000UL;
 150         apcb_s->adm[0] = apcb_h->adm[0] & tmp.adm[0] & 0xffff000000000000UL;
 151
 152         return 0;
 153
 154 }
 155
 156 /**
 157  * setup_apcb00 - Copy to APCB FORMAT0 from APCB FORMAT0
 158  * @vcpu: pointer to the virtual CPU
 159  * @apcb_s: pointer to start of apcb in the shadow crycb
 160  * @apcb_o: pointer to start of original apcb in the guest2
 161  * @apcb_h: pointer to start of apcb in the guest1
 162  *
 163  * Returns 0 and -EFAULT on error reading guest apcb
 164  */
 165 static int setup_apcb00(struct kvm_vcpu *vcpu, unsigned long *apcb_s,
 166                         unsigned long apcb_o, unsigned long *apcb_h)
 167 {
 168         if (read_guest_real(vcpu, apcb_o, apcb_s,
 169                             sizeof(struct kvm_s390_apcb0)))
 170                 return -EFAULT;
 171
 172         bitmap_and(apcb_s, apcb_s, apcb_h, sizeof(struct kvm_s390_apcb0));
 173
 174         return 0;
 175 }
 176
 177 /**
 178  * setup_apcb11 - Copy the FORMAT1 APCB from the guest to the shadow CRYCB
 179  * @vcpu: pointer to the virtual CPU
 180  * @apcb_s: pointer to start of apcb in the shadow crycb
 181  * @apcb_o: pointer to start of original guest apcb
 182  * @apcb_h: pointer to start of apcb in the host
 183  *
 184  * Returns 0 and -EFAULT on error reading guest apcb
 185  */
 186 static int setup_apcb11(struct kvm_vcpu *vcpu, unsigned long *apcb_s,
 187                         unsigned long apcb_o,
 188                         unsigned long *apcb_h)
 189 {
 190         if (read_guest_real(vcpu, apcb_o, apcb_s,
 191                             sizeof(struct kvm_s390_apcb1)))
 192                 return -EFAULT;
 193
 194         bitmap_and(apcb_s, apcb_s, apcb_h, sizeof(struct kvm_s390_apcb1));
 195
 196         return 0;
 197 }
 198
 199 /**
 200  * setup_apcb - Create a shadow copy of the apcb.
 201  * @vcpu: pointer to the virtual CPU
 202  * @crycb_s: pointer to shadow crycb
 203  * @crycb_o: pointer to original guest crycb
 204  * @crycb_h: pointer to the host crycb
 205  * @fmt_o: format of the original guest crycb.
 206  * @fmt_h: format of the host crycb.
 207  *
 208  * Checks the compatibility between the guest and host crycb and calls the
 209  * appropriate copy function.
 210  *
 211  * Return 0 or an error number if the guest and host crycb are incompatible.
 212  */
 213 static int setup_apcb(struct kvm_vcpu *vcpu, struct kvm_s390_crypto_cb *crycb_s,
 214                const u32 crycb_o,
 215                struct kvm_s390_crypto_cb *crycb_h,
 216                int fmt_o, int fmt_h)
 217 {
 218         struct kvm_s390_crypto_cb *crycb;
 219
 220         crycb = (struct kvm_s390_crypto_cb *) (unsigned long)crycb_o;
 221
 222         switch (fmt_o) {
 223         case CRYCB_FORMAT2:
 224                 if ((crycb_o & PAGE_MASK) != ((crycb_o + 256) & PAGE_MASK))
 225                         return -EACCES;
 226                 if (fmt_h != CRYCB_FORMAT2)
 227                         return -EINVAL;
 228                 return setup_apcb11(vcpu, (unsigned long *)&crycb_s->apcb1,
 229                                     (unsigned long) &crycb->apcb1,
 230                                     (unsigned long *)&crycb_h->apcb1);
 231         case CRYCB_FORMAT1:
 232                 switch (fmt_h) {
 233                 case CRYCB_FORMAT2:
 234                         return setup_apcb10(vcpu, &crycb_s->apcb1,
 235                                             (unsigned long) &crycb->apcb0,
 236                                             &crycb_h->apcb1);
 237                 case CRYCB_FORMAT1:
 238                         return setup_apcb00(vcpu,
 239                                             (unsigned long *) &crycb_s->apcb0,
 240                                             (unsigned long) &crycb->apcb0,
 241                                             (unsigned long *) &crycb_h->apcb0);
 242                 }
 243                 break;
 244         case CRYCB_FORMAT0:
 245                 if ((crycb_o & PAGE_MASK) != ((crycb_o + 32) & PAGE_MASK))
 246                         return -EACCES;
 247
 248                 switch (fmt_h) {
 249                 case CRYCB_FORMAT2:
 250                         return setup_apcb10(vcpu, &crycb_s->apcb1,
 251                                             (unsigned long) &crycb->apcb0,
 252                                             &crycb_h->apcb1);
 253                 case CRYCB_FORMAT1:
 254                 case CRYCB_FORMAT0:
 255                         return setup_apcb00(vcpu,
 256                                             (unsigned long *) &crycb_s->apcb0,
 257                                             (unsigned long) &crycb->apcb0,
 258                                             (unsigned long *) &crycb_h->apcb0);
 259                 }
 260         }
 261         return -EINVAL;
 262 }
 263
 264 /**
 265  * shadow_crycb - Create a shadow copy of the crycb block
 266  * @vcpu: a pointer to the virtual CPU
 267  * @vsie_page: a pointer to internal date used for the vSIE
 268  *
 269  * Create a shadow copy of the crycb block and setup key wrapping, if
 270  * requested for guest 3 and enabled for guest 2.
 271  *
 272  * We accept format-1 or format-2, but we convert format-1 into format-2
 273  * in the shadow CRYCB.
 274  * Using format-2 enables the firmware to choose the right format when
 275  * scheduling the SIE.
 276  * There is nothing to do for format-0.
 277  *
 278  * This function centralize the issuing of set_validity_icpt() for all
 279  * the subfunctions working on the crycb.
