arch/x86/kvm/svm/sev.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Kernel-based Virtual Machine driver for Linux
   4  *
   5  * AMD SVM-SEV support
   6  *
   7  * Copyright 2010 Red Hat, Inc. and/or its affiliates.
   8  */
   9
  10 #include <linux/kvm_types.h>
  11 #include <linux/kvm_host.h>
  12 #include <linux/kernel.h>
  13 #include <linux/highmem.h>
  14 #include <linux/psp-sev.h>
  15 #include <linux/pagemap.h>
  16 #include <linux/swap.h>
  17 #include <linux/misc_cgroup.h>
  18 #include <linux/processor.h>
  19 #include <linux/trace_events.h>
  20
  21 #include <asm/pkru.h>
  22 #include <asm/trapnr.h>
  23 #include <asm/fpu/xcr.h>
  24
  25 #include "x86.h"
  26 #include "svm.h"
  27 #include "svm_ops.h"
  28 #include "cpuid.h"
  29 #include "trace.h"
  30
  31 #ifndef CONFIG_KVM_AMD_SEV
  32 /*
  33  * When this config is not defined, SEV feature is not supported and APIs in
  34  * this file are not used but this file still gets compiled into the KVM AMD
  35  * module.
  36  *
  37  * We will not have MISC_CG_RES_SEV and MISC_CG_RES_SEV_ES entries in the enum
  38  * misc_res_type {} defined in linux/misc_cgroup.h.
  39  *
  40  * Below macros allow compilation to succeed.
  41  */
  42 #define MISC_CG_RES_SEV MISC_CG_RES_TYPES
  43 #define MISC_CG_RES_SEV_ES MISC_CG_RES_TYPES
  44 #endif
  45
  46 #ifdef CONFIG_KVM_AMD_SEV
  47 /* enable/disable SEV support */
  48 static bool sev_enabled = true;
  49 module_param_named(sev, sev_enabled, bool, 0444);
  50
  51 /* enable/disable SEV-ES support */
  52 static bool sev_es_enabled = true;
  53 module_param_named(sev_es, sev_es_enabled, bool, 0444);
  54 #else
  55 #define sev_enabled false
  56 #define sev_es_enabled false
  57 #endif /* CONFIG_KVM_AMD_SEV */
  58
  59 static u8 sev_enc_bit;
  60 static DECLARE_RWSEM(sev_deactivate_lock);
  61 static DEFINE_MUTEX(sev_bitmap_lock);
  62 unsigned int max_sev_asid;
  63 static unsigned int min_sev_asid;
  64 static unsigned long sev_me_mask;
  65 static unsigned int nr_asids;
  66 static unsigned long *sev_asid_bitmap;
  67 static unsigned long *sev_reclaim_asid_bitmap;
  68
  69 struct enc_region {
  70         struct list_head list;
  71         unsigned long npages;
  72         struct page **pages;
  73         unsigned long uaddr;
  74         unsigned long size;
  75 };
  76
  77 /* Called with the sev_bitmap_lock held, or on shutdown  */
  78 static int sev_flush_asids(int min_asid, int max_asid)
  79 {
  80         int ret, asid, error = 0;
  81
  82         /* Check if there are any ASIDs to reclaim before performing a flush */
  83         asid = find_next_bit(sev_reclaim_asid_bitmap, nr_asids, min_asid);
  84         if (asid > max_asid)
  85                 return -EBUSY;
  86
  87         /*
  88          * DEACTIVATE will clear the WBINVD indicator causing DF_FLUSH to fail,
  89          * so it must be guarded.
  90          */
  91         down_write(&sev_deactivate_lock);
  92
  93         wbinvd_on_all_cpus();
  94         ret = sev_guest_df_flush(&error);
  95
  96         up_write(&sev_deactivate_lock);
  97
  98         if (ret)
  99                 pr_err("SEV: DF_FLUSH failed, ret=%d, error=%#x\n", ret, error);
 100
 101         return ret;
 102 }
 103
 104 static inline bool is_mirroring_enc_context(struct kvm *kvm)
 105 {
 106         return !!to_kvm_svm(kvm)->sev_info.enc_context_owner;
 107 }
 108
 109 /* Must be called with the sev_bitmap_lock held */
 110 static bool __sev_recycle_asids(int min_asid, int max_asid)
 111 {
 112         if (sev_flush_asids(min_asid, max_asid))
 113                 return false;
 114
 115         /* The flush process will flush all reclaimable SEV and SEV-ES ASIDs */
 116         bitmap_xor(sev_asid_bitmap, sev_asid_bitmap, sev_reclaim_asid_bitmap,
 117                    nr_asids);
 118         bitmap_zero(sev_reclaim_asid_bitmap, nr_asids);
 119
 120         return true;
 121 }
 122
 123 static int sev_misc_cg_try_charge(struct kvm_sev_info *sev)
 124 {
 125         enum misc_res_type type = sev->es_active ? MISC_CG_RES_SEV_ES : MISC_CG_RES_SEV;
 126         return misc_cg_try_charge(type, sev->misc_cg, 1);
 127 }
 128
 129 static void sev_misc_cg_uncharge(struct kvm_sev_info *sev)
 130 {
 131         enum misc_res_type type = sev->es_active ? MISC_CG_RES_SEV_ES : MISC_CG_RES_SEV;
 132         misc_cg_uncharge(type, sev->misc_cg, 1);
 133 }
 134
 135 static int sev_asid_new(struct kvm_sev_info *sev)
 136 {
 137         int asid, min_asid, max_asid, ret;
 138         bool retry = true;
 139
 140         WARN_ON(sev->misc_cg);
 141         sev->misc_cg = get_current_misc_cg();
 142         ret = sev_misc_cg_try_charge(sev);
 143         if (ret) {
 144                 put_misc_cg(sev->misc_cg);
 145                 sev->misc_cg = NULL;
 146                 return ret;
 147         }
 148
 149         mutex_lock(&sev_bitmap_lock);
 150
 151         /*
 152          * SEV-enabled guests must use asid from min_sev_asid to max_sev_asid.
 153          * SEV-ES-enabled guest can use from 1 to min_sev_asid - 1.
 154          */
 155         min_asid = sev->es_active ? 1 : min_sev_asid;
 156         max_asid = sev->es_active ? min_sev_asid - 1 : max_sev_asid;
 157 again:
 158         asid = find_next_zero_bit(sev_asid_bitmap, max_asid + 1, min_asid);
 159         if (asid > max_asid) {
 160                 if (retry && __sev_recycle_asids(min_asid, max_asid)) {
 161                         retry = false;
 162                         goto again;
 163                 }
 164                 mutex_unlock(&sev_bitmap_lock);
 165                 ret = -EBUSY;
 166                 goto e_uncharge;
 167         }
 168
 169         __set_bit(asid, sev_asid_bitmap);
 170
 171         mutex_unlock(&sev_bitmap_lock);
 172
 173         return asid;
 174 e_uncharge:
 175         sev_misc_cg_uncharge(sev);
 176         put_misc_cg(sev->misc_cg);
 177         sev->misc_cg = NULL;
 178         return ret;
 179 }
 180
 181 static int sev_get_asid(struct kvm *kvm)
 182 {
 183         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 184
 185         return sev->asid;
 186 }
 187
 188 static void sev_asid_free(struct kvm_sev_info *sev)
 189 {
 190         struct svm_cpu_data *sd;
 191         int cpu;
 192
 193         mutex_lock(&sev_bitmap_lock);
 194
 195         __set_bit(sev->asid, sev_reclaim_asid_bitmap);
 196
 197         for_each_possible_cpu(cpu) {
 198                 sd = per_cpu(svm_data, cpu);
 199                 sd->sev_vmcbs[sev->asid] = NULL;
 200         }
 201
 202         mutex_unlock(&sev_bitmap_lock);
 203
 204         sev_misc_cg_uncharge(sev);
 205         put_misc_cg(sev->misc_cg);
 206         sev->misc_cg = NULL;
 207 }
 208
 209 static void sev_decommission(unsigned int handle)
 210 {
 211         struct sev_data_decommission decommission;
 212
 213         if (!handle)
 214                 return;
 215
 216         decommission.handle = handle;
 217         sev_guest_decommission(&decommission, NULL);
 218 }
 219
 220 static void sev_unbind_asid(struct kvm *kvm, unsigned int handle)
 221 {
 222         struct sev_data_deactivate deactivate;
 223
 224         if (!handle)
 225                 return;
 226
 227         deactivate.handle = handle;
 228
 229         /* Guard DEACTIVATE against WBINVD/DF_FLUSH used in ASID recycling */
 230         down_read(&sev_deactivate_lock);
 231         sev_guest_deactivate(&deactivate, NULL);
 232         up_read(&sev_deactivate_lock);
 233
 234         sev_decommission(handle);
 235 }
 236
 237 static int sev_guest_init(struct kvm *kvm, struct kvm_sev_cmd *argp)
 238 {
 239         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 240         int asid, ret;
 241
 242         if (kvm->created_vcpus)
 243                 return -EINVAL;
 244
 245         ret = -EBUSY;
 246         if (unlikely(sev->active))
 247                 return ret;
 248
 249         sev->active = true;
 250         sev->es_active = argp->id == KVM_SEV_ES_INIT;
 251         asid = sev_asid_new(sev);
 252         if (asid < 0)
 253                 goto e_no_asid;
 254         sev->asid = asid;
 255
 256         ret = sev_platform_init(&argp->error);
 257         if (ret)
 258                 goto e_free;
 259
 260         INIT_LIST_HEAD(&sev->regions_list);
 261
 262         return 0;
 263
 264 e_free:
 265         sev_asid_free(sev);
 266         sev->asid = 0;
 267 e_no_asid:
 268         sev->es_active = false;
 269         sev->active = false;
 270         return ret;
 271 }
 272
 273 static int sev_bind_asid(struct kvm *kvm, unsigned int handle, int *error)
 274 {
 275         struct sev_data_activate activate;
 276         int asid = sev_get_asid(kvm);
 277         int ret;
 278
 279         /* activate ASID on the given handle */
 280         activate.handle = handle;
 281         activate.asid   = asid;
 282         ret = sev_guest_activate(&activate, error);
 283
 284         return ret;
 285 }
 286
 287 static int __sev_issue_cmd(int fd, int id, void *data, int *error)
 288 {
 289         struct fd f;
 290         int ret;
 291
 292         f = fdget(fd);
 293         if (!f.file)
 294                 return -EBADF;
 295
 296         ret = sev_issue_cmd_external_user(f.file, id, data, error);
 297
 298         fdput(f);
 299         return ret;
 300 }
 301
 302 static int sev_issue_cmd(struct kvm *kvm, int id, void *data, int *error)
 303 {
 304         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 305
 306         return __sev_issue_cmd(sev->fd, id, data, error);
 307 }
 308
 309 static int sev_launch_start(struct kvm *kvm, struct kvm_sev_cmd *argp)
 310 {
 311         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 312         struct sev_data_launch_start start;
 313         struct kvm_sev_launch_start params;
 314         void *dh_blob, *session_blob;
 315         int *error = &argp->error;
 316         int ret;
 317
 318         if (!sev_guest(kvm))
 319                 return -ENOTTY;
 320
 321         if (copy_from_user(&params, (void __user *)(uintptr_t)argp->data, sizeof(params)))
 322                 return -EFAULT;
 323
 324         memset(&start, 0, sizeof(start));
 325
 326         dh_blob = NULL;
 327         if (params.dh_uaddr) {
 328                 dh_blob = psp_copy_user_blob(params.dh_uaddr, params.dh_len);
 329                 if (IS_ERR(dh_blob))
 330                         return PTR_ERR(dh_blob);
 331
 332                 start.dh_cert_address = __sme_set(__pa(dh_blob));
 333                 start.dh_cert_len = params.dh_len;
 334         }
 335
 336         session_blob = NULL;
 337         if (params.session_uaddr) {
 338                 session_blob = psp_copy_user_blob(params.session_uaddr, params.session_len);
 339                 if (IS_ERR(session_blob)) {
 340                         ret = PTR_ERR(session_blob);
 341                         goto e_free_dh;
 342                 }
 343
 344                 start.session_address = __sme_set(__pa(session_blob));
 345                 start.session_len = params.session_len;
 346         }
 347
 348         start.handle = params.handle;
 349         start.policy = params.policy;
 350
 351         /* create memory encryption context */
 352         ret = __sev_issue_cmd(argp->sev_fd, SEV_CMD_LAUNCH_START, &start, error);
 353         if (ret)
 354                 goto e_free_session;
 355
 356         /* Bind ASID to this guest */
 357         ret = sev_bind_asid(kvm, start.handle, error);
 358         if (ret) {
 359                 sev_decommission(start.handle);
 360                 goto e_free_session;
 361         }
 362
 363         /* return handle to userspace */
 364         params.handle = start.handle;
 365         if (copy_to_user((void __user *)(uintptr_t)argp->data, &params, sizeof(params))) {
 366                 sev_unbind_asid(kvm, start.handle);
 367                 ret = -EFAULT;
 368                 goto e_free_session;
 369         }
 370
 371         sev->handle = start.handle;
 372         sev->fd = argp->sev_fd;
 373
 374 e_free_session:
 375         kfree(session_blob);
 376 e_free_dh:
 377         kfree(dh_blob);
 378         return ret;
 379 }
 380
 381 static struct page **sev_pin_memory(struct kvm *kvm, unsigned long uaddr,
 382                                     unsigned long ulen, unsigned long *n,
 383                                     int write)
 384 {
 385         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 386         unsigned long npages, size;
 387         int npinned;
 388         unsigned long locked, lock_limit;
 389         struct page **pages;
 390         unsigned long first, last;
 391         int ret;
 392
 393         lockdep_assert_held(&kvm->lock);
 394
 395         if (ulen == 0 || uaddr + ulen < uaddr)
 396                 return ERR_PTR(-EINVAL);
 397
 398         /* Calculate number of pages. */
 399         first = (uaddr & PAGE_MASK) >> PAGE_SHIFT;
 400         last = ((uaddr + ulen - 1) & PAGE_MASK) >> PAGE_SHIFT;
 401         npages = (last - first + 1);
 402
 403         locked = sev->pages_locked + npages;
 404         lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
 405         if (locked > lock_limit && !capable(CAP_IPC_LOCK)) {
 406                 pr_err("SEV: %lu locked pages exceed the lock limit of %lu.\n", locked, lock_limit);
 407                 return ERR_PTR(-ENOMEM);
 408         }
 409
 410         if (WARN_ON_ONCE(npages > INT_MAX))
 411                 return ERR_PTR(-EINVAL);
 412
 413         /* Avoid using vmalloc for smaller buffers. */
 414         size = npages * sizeof(struct page *);
 415         if (size > PAGE_SIZE)
 416                 pages = __vmalloc(size, GFP_KERNEL_ACCOUNT | __GFP_ZERO);
 417         else
 418                 pages = kmalloc(size, GFP_KERNEL_ACCOUNT);
 419
 420         if (!pages)
 421                 return ERR_PTR(-ENOMEM);
 422
 423         /* Pin the user virtual address. */
 424         npinned = pin_user_pages_fast(uaddr, npages, write ? FOLL_WRITE : 0, pages);
 425         if (npinned != npages) {
 426                 pr_err("SEV: Failure locking %lu pages.\n", npages);
 427                 ret = -ENOMEM;
 428                 goto err;
 429         }
 430
 431         *n = npages;
 432         sev->pages_locked = locked;
 433
 434         return pages;
 435
 436 err:
 437         if (npinned > 0)
 438                 unpin_user_pages(pages, npinned);
 439
 440         kvfree(pages);
 441         return ERR_PTR(ret);
 442 }
 443
 444 static void sev_unpin_memory(struct kvm *kvm, struct page **pages,
 445                              unsigned long npages)
 446 {
 447         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 448
 449         unpin_user_pages(pages, npages);
 450         kvfree(pages);
 451         sev->pages_locked -= npages;
 452 }
 453
 454 static void sev_clflush_pages(struct page *pages[], unsigned long npages)
 455 {
 456         uint8_t *page_virtual;
 457         unsigned long i;
 458
 459         if (this_cpu_has(X86_FEATURE_SME_COHERENT) || npages == 0 ||
 460             pages == NULL)
 461                 return;
 462
 463         for (i = 0; i < npages; i++) {
 464                 page_virtual = kmap_atomic(pages[i]);
 465                 clflush_cache_range(page_virtual, PAGE_SIZE);
 466                 kunmap_atomic(page_virtual);
 467         }
 468 }
 469
 470 static unsigned long get_num_contig_pages(unsigned long idx,
 471                                 struct page **inpages, unsigned long npages)
 472 {
 473         unsigned long paddr, next_paddr;
 474         unsigned long i = idx + 1, pages = 1;
 475
 476         /* find the number of contiguous pages starting from idx */
 477         paddr = __sme_page_pa(inpages[idx]);
 478         while (i < npages) {
 479                 next_paddr = __sme_page_pa(inpages[i++]);
 480                 if ((paddr + PAGE_SIZE) == next_paddr) {
 481                         pages++;
 482                         paddr = next_paddr;
 483                         continue;
 484                 }
 485                 break;
 486         }
 487
 488         return pages;
 489 }
 490
 491 static int sev_launch_update_data(struct kvm *kvm, struct kvm_sev_cmd *argp)
 492 {
 493         unsigned long vaddr, vaddr_end, next_vaddr, npages, pages, size, i;
 494         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 495         struct kvm_sev_launch_update_data params;
 496         struct sev_data_launch_update_data data;
 497         struct page **inpages;
 498         int ret;
 499
 500         if (!sev_guest(kvm))
 501                 return -ENOTTY;
 502
 503         if (copy_from_user(&params, (void __user *)(uintptr_t)argp->data, sizeof(params)))
 504                 return -EFAULT;
 505
 506         vaddr = params.uaddr;
 507         size = params.len;
 508         vaddr_end = vaddr + size;
 509
 510         /* Lock the user memory. */
 511         inpages = sev_pin_memory(kvm, vaddr, size, &npages, 1);
 512         if (IS_ERR(inpages))
 513                 return PTR_ERR(inpages);
 514
 515         /*
 516          * Flush (on non-coherent CPUs) before LAUNCH_UPDATE encrypts pages in
 517          * place; the cache may contain the data that was written unencrypted.
 518          */
 519         sev_clflush_pages(inpages, npages);
 520
 521         data.reserved = 0;
 522         data.handle = sev->handle;
 523
 524         for (i = 0; vaddr < vaddr_end; vaddr = next_vaddr, i += pages) {
 525                 int offset, len;
 526
 527                 /*
 528                  * If the user buffer is not page-aligned, calculate the offset
 529                  * within the page.
 530                  */
 531                 offset = vaddr & (PAGE_SIZE - 1);
 532
 533                 /* Calculate the number of pages that can be encrypted in one go. */
 534                 pages = get_num_contig_pages(i, inpages, npages);
 535
 536                 len = min_t(size_t, ((pages * PAGE_SIZE) - offset), size);
 537
 538                 data.len = len;
 539                 data.address = __sme_page_pa(inpages[i]) + offset;
 540                 ret = sev_issue_cmd(kvm, SEV_CMD_LAUNCH_UPDATE_DATA, &data, &argp->error);
 541                 if (ret)
 542                         goto e_unpin;
 543
 544                 size -= len;
 545                 next_vaddr = vaddr + len;
 546         }
 547
 548 e_unpin:
 549         /* content of memory is updated, mark pages dirty */
 550         for (i = 0; i < npages; i++) {
 551                 set_page_dirty_lock(inpages[i]);
 552                 mark_page_accessed(inpages[i]);
 553         }
 554         /* unlock the user pages */
 555         sev_unpin_memory(kvm, inpages, npages);
 556         return ret;
 557 }
 558
 559 static int sev_es_sync_vmsa(struct vcpu_svm *svm)
 560 {
 561         struct vmcb_save_area *save = &svm->vmcb->save;
 562
 563         /* Check some debug related fields before encrypting the VMSA */
 564         if (svm->vcpu.guest_debug || (save->dr7 & ~DR7_FIXED_1))
 565                 return -EINVAL;
 566
 567         /* Sync registgers */
 568         save->rax = svm->vcpu.arch.regs[VCPU_REGS_RAX];
 569         save->rbx = svm->vcpu.arch.regs[VCPU_REGS_RBX];
 570         save->rcx = svm->vcpu.arch.regs[VCPU_REGS_RCX];
 571         save->rdx = svm->vcpu.arch.regs[VCPU_REGS_RDX];
 572         save->rsp = svm->vcpu.arch.regs[VCPU_REGS_RSP];
 573         save->rbp = svm->vcpu.arch.regs[VCPU_REGS_RBP];
 574         save->rsi = svm->vcpu.arch.regs[VCPU_REGS_RSI];
 575         save->rdi = svm->vcpu.arch.regs[VCPU_REGS_RDI];
 576 #ifdef CONFIG_X86_64
 577         save->r8  = svm->vcpu.arch.regs[VCPU_REGS_R8];
 578         save->r9  = svm->vcpu.arch.regs[VCPU_REGS_R9];
 579         save->r10 = svm->vcpu.arch.regs[VCPU_REGS_R10];
 580         save->r11 = svm->vcpu.arch.regs[VCPU_REGS_R11];
 581         save->r12 = svm->vcpu.arch.regs[VCPU_REGS_R12];
 582         save->r13 = svm->vcpu.arch.regs[VCPU_REGS_R13];
 583         save->r14 = svm->vcpu.arch.regs[VCPU_REGS_R14];
 584         save->r15 = svm->vcpu.arch.regs[VCPU_REGS_R15];
 585 #endif
 586         save->rip = svm->vcpu.arch.regs[VCPU_REGS_RIP];
 587
 588         /* Sync some non-GPR registers before encrypting */
 589         save->xcr0 = svm->vcpu.arch.xcr0;
 590         save->pkru = svm->vcpu.arch.pkru;
 591         save->xss  = svm->vcpu.arch.ia32_xss;
 592         save->dr6  = svm->vcpu.arch.dr6;
 593
 594         /*
 595          * SEV-ES will use a VMSA that is pointed to by the VMCB, not
 596          * the traditional VMSA that is part of the VMCB. Copy the
 597          * traditional VMSA as it has been built so far (in prep
 598          * for LAUNCH_UPDATE_VMSA) to be the initial SEV-ES state.
 599          */
 600         memcpy(svm->sev_es.vmsa, save, sizeof(*save));
 601
 602         return 0;
 603 }
 604
 605 static int __sev_launch_update_vmsa(struct kvm *kvm, struct kvm_vcpu *vcpu,
 606                                     int *error)
 607 {
 608         struct sev_data_launch_update_vmsa vmsa;
 609         struct vcpu_svm *svm = to_svm(vcpu);
 610         int ret;
 611
 612         /* Perform some pre-encryption checks against the VMSA */
 613         ret = sev_es_sync_vmsa(svm);
 614         if (ret)
 615                 return ret;
 616
 617         /*
 618          * The LAUNCH_UPDATE_VMSA command will perform in-place encryption of
 619          * the VMSA memory content (i.e it will write the same memory region
 620          * with the guest's key), so invalidate it first.
 621          */
 622         clflush_cache_range(svm->sev_es.vmsa, PAGE_SIZE);
 623
 624         vmsa.reserved = 0;
 625         vmsa.handle = to_kvm_svm(kvm)->sev_info.handle;
 626         vmsa.address = __sme_pa(svm->sev_es.vmsa);
 627         vmsa.len = PAGE_SIZE;
 628         ret = sev_issue_cmd(kvm, SEV_CMD_LAUNCH_UPDATE_VMSA, &vmsa, error);
 629         if (ret)
 630           return ret;
 631
 632         vcpu->arch.guest_state_protected = true;
 633         return 0;
 634 }
 635
 636 static int sev_launch_update_vmsa(struct kvm *kvm, struct kvm_sev_cmd *argp)
 637 {
 638         struct kvm_vcpu *vcpu;
 639         int i, ret;
 640
 641         if (!sev_es_guest(kvm))
 642                 return -ENOTTY;
 643
 644         kvm_for_each_vcpu(i, vcpu, kvm) {
 645                 ret = mutex_lock_killable(&vcpu->mutex);
 646                 if (ret)
 647                         return ret;
 648
 649                 ret = __sev_launch_update_vmsa(kvm, vcpu, &argp->error);
 650
 651                 mutex_unlock(&vcpu->mutex);
 652                 if (ret)
 653                         return ret;
 654         }
 655
 656         return 0;
 657 }
 658
 659 static int sev_launch_measure(struct kvm *kvm, struct kvm_sev_cmd *argp)
 660 {
 661         void __user *measure = (void __user *)(uintptr_t)argp->data;
 662         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 663         struct sev_data_launch_measure data;
 664         struct kvm_sev_launch_measure params;
 665         void __user *p = NULL;
 666         void *blob = NULL;
 667         int ret;
 668
 669         if (!sev_guest(kvm))
 670                 return -ENOTTY;
 671
 672         if (copy_from_user(&params, measure, sizeof(params)))
 673                 return -EFAULT;
 674
 675         memset(&data, 0, sizeof(data));
 676
 677         /* User wants to query the blob length */
 678         if (!params.len)
 679                 goto cmd;
 680
 681         p = (void __user *)(uintptr_t)params.uaddr;
 682         if (p) {
 683                 if (params.len > SEV_FW_BLOB_MAX_SIZE)
 684                         return -EINVAL;
 685
 686                 blob = kmalloc(params.len, GFP_KERNEL_ACCOUNT);
 687                 if (!blob)
 688                         return -ENOMEM;
 689
 690                 data.address = __psp_pa(blob);
 691                 data.len = params.len;
 692         }
 693
 694 cmd:
 695         data.handle = sev->handle;
 696         ret = sev_issue_cmd(kvm, SEV_CMD_LAUNCH_MEASURE, &data, &argp->error);
 697
 698         /*
 699          * If we query the session length, FW responded with expected data.