 280  *
 281  * Returns: - 0 if shadowed or nothing to do
 282  *          - > 0 if control has to be given to guest 2
 283  */
 284 static int shadow_crycb(struct kvm_vcpu *vcpu, struct vsie_page *vsie_page)
 285 {
 286         struct kvm_s390_sie_block *scb_s = &vsie_page->scb_s;
 287         struct kvm_s390_sie_block *scb_o = vsie_page->scb_o;
 288         const uint32_t crycbd_o = READ_ONCE(scb_o->crycbd);
 289         const u32 crycb_addr = crycbd_o & 0x7ffffff8U;
 290         unsigned long *b1, *b2;
 291         u8 ecb3_flags;
 292         u32 ecd_flags;
 293         int apie_h;
 294         int apie_s;
 295         int key_msk = test_kvm_facility(vcpu->kvm, 76);
 296         int fmt_o = crycbd_o & CRYCB_FORMAT_MASK;
 297         int fmt_h = vcpu->arch.sie_block->crycbd & CRYCB_FORMAT_MASK;
 298         int ret = 0;
 299
 300         scb_s->crycbd = 0;
 301
 302         apie_h = vcpu->arch.sie_block->eca & ECA_APIE;
 303         apie_s = apie_h & scb_o->eca;
 304         if (!apie_s && (!key_msk || (fmt_o == CRYCB_FORMAT0)))
 305                 return 0;
 306
 307         if (!crycb_addr)
 308                 return set_validity_icpt(scb_s, 0x0039U);
 309
 310         if (fmt_o == CRYCB_FORMAT1)
 311                 if ((crycb_addr & PAGE_MASK) !=
 312                     ((crycb_addr + 128) & PAGE_MASK))
 313                         return set_validity_icpt(scb_s, 0x003CU);
 314
 315         if (apie_s) {
 316                 ret = setup_apcb(vcpu, &vsie_page->crycb, crycb_addr,
 317                                  vcpu->kvm->arch.crypto.crycb,
 318                                  fmt_o, fmt_h);
 319                 if (ret)
 320                         goto end;
 321                 scb_s->eca |= scb_o->eca & ECA_APIE;
 322         }
 323
 324         /* we may only allow it if enabled for guest 2 */
 325         ecb3_flags = scb_o->ecb3 & vcpu->arch.sie_block->ecb3 &
 326                      (ECB3_AES | ECB3_DEA);
 327         ecd_flags = scb_o->ecd & vcpu->arch.sie_block->ecd & ECD_ECC;
 328         if (!ecb3_flags && !ecd_flags)
 329                 goto end;
 330
 331         /* copy only the wrapping keys */
 332         if (read_guest_real(vcpu, crycb_addr + 72,
 333                             vsie_page->crycb.dea_wrapping_key_mask, 56))
 334                 return set_validity_icpt(scb_s, 0x0035U);
 335
 336         scb_s->ecb3 |= ecb3_flags;
 337         scb_s->ecd |= ecd_flags;
 338
 339         /* xor both blocks in one run */
 340         b1 = (unsigned long *) vsie_page->crycb.dea_wrapping_key_mask;
 341         b2 = (unsigned long *)
 342                             vcpu->kvm->arch.crypto.crycb->dea_wrapping_key_mask;
 343         /* as 56%8 == 0, bitmap_xor won't overwrite any data */
 344         bitmap_xor(b1, b1, b2, BITS_PER_BYTE * 56);
 345 end:
 346         switch (ret) {
 347         case -EINVAL:
 348                 return set_validity_icpt(scb_s, 0x0022U);
 349         case -EFAULT:
 350                 return set_validity_icpt(scb_s, 0x0035U);
 351         case -EACCES:
 352                 return set_validity_icpt(scb_s, 0x003CU);
 353         }
 354         scb_s->crycbd = ((__u32)(__u64) &vsie_page->crycb) | CRYCB_FORMAT2;
 355         return 0;
 356 }
 357
 358 /* shadow (round up/down) the ibc to avoid validity icpt */
 359 static void prepare_ibc(struct kvm_vcpu *vcpu, struct vsie_page *vsie_page)
 360 {
 361         struct kvm_s390_sie_block *scb_s = &vsie_page->scb_s;
 362         struct kvm_s390_sie_block *scb_o = vsie_page->scb_o;
 363         /* READ_ONCE does not work on bitfields - use a temporary variable */
 364         const uint32_t __new_ibc = scb_o->ibc;
 365         const uint32_t new_ibc = READ_ONCE(__new_ibc) & 0x0fffU;
 366         __u64 min_ibc = (sclp.ibc >> 16) & 0x0fffU;
 367
 368         scb_s->ibc = 0;
 369         /* ibc installed in g2 and requested for g3 */
 370         if (vcpu->kvm->arch.model.ibc && new_ibc) {
 371                 scb_s->ibc = new_ibc;
 372                 /* takte care of the minimum ibc level of the machine */
 373                 if (scb_s->ibc < min_ibc)
 374                         scb_s->ibc = min_ibc;
 375                 /* take care of the maximum ibc level set for the guest */
 376                 if (scb_s->ibc > vcpu->kvm->arch.model.ibc)
 377                         scb_s->ibc = vcpu->kvm->arch.model.ibc;
 378         }
 379 }
 380
 381 /* unshadow the scb, copying parameters back to the real scb */
 382 static void unshadow_scb(struct kvm_vcpu *vcpu, struct vsie_page *vsie_page)
 383 {
 384         struct kvm_s390_sie_block *scb_s = &vsie_page->scb_s;
 385         struct kvm_s390_sie_block *scb_o = vsie_page->scb_o;
 386
 387         /* interception */
 388         scb_o->icptcode = scb_s->icptcode;
 389         scb_o->icptstatus = scb_s->icptstatus;
 390         scb_o->ipa = scb_s->ipa;
 391         scb_o->ipb = scb_s->ipb;
 392         scb_o->gbea = scb_s->gbea;
 393
 394         /* timer */
 395         scb_o->cputm = scb_s->cputm;
 396         scb_o->ckc = scb_s->ckc;
 397         scb_o->todpr = scb_s->todpr;
 398
 399         /* guest state */
 400         scb_o->gpsw = scb_s->gpsw;
 401         scb_o->gg14 = scb_s->gg14;
 402         scb_o->gg15 = scb_s->gg15;
 403         memcpy(scb_o->gcr, scb_s->gcr, 128);
 404         scb_o->pp = scb_s->pp;
 405
 406         /* branch prediction */
 407         if (test_kvm_facility(vcpu->kvm, 82)) {
 408                 scb_o->fpf &= ~FPF_BPBC;
 409                 scb_o->fpf |= scb_s->fpf & FPF_BPBC;
 410         }
 411
 412         /* interrupt intercept */
 413         switch (scb_s->icptcode) {
 414         case ICPT_PROGI:
 415         case ICPT_INSTPROGI:
 416         case ICPT_EXTINT:
 417                 memcpy((void *)((u64)scb_o + 0xc0),
 418                        (void *)((u64)scb_s + 0xc0), 0xf0 - 0xc0);
 419                 break;
 420         case ICPT_PARTEXEC:
 421                 /* MVPG only */
 422                 memcpy((void *)((u64)scb_o + 0xc0),
 423                        (void *)((u64)scb_s + 0xc0), 0xd0 - 0xc0);
 424                 break;
 425         }
 426
 427         if (scb_s->ihcpu != 0xffffU)
 428                 scb_o->ihcpu = scb_s->ihcpu;
 429 }
 430
 431 /*
 432  * Setup the shadow scb by copying and checking the relevant parts of the g2
 433  * provided scb.
 434  *
 435  * Returns: - 0 if the scb has been shadowed
 436  *          - > 0 if control has to be given to guest 2
 437  */
 438 static int shadow_scb(struct kvm_vcpu *vcpu, struct vsie_page *vsie_page)
 439 {
 440         struct kvm_s390_sie_block *scb_o = vsie_page->scb_o;
 441         struct kvm_s390_sie_block *scb_s = &vsie_page->scb_s;
 442         /* READ_ONCE does not work on bitfields - use a temporary variable */
 443         const uint32_t __new_prefix = scb_o->prefix;
 444         const uint32_t new_prefix = READ_ONCE(__new_prefix);
 445         const bool wants_tx = READ_ONCE(scb_o->ecb) & ECB_TE;
 446         bool had_tx = scb_s->ecb & ECB_TE;
 447         unsigned long new_mso = 0;
 448         int rc;
 449
 450         /* make sure we don't have any leftovers when reusing the scb */
 451         scb_s->icptcode = 0;
 452         scb_s->eca = 0;
 453         scb_s->ecb = 0;
 454         scb_s->ecb2 = 0;
 455         scb_s->ecb3 = 0;
 456         scb_s->ecd = 0;
 457         scb_s->fac = 0;
 458         scb_s->fpf = 0;
 459
 460         rc = prepare_cpuflags(vcpu, vsie_page);
 461         if (rc)
 462                 goto out;
 463
 464         /* timer */
 465         scb_s->cputm = scb_o->cputm;
 466         scb_s->ckc = scb_o->ckc;
 467         scb_s->todpr = scb_o->todpr;
 468         scb_s->epoch = scb_o->epoch;
 469
 470         /* guest state */
 471         scb_s->gpsw = scb_o->gpsw;
 472         scb_s->gg14 = scb_o->gg14;
 473         scb_s->gg15 = scb_o->gg15;
 474         memcpy(scb_s->gcr, scb_o->gcr, 128);
 475         scb_s->pp = scb_o->pp;
 476
 477         /* interception / execution handling */
 478         scb_s->gbea = scb_o->gbea;
 479         scb_s->lctl = scb_o->lctl;
 480         scb_s->svcc = scb_o->svcc;
 481         scb_s->ictl = scb_o->ictl;
 482         /*
 483          * SKEY handling functions can't deal with false setting of PTE invalid
 484          * bits. Therefore we cannot provide interpretation and would later
 485          * have to provide own emulation handlers.
 486          */
 487         if (!(atomic_read(&scb_s->cpuflags) & CPUSTAT_KSS))
 488                 scb_s->ictl |= ICTL_ISKE | ICTL_SSKE | ICTL_RRBE;
 489
 490         scb_s->icpua = scb_o->icpua;
 491
 492         if (!(atomic_read(&scb_s->cpuflags) & CPUSTAT_SM))
 493                 new_mso = READ_ONCE(scb_o->mso) & 0xfffffffffff00000UL;
 494         /* if the hva of the prefix changes, we have to remap the prefix */
 495         if (scb_s->mso != new_mso || scb_s->prefix != new_prefix)
 496                 prefix_unmapped(vsie_page);
 497          /* SIE will do mso/msl validity and exception checks for us */
 498         scb_s->msl = scb_o->msl & 0xfffffffffff00000UL;
 499         scb_s->mso = new_mso;
 500         scb_s->prefix = new_prefix;
 501
 502         /* We have to definetly flush the tlb if this scb never ran */
 503         if (scb_s->ihcpu != 0xffffU)
 504                 scb_s->ihcpu = scb_o->ihcpu;
 505
 506         /* MVPG and Protection Exception Interpretation are always available */
 507         scb_s->eca |= scb_o->eca & (ECA_MVPGI | ECA_PROTEXCI);
 508         /* Host-protection-interruption introduced with ESOP */
 509         if (test_kvm_cpu_feat(vcpu->kvm, KVM_S390_VM_CPU_FEAT_ESOP))
 510                 scb_s->ecb |= scb_o->ecb & ECB_HOSTPROTINT;
 511         /* transactional execution */
 512         if (test_kvm_facility(vcpu->kvm, 73) && wants_tx) {
 513                 /* remap the prefix is tx is toggled on */
 514                 if (!had_tx)
 515                         prefix_unmapped(vsie_page);
 516                 scb_s->ecb |= ECB_TE;
 517         }
 518         /* branch prediction */
 519         if (test_kvm_facility(vcpu->kvm, 82))
 520                 scb_s->fpf |= scb_o->fpf & FPF_BPBC;
 521         /* SIMD */
 522         if (test_kvm_facility(vcpu->kvm, 129)) {
 523                 scb_s->eca |= scb_o->eca & ECA_VX;
 524                 scb_s->ecd |= scb_o->ecd & ECD_HOSTREGMGMT;
 525         }
 526         /* Run-time-Instrumentation */
 527         if (test_kvm_facility(vcpu->kvm, 64))
 528                 scb_s->ecb3 |= scb_o->ecb3 & ECB3_RI;
 529         /* Instruction Execution Prevention */
 530         if (test_kvm_facility(vcpu->kvm, 130))
 531                 scb_s->ecb2 |= scb_o->ecb2 & ECB2_IEP;
 532         /* Guarded Storage */
 533         if (test_kvm_facility(vcpu->kvm, 133)) {
 534                 scb_s->ecb |= scb_o->ecb & ECB_GS;
 535                 scb_s->ecd |= scb_o->ecd & ECD_HOSTREGMGMT;
 536         }
 537         if (test_kvm_cpu_feat(vcpu->kvm, KVM_S390_VM_CPU_FEAT_SIIF))
 538                 scb_s->eca |= scb_o->eca & ECA_SII;
 539         if (test_kvm_cpu_feat(vcpu->kvm, KVM_S390_VM_CPU_FEAT_IB))
 540                 scb_s->eca |= scb_o->eca & ECA_IB;
 541         if (test_kvm_cpu_feat(vcpu->kvm, KVM_S390_VM_CPU_FEAT_CEI))
 542                 scb_s->eca |= scb_o->eca & ECA_CEI;
 543         /* Epoch Extension */
 544         if (test_kvm_facility(vcpu->kvm, 139))
 545                 scb_s->ecd |= scb_o->ecd & ECD_MEF;
 546
 547         /* etoken */
 548         if (test_kvm_facility(vcpu->kvm, 156))
 549                 scb_s->ecd |= scb_o->ecd & ECD_ETOKENF;
 550
 551         scb_s->hpid = HPID_VSIE;
 552         scb_s->cpnc = scb_o->cpnc;
 553
 554         prepare_ibc(vcpu, vsie_page);
 555         rc = shadow_crycb(vcpu, vsie_page);
 556 out:
 557         if (rc)
 558                 unshadow_scb(vcpu, vsie_page);
 559         return rc;
 560 }
 561
 562 void kvm_s390_vsie_gmap_notifier(struct gmap *gmap, unsigned long start,
 563                                  unsigned long end)
 564 {
 565         struct kvm *kvm = gmap->private;
 566         struct vsie_page *cur;
 567         unsigned long prefix;
 568         struct page *page;
 569         int i;
 570
 571         if (!gmap_is_shadow(gmap))
 572                 return;
 573         if (start >= 1UL << 31)
 574                 /* We are only interested in prefix pages */
 575                 return;
 576
 577         /*
 578          * Only new shadow blocks are added to the list during runtime,
 579          * therefore we can safely reference them all the time.