 700          */
 701         if (!params.len)
 702                 goto done;
 703
 704         if (ret)
 705                 goto e_free_blob;
 706
 707         if (blob) {
 708                 if (copy_to_user(p, blob, params.len))
 709                         ret = -EFAULT;
 710         }
 711
 712 done:
 713         params.len = data.len;
 714         if (copy_to_user(measure, &params, sizeof(params)))
 715                 ret = -EFAULT;
 716 e_free_blob:
 717         kfree(blob);
 718         return ret;
 719 }
 720
 721 static int sev_launch_finish(struct kvm *kvm, struct kvm_sev_cmd *argp)
 722 {
 723         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 724         struct sev_data_launch_finish data;
 725
 726         if (!sev_guest(kvm))
 727                 return -ENOTTY;
 728
 729         data.handle = sev->handle;
 730         return sev_issue_cmd(kvm, SEV_CMD_LAUNCH_FINISH, &data, &argp->error);
 731 }
 732
 733 static int sev_guest_status(struct kvm *kvm, struct kvm_sev_cmd *argp)
 734 {
 735         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 736         struct kvm_sev_guest_status params;
 737         struct sev_data_guest_status data;
 738         int ret;
 739
 740         if (!sev_guest(kvm))
 741                 return -ENOTTY;
 742
 743         memset(&data, 0, sizeof(data));
 744
 745         data.handle = sev->handle;
 746         ret = sev_issue_cmd(kvm, SEV_CMD_GUEST_STATUS, &data, &argp->error);
 747         if (ret)
 748                 return ret;
 749
 750         params.policy = data.policy;
 751         params.state = data.state;
 752         params.handle = data.handle;
 753
 754         if (copy_to_user((void __user *)(uintptr_t)argp->data, &params, sizeof(params)))
 755                 ret = -EFAULT;
 756
 757         return ret;
 758 }
 759
 760 static int __sev_issue_dbg_cmd(struct kvm *kvm, unsigned long src,
 761                                unsigned long dst, int size,
 762                                int *error, bool enc)
 763 {
 764         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 765         struct sev_data_dbg data;
 766
 767         data.reserved = 0;
 768         data.handle = sev->handle;
 769         data.dst_addr = dst;
 770         data.src_addr = src;
 771         data.len = size;
 772
 773         return sev_issue_cmd(kvm,
 774                              enc ? SEV_CMD_DBG_ENCRYPT : SEV_CMD_DBG_DECRYPT,
 775                              &data, error);
 776 }
 777
 778 static int __sev_dbg_decrypt(struct kvm *kvm, unsigned long src_paddr,
 779                              unsigned long dst_paddr, int sz, int *err)
 780 {
 781         int offset;
 782
 783         /*
 784          * Its safe to read more than we are asked, caller should ensure that
 785          * destination has enough space.
 786          */
 787         offset = src_paddr & 15;
 788         src_paddr = round_down(src_paddr, 16);
 789         sz = round_up(sz + offset, 16);
 790
 791         return __sev_issue_dbg_cmd(kvm, src_paddr, dst_paddr, sz, err, false);
 792 }
 793
 794 static int __sev_dbg_decrypt_user(struct kvm *kvm, unsigned long paddr,
 795                                   void __user *dst_uaddr,
 796                                   unsigned long dst_paddr,
 797                                   int size, int *err)
 798 {
 799         struct page *tpage = NULL;
 800         int ret, offset;
 801
 802         /* if inputs are not 16-byte then use intermediate buffer */
 803         if (!IS_ALIGNED(dst_paddr, 16) ||
 804             !IS_ALIGNED(paddr,     16) ||
 805             !IS_ALIGNED(size,      16)) {
 806                 tpage = (void *)alloc_page(GFP_KERNEL);
 807                 if (!tpage)
 808                         return -ENOMEM;
 809
 810                 dst_paddr = __sme_page_pa(tpage);
 811         }
 812
 813         ret = __sev_dbg_decrypt(kvm, paddr, dst_paddr, size, err);
 814         if (ret)
 815                 goto e_free;
 816
 817         if (tpage) {
 818                 offset = paddr & 15;
 819                 if (copy_to_user(dst_uaddr, page_address(tpage) + offset, size))
 820                         ret = -EFAULT;
 821         }
 822
 823 e_free:
 824         if (tpage)
 825                 __free_page(tpage);
 826
 827         return ret;
 828 }
 829
 830 static int __sev_dbg_encrypt_user(struct kvm *kvm, unsigned long paddr,
 831                                   void __user *vaddr,
 832                                   unsigned long dst_paddr,
 833                                   void __user *dst_vaddr,
 834                                   int size, int *error)
 835 {
 836         struct page *src_tpage = NULL;
 837         struct page *dst_tpage = NULL;
 838         int ret, len = size;
 839
 840         /* If source buffer is not aligned then use an intermediate buffer */
 841         if (!IS_ALIGNED((unsigned long)vaddr, 16)) {
 842                 src_tpage = alloc_page(GFP_KERNEL);
 843                 if (!src_tpage)
 844                         return -ENOMEM;
 845
 846                 if (copy_from_user(page_address(src_tpage), vaddr, size)) {
 847                         __free_page(src_tpage);
 848                         return -EFAULT;
 849                 }
 850
 851                 paddr = __sme_page_pa(src_tpage);
 852         }
 853
 854         /*
 855          *  If destination buffer or length is not aligned then do read-modify-write:
 856          *   - decrypt destination in an intermediate buffer
 857          *   - copy the source buffer in an intermediate buffer
 858          *   - use the intermediate buffer as source buffer
 859          */
 860         if (!IS_ALIGNED((unsigned long)dst_vaddr, 16) || !IS_ALIGNED(size, 16)) {
 861                 int dst_offset;
 862
 863                 dst_tpage = alloc_page(GFP_KERNEL);
 864                 if (!dst_tpage) {
 865                         ret = -ENOMEM;
 866                         goto e_free;
 867                 }
 868
 869                 ret = __sev_dbg_decrypt(kvm, dst_paddr,
 870                                         __sme_page_pa(dst_tpage), size, error);
 871                 if (ret)
 872                         goto e_free;
 873
 874                 /*
 875                  *  If source is kernel buffer then use memcpy() otherwise
 876                  *  copy_from_user().
 877                  */
 878                 dst_offset = dst_paddr & 15;
 879
 880                 if (src_tpage)
 881                         memcpy(page_address(dst_tpage) + dst_offset,
 882                                page_address(src_tpage), size);
 883                 else {
 884                         if (copy_from_user(page_address(dst_tpage) + dst_offset,
 885                                            vaddr, size)) {
 886                                 ret = -EFAULT;
 887                                 goto e_free;
 888                         }
 889                 }
 890
 891                 paddr = __sme_page_pa(dst_tpage);
 892                 dst_paddr = round_down(dst_paddr, 16);
 893                 len = round_up(size, 16);
 894         }
 895
 896         ret = __sev_issue_dbg_cmd(kvm, paddr, dst_paddr, len, error, true);
 897
 898 e_free:
 899         if (src_tpage)
 900                 __free_page(src_tpage);
 901         if (dst_tpage)
 902                 __free_page(dst_tpage);
 903         return ret;
 904 }
 905
 906 static int sev_dbg_crypt(struct kvm *kvm, struct kvm_sev_cmd *argp, bool dec)
 907 {
 908         unsigned long vaddr, vaddr_end, next_vaddr;
 909         unsigned long dst_vaddr;
 910         struct page **src_p, **dst_p;
 911         struct kvm_sev_dbg debug;
 912         unsigned long n;
 913         unsigned int size;
 914         int ret;
 915
 916         if (!sev_guest(kvm))
 917                 return -ENOTTY;
 918
 919         if (copy_from_user(&debug, (void __user *)(uintptr_t)argp->data, sizeof(debug)))
 920                 return -EFAULT;
 921
 922         if (!debug.len || debug.src_uaddr + debug.len < debug.src_uaddr)
 923                 return -EINVAL;
 924         if (!debug.dst_uaddr)
 925                 return -EINVAL;
 926
 927         vaddr = debug.src_uaddr;
 928         size = debug.len;
 929         vaddr_end = vaddr + size;
 930         dst_vaddr = debug.dst_uaddr;
 931
 932         for (; vaddr < vaddr_end; vaddr = next_vaddr) {
 933                 int len, s_off, d_off;
 934
 935                 /* lock userspace source and destination page */
 936                 src_p = sev_pin_memory(kvm, vaddr & PAGE_MASK, PAGE_SIZE, &n, 0);
 937                 if (IS_ERR(src_p))
 938                         return PTR_ERR(src_p);
 939
 940                 dst_p = sev_pin_memory(kvm, dst_vaddr & PAGE_MASK, PAGE_SIZE, &n, 1);
 941                 if (IS_ERR(dst_p)) {
 942                         sev_unpin_memory(kvm, src_p, n);
 943                         return PTR_ERR(dst_p);
 944                 }
 945
 946                 /*
 947                  * Flush (on non-coherent CPUs) before DBG_{DE,EN}CRYPT read or modify
 948                  * the pages; flush the destination too so that future accesses do not
 949                  * see stale data.
 950                  */
 951                 sev_clflush_pages(src_p, 1);
 952                 sev_clflush_pages(dst_p, 1);
 953
 954                 /*
 955                  * Since user buffer may not be page aligned, calculate the
 956                  * offset within the page.
 957                  */
 958                 s_off = vaddr & ~PAGE_MASK;
 959                 d_off = dst_vaddr & ~PAGE_MASK;
 960                 len = min_t(size_t, (PAGE_SIZE - s_off), size);
 961
 962                 if (dec)
 963                         ret = __sev_dbg_decrypt_user(kvm,
 964                                                      __sme_page_pa(src_p[0]) + s_off,
 965                                                      (void __user *)dst_vaddr,
 966                                                      __sme_page_pa(dst_p[0]) + d_off,
 967                                                      len, &argp->error);
 968                 else
 969                         ret = __sev_dbg_encrypt_user(kvm,
 970                                                      __sme_page_pa(src_p[0]) + s_off,
 971                                                      (void __user *)vaddr,
 972                                                      __sme_page_pa(dst_p[0]) + d_off,
 973                                                      (void __user *)dst_vaddr,
 974                                                      len, &argp->error);
 975
 976                 sev_unpin_memory(kvm, src_p, n);
 977                 sev_unpin_memory(kvm, dst_p, n);
 978
 979                 if (ret)
 980                         goto err;
 981
 982                 next_vaddr = vaddr + len;
 983                 dst_vaddr = dst_vaddr + len;
 984                 size -= len;
 985         }
 986 err:
 987         return ret;
 988 }
 989
 990 static int sev_launch_secret(struct kvm *kvm, struct kvm_sev_cmd *argp)
 991 {
 992         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
 993         struct sev_data_launch_secret data;
 994         struct kvm_sev_launch_secret params;
 995         struct page **pages;
 996         void *blob, *hdr;
 997         unsigned long n, i;
 998         int ret, offset;
 999
1000         if (!sev_guest(kvm))
1001                 return -ENOTTY;
1002
1003         if (copy_from_user(&params, (void __user *)(uintptr_t)argp->data, sizeof(params)))
1004                 return -EFAULT;
1005
1006         pages = sev_pin_memory(kvm, params.guest_uaddr, params.guest_len, &n, 1);
1007         if (IS_ERR(pages))
1008                 return PTR_ERR(pages);
1009
1010         /*
1011          * Flush (on non-coherent CPUs) before LAUNCH_SECRET encrypts pages in
1012          * place; the cache may contain the data that was written unencrypted.
1013          */
1014         sev_clflush_pages(pages, n);
1015
1016         /*
1017          * The secret must be copied into contiguous memory region, lets verify
1018          * that userspace memory pages are contiguous before we issue command.