 580          */
 581         for (i = 0; i < kvm->arch.vsie.page_count; i++) {
 582                 page = READ_ONCE(kvm->arch.vsie.pages[i]);
 583                 if (!page)
 584                         continue;
 585                 cur = page_to_virt(page);
 586                 if (READ_ONCE(cur->gmap) != gmap)
 587                         continue;
 588                 prefix = cur->scb_s.prefix << GUEST_PREFIX_SHIFT;
 589                 /* with mso/msl, the prefix lies at an offset */
 590                 prefix += cur->scb_s.mso;
 591                 if (prefix <= end && start <= prefix + 2 * PAGE_SIZE - 1)
 592                         prefix_unmapped_sync(cur);
 593         }
 594 }
 595
 596 /*
 597  * Map the first prefix page and if tx is enabled also the second prefix page.
 598  *
 599  * The prefix will be protected, a gmap notifier will inform about unmaps.
 600  * The shadow scb must not be executed until the prefix is remapped, this is
 601  * guaranteed by properly handling PROG_REQUEST.
 602  *
 603  * Returns: - 0 on if successfully mapped or already mapped
 604  *          - > 0 if control has to be given to guest 2
 605  *          - -EAGAIN if the caller can retry immediately
 606  *          - -ENOMEM if out of memory
 607  */
 608 static int map_prefix(struct kvm_vcpu *vcpu, struct vsie_page *vsie_page)
 609 {
 610         struct kvm_s390_sie_block *scb_s = &vsie_page->scb_s;
 611         u64 prefix = scb_s->prefix << GUEST_PREFIX_SHIFT;
 612         int rc;
 613
 614         if (prefix_is_mapped(vsie_page))
 615                 return 0;
 616
 617         /* mark it as mapped so we can catch any concurrent unmappers */
 618         prefix_mapped(vsie_page);
 619
 620         /* with mso/msl, the prefix lies at offset *mso* */
 621         prefix += scb_s->mso;
 622
 623         rc = kvm_s390_shadow_fault(vcpu, vsie_page->gmap, prefix);
 624         if (!rc && (scb_s->ecb & ECB_TE))
 625                 rc = kvm_s390_shadow_fault(vcpu, vsie_page->gmap,
 626                                            prefix + PAGE_SIZE);
 627         /*
 628          * We don't have to mprotect, we will be called for all unshadows.
 629          * SIE will detect if protection applies and trigger a validity.
 630          */
 631         if (rc)
 632                 prefix_unmapped(vsie_page);
 633         if (rc > 0 || rc == -EFAULT)
 634                 rc = set_validity_icpt(scb_s, 0x0037U);
 635         return rc;
 636 }
 637
 638 /*
 639  * Pin the guest page given by gpa and set hpa to the pinned host address.
 640  * Will always be pinned writable.
 641  *
 642  * Returns: - 0 on success
 643  *          - -EINVAL if the gpa is not valid guest storage
 644  */
 645 static int pin_guest_page(struct kvm *kvm, gpa_t gpa, hpa_t *hpa)
 646 {
 647         struct page *page;
 648
 649         page = gfn_to_page(kvm, gpa_to_gfn(gpa));
 650         if (is_error_page(page))
 651                 return -EINVAL;
 652         *hpa = (hpa_t) page_to_virt(page) + (gpa & ~PAGE_MASK);
 653         return 0;
 654 }
 655
 656 /* Unpins a page previously pinned via pin_guest_page, marking it as dirty. */
 657 static void unpin_guest_page(struct kvm *kvm, gpa_t gpa, hpa_t hpa)
 658 {
 659         kvm_release_pfn_dirty(hpa >> PAGE_SHIFT);
 660         /* mark the page always as dirty for migration */
 661         mark_page_dirty(kvm, gpa_to_gfn(gpa));
 662 }
 663
 664 /* unpin all blocks previously pinned by pin_blocks(), marking them dirty */
 665 static void unpin_blocks(struct kvm_vcpu *vcpu, struct vsie_page *vsie_page)
 666 {
 667         struct kvm_s390_sie_block *scb_s = &vsie_page->scb_s;
 668         hpa_t hpa;
 669
 670         hpa = (u64) scb_s->scaoh << 32 | scb_s->scaol;
 671         if (hpa) {
 672                 unpin_guest_page(vcpu->kvm, vsie_page->sca_gpa, hpa);
 673                 vsie_page->sca_gpa = 0;
 674                 scb_s->scaol = 0;
 675                 scb_s->scaoh = 0;
 676         }
 677
 678         hpa = scb_s->itdba;
 679         if (hpa) {
 680                 unpin_guest_page(vcpu->kvm, vsie_page->itdba_gpa, hpa);
 681                 vsie_page->itdba_gpa = 0;
 682                 scb_s->itdba = 0;
 683         }
 684
 685         hpa = scb_s->gvrd;
 686         if (hpa) {
 687                 unpin_guest_page(vcpu->kvm, vsie_page->gvrd_gpa, hpa);
 688                 vsie_page->gvrd_gpa = 0;
 689                 scb_s->gvrd = 0;
 690         }
 691
 692         hpa = scb_s->riccbd;
 693         if (hpa) {
 694                 unpin_guest_page(vcpu->kvm, vsie_page->riccbd_gpa, hpa);
 695                 vsie_page->riccbd_gpa = 0;
 696                 scb_s->riccbd = 0;
 697         }
 698
 699         hpa = scb_s->sdnxo;
 700         if (hpa) {
 701                 unpin_guest_page(vcpu->kvm, vsie_page->sdnx_gpa, hpa);
 702                 vsie_page->sdnx_gpa = 0;
 703                 scb_s->sdnxo = 0;
 704         }
 705 }
 706
 707 /*
 708  * Instead of shadowing some blocks, we can simply forward them because the
 709  * addresses in the scb are 64 bit long.
 710  *
 711  * This works as long as the data lies in one page. If blocks ever exceed one
 712  * page, we have to fall back to shadowing.
 713  *
 714  * As we reuse the sca, the vcpu pointers contained in it are invalid. We must
 715  * therefore not enable any facilities that access these pointers (e.g. SIGPIF).
 716  *
 717  * Returns: - 0 if all blocks were pinned.
 718  *          - > 0 if control has to be given to guest 2
 719  *          - -ENOMEM if out of memory
 720  */
 721 static int pin_blocks(struct kvm_vcpu *vcpu, struct vsie_page *vsie_page)
 722 {
 723         struct kvm_s390_sie_block *scb_o = vsie_page->scb_o;
 724         struct kvm_s390_sie_block *scb_s = &vsie_page->scb_s;
 725         hpa_t hpa;
 726         gpa_t gpa;
 727         int rc = 0;
 728
 729         gpa = READ_ONCE(scb_o->scaol) & ~0xfUL;
 730         if (test_kvm_cpu_feat(vcpu->kvm, KVM_S390_VM_CPU_FEAT_64BSCAO))
 731                 gpa |= (u64) READ_ONCE(scb_o->scaoh) << 32;
 732         if (gpa) {
 733                 if (gpa < 2 * PAGE_SIZE)
 734                         rc = set_validity_icpt(scb_s, 0x0038U);
 735                 else if ((gpa & ~0x1fffUL) == kvm_s390_get_prefix(vcpu))
 736                         rc = set_validity_icpt(scb_s, 0x0011U);
 737                 else if ((gpa & PAGE_MASK) !=
 738                          ((gpa + sizeof(struct bsca_block) - 1) & PAGE_MASK))
 739                         rc = set_validity_icpt(scb_s, 0x003bU);
 740                 if (!rc) {
 741                         rc = pin_guest_page(vcpu->kvm, gpa, &hpa);
 742                         if (rc)
 743                                 rc = set_validity_icpt(scb_s, 0x0034U);
 744                 }
 745                 if (rc)
 746                         goto unpin;
 747                 vsie_page->sca_gpa = gpa;
 748                 scb_s->scaoh = (u32)((u64)hpa >> 32);
 749                 scb_s->scaol = (u32)(u64)hpa;
 750         }
 751
 752         gpa = READ_ONCE(scb_o->itdba) & ~0xffUL;
 753         if (gpa && (scb_s->ecb & ECB_TE)) {
 754                 if (gpa < 2 * PAGE_SIZE) {
 755                         rc = set_validity_icpt(scb_s, 0x0080U);
 756                         goto unpin;
 757                 }
 758                 /* 256 bytes cannot cross page boundaries */
 759                 rc = pin_guest_page(vcpu->kvm, gpa, &hpa);
 760                 if (rc) {
 761                         rc = set_validity_icpt(scb_s, 0x0080U);
 762                         goto unpin;
 763                 }
 764                 vsie_page->itdba_gpa = gpa;
 765                 scb_s->itdba = hpa;
 766         }
 767
 768         gpa = READ_ONCE(scb_o->gvrd) & ~0x1ffUL;
 769         if (gpa && (scb_s->eca & ECA_VX) && !(scb_s->ecd & ECD_HOSTREGMGMT)) {
 770                 if (gpa < 2 * PAGE_SIZE) {
 771                         rc = set_validity_icpt(scb_s, 0x1310U);
 772                         goto unpin;
 773                 }
 774                 /*
 775                  * 512 bytes vector registers cannot cross page boundaries
 776                  * if this block gets bigger, we have to shadow it.
 777                  */
 778                 rc = pin_guest_page(vcpu->kvm, gpa, &hpa);
 779                 if (rc) {
 780                         rc = set_validity_icpt(scb_s, 0x1310U);
 781                         goto unpin;
 782                 }
 783                 vsie_page->gvrd_gpa = gpa;
 784                 scb_s->gvrd = hpa;
 785         }
 786
 787         gpa = READ_ONCE(scb_o->riccbd) & ~0x3fUL;
 788         if (gpa && (scb_s->ecb3 & ECB3_RI)) {
 789                 if (gpa < 2 * PAGE_SIZE) {
 790                         rc = set_validity_icpt(scb_s, 0x0043U);
 791                         goto unpin;
 792                 }
 793                 /* 64 bytes cannot cross page boundaries */
 794                 rc = pin_guest_page(vcpu->kvm, gpa, &hpa);
 795                 if (rc) {
 796                         rc = set_validity_icpt(scb_s, 0x0043U);
 797                         goto unpin;
 798                 }
 799                 /* Validity 0x0044 will be checked by SIE */
 800                 vsie_page->riccbd_gpa = gpa;
 801                 scb_s->riccbd = hpa;
 802         }
 803         if (((scb_s->ecb & ECB_GS) && !(scb_s->ecd & ECD_HOSTREGMGMT)) ||
 804             (scb_s->ecd & ECD_ETOKENF)) {
 805                 unsigned long sdnxc;
 806
 807                 gpa = READ_ONCE(scb_o->sdnxo) & ~0xfUL;
 808                 sdnxc = READ_ONCE(scb_o->sdnxo) & 0xfUL;
 809                 if (!gpa || gpa < 2 * PAGE_SIZE) {
 810                         rc = set_validity_icpt(scb_s, 0x10b0U);
 811                         goto unpin;
 812                 }
 813                 if (sdnxc < 6 || sdnxc > 12) {
 814                         rc = set_validity_icpt(scb_s, 0x10b1U);
 815                         goto unpin;
 816                 }
 817                 if (gpa & ((1 << sdnxc) - 1)) {
 818                         rc = set_validity_icpt(scb_s, 0x10b2U);
 819                         goto unpin;
 820                 }
 821                 /* Due to alignment rules (checked above) this cannot
 822                  * cross page boundaries
 823                  */
 824                 rc = pin_guest_page(vcpu->kvm, gpa, &hpa);
 825                 if (rc) {
 826                         rc = set_validity_icpt(scb_s, 0x10b0U);
 827                         goto unpin;
 828                 }
 829                 vsie_page->sdnx_gpa = gpa;
 830                 scb_s->sdnxo = hpa | sdnxc;
 831         }
 832         return 0;
 833 unpin:
 834         unpin_blocks(vcpu, vsie_page);
 835         return rc;
 836 }
 837
 838 /* unpin the scb provided by guest 2, marking it as dirty */
 839 static void unpin_scb(struct kvm_vcpu *vcpu, struct vsie_page *vsie_page,
 840                       gpa_t gpa)
 841 {
 842         hpa_t hpa = (hpa_t) vsie_page->scb_o;
 843
 844         if (hpa)
 845                 unpin_guest_page(vcpu->kvm, gpa, hpa);
 846         vsie_page->scb_o = NULL;
 847 }
 848
 849 /*
 850  * Pin the scb at gpa provided by guest 2 at vsie_page->scb_o.
 851  *
 852  * Returns: - 0 if the scb was pinned.
 853  *          - > 0 if control has to be given to guest 2
 854  */
 855 static int pin_scb(struct kvm_vcpu *vcpu, struct vsie_page *vsie_page,
 856                    gpa_t gpa)
 857 {
 858         hpa_t hpa;
 859         int rc;
 860
 861         rc = pin_guest_page(vcpu->kvm, gpa, &hpa);
 862         if (rc) {
 863                 rc = kvm_s390_inject_program_int(vcpu, PGM_ADDRESSING);
 864                 WARN_ON_ONCE(rc);
 865                 return 1;
 866         }
 867         vsie_page->scb_o = (struct kvm_s390_sie_block *) hpa;
 868         return 0;
 869 }
 870
 871 /*
 872  * Inject a fault into guest 2.
 873  *
 874  * Returns: - > 0 if control has to be given to guest 2
 875  *            < 0 if an error occurred during injection.
 876  */
 877 static int inject_fault(struct kvm_vcpu *vcpu, __u16 code, __u64 vaddr,
 878                         bool write_flag)
 879 {
 880         struct kvm_s390_pgm_info pgm = {
 881                 .code = code,
 882                 .trans_exc_code =
 883                         /* 0-51: virtual address */
 884                         (vaddr & 0xfffffffffffff000UL) |
 885                         /* 52-53: store / fetch */
 886                         (((unsigned int) !write_flag) + 1) << 10,
 887                         /* 62-63: asce id (alway primary == 0) */
 888                 .exc_access_id = 0, /* always primary */
 889                 .op_access_id = 0, /* not MVPG */
 890         };
 891         int rc;
 892
 893         if (code == PGM_PROTECTION)
 894                 pgm.trans_exc_code |= 0x4UL;
 895
 896         rc = kvm_s390_inject_prog_irq(vcpu, &pgm);
 897         return rc ? rc : 1;
 898 }
 899
 900 /*
 901  * Handle a fault during vsie execution on a gmap shadow.