1019          */
1020         if (get_num_contig_pages(0, pages, n) != n) {
1021                 ret = -EINVAL;
1022                 goto e_unpin_memory;
1023         }
1024
1025         memset(&data, 0, sizeof(data));
1026
1027         offset = params.guest_uaddr & (PAGE_SIZE - 1);
1028         data.guest_address = __sme_page_pa(pages[0]) + offset;
1029         data.guest_len = params.guest_len;
1030
1031         blob = psp_copy_user_blob(params.trans_uaddr, params.trans_len);
1032         if (IS_ERR(blob)) {
1033                 ret = PTR_ERR(blob);
1034                 goto e_unpin_memory;
1035         }
1036
1037         data.trans_address = __psp_pa(blob);
1038         data.trans_len = params.trans_len;
1039
1040         hdr = psp_copy_user_blob(params.hdr_uaddr, params.hdr_len);
1041         if (IS_ERR(hdr)) {
1042                 ret = PTR_ERR(hdr);
1043                 goto e_free_blob;
1044         }
1045         data.hdr_address = __psp_pa(hdr);
1046         data.hdr_len = params.hdr_len;
1047
1048         data.handle = sev->handle;
1049         ret = sev_issue_cmd(kvm, SEV_CMD_LAUNCH_UPDATE_SECRET, &data, &argp->error);
1050
1051         kfree(hdr);
1052
1053 e_free_blob:
1054         kfree(blob);
1055 e_unpin_memory:
1056         /* content of memory is updated, mark pages dirty */
1057         for (i = 0; i < n; i++) {
1058                 set_page_dirty_lock(pages[i]);
1059                 mark_page_accessed(pages[i]);
1060         }
1061         sev_unpin_memory(kvm, pages, n);
1062         return ret;
1063 }
1064
1065 static int sev_get_attestation_report(struct kvm *kvm, struct kvm_sev_cmd *argp)
1066 {
1067         void __user *report = (void __user *)(uintptr_t)argp->data;
1068         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1069         struct sev_data_attestation_report data;
1070         struct kvm_sev_attestation_report params;
1071         void __user *p;
1072         void *blob = NULL;
1073         int ret;
1074
1075         if (!sev_guest(kvm))
1076                 return -ENOTTY;
1077
1078         if (copy_from_user(&params, (void __user *)(uintptr_t)argp->data, sizeof(params)))
1079                 return -EFAULT;
1080
1081         memset(&data, 0, sizeof(data));
1082
1083         /* User wants to query the blob length */
1084         if (!params.len)
1085                 goto cmd;
1086
1087         p = (void __user *)(uintptr_t)params.uaddr;
1088         if (p) {
1089                 if (params.len > SEV_FW_BLOB_MAX_SIZE)
1090                         return -EINVAL;
1091
1092                 blob = kmalloc(params.len, GFP_KERNEL_ACCOUNT);
1093                 if (!blob)
1094                         return -ENOMEM;
1095
1096                 data.address = __psp_pa(blob);
1097                 data.len = params.len;
1098                 memcpy(data.mnonce, params.mnonce, sizeof(params.mnonce));
1099         }
1100 cmd:
1101         data.handle = sev->handle;
1102         ret = sev_issue_cmd(kvm, SEV_CMD_ATTESTATION_REPORT, &data, &argp->error);
1103         /*
1104          * If we query the session length, FW responded with expected data.
1105          */
1106         if (!params.len)
1107                 goto done;
1108
1109         if (ret)
1110                 goto e_free_blob;
1111
1112         if (blob) {
1113                 if (copy_to_user(p, blob, params.len))
1114                         ret = -EFAULT;
1115         }
1116
1117 done:
1118         params.len = data.len;
1119         if (copy_to_user(report, &params, sizeof(params)))
1120                 ret = -EFAULT;
1121 e_free_blob:
1122         kfree(blob);
1123         return ret;
1124 }
1125
1126 /* Userspace wants to query session length. */
1127 static int
1128 __sev_send_start_query_session_length(struct kvm *kvm, struct kvm_sev_cmd *argp,
1129                                       struct kvm_sev_send_start *params)
1130 {
1131         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1132         struct sev_data_send_start data;
1133         int ret;
1134
1135         memset(&data, 0, sizeof(data));
1136         data.handle = sev->handle;
1137         ret = sev_issue_cmd(kvm, SEV_CMD_SEND_START, &data, &argp->error);
1138
1139         params->session_len = data.session_len;
1140         if (copy_to_user((void __user *)(uintptr_t)argp->data, params,
1141                                 sizeof(struct kvm_sev_send_start)))
1142                 ret = -EFAULT;
1143
1144         return ret;
1145 }
1146
1147 static int sev_send_start(struct kvm *kvm, struct kvm_sev_cmd *argp)
1148 {
1149         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1150         struct sev_data_send_start data;
1151         struct kvm_sev_send_start params;
1152         void *amd_certs, *session_data;
1153         void *pdh_cert, *plat_certs;
1154         int ret;
1155
1156         if (!sev_guest(kvm))
1157                 return -ENOTTY;
1158
1159         if (copy_from_user(&params, (void __user *)(uintptr_t)argp->data,
1160                                 sizeof(struct kvm_sev_send_start)))
1161                 return -EFAULT;
1162
1163         /* if session_len is zero, userspace wants to query the session length */
1164         if (!params.session_len)
1165                 return __sev_send_start_query_session_length(kvm, argp,
1166                                 &params);
1167
1168         /* some sanity checks */
1169         if (!params.pdh_cert_uaddr || !params.pdh_cert_len ||
1170             !params.session_uaddr || params.session_len > SEV_FW_BLOB_MAX_SIZE)
1171                 return -EINVAL;
1172
1173         /* allocate the memory to hold the session data blob */
1174         session_data = kmalloc(params.session_len, GFP_KERNEL_ACCOUNT);
1175         if (!session_data)
1176                 return -ENOMEM;
1177
1178         /* copy the certificate blobs from userspace */
1179         pdh_cert = psp_copy_user_blob(params.pdh_cert_uaddr,
1180                                 params.pdh_cert_len);
1181         if (IS_ERR(pdh_cert)) {
1182                 ret = PTR_ERR(pdh_cert);
1183                 goto e_free_session;
1184         }
1185
1186         plat_certs = psp_copy_user_blob(params.plat_certs_uaddr,
1187                                 params.plat_certs_len);
1188         if (IS_ERR(plat_certs)) {
1189                 ret = PTR_ERR(plat_certs);
1190                 goto e_free_pdh;
1191         }
1192
1193         amd_certs = psp_copy_user_blob(params.amd_certs_uaddr,
1194                                 params.amd_certs_len);
1195         if (IS_ERR(amd_certs)) {
1196                 ret = PTR_ERR(amd_certs);
1197                 goto e_free_plat_cert;
1198         }
1199
1200         /* populate the FW SEND_START field with system physical address */
1201         memset(&data, 0, sizeof(data));
1202         data.pdh_cert_address = __psp_pa(pdh_cert);
1203         data.pdh_cert_len = params.pdh_cert_len;
1204         data.plat_certs_address = __psp_pa(plat_certs);
1205         data.plat_certs_len = params.plat_certs_len;
1206         data.amd_certs_address = __psp_pa(amd_certs);
1207         data.amd_certs_len = params.amd_certs_len;
1208         data.session_address = __psp_pa(session_data);
1209         data.session_len = params.session_len;
1210         data.handle = sev->handle;
1211
1212         ret = sev_issue_cmd(kvm, SEV_CMD_SEND_START, &data, &argp->error);
1213
1214         if (!ret && copy_to_user((void __user *)(uintptr_t)params.session_uaddr,
1215                         session_data, params.session_len)) {
1216                 ret = -EFAULT;
1217                 goto e_free_amd_cert;
1218         }
1219
1220         params.policy = data.policy;
1221         params.session_len = data.session_len;
1222         if (copy_to_user((void __user *)(uintptr_t)argp->data, &params,
1223                                 sizeof(struct kvm_sev_send_start)))
1224                 ret = -EFAULT;
1225
1226 e_free_amd_cert:
1227         kfree(amd_certs);
1228 e_free_plat_cert:
1229         kfree(plat_certs);
1230 e_free_pdh:
1231         kfree(pdh_cert);
1232 e_free_session:
1233         kfree(session_data);
1234         return ret;
1235 }
1236
1237 /* Userspace wants to query either header or trans length. */
1238 static int
1239 __sev_send_update_data_query_lengths(struct kvm *kvm, struct kvm_sev_cmd *argp,
1240                                      struct kvm_sev_send_update_data *params)
1241 {
1242         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1243         struct sev_data_send_update_data data;
1244         int ret;
1245
1246         memset(&data, 0, sizeof(data));
1247         data.handle = sev->handle;
1248         ret = sev_issue_cmd(kvm, SEV_CMD_SEND_UPDATE_DATA, &data, &argp->error);
1249
1250         params->hdr_len = data.hdr_len;
1251         params->trans_len = data.trans_len;
1252
1253         if (copy_to_user((void __user *)(uintptr_t)argp->data, params,
1254                          sizeof(struct kvm_sev_send_update_data)))
1255                 ret = -EFAULT;
1256
1257         return ret;
1258 }
1259
1260 static int sev_send_update_data(struct kvm *kvm, struct kvm_sev_cmd *argp)
1261 {
1262         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1263         struct sev_data_send_update_data data;
1264         struct kvm_sev_send_update_data params;
1265         void *hdr, *trans_data;
1266         struct page **guest_page;
1267         unsigned long n;
1268         int ret, offset;
1269
1270         if (!sev_guest(kvm))
1271                 return -ENOTTY;
1272
1273         if (copy_from_user(&params, (void __user *)(uintptr_t)argp->data,
1274                         sizeof(struct kvm_sev_send_update_data)))
1275                 return -EFAULT;
1276
1277         /* userspace wants to query either header or trans length */
1278         if (!params.trans_len || !params.hdr_len)
1279                 return __sev_send_update_data_query_lengths(kvm, argp, &params);
1280
1281         if (!params.trans_uaddr || !params.guest_uaddr ||
1282             !params.guest_len || !params.hdr_uaddr)
1283                 return -EINVAL;
1284
1285         /* Check if we are crossing the page boundary */
1286         offset = params.guest_uaddr & (PAGE_SIZE - 1);
1287         if ((params.guest_len + offset > PAGE_SIZE))
1288                 return -EINVAL;
1289
1290         /* Pin guest memory */
1291         guest_page = sev_pin_memory(kvm, params.guest_uaddr & PAGE_MASK,
1292                                     PAGE_SIZE, &n, 0);
1293         if (IS_ERR(guest_page))
1294                 return PTR_ERR(guest_page);
1295
1296         /* allocate memory for header and transport buffer */
1297         ret = -ENOMEM;
1298         hdr = kmalloc(params.hdr_len, GFP_KERNEL_ACCOUNT);
1299         if (!hdr)
1300                 goto e_unpin;
1301
1302         trans_data = kmalloc(params.trans_len, GFP_KERNEL_ACCOUNT);
1303         if (!trans_data)
1304                 goto e_free_hdr;
1305
1306         memset(&data, 0, sizeof(data));
1307         data.hdr_address = __psp_pa(hdr);
1308         data.hdr_len = params.hdr_len;
1309         data.trans_address = __psp_pa(trans_data);
1310         data.trans_len = params.trans_len;
1311
1312         /* The SEND_UPDATE_DATA command requires C-bit to be always set. */
1313         data.guest_address = (page_to_pfn(guest_page[0]) << PAGE_SHIFT) + offset;
1314         data.guest_address |= sev_me_mask;
1315         data.guest_len = params.guest_len;
1316         data.handle = sev->handle;
1317
1318         ret = sev_issue_cmd(kvm, SEV_CMD_SEND_UPDATE_DATA, &data, &argp->error);
1319
1320         if (ret)
1321                 goto e_free_trans_data;
1322
1323         /* copy transport buffer to user space */
1324         if (copy_to_user((void __user *)(uintptr_t)params.trans_uaddr,
1325                          trans_data, params.trans_len)) {
1326                 ret = -EFAULT;
1327                 goto e_free_trans_data;
1328         }
1329
1330         /* Copy packet header to userspace. */
1331         if (copy_to_user((void __user *)(uintptr_t)params.hdr_uaddr, hdr,
1332                          params.hdr_len))
1333                 ret = -EFAULT;
1334
1335 e_free_trans_data:
1336         kfree(trans_data);
1337 e_free_hdr:
1338         kfree(hdr);
1339 e_unpin:
1340         sev_unpin_memory(kvm, guest_page, n);
1341
1342         return ret;
1343 }
1344
1345 static int sev_send_finish(struct kvm *kvm, struct kvm_sev_cmd *argp)
1346 {
1347         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1348         struct sev_data_send_finish data;
1349
1350         if (!sev_guest(kvm))
1351                 return -ENOTTY;
1352
1353         data.handle = sev->handle;
1354         return sev_issue_cmd(kvm, SEV_CMD_SEND_FINISH, &data, &argp->error);
1355 }
1356
1357 static int sev_send_cancel(struct kvm *kvm, struct kvm_sev_cmd *argp)
1358 {
1359         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1360         struct sev_data_send_cancel data;
1361
1362         if (!sev_guest(kvm))
1363                 return -ENOTTY;
1364
1365         data.handle = sev->handle;
1366         return sev_issue_cmd(kvm, SEV_CMD_SEND_CANCEL, &data, &argp->error);
1367 }
1368
1369 static int sev_receive_start(struct kvm *kvm, struct kvm_sev_cmd *argp)
1370 {
1371         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1372         struct sev_data_receive_start start;
1373         struct kvm_sev_receive_start params;
1374         int *error = &argp->error;
1375         void *session_data;
1376         void *pdh_data;
1377         int ret;
1378
1379         if (!sev_guest(kvm))
1380                 return -ENOTTY;
1381
1382         /* Get parameter from the userspace */
1383         if (copy_from_user(&params, (void __user *)(uintptr_t)argp->data,
1384                         sizeof(struct kvm_sev_receive_start)))
1385                 return -EFAULT;
1386
1387         /* some sanity checks */
1388         if (!params.pdh_uaddr || !params.pdh_len ||
1389             !params.session_uaddr || !params.session_len)
1390                 return -EINVAL;
1391
1392         pdh_data = psp_copy_user_blob(params.pdh_uaddr, params.pdh_len);
1393         if (IS_ERR(pdh_data))
1394                 return PTR_ERR(pdh_data);
1395
1396         session_data = psp_copy_user_blob(params.session_uaddr,
1397                         params.session_len);
1398         if (IS_ERR(session_data)) {
1399                 ret = PTR_ERR(session_data);
1400                 goto e_free_pdh;
1401         }
1402
1403         memset(&start, 0, sizeof(start));
1404         start.handle = params.handle;
1405         start.policy = params.policy;
1406         start.pdh_cert_address = __psp_pa(pdh_data);
1407         start.pdh_cert_len = params.pdh_len;
1408         start.session_address = __psp_pa(session_data);
1409         start.session_len = params.session_len;
1410
1411         /* create memory encryption context */
1412         ret = __sev_issue_cmd(argp->sev_fd, SEV_CMD_RECEIVE_START, &start,
1413                                 error);
1414         if (ret)
1415                 goto e_free_session;
1416
1417         /* Bind ASID to this guest */
1418         ret = sev_bind_asid(kvm, start.handle, error);
1419         if (ret) {
1420                 sev_decommission(start.handle);
1421                 goto e_free_session;
1422         }
1423
1424         params.handle = start.handle;
1425         if (copy_to_user((void __user *)(uintptr_t)argp->data,
1426                          &params, sizeof(struct kvm_sev_receive_start))) {
1427                 ret = -EFAULT;
1428                 sev_unbind_asid(kvm, start.handle);
1429                 goto e_free_session;
1430         }
1431
1432         sev->handle = start.handle;
1433         sev->fd = argp->sev_fd;
1434
1435 e_free_session:
1436         kfree(session_data);
1437 e_free_pdh:
1438         kfree(pdh_data);
1439
1440         return ret;
1441 }
1442
1443 static int sev_receive_update_data(struct kvm *kvm, struct kvm_sev_cmd *argp)
1444 {
1445         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1446         struct kvm_sev_receive_update_data params;
1447         struct sev_data_receive_update_data data;
1448         void *hdr = NULL, *trans = NULL;
1449         struct page **guest_page;
1450         unsigned long n;
1451         int ret, offset;
1452
1453         if (!sev_guest(kvm))
1454                 return -EINVAL;
1455
1456         if (copy_from_user(&params, (void __user *)(uintptr_t)argp->data,
1457                         sizeof(struct kvm_sev_receive_update_data)))
1458                 return -EFAULT;
1459
1460         if (!params.hdr_uaddr || !params.hdr_len ||
1461             !params.guest_uaddr || !params.guest_len ||
1462             !params.trans_uaddr || !params.trans_len)
1463                 return -EINVAL;
1464
1465         /* Check if we are crossing the page boundary */
1466         offset = params.guest_uaddr & (PAGE_SIZE - 1);
1467         if ((params.guest_len + offset > PAGE_SIZE))
1468                 return -EINVAL;
1469
1470         hdr = psp_copy_user_blob(params.hdr_uaddr, params.hdr_len);
1471         if (IS_ERR(hdr))
1472                 return PTR_ERR(hdr);
1473
1474         trans = psp_copy_user_blob(params.trans_uaddr, params.trans_len);
1475         if (IS_ERR(trans)) {
1476                 ret = PTR_ERR(trans);
1477                 goto e_free_hdr;
1478         }
1479
1480         memset(&data, 0, sizeof(data));
1481         data.hdr_address = __psp_pa(hdr);
1482         data.hdr_len = params.hdr_len;
1483         data.trans_address = __psp_pa(trans);
1484         data.trans_len = params.trans_len;
1485
1486         /* Pin guest memory */
1487         guest_page = sev_pin_memory(kvm, params.guest_uaddr & PAGE_MASK,
1488                                     PAGE_SIZE, &n, 1);
1489         if (IS_ERR(guest_page)) {
1490                 ret = PTR_ERR(guest_page);
1491                 goto e_free_trans;
1492         }
1493
1494         /*
1495          * Flush (on non-coherent CPUs) before RECEIVE_UPDATE_DATA, the PSP
1496          * encrypts the written data with the guest's key, and the cache may
1497          * contain dirty, unencrypted data.
1498          */
1499         sev_clflush_pages(guest_page, n);
1500
1501         /* The RECEIVE_UPDATE_DATA command requires C-bit to be always set. */
1502         data.guest_address = (page_to_pfn(guest_page[0]) << PAGE_SHIFT) + offset;
1503         data.guest_address |= sev_me_mask;
1504         data.guest_len = params.guest_len;
1505         data.handle = sev->handle;
1506
1507         ret = sev_issue_cmd(kvm, SEV_CMD_RECEIVE_UPDATE_DATA, &data,
1508                                 &argp->error);
1509
1510         sev_unpin_memory(kvm, guest_page, n);
1511
1512 e_free_trans:
1513         kfree(trans);
1514 e_free_hdr:
1515         kfree(hdr);
1516
1517         return ret;
1518 }
1519
1520 static int sev_receive_finish(struct kvm *kvm, struct kvm_sev_cmd *argp)
1521 {
1522         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1523         struct sev_data_receive_finish data;
1524
1525         if (!sev_guest(kvm))
1526                 return -ENOTTY;
1527
1528         data.handle = sev->handle;
1529         return sev_issue_cmd(kvm, SEV_CMD_RECEIVE_FINISH, &data, &argp->error);
1530 }
1531
1532 static bool is_cmd_allowed_from_mirror(u32 cmd_id)
1533 {
1534         /*
1535          * Allow mirrors VM to call KVM_SEV_LAUNCH_UPDATE_VMSA to enable SEV-ES
1536          * active mirror VMs. Also allow the debugging and status commands.
1537          */
1538         if (cmd_id == KVM_SEV_LAUNCH_UPDATE_VMSA ||
1539             cmd_id == KVM_SEV_GUEST_STATUS || cmd_id == KVM_SEV_DBG_DECRYPT ||
1540             cmd_id == KVM_SEV_DBG_ENCRYPT)
1541                 return true;
1542
1543         return false;
1544 }
1545
1546 static int sev_lock_two_vms(struct kvm *dst_kvm, struct kvm *src_kvm)
1547 {
1548         struct kvm_sev_info *dst_sev = &to_kvm_svm(dst_kvm)->sev_info;
1549         struct kvm_sev_info *src_sev = &to_kvm_svm(src_kvm)->sev_info;
1550         int r = -EBUSY;
1551
1552         if (dst_kvm == src_kvm)
1553                 return -EINVAL;
1554
1555         /*
1556          * Bail if these VMs are already involved in a migration to avoid
1557          * deadlock between two VMs trying to migrate to/from each other.
1558          */
1559         if (atomic_cmpxchg_acquire(&dst_sev->migration_in_progress, 0, 1))
1560                 return -EBUSY;
1561
1562         if (atomic_cmpxchg_acquire(&src_sev->migration_in_progress, 0, 1))
1563                 goto release_dst;
1564
1565         r = -EINTR;
1566         if (mutex_lock_killable(&dst_kvm->lock))
1567                 goto release_src;
1568         if (mutex_lock_killable(&src_kvm->lock))
1569                 goto unlock_dst;
1570         return 0;
1571
1572 unlock_dst:
1573         mutex_unlock(&dst_kvm->lock);
1574 release_src:
1575         atomic_set_release(&src_sev->migration_in_progress, 0);
1576 release_dst:
1577         atomic_set_release(&dst_sev->migration_in_progress, 0);
1578         return r;
1579 }
1580
1581 static void sev_unlock_two_vms(struct kvm *dst_kvm, struct kvm *src_kvm)
1582 {
1583         struct kvm_sev_info *dst_sev = &to_kvm_svm(dst_kvm)->sev_info;
1584         struct kvm_sev_info *src_sev = &to_kvm_svm(src_kvm)->sev_info;
1585
1586         mutex_unlock(&dst_kvm->lock);
1587         mutex_unlock(&src_kvm->lock);
1588         atomic_set_release(&dst_sev->migration_in_progress, 0);
1589         atomic_set_release(&src_sev->migration_in_progress, 0);
1590 }
1591
1592
1593 static int sev_lock_vcpus_for_migration(struct kvm *kvm)
1594 {
1595         struct kvm_vcpu *vcpu;
1596         int i, j;
1597
1598         kvm_for_each_vcpu(i, vcpu, kvm) {
1599                 if (mutex_lock_killable(&vcpu->mutex))
1600                         goto out_unlock;
1601         }
1602
1603         return 0;
1604
1605 out_unlock:
1606         kvm_for_each_vcpu(j, vcpu, kvm) {
1607                 if (i == j)
1608                         break;
1609
1610                 mutex_unlock(&vcpu->mutex);
1611         }
1612         return -EINTR;
1613 }
1614
1615 static void sev_unlock_vcpus_for_migration(struct kvm *kvm)
1616 {
1617         struct kvm_vcpu *vcpu;
1618         int i;
1619
1620         kvm_for_each_vcpu(i, vcpu, kvm) {
1621                 mutex_unlock(&vcpu->mutex);
1622         }
1623 }
1624
1625 static void sev_migrate_from(struct kvm_sev_info *dst,
1626                               struct kvm_sev_info *src)
1627 {
1628         dst->active = true;
1629         dst->asid = src->asid;
1630         dst->handle = src->handle;
1631         dst->pages_locked = src->pages_locked;
1632         dst->enc_context_owner = src->enc_context_owner;
1633
1634         src->asid = 0;
1635         src->active = false;
1636         src->handle = 0;
1637         src->pages_locked = 0;
1638         src->enc_context_owner = NULL;
1639
1640         list_cut_before(&dst->regions_list, &src->regions_list, &src->regions_list);
1641 }
1642
1643 static int sev_es_migrate_from(struct kvm *dst, struct kvm *src)
1644 {
1645         int i;
1646         struct kvm_vcpu *dst_vcpu, *src_vcpu;
1647         struct vcpu_svm *dst_svm, *src_svm;
1648
1649         if (atomic_read(&src->online_vcpus) != atomic_read(&dst->online_vcpus))
1650                 return -EINVAL;
1651
1652         kvm_for_each_vcpu(i, src_vcpu, src) {
1653                 if (!src_vcpu->arch.guest_state_protected)
1654                         return -EINVAL;
1655         }
1656
1657         kvm_for_each_vcpu(i, src_vcpu, src) {
1658                 src_svm = to_svm(src_vcpu);
1659                 dst_vcpu = kvm_get_vcpu(dst, i);
1660                 dst_svm = to_svm(dst_vcpu);
1661
1662                 /*
1663                  * Transfer VMSA and GHCB state to the destination.  Nullify and
1664                  * clear source fields as appropriate, the state now belongs to
1665                  * the destination.
1666                  */
1667                 memcpy(&dst_svm->sev_es, &src_svm->sev_es, sizeof(src_svm->sev_es));
1668                 dst_svm->vmcb->control.ghcb_gpa = src_svm->vmcb->control.ghcb_gpa;
1669                 dst_svm->vmcb->control.vmsa_pa = src_svm->vmcb->control.vmsa_pa;
1670                 dst_vcpu->arch.guest_state_protected = true;
1671
1672                 memset(&src_svm->sev_es, 0, sizeof(src_svm->sev_es));
1673                 src_svm->vmcb->control.ghcb_gpa = INVALID_PAGE;
1674                 src_svm->vmcb->control.vmsa_pa = INVALID_PAGE;
1675                 src_vcpu->arch.guest_state_protected = false;
1676         }
1677         to_kvm_svm(src)->sev_info.es_active = false;
1678         to_kvm_svm(dst)->sev_info.es_active = true;
1679
1680         return 0;
1681 }
1682
1683 int svm_vm_migrate_from(struct kvm *kvm, unsigned int source_fd)
1684 {
1685         struct kvm_sev_info *dst_sev = &to_kvm_svm(kvm)->sev_info;
1686         struct kvm_sev_info *src_sev, *cg_cleanup_sev;
1687         struct file *source_kvm_file;
1688         struct kvm *source_kvm;
1689         bool charged = false;
1690         int ret;
1691
1692         source_kvm_file = fget(source_fd);
1693         if (!file_is_kvm(source_kvm_file)) {
1694                 ret = -EBADF;
1695                 goto out_fput;
1696         }
1697
1698         source_kvm = source_kvm_file->private_data;
1699         ret = sev_lock_two_vms(kvm, source_kvm);
1700         if (ret)
1701                 goto out_fput;
1702
1703         if (sev_guest(kvm) || !sev_guest(source_kvm)) {
1704                 ret = -EINVAL;
1705                 goto out_unlock;
1706         }
1707
1708         src_sev = &to_kvm_svm(source_kvm)->sev_info;
1709
1710         /*
1711          * VMs mirroring src's encryption context rely on it to keep the
1712          * ASID allocated, but below we are clearing src_sev->asid.
1713          */
1714         if (src_sev->num_mirrored_vms) {
1715                 ret = -EBUSY;
1716                 goto out_unlock;
1717         }
1718
1719         dst_sev->misc_cg = get_current_misc_cg();
1720         cg_cleanup_sev = dst_sev;
1721         if (dst_sev->misc_cg != src_sev->misc_cg) {
1722                 ret = sev_misc_cg_try_charge(dst_sev);
1723                 if (ret)
1724                         goto out_dst_cgroup;
1725                 charged = true;
1726         }
1727
1728         ret = sev_lock_vcpus_for_migration(kvm);
1729         if (ret)
1730                 goto out_dst_cgroup;
1731         ret = sev_lock_vcpus_for_migration(source_kvm);
1732         if (ret)
1733                 goto out_dst_vcpu;
1734
1735         if (sev_es_guest(source_kvm)) {
1736                 ret = sev_es_migrate_from(kvm, source_kvm);
1737                 if (ret)
1738                         goto out_source_vcpu;
1739         }
1740         sev_migrate_from(dst_sev, src_sev);
1741         kvm_vm_dead(source_kvm);
1742         cg_cleanup_sev = src_sev;
1743         ret = 0;
1744
1745 out_source_vcpu:
1746         sev_unlock_vcpus_for_migration(source_kvm);
1747 out_dst_vcpu:
1748         sev_unlock_vcpus_for_migration(kvm);
1749 out_dst_cgroup:
1750         /* Operates on the source on success, on the destination on failure.  */
1751         if (charged)
1752                 sev_misc_cg_uncharge(cg_cleanup_sev);
1753         put_misc_cg(cg_cleanup_sev->misc_cg);
1754         cg_cleanup_sev->misc_cg = NULL;
1755 out_unlock:
1756         sev_unlock_two_vms(kvm, source_kvm);
1757 out_fput:
1758         if (source_kvm_file)
1759                 fput(source_kvm_file);
1760         return ret;
1761 }
1762
1763 int svm_mem_enc_op(struct kvm *kvm, void __user *argp)
1764 {
1765         struct kvm_sev_cmd sev_cmd;
1766         int r;
1767
1768         if (!sev_enabled)
1769                 return -ENOTTY;
1770
1771         if (!argp)
1772                 return 0;
1773
1774         if (copy_from_user(&sev_cmd, argp, sizeof(struct kvm_sev_cmd)))
1775                 return -EFAULT;
1776
1777         mutex_lock(&kvm->lock);
1778
1779         /* Only the enc_context_owner handles some memory enc operations. */
1780         if (is_mirroring_enc_context(kvm) &&
1781             !is_cmd_allowed_from_mirror(sev_cmd.id)) {
1782                 r = -EINVAL;
1783                 goto out;
1784         }
1785
1786         switch (sev_cmd.id) {
1787         case KVM_SEV_ES_INIT:
1788                 if (!sev_es_enabled) {
1789                         r = -ENOTTY;
1790                         goto out;
1791                 }
1792                 fallthrough;
1793         case KVM_SEV_INIT:
1794                 r = sev_guest_init(kvm, &sev_cmd);
1795                 break;
1796         case KVM_SEV_LAUNCH_START:
1797                 r = sev_launch_start(kvm, &sev_cmd);
1798                 break;
1799         case KVM_SEV_LAUNCH_UPDATE_DATA:
1800                 r = sev_launch_update_data(kvm, &sev_cmd);
1801                 break;
1802         case KVM_SEV_LAUNCH_UPDATE_VMSA:
1803                 r = sev_launch_update_vmsa(kvm, &sev_cmd);
1804                 break;
1805         case KVM_SEV_LAUNCH_MEASURE:
1806                 r = sev_launch_measure(kvm, &sev_cmd);
1807                 break;
1808         case KVM_SEV_LAUNCH_FINISH:
1809                 r = sev_launch_finish(kvm, &sev_cmd);
1810                 break;
1811         case KVM_SEV_GUEST_STATUS:
1812                 r = sev_guest_status(kvm, &sev_cmd);
1813                 break;
1814         case KVM_SEV_DBG_DECRYPT:
1815                 r = sev_dbg_crypt(kvm, &sev_cmd, true);
1816                 break;
1817         case KVM_SEV_DBG_ENCRYPT:
1818                 r = sev_dbg_crypt(kvm, &sev_cmd, false);
1819                 break;
1820         case KVM_SEV_LAUNCH_SECRET:
1821                 r = sev_launch_secret(kvm, &sev_cmd);
1822                 break;
1823         case KVM_SEV_GET_ATTESTATION_REPORT:
1824                 r = sev_get_attestation_report(kvm, &sev_cmd);
1825                 break;
1826         case KVM_SEV_SEND_START:
1827                 r = sev_send_start(kvm, &sev_cmd);
1828                 break;
1829         case KVM_SEV_SEND_UPDATE_DATA:
1830                 r = sev_send_update_data(kvm, &sev_cmd);
1831                 break;
1832         case KVM_SEV_SEND_FINISH:
1833                 r = sev_send_finish(kvm, &sev_cmd);
1834                 break;
1835         case KVM_SEV_SEND_CANCEL:
1836                 r = sev_send_cancel(kvm, &sev_cmd);
1837                 break;
1838         case KVM_SEV_RECEIVE_START:
1839                 r = sev_receive_start(kvm, &sev_cmd);
1840                 break;
1841         case KVM_SEV_RECEIVE_UPDATE_DATA:
1842                 r = sev_receive_update_data(kvm, &sev_cmd);
1843                 break;
1844         case KVM_SEV_RECEIVE_FINISH:
1845                 r = sev_receive_finish(kvm, &sev_cmd);
1846                 break;
1847         default:
1848                 r = -EINVAL;
1849                 goto out;
1850         }
1851
1852         if (copy_to_user(argp, &sev_cmd, sizeof(struct kvm_sev_cmd)))
1853                 r = -EFAULT;
1854
1855 out:
1856         mutex_unlock(&kvm->lock);
1857         return r;
1858 }
1859
1860 int svm_register_enc_region(struct kvm *kvm,
1861                             struct kvm_enc_region *range)
1862 {
1863         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1864         struct enc_region *region;
1865         int ret = 0;
1866
1867         if (!sev_guest(kvm))
1868                 return -ENOTTY;
1869
1870         /* If kvm is mirroring encryption context it isn't responsible for it */
1871         if (is_mirroring_enc_context(kvm))
1872                 return -EINVAL;
1873
1874         if (range->addr > ULONG_MAX || range->size > ULONG_MAX)
1875                 return -EINVAL;
1876
1877         region = kzalloc(sizeof(*region), GFP_KERNEL_ACCOUNT);
1878         if (!region)
1879                 return -ENOMEM;
1880
1881         mutex_lock(&kvm->lock);
1882         region->pages = sev_pin_memory(kvm, range->addr, range->size, &region->npages, 1);
1883         if (IS_ERR(region->pages)) {
1884                 ret = PTR_ERR(region->pages);
1885                 mutex_unlock(&kvm->lock);
1886                 goto e_free;
1887         }
1888
1889         region->uaddr = range->addr;
1890         region->size = range->size;
1891
1892         list_add_tail(&region->list, &sev->regions_list);
1893         mutex_unlock(&kvm->lock);
1894
1895         /*
1896          * The guest may change the memory encryption attribute from C=0 -> C=1
1897          * or vice versa for this memory range. Lets make sure caches are
1898          * flushed to ensure that guest data gets written into memory with
1899          * correct C-bit.
1900          */
1901         sev_clflush_pages(region->pages, region->npages);
1902
1903         return ret;
1904
1905 e_free:
1906         kfree(region);
1907         return ret;
1908 }
1909
1910 static struct enc_region *
1911 find_enc_region(struct kvm *kvm, struct kvm_enc_region *range)
1912 {
1913         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
1914         struct list_head *head = &sev->regions_list;
1915         struct enc_region *i;
1916
1917         list_for_each_entry(i, head, list) {
1918                 if (i->uaddr == range->addr &&
1919                     i->size == range->size)
1920                         return i;
1921         }
1922
1923         return NULL;
1924 }
1925
1926 static void __unregister_enc_region_locked(struct kvm *kvm,
1927                                            struct enc_region *region)
1928 {
1929         sev_unpin_memory(kvm, region->pages, region->npages);
1930         list_del(&region->list);
1931         kfree(region);
1932 }
1933
1934 int svm_unregister_enc_region(struct kvm *kvm,
1935                               struct kvm_enc_region *range)
1936 {
1937         struct enc_region *region;
1938         int ret;
1939
1940         /* If kvm is mirroring encryption context it isn't responsible for it */
1941         if (is_mirroring_enc_context(kvm))
1942                 return -EINVAL;
1943
1944         mutex_lock(&kvm->lock);
1945
1946         if (!sev_guest(kvm)) {
1947                 ret = -ENOTTY;
1948                 goto failed;
1949         }
1950
1951         region = find_enc_region(kvm, range);
1952         if (!region) {
1953                 ret = -EINVAL;
1954                 goto failed;
1955         }
1956
1957         /*
1958          * Ensure that all guest tagged cache entries are flushed before
1959          * releasing the pages back to the system for use. CLFLUSH will
1960          * not do this, so issue a WBINVD.
1961          */
1962         wbinvd_on_all_cpus();
1963
1964         __unregister_enc_region_locked(kvm, region);
1965
1966         mutex_unlock(&kvm->lock);
1967         return 0;
1968
1969 failed:
1970         mutex_unlock(&kvm->lock);
1971         return ret;
1972 }
1973
1974 int svm_vm_copy_asid_from(struct kvm *kvm, unsigned int source_fd)
1975 {
1976         struct file *source_kvm_file;
1977         struct kvm *source_kvm;
1978         struct kvm_sev_info *source_sev, *mirror_sev;
1979         int ret;
1980
1981         source_kvm_file = fget(source_fd);
1982         if (!file_is_kvm(source_kvm_file)) {
1983                 ret = -EBADF;
1984                 goto e_source_fput;
1985         }
1986
1987         source_kvm = source_kvm_file->private_data;
1988         ret = sev_lock_two_vms(kvm, source_kvm);
1989         if (ret)
1990                 goto e_source_fput;
1991
1992         /*
1993          * Mirrors of mirrors should work, but let's not get silly.  Also
1994          * disallow out-of-band SEV/SEV-ES init if the target is already an
1995          * SEV guest, or if vCPUs have been created.  KVM relies on vCPUs being
1996          * created after SEV/SEV-ES initialization, e.g. to init intercepts.
1997          */
1998         if (sev_guest(kvm) || !sev_guest(source_kvm) ||
1999             is_mirroring_enc_context(source_kvm) || kvm->created_vcpus) {
2000                 ret = -EINVAL;
2001                 goto e_unlock;
2002         }
2003
2004         /*
2005          * The mirror kvm holds an enc_context_owner ref so its asid can't
2006          * disappear until we're done with it
2007          */
2008         source_sev = &to_kvm_svm(source_kvm)->sev_info;
2009         kvm_get_kvm(source_kvm);
2010         source_sev->num_mirrored_vms++;
2011
2012         /* Set enc_context_owner and copy its encryption context over */
2013         mirror_sev = &to_kvm_svm(kvm)->sev_info;
2014         mirror_sev->enc_context_owner = source_kvm;
2015         mirror_sev->active = true;
2016         mirror_sev->asid = source_sev->asid;
2017         mirror_sev->fd = source_sev->fd;
2018         mirror_sev->es_active = source_sev->es_active;
2019         mirror_sev->handle = source_sev->handle;
2020         INIT_LIST_HEAD(&mirror_sev->regions_list);
2021         ret = 0;
2022
2023         /*
2024          * Do not copy ap_jump_table. Since the mirror does not share the same
2025          * KVM contexts as the original, and they may have different
2026          * memory-views.
2027          */
2028
2029 e_unlock:
2030         sev_unlock_two_vms(kvm, source_kvm);
2031 e_source_fput:
2032         if (source_kvm_file)
2033                 fput(source_kvm_file);
2034         return ret;
2035 }
2036
2037 void sev_vm_destroy(struct kvm *kvm)
2038 {
2039         struct kvm_sev_info *sev = &to_kvm_svm(kvm)->sev_info;
2040         struct list_head *head = &sev->regions_list;
2041         struct list_head *pos, *q;
2042
2043         WARN_ON(sev->num_mirrored_vms);
2044
2045         if (!sev_guest(kvm))
2046                 return;
2047
2048         /* If this is a mirror_kvm release the enc_context_owner and skip sev cleanup */
2049         if (is_mirroring_enc_context(kvm)) {
2050                 struct kvm *owner_kvm = sev->enc_context_owner;
2051                 struct kvm_sev_info *owner_sev = &to_kvm_svm(owner_kvm)->sev_info;
2052
2053                 mutex_lock(&owner_kvm->lock);
2054                 if (!WARN_ON(!owner_sev->num_mirrored_vms))
2055                         owner_sev->num_mirrored_vms--;
2056                 mutex_unlock(&owner_kvm->lock);
2057                 kvm_put_kvm(owner_kvm);
2058                 return;
2059         }
2060
2061         /*
2062          * Ensure that all guest tagged cache entries are flushed before
2063          * releasing the pages back to the system for use. CLFLUSH will
2064          * not do this, so issue a WBINVD.
2065          */
2066         wbinvd_on_all_cpus();
2067
2068         /*
2069          * if userspace was terminated before unregistering the memory regions
2070          * then lets unpin all the registered memory.
2071          */
2072         if (!list_empty(head)) {
2073                 list_for_each_safe(pos, q, head) {
2074                         __unregister_enc_region_locked(kvm,
2075                                 list_entry(pos, struct enc_region, list));
2076                         cond_resched();
2077                 }
2078         }
2079
2080         sev_unbind_asid(kvm, sev->handle);
2081         sev_asid_free(sev);
2082 }
2083
2084 void __init sev_set_cpu_caps(void)
2085 {
2086         if (!sev_enabled)
2087                 kvm_cpu_cap_clear(X86_FEATURE_SEV);
2088         if (!sev_es_enabled)
2089                 kvm_cpu_cap_clear(X86_FEATURE_SEV_ES);
2090 }
2091
2092 void __init sev_hardware_setup(void)
2093 {
2094 #ifdef CONFIG_KVM_AMD_SEV
2095         unsigned int eax, ebx, ecx, edx, sev_asid_count, sev_es_asid_count;
2096         bool sev_es_supported = false;
2097         bool sev_supported = false;
2098
2099         if (!sev_enabled || !npt_enabled)
2100                 goto out;
2101
2102         /* Does the CPU support SEV? */
2103         if (!boot_cpu_has(X86_FEATURE_SEV))
2104                 goto out;
2105
2106         /* Retrieve SEV CPUID information */
2107         cpuid(0x8000001f, &eax, &ebx, &ecx, &edx);
2108
2109         /* Set encryption bit location for SEV-ES guests */
2110         sev_enc_bit = ebx & 0x3f;
2111
2112         /* Maximum number of encrypted guests supported simultaneously */
2113         max_sev_asid = ecx;
2114         if (!max_sev_asid)
2115                 goto out;
2116
2117         /* Minimum ASID value that should be used for SEV guest */
2118         min_sev_asid = edx;
2119         sev_me_mask = 1UL << (ebx & 0x3f);
2120
2121         /*
2122          * Initialize SEV ASID bitmaps. Allocate space for ASID 0 in the bitmap,
2123          * even though it's never used, so that the bitmap is indexed by the
2124          * actual ASID.
2125          */
2126         nr_asids = max_sev_asid + 1;
2127         sev_asid_bitmap = bitmap_zalloc(nr_asids, GFP_KERNEL);
2128         if (!sev_asid_bitmap)
2129                 goto out;
2130
2131         sev_reclaim_asid_bitmap = bitmap_zalloc(nr_asids, GFP_KERNEL);
2132         if (!sev_reclaim_asid_bitmap) {
2133                 bitmap_free(sev_asid_bitmap);
2134                 sev_asid_bitmap = NULL;
2135                 goto out;
2136         }
2137
2138         sev_asid_count = max_sev_asid - min_sev_asid + 1;
2139         if (misc_cg_set_capacity(MISC_CG_RES_SEV, sev_asid_count))
2140                 goto out;
2141
2142         pr_info("SEV supported: %u ASIDs\n", sev_asid_count);
2143         sev_supported = true;
2144
2145         /* SEV-ES support requested? */
2146         if (!sev_es_enabled)
2147                 goto out;
2148
2149         /* Does the CPU support SEV-ES? */
2150         if (!boot_cpu_has(X86_FEATURE_SEV_ES))
2151                 goto out;
2152
2153         /* Has the system been allocated ASIDs for SEV-ES? */
2154         if (min_sev_asid == 1)
2155                 goto out;
2156
2157         sev_es_asid_count = min_sev_asid - 1;
2158         if (misc_cg_set_capacity(MISC_CG_RES_SEV_ES, sev_es_asid_count))
2159                 goto out;
2160
2161         pr_info("SEV-ES supported: %u ASIDs\n", sev_es_asid_count);
2162         sev_es_supported = true;
2163
2164 out:
2165         sev_enabled = sev_supported;
2166         sev_es_enabled = sev_es_supported;
2167 #endif
2168 }
2169
2170 void sev_hardware_teardown(void)
2171 {
2172         if (!sev_enabled)
2173                 return;
2174
2175         /* No need to take sev_bitmap_lock, all VMs have been destroyed. */
2176         sev_flush_asids(1, max_sev_asid);
2177
2178         bitmap_free(sev_asid_bitmap);
2179         bitmap_free(sev_reclaim_asid_bitmap);
2180
2181         misc_cg_set_capacity(MISC_CG_RES_SEV, 0);
2182         misc_cg_set_capacity(MISC_CG_RES_SEV_ES, 0);
2183 }
2184
2185 int sev_cpu_init(struct svm_cpu_data *sd)
2186 {
2187         if (!sev_enabled)
2188                 return 0;
2189
2190         sd->sev_vmcbs = kcalloc(nr_asids, sizeof(void *), GFP_KERNEL);
2191         if (!sd->sev_vmcbs)
2192                 return -ENOMEM;
2193
2194         return 0;
2195 }
2196
2197 /*
2198  * Pages used by hardware to hold guest encrypted state must be flushed before
2199  * returning them to the system.
2200  */
2201 static void sev_flush_guest_memory(struct vcpu_svm *svm, void *va,
2202                                    unsigned long len)
2203 {
2204         /*
2205          * If hardware enforced cache coherency for encrypted mappings of the
2206          * same physical page is supported, nothing to do.
2207          */
2208         if (boot_cpu_has(X86_FEATURE_SME_COHERENT))
2209                 return;
2210
2211         /*
2212          * If the VM Page Flush MSR is supported, use it to flush the page
2213          * (using the page virtual address and the guest ASID).
2214          */
2215         if (boot_cpu_has(X86_FEATURE_VM_PAGE_FLUSH)) {
2216                 struct kvm_sev_info *sev;
2217                 unsigned long va_start;
2218                 u64 start, stop;
2219
2220                 /* Align start and stop to page boundaries. */
2221                 va_start = (unsigned long)va;
2222                 start = (u64)va_start & PAGE_MASK;
2223                 stop = PAGE_ALIGN((u64)va_start + len);
2224
2225                 if (start < stop) {
2226                         sev = &to_kvm_svm(svm->vcpu.kvm)->sev_info;
2227
2228                         while (start < stop) {
2229                                 wrmsrl(MSR_AMD64_VM_PAGE_FLUSH,
2230                                        start | sev->asid);
2231
2232                                 start += PAGE_SIZE;
2233                         }
2234
2235                         return;
2236                 }
2237
2238                 WARN(1, "Address overflow, using WBINVD\n");
2239         }
2240
2241         /*
2242          * Hardware should always have one of the above features,
2243          * but if not, use WBINVD and issue a warning.
2244          */
2245         WARN_ONCE(1, "Using WBINVD to flush guest memory\n");
2246         wbinvd_on_all_cpus();
2247 }
2248
2249 void sev_free_vcpu(struct kvm_vcpu *vcpu)
2250 {
2251         struct vcpu_svm *svm;
2252
2253         if (!sev_es_guest(vcpu->kvm))
2254                 return;
2255
2256         svm = to_svm(vcpu);
2257
2258         if (vcpu->arch.guest_state_protected)
2259                 sev_flush_guest_memory(svm, svm->sev_es.vmsa, PAGE_SIZE);
2260         __free_page(virt_to_page(svm->sev_es.vmsa));
2261
2262         if (svm->sev_es.ghcb_sa_free)
2263                 kvfree(svm->sev_es.ghcb_sa);
2264 }
2265
2266 static void dump_ghcb(struct vcpu_svm *svm)
2267 {
2268         struct ghcb *ghcb = svm->sev_es.ghcb;
2269         unsigned int nbits;
2270
2271         /* Re-use the dump_invalid_vmcb module parameter */
2272         if (!dump_invalid_vmcb) {
2273                 pr_warn_ratelimited("set kvm_amd.dump_invalid_vmcb=1 to dump internal KVM state.\n");
2274                 return;
2275         }
2276
2277         nbits = sizeof(ghcb->save.valid_bitmap) * 8;
2278
2279         pr_err("GHCB (GPA=%016llx):\n", svm->vmcb->control.ghcb_gpa);
2280         pr_err("%-20s%016llx is_valid: %u\n", "sw_exit_code",
2281                ghcb->save.sw_exit_code, ghcb_sw_exit_code_is_valid(ghcb));
2282         pr_err("%-20s%016llx is_valid: %u\n", "sw_exit_info_1",
2283                ghcb->save.sw_exit_info_1, ghcb_sw_exit_info_1_is_valid(ghcb));
2284         pr_err("%-20s%016llx is_valid: %u\n", "sw_exit_info_2",
2285                ghcb->save.sw_exit_info_2, ghcb_sw_exit_info_2_is_valid(ghcb));
2286         pr_err("%-20s%016llx is_valid: %u\n", "sw_scratch",
2287                ghcb->save.sw_scratch, ghcb_sw_scratch_is_valid(ghcb));
2288         pr_err("%-20s%*pb\n", "valid_bitmap", nbits, ghcb->save.valid_bitmap);
2289 }
2290
2291 static void sev_es_sync_to_ghcb(struct vcpu_svm *svm)
2292 {
2293         struct kvm_vcpu *vcpu = &svm->vcpu;
2294         struct ghcb *ghcb = svm->sev_es.ghcb;
2295
2296         /*
2297          * The GHCB protocol so far allows for the following data
2298          * to be returned:
2299          *   GPRs RAX, RBX, RCX, RDX
2300          *
2301          * Copy their values, even if they may not have been written during the
2302          * VM-Exit.  It's the guest's responsibility to not consume random data.
2303          */
2304         ghcb_set_rax(ghcb, vcpu->arch.regs[VCPU_REGS_RAX]);
2305         ghcb_set_rbx(ghcb, vcpu->arch.regs[VCPU_REGS_RBX]);
2306         ghcb_set_rcx(ghcb, vcpu->arch.regs[VCPU_REGS_RCX]);
2307         ghcb_set_rdx(ghcb, vcpu->arch.regs[VCPU_REGS_RDX]);
2308 }
2309
2310 static void sev_es_sync_from_ghcb(struct vcpu_svm *svm)
2311 {
2312         struct vmcb_control_area *control = &svm->vmcb->control;
2313         struct kvm_vcpu *vcpu = &svm->vcpu;
2314         struct ghcb *ghcb = svm->sev_es.ghcb;
2315         u64 exit_code;
2316
2317         /*
2318          * The GHCB protocol so far allows for the following data
2319          * to be supplied:
2320          *   GPRs RAX, RBX, RCX, RDX
2321          *   XCR0
2322          *   CPL
2323          *
2324          * VMMCALL allows the guest to provide extra registers. KVM also
2325          * expects RSI for hypercalls, so include that, too.
2326          *
2327          * Copy their values to the appropriate location if supplied.
2328          */
2329         memset(vcpu->arch.regs, 0, sizeof(vcpu->arch.regs));
2330
2331         vcpu->arch.regs[VCPU_REGS_RAX] = ghcb_get_rax_if_valid(ghcb);
2332         vcpu->arch.regs[VCPU_REGS_RBX] = ghcb_get_rbx_if_valid(ghcb);
2333         vcpu->arch.regs[VCPU_REGS_RCX] = ghcb_get_rcx_if_valid(ghcb);
2334         vcpu->arch.regs[VCPU_REGS_RDX] = ghcb_get_rdx_if_valid(ghcb);
2335         vcpu->arch.regs[VCPU_REGS_RSI] = ghcb_get_rsi_if_valid(ghcb);
2336
2337         svm->vmcb->save.cpl = ghcb_get_cpl_if_valid(ghcb);
2338
2339         if (ghcb_xcr0_is_valid(ghcb)) {
2340                 vcpu->arch.xcr0 = ghcb_get_xcr0(ghcb);
2341                 kvm_update_cpuid_runtime(vcpu);
2342         }
2343
2344         /* Copy the GHCB exit information into the VMCB fields */
2345         exit_code = ghcb_get_sw_exit_code(ghcb);
2346         control->exit_code = lower_32_bits(exit_code);
2347         control->exit_code_hi = upper_32_bits(exit_code);
2348         control->exit_info_1 = ghcb_get_sw_exit_info_1(ghcb);
2349         control->exit_info_2 = ghcb_get_sw_exit_info_2(ghcb);
2350
2351         /* Clear the valid entries fields */
2352         memset(ghcb->save.valid_bitmap, 0, sizeof(ghcb->save.valid_bitmap));
2353 }
2354
2355 static int sev_es_validate_vmgexit(struct vcpu_svm *svm)
2356 {
2357         struct kvm_vcpu *vcpu;
2358         struct ghcb *ghcb;
2359         u64 exit_code = 0;
2360
2361         ghcb = svm->sev_es.ghcb;
2362
2363         /* Only GHCB Usage code 0 is supported */
2364         if (ghcb->ghcb_usage)
2365                 goto vmgexit_err;
2366
2367         /*
2368          * Retrieve the exit code now even though is may not be marked valid
2369          * as it could help with debugging.
2370          */
2371         exit_code = ghcb_get_sw_exit_code(ghcb);
2372
2373         if (!ghcb_sw_exit_code_is_valid(ghcb) ||
2374             !ghcb_sw_exit_info_1_is_valid(ghcb) ||
2375             !ghcb_sw_exit_info_2_is_valid(ghcb))
2376                 goto vmgexit_err;
2377
2378         switch (ghcb_get_sw_exit_code(ghcb)) {
2379         case SVM_EXIT_READ_DR7:
2380                 break;
2381         case SVM_EXIT_WRITE_DR7:
2382                 if (!ghcb_rax_is_valid(ghcb))
2383                         goto vmgexit_err;
2384                 break;
2385         case SVM_EXIT_RDTSC:
2386                 break;
2387         case SVM_EXIT_RDPMC:
2388                 if (!ghcb_rcx_is_valid(ghcb))
2389                         goto vmgexit_err;
2390                 break;
2391         case SVM_EXIT_CPUID:
2392                 if (!ghcb_rax_is_valid(ghcb) ||
2393                     !ghcb_rcx_is_valid(ghcb))
2394                         goto vmgexit_err;
2395                 if (ghcb_get_rax(ghcb) == 0xd)
2396                         if (!ghcb_xcr0_is_valid(ghcb))
2397                                 goto vmgexit_err;
2398                 break;
2399         case SVM_EXIT_INVD:
2400                 break;
2401         case SVM_EXIT_IOIO:
2402                 if (ghcb_get_sw_exit_info_1(ghcb) & SVM_IOIO_STR_MASK) {
2403                         if (!ghcb_sw_scratch_is_valid(ghcb))
2404                                 goto vmgexit_err;
2405                 } else {
2406                         if (!(ghcb_get_sw_exit_info_1(ghcb) & SVM_IOIO_TYPE_MASK))
2407                                 if (!ghcb_rax_is_valid(ghcb))
2408                                         goto vmgexit_err;
2409                 }
2410                 break;
2411         case SVM_EXIT_MSR:
2412                 if (!ghcb_rcx_is_valid(ghcb))
2413                         goto vmgexit_err;
2414                 if (ghcb_get_sw_exit_info_1(ghcb)) {
2415                         if (!ghcb_rax_is_valid(ghcb) ||
2416                             !ghcb_rdx_is_valid(ghcb))
2417                                 goto vmgexit_err;
2418                 }
2419                 break;
2420         case SVM_EXIT_VMMCALL:
2421                 if (!ghcb_rax_is_valid(ghcb) ||
2422                     !ghcb_cpl_is_valid(ghcb))
2423                         goto vmgexit_err;
2424                 break;
2425         case SVM_EXIT_RDTSCP:
2426                 break;
2427         case SVM_EXIT_WBINVD:
2428                 break;
2429         case SVM_EXIT_MONITOR:
2430                 if (!ghcb_rax_is_valid(ghcb) ||
2431                     !ghcb_rcx_is_valid(ghcb) ||
2432                     !ghcb_rdx_is_valid(ghcb))
2433                         goto vmgexit_err;
2434                 break;
2435         case SVM_EXIT_MWAIT:
2436                 if (!ghcb_rax_is_valid(ghcb) ||
2437                     !ghcb_rcx_is_valid(ghcb))
2438                         goto vmgexit_err;
2439                 break;
2440         case SVM_VMGEXIT_MMIO_READ:
2441         case SVM_VMGEXIT_MMIO_WRITE:
2442                 if (!ghcb_sw_scratch_is_valid(ghcb))
2443                         goto vmgexit_err;
2444                 break;
2445         case SVM_VMGEXIT_NMI_COMPLETE:
2446         case SVM_VMGEXIT_AP_HLT_LOOP:
2447         case SVM_VMGEXIT_AP_JUMP_TABLE:
2448         case SVM_VMGEXIT_UNSUPPORTED_EVENT:
2449                 break;
2450         default:
2451                 goto vmgexit_err;
2452         }
2453
2454         return 0;
2455
2456 vmgexit_err:
2457         vcpu = &svm->vcpu;
2458
2459         if (ghcb->ghcb_usage) {
2460                 vcpu_unimpl(vcpu, "vmgexit: ghcb usage %#x is not valid\n",
2461                             ghcb->ghcb_usage);
2462         } else {
2463                 vcpu_unimpl(vcpu, "vmgexit: exit reason %#llx is not valid\n",
2464                             exit_code);
2465                 dump_ghcb(svm);
2466         }
2467
2468         vcpu->run->exit_reason = KVM_EXIT_INTERNAL_ERROR;
2469         vcpu->run->internal.suberror = KVM_INTERNAL_ERROR_UNEXPECTED_EXIT_REASON;
2470         vcpu->run->internal.ndata = 2;
2471         vcpu->run->internal.data[0] = exit_code;
2472         vcpu->run->internal.data[1] = vcpu->arch.last_vmentry_cpu;
2473
2474         return -EINVAL;
2475 }
2476
2477 void sev_es_unmap_ghcb(struct vcpu_svm *svm)
2478 {
2479         if (!svm->sev_es.ghcb)
2480                 return;
2481
2482         if (svm->sev_es.ghcb_sa_free) {
2483                 /*
2484                  * The scratch area lives outside the GHCB, so there is a
2485                  * buffer that, depending on the operation performed, may
2486                  * need to be synced, then freed.
2487                  */
2488                 if (svm->sev_es.ghcb_sa_sync) {
2489                         kvm_write_guest(svm->vcpu.kvm,
2490                                         ghcb_get_sw_scratch(svm->sev_es.ghcb),
2491                                         svm->sev_es.ghcb_sa,
2492                                         svm->sev_es.ghcb_sa_len);
2493                         svm->sev_es.ghcb_sa_sync = false;
2494                 }
2495
2496                 kvfree(svm->sev_es.ghcb_sa);
2497                 svm->sev_es.ghcb_sa = NULL;
2498                 svm->sev_es.ghcb_sa_free = false;
2499         }
2500
2501         trace_kvm_vmgexit_exit(svm->vcpu.vcpu_id, svm->sev_es.ghcb);
2502
2503         sev_es_sync_to_ghcb(svm);
2504
2505         kvm_vcpu_unmap(&svm->vcpu, &svm->sev_es.ghcb_map, true);
2506         svm->sev_es.ghcb = NULL;
2507 }
2508
2509 void pre_sev_run(struct vcpu_svm *svm, int cpu)
2510 {
2511         struct svm_cpu_data *sd = per_cpu(svm_data, cpu);
2512         int asid = sev_get_asid(svm->vcpu.kvm);
2513
2514         /* Assign the asid allocated with this SEV guest */
2515         svm->asid = asid;
2516
2517         /*
2518          * Flush guest TLB:
2519          *
2520          * 1) when different VMCB for the same ASID is to be run on the same host CPU.
2521          * 2) or this VMCB was executed on different host CPU in previous VMRUNs.
2522          */
2523         if (sd->sev_vmcbs[asid] == svm->vmcb &&
2524             svm->vcpu.arch.last_vmentry_cpu == cpu)
2525                 return;
2526
2527         sd->sev_vmcbs[asid] = svm->vmcb;
2528         svm->vmcb->control.tlb_ctl = TLB_CONTROL_FLUSH_ASID;
2529         vmcb_mark_dirty(svm->vmcb, VMCB_ASID);
2530 }
2531
2532 #define GHCB_SCRATCH_AREA_LIMIT         (16ULL * PAGE_SIZE)
2533 static int setup_vmgexit_scratch(struct vcpu_svm *svm, bool sync, u64 len)
2534 {
2535         struct vmcb_control_area *control = &svm->vmcb->control;
2536         struct ghcb *ghcb = svm->sev_es.ghcb;
2537         u64 ghcb_scratch_beg, ghcb_scratch_end;
2538         u64 scratch_gpa_beg, scratch_gpa_end;
2539         void *scratch_va;
2540
2541         scratch_gpa_beg = ghcb_get_sw_scratch(ghcb);
2542         if (!scratch_gpa_beg) {
2543                 pr_err("vmgexit: scratch gpa not provided\n");
2544                 return -EINVAL;
2545         }
2546
2547         scratch_gpa_end = scratch_gpa_beg + len;
2548         if (scratch_gpa_end < scratch_gpa_beg) {
2549                 pr_err("vmgexit: scratch length (%#llx) not valid for scratch address (%#llx)\n",
2550                        len, scratch_gpa_beg);
2551                 return -EINVAL;
2552         }
2553
2554         if ((scratch_gpa_beg & PAGE_MASK) == control->ghcb_gpa) {
2555                 /* Scratch area begins within GHCB */
2556                 ghcb_scratch_beg = control->ghcb_gpa +
2557                                    offsetof(struct ghcb, shared_buffer);
2558                 ghcb_scratch_end = control->ghcb_gpa +
2559                                    offsetof(struct ghcb, reserved_1);
2560
2561                 /*
2562                  * If the scratch area begins within the GHCB, it must be
2563                  * completely contained in the GHCB shared buffer area.
2564                  */
2565                 if (scratch_gpa_beg < ghcb_scratch_beg ||
2566                     scratch_gpa_end > ghcb_scratch_end) {
2567                         pr_err("vmgexit: scratch area is outside of GHCB shared buffer area (%#llx - %#llx)\n",
2568                                scratch_gpa_beg, scratch_gpa_end);
2569                         return -EINVAL;
2570                 }
2571
2572                 scratch_va = (void *)svm->sev_es.ghcb;
2573                 scratch_va += (scratch_gpa_beg - control->ghcb_gpa);
2574         } else {
2575                 /*
2576                  * The guest memory must be read into a kernel buffer, so
2577                  * limit the size
2578                  */
2579                 if (len > GHCB_SCRATCH_AREA_LIMIT) {
2580                         pr_err("vmgexit: scratch area exceeds KVM limits (%#llx requested, %#llx limit)\n",
2581                                len, GHCB_SCRATCH_AREA_LIMIT);
2582                         return -EINVAL;
2583                 }
2584                 scratch_va = kvzalloc(len, GFP_KERNEL_ACCOUNT);
2585                 if (!scratch_va)
2586                         return -ENOMEM;
2587
2588                 if (kvm_read_guest(svm->vcpu.kvm, scratch_gpa_beg, scratch_va, len)) {
2589                         /* Unable to copy scratch area from guest */
2590                         pr_err("vmgexit: kvm_read_guest for scratch area failed\n");
2591
2592                         kvfree(scratch_va);
2593                         return -EFAULT;
2594                 }
2595
2596                 /*
2597                  * The scratch area is outside the GHCB. The operation will
2598                  * dictate whether the buffer needs to be synced before running
2599                  * the vCPU next time (i.e. a read was requested so the data
2600                  * must be written back to the guest memory).
2601                  */
2602                 svm->sev_es.ghcb_sa_sync = sync;
2603                 svm->sev_es.ghcb_sa_free = true;
2604         }
2605
2606         svm->sev_es.ghcb_sa = scratch_va;
2607         svm->sev_es.ghcb_sa_len = len;
2608
2609         return 0;
2610 }
2611
2612 static void set_ghcb_msr_bits(struct vcpu_svm *svm, u64 value, u64 mask,
2613                               unsigned int pos)
2614 {
2615         svm->vmcb->control.ghcb_gpa &= ~(mask << pos);
2616         svm->vmcb->control.ghcb_gpa |= (value & mask) << pos;
2617 }
2618
2619 static u64 get_ghcb_msr_bits(struct vcpu_svm *svm, u64 mask, unsigned int pos)
2620 {
2621         return (svm->vmcb->control.ghcb_gpa >> pos) & mask;
2622 }
2623
2624 static void set_ghcb_msr(struct vcpu_svm *svm, u64 value)
2625 {
2626         svm->vmcb->control.ghcb_gpa = value;
2627 }
2628
2629 static int sev_handle_vmgexit_msr_protocol(struct vcpu_svm *svm)
2630 {
2631         struct vmcb_control_area *control = &svm->vmcb->control;
2632         struct kvm_vcpu *vcpu = &svm->vcpu;
2633         u64 ghcb_info;
2634         int ret = 1;
2635
2636         ghcb_info = control->ghcb_gpa & GHCB_MSR_INFO_MASK;
2637
2638         trace_kvm_vmgexit_msr_protocol_enter(svm->vcpu.vcpu_id,
2639                                              control->ghcb_gpa);
2640
2641         switch (ghcb_info) {
2642         case GHCB_MSR_SEV_INFO_REQ:
2643                 set_ghcb_msr(svm, GHCB_MSR_SEV_INFO(GHCB_VERSION_MAX,
2644                                                     GHCB_VERSION_MIN,
2645                                                     sev_enc_bit));
2646                 break;
2647         case GHCB_MSR_CPUID_REQ: {
2648                 u64 cpuid_fn, cpuid_reg, cpuid_value;
2649
2650                 cpuid_fn = get_ghcb_msr_bits(svm,
2651                                              GHCB_MSR_CPUID_FUNC_MASK,
2652                                              GHCB_MSR_CPUID_FUNC_POS);
2653
2654                 /* Initialize the registers needed by the CPUID intercept */
2655                 vcpu->arch.regs[VCPU_REGS_RAX