 902  *
 903  * Returns: - 0 if the fault was resolved
 904  *          - > 0 if control has to be given to guest 2
 905  *          - < 0 if an error occurred
 906  */
 907 static int handle_fault(struct kvm_vcpu *vcpu, struct vsie_page *vsie_page)
 908 {
 909         int rc;
 910
 911         if (current->thread.gmap_int_code == PGM_PROTECTION)
 912                 /* we can directly forward all protection exceptions */
 913                 return inject_fault(vcpu, PGM_PROTECTION,
 914                                     current->thread.gmap_addr, 1);
 915
 916         rc = kvm_s390_shadow_fault(vcpu, vsie_page->gmap,
 917                                    current->thread.gmap_addr);
 918         if (rc > 0) {
 919                 rc = inject_fault(vcpu, rc,
 920                                   current->thread.gmap_addr,
 921                                   current->thread.gmap_write_flag);
 922                 if (rc >= 0)
 923                         vsie_page->fault_addr = current->thread.gmap_addr;
 924         }
 925         return rc;
 926 }
 927
 928 /*
 929  * Retry the previous fault that required guest 2 intervention. This avoids
 930  * one superfluous SIE re-entry and direct exit.
 931  *
 932  * Will ignore any errors. The next SIE fault will do proper fault handling.
 933  */
 934 static void handle_last_fault(struct kvm_vcpu *vcpu,
 935                               struct vsie_page *vsie_page)
 936 {
 937         if (vsie_page->fault_addr)
 938                 kvm_s390_shadow_fault(vcpu, vsie_page->gmap,
 939                                       vsie_page->fault_addr);
 940         vsie_page->fault_addr = 0;
 941 }
 942
 943 static inline void clear_vsie_icpt(struct vsie_page *vsie_page)
 944 {
 945         vsie_page->scb_s.icptcode = 0;
 946 }
 947
 948 /* rewind the psw and clear the vsie icpt, so we can retry execution */
 949 static void retry_vsie_icpt(struct vsie_page *vsie_page)
 950 {
 951         struct kvm_s390_sie_block *scb_s = &vsie_page->scb_s;
 952         int ilen = insn_length(scb_s->ipa >> 8);
 953
 954         /* take care of EXECUTE instructions */
 955         if (scb_s->icptstatus & 1) {
 956                 ilen = (scb_s->icptstatus >> 4) & 0x6;
 957                 if (!ilen)
 958                         ilen = 4;
 959         }
 960         scb_s->gpsw.addr = __rewind_psw(scb_s->gpsw, ilen);
 961         clear_vsie_icpt(vsie_page);
 962 }
 963
 964 /*
 965  * Try to shadow + enable the guest 2 provided facility list.
 966  * Retry instruction execution if enabled for and provided by guest 2.
 967  *
 968  * Returns: - 0 if handled (retry or guest 2 icpt)
 969  *          - > 0 if control has to be given to guest 2
 970  */
 971 static int handle_stfle(struct kvm_vcpu *vcpu, struct vsie_page *vsie_page)
 972 {
 973         struct kvm_s390_sie_block *scb_s = &vsie_page->scb_s;
 974         __u32 fac = READ_ONCE(vsie_page->scb_o->fac) & 0x7ffffff8U;
 975
 976         if (fac && test_kvm_facility(vcpu->kvm, 7)) {
 977                 retry_vsie_icpt(vsie_page);
 978                 if (read_guest_real(vcpu, fac, &vsie_page->fac,
 979                                     sizeof(vsie_page->fac)))
 980                         return set_validity_icpt(scb_s, 0x1090U);
 981                 scb_s->fac = (__u32)(__u64) &vsie_page->fac;
 982         }
 983         return 0;
 984 }
 985
 986 /*
 987  * Run the vsie on a shadow scb and a shadow gmap, without any further
 988  * sanity checks, handling SIE faults.
 989  *
 990  * Returns: - 0 everything went fine
 991  *          - > 0 if control has to be given to guest 2
 992  *          - < 0 if an error occurred
 993  */
 994 static int do_vsie_run(struct kvm_vcpu *vcpu, struct vsie_page *vsie_page)
 995         __releases(vcpu->kvm->srcu)
 996         __acquires(vcpu->kvm->srcu)
 997 {
 998         struct kvm_s390_sie_block *scb_s = &vsie_page->scb_s;
 999         struct kvm_s390_sie_block *scb_o = vsie_page->scb_o;
1000         int guest_bp_isolation;
1001         int rc = 0;
1002
1003         handle_last_fault(vcpu, vsie_page);
1004
1005         srcu_read_unlock(&vcpu->kvm->srcu, vcpu->srcu_idx);
1006
1007         /* save current guest state of bp isolation override */
1008         guest_bp_isolation = test_thread_flag(TIF_ISOLATE_BP_GUEST);
1009
1010         /*
1011          * The guest is running with BPBC, so we have to force it on for our
1012          * nested guest. This is done by enabling BPBC globally, so the BPBC
1013          * control in the SCB (which the nested guest can modify) is simply
1014          * ignored.
1015          */
1016         if (test_kvm_facility(vcpu->kvm, 82) &&
1017             vcpu->arch.sie_block->fpf & FPF_BPBC)
1018                 set_thread_flag(TIF_ISOLATE_BP_GUEST);
1019
1020         local_irq_disable();
1021         guest_enter_irqoff();
1022         local_irq_enable();
1023
1024         /*
1025          * Simulate a SIE entry of the VCPU (see sie64a), so VCPU blocking
1026          * and VCPU requests also hinder the vSIE from running and lead
1027          * to an immediate exit. kvm_s390_vsie_kick() has to be used to
1028          * also kick the vSIE.
1029          */
1030         vcpu->arch.sie_block->prog0c |= PROG_IN_SIE;
1031         barrier();
1032         if (test_cpu_flag(CIF_FPU))
1033                 load_fpu_regs();
1034         if (!kvm_s390_vcpu_sie_inhibited(vcpu))
1035                 rc = sie64a(scb_s, vcpu->run->s.regs.gprs);
1036         barrier();
1037         vcpu->arch.sie_block->prog0c &= ~PROG_IN_SIE;
1038
1039         local_irq_disable();
1040         guest_exit_irqoff();
1041         local_irq_enable();
1042
1043         /* restore guest state for bp isolation override */
1044         if (!guest_bp_isolation)
1045                 clear_thread_flag(TIF_ISOLATE_BP_GUEST);
1046
1047         vcpu->srcu_idx = srcu_read_lock(&vcpu->kvm->srcu);
1048
1049         if (rc == -EINTR) {
1050                 VCPU_EVENT(vcpu, 3, "%s", "machine check");
1051                 kvm_s390_reinject_machine_check(vcpu, &vsie_page->mcck_info);
1052                 return 0;
1053         }
1054
1055         if (rc > 0)
1056                 rc = 0; /* we could still have an icpt */
1057         else if (rc == -EFAULT)
1058                 return handle_fault(vcpu, vsie_page);
1059
1060         switch (scb_s->icptcode) {
1061         case ICPT_INST:
1062                 if (scb_s->ipa == 0xb2b0)
1063                         rc = handle_stfle(vcpu, vsie_page);
1064                 break;
1065         case ICPT_STOP:
1066                 /* stop not requested by g2 - must have been a kick */
1067                 if (!(atomic_read(&scb_o->cpuflags) & CPUSTAT_STOP_INT))
1068                         clear_vsie_icpt(vsie_page);
1069                 break;
1070         case ICPT_VALIDITY:
1071                 if ((scb_s->ipa & 0xf000) != 0xf000)
1072                         scb_s->ipa += 0x1000;
1073                 break;
1074         }
1075         return rc;
1076 }
1077
1078 static void release_gmap_shadow(struct vsie_page *vsie_page)
1079 {
1080         if (vsie_page->gmap)
1081                 gmap_put(vsie_page->gmap);
1082         WRITE_ONCE(vsie_page->gmap, NULL);
1083         prefix_unmapped(vsie_page);
1084 }
1085
1086 static int acquire_gmap_shadow(struct kvm_vcpu *vcpu,
1087                                struct vsie_page *vsie_page)
1088 {
1089         unsigned long asce;
1090         union ctlreg0 cr0;
1091         struct gmap *gmap;
1092         int edat;
1093
1094         asce = vcpu->arch.sie_block->gcr[1];
1095         cr0.val = vcpu->arch.sie_block->gcr[0];
1096         edat = cr0.edat && test_kvm_facility(vcpu->kvm, 8);
1097         edat += edat && test_kvm_facility(vcpu->kvm, 78);
1098
1099         /*
1100          * ASCE or EDAT could have changed since last icpt, or the gmap
1101          * we're holding has been unshadowed. If the gmap is still valid,
1102          * we can safely reuse it.
1103          */
1104         if (vsie_page->gmap && gmap_shadow_valid(vsie_page->gmap, asce, edat))
1105                 return 0;
1106
1107         /* release the old shadow - if any, and mark the prefix as unmapped */
1108         release_gmap_shadow(vsie_page);
1109         gmap = gmap_shadow(vcpu->arch.gmap, asce, edat);
1110         if (IS_ERR(gmap))
1111                 return PTR_ERR(gmap);
1112         gmap->private = vcpu->kvm;
1113         WRITE_ONCE(vsie_page->gmap, gmap);
1114         return 0;
1115 }
1116
1117 /*
1118  * Register the shadow scb at the VCPU, e.g. for kicking out of vsie.
1119  */
1120 static void register_shadow_scb(struct kvm_vcpu *vcpu,
1121                                 struct vsie_page *vsie_page)
1122 {
1123         struct kvm_s390_sie_block *scb_s = &vsie_page->scb_s;
1124
1125         WRITE_ONCE(vcpu->arch.vsie_block, &vsie_page->scb_s);
1126         /*
1127          * External calls have to lead to a kick of the vcpu and
1128          * therefore the vsie -> Simulate Wait state.
1129          */
1130         kvm_s390_set_cpuflags(vcpu, CPUSTAT_WAIT);
1131         /*
1132          * We have to adjust the g3 epoch by the g2 epoch. The epoch will
1133          * automatically be adjusted on tod clock changes via kvm_sync_clock.
1134          */
1135         preempt_disable();
1136         scb_s->epoch += vcpu->kvm->arch.epoch;
1137
1138         if (scb_s->ecd & ECD_MEF) {
1139                 scb_s->epdx += vcpu->kvm->arch.epdx;
1140                 if (scb_s->epoch < vcpu->kvm->arch.epoch)
1141                         scb_s->epdx += 1;
1142         }
1143
1144         preempt_enable();
1145 }
1146
1147 /*
1148  * Unregister a shadow scb from a VCPU.
1149  */
1150 static void unregister_shadow_scb(struct kvm_vcpu *vcpu)
1151 {
1152         kvm_s390_clear_cpuflags(vcpu, CPUSTAT_WAIT);
1153         WRITE_ONCE(vcpu->arch.vsie_block, NULL);
1154 }
1155
1156 /*
1157  * Run the vsie on a shadowed scb, managing the gmap shadow, handling
1158  * prefix pages and faults.
1159  *
1160  * Returns: - 0 if no errors occurred
1161  *          - > 0 if control has to be given to guest 2
1162  *          - -ENOMEM if out of memory
1163  */
1164 static int vsie_run(struct kvm_vcpu *vcpu, struct vsie_page *vsie_page)
1165 {
1166         struct kvm_s390_sie_block *scb_s = &vsie_page->scb_s;
1167         int rc = 0;
1168
1169         while (1) {
1170                 rc = acquire_gmap_shadow(vcpu, vsie_page);
1171                 if (!rc)
1172                         rc = map_prefix(vcpu, vsie_page);
1173                 if (!rc) {
1174                         gmap_enable(vsie_page->gmap);
1175                         update_intervention_requests(vsie_page);
1176                         rc = do_vsie_run(vcpu, vsie_page);
1177                         gmap_enable(vcpu->arch.gmap);
1178                 }
1179                 atomic_andnot(PROG_BLOCK_SIE, &scb_s->prog20);
1180
1181                 if (rc == -EAGAIN)
1182                         rc = 0;
1183                 if (rc || scb_s->icptcode || signal_pending(current) ||
1184                     kvm_s390_vcpu_has_irq(vcpu, 0) ||
1185                     kvm_s390_vcpu_sie_inhibited(vcpu))
1186                         break;
1187                 cond_resched();
1188         }
1189
1190         if (rc == -EFAULT) {
1191                 /*
1192                  * Addressing exceptions are always presentes as intercepts.
1193                  * As addressing exceptions are suppressing and our guest 3 PSW
1194                  * points at the responsible instruction, we have to
1195                  * forward the PSW and set the ilc. If we can't read guest 3
1196                  * instruction, we can use an arbitrary ilc. Let's always use
1197                  * ilen = 4 for now, so we can avoid reading in guest 3 virtual
1198                  * memory. (we could also fake the shadow so the hardware
1199                  * handles it).
1200                  */
1201                 scb_s->icptcode = ICPT_PROGI;
1202                 scb_s->iprcc = PGM_ADDRESSING;
1203                 scb_s->pgmilc = 4;
1204                 scb_s->gpsw.addr = __rewind_psw(scb_s->gpsw, 4);
1205                 rc = 1;
1206         }
1207         return rc;
1208 }
1209
1210 /*
1211  * Get or create a vsie page for a scb address.
1212  *
1213  * Returns: - address of a vsie page (cached or new one)
1214  *          - NULL if the same scb address is already used by another VCPU
1215  *          - ERR_PTR(-ENOMEM) if out of memory
1216  */
1217 static struct vsie_page *get_vsie_page(struct kvm *kvm, unsigned long addr)
1218 {
1219         struct vsie_page *vsie_page;
1220         struct page *page;
1221         int nr_vcpus;
1222
1223         rcu_read_lock();
1224         page = radix_tree_lookup(&kvm->arch.vsie.addr_to_page, addr >> 9);
1225         rcu_read_unlock();
1226         if (page) {
1227                 if (page_ref_inc_return(page) == 2)
1228                         return page_to_virt(page);
1229                 page_ref_dec(page);
1230         }
1231
1232         /*
1233          * We want at least #online_vcpus shadows, so every VCPU can execute
1234          * the VSIE in parallel.
1235          */
1236         nr_vcpus = atomic_read(&kvm->online_vcpus);
1237
1238         mutex_lock(&kvm->arch.vsie.mutex);
1239         if (kvm->arch.vsie.page_count < nr_vcpus) {
1240                 page = alloc_page(GFP_KERNEL_ACCOUNT | __GFP_ZERO | GFP_DMA);
1241                 if (!page) {
1242                         mutex_unlock(&kvm->arch.vsie.mutex);
1243                         return ERR_PTR(-ENOMEM);
1244                 }
1245                 page_ref_inc(page);
1246                 kvm->arch.vsie.pages[kvm->arch.vsie.page_count] = page;
1247                 kvm->arch.vsie.page_count++;
1248         } else {
1249                 /* reuse an existing entry that belongs to nobody */
1250                 while (true) {
1251                         page = kvm->arch.vsie.pages[kvm->arch.vsie.next];
1252                         if (page_ref_inc_return(page) == 2)
1253                                 break;
1254                         page_ref_dec(page);
1255                         kvm->arch.vsie.next++;
1256                         kvm->arch.vsie.next %= nr_vcpus;
1257                 }
1258                 radix_tree_delete(&kvm->arch.vsie.addr_to_page, page->index >> 9);
1259         }
1260         page->index = addr;
1261         /* double use of the same address */
1262         if (radix_tree_insert(&kvm->arch.vsie.addr_to_page, addr >> 9, page)) {
1263                 page_ref_dec(page);
1264                 mutex_unlock(&kvm->arch.vsie.mutex);
1265                 return NULL;
1266         }
1267         mutex_unlock(&kvm->arch.vsie.mutex);
1268
1269         vsie_page = page_to_virt(page);
1270         memset(&vsie_page->scb_s, 0, sizeof(struct kvm_s390_sie_block));
1271         release_gmap_shadow(vsie_page);
1272         vsie_page->fault_addr = 0;
1273         vsie_page->scb_s.ihcpu = 0xffffU;
1274         return vsie_page;
1275 }
1276
1277 /* put a vsie page acquired via get_vsie_page */
1278 static void put_vsie_page(struct kvm *kvm, struct vsie_page *vsie_page)
1279 {
1280         struct page *page = pfn_to_page(__pa(vsie_page) >> PAGE_SHIFT);
1281
1282         page_ref_dec(page);
1283 }
1284
1285 int kvm_s390_handle_vsie(struct kvm_vcpu *vcpu)
1286 {
1287         struct vsie_page *vsie_page;
1288         unsigned long scb_addr;
1289         int rc;
1290
1291         vcpu->stat.instruction_sie++;
1292         if (!test_kvm_cpu_feat(vcpu->kvm, KVM_S390_VM_CPU_FEAT_SIEF2))
1293                 return -EOPNOTSUPP;
1294         if (vcpu->arch.sie_block->gpsw.mask & PSW_MASK_PSTATE)
1295                 return kvm_s390_inject_program_int(vcpu, PGM_PRIVILEGED_OP);
1296
1297         BUILD_BUG_ON(sizeof(struct vsie_page) != PAGE_SIZE);
1298         scb_addr = kvm_s390_get_base_disp_s(vcpu, NULL);
1299
1300         /* 512 byte alignment */
1301         if (unlikely(scb_addr & 0x1ffUL))
1302                 return kvm_s390_inject_program_int(vcpu, PGM_SPECIFICATION);
1303
1304         if (signal_pending(current) || kvm_s390_vcpu_has_irq(vcpu, 0) ||
1305             kvm_s390_vcpu_sie_inhibited(vcpu))
1306                 return 0;
1307
1308         vsie_page = get_vsie_page(vcpu->kvm, scb_addr);
1309         if (IS_ERR(vsie_page))
1310                 return PTR_ERR(vsie_page);
1311         else if (!vsie_page)
1312                 /* double use of sie control block - simply do nothing */
1313                 return 0;
1314
1315         rc = pin_scb(vcpu, vsie_page, scb_addr);
1316         if (rc)
1317                 goto out_put;
1318         rc = shadow_scb(vcpu, vsie_page);
1319         if (rc)
1320                 goto out_unpin_scb;
1321         rc = pin_blocks(vcpu, vsie_page);
1322         if (rc)
1323                 goto out_unshadow;
1324         register_shadow_scb(vcpu, vsie_page);
1325         rc = vsie_run(vcpu, vsie_page);
1326         unregister_shadow_scb(vcpu);
1327         unpin_blocks(vcpu, vsie_page);
1328 out_unshadow:
1329         unshadow_scb(vcpu, vsie_page);
1330 out_unpin_scb:
1331         unpin_scb(vcpu, vsie_page, scb_addr);
1332 out_put:
1333         put_vsie_page(vcpu->kvm, vsie_page);
1334
1335         return rc < 0 ? rc : 0;
1336 }
1337
1338 /* Init the vsie data structures. To be called when a vm is initialized. */
1339 void kvm_s390_vsie_init(struct kvm *kvm)
1340 {
1341         mutex_init(&kvm->arch.vsie.mutex);
1342         INIT_RADIX_TREE(&kvm->arch.vsie.addr_to_page, GFP_KERNEL_ACCOUNT);
1343 }
1344
1345 /* Destroy the vsie data structures. To be called when a vm is destroyed. */
1346 void kvm_s390_vsie_destroy(struct kvm *kvm)
1347 {
1348         struct vsie_page *vsie_page;
1349         struct page *page;
1350         int i;
1351
1352         mutex_lock(&kvm->arch.vsie.mutex);
1353         for (i = 0; i < kvm->arch.vsie.page_count; i++) {
1354                 page = kvm->arch.vsie.pages[i];
1355                 kvm->arch.vsie.pages[i] = NULL;
1356                 vsie_page = page_to_virt(page);
1357                 release_gmap_shadow(vsie_page);
1358                 /* free the radix tree entry */
1359                 radix_tree_delete(&kvm->arch.vsie.addr_to_page, page->index >> 9);
1360                 __free_page(page);
1361         }
1362         kvm->arch.vsie.page_count = 0;
1363         mutex_unlock(&kvm->arch.vsie.mutex);
1364 }
1365
1366 void kvm_s390_vsie_kick(struct kvm_vcpu *vcpu)
1367 {
1368         struct kvm_s390_sie_block *scb = READ_ONCE(vcpu->arch.vsie_block);
1369
1370         /*
1371          * Even if the VCPU lets go of the shadow sie block reference, it is
1372          * still valid in the cache. So we can safely kick it.
1373          */
1374         if (scb) {
1375                 atomic_or(PROG_BLOCK_SIE, &scb->prog20);
1376                 if (scb->prog0c & PROG_IN_SIE)
1377                         atomic_or(CPUSTAT_STOP_INT, &scb->cpuflags);
1378         }
1379 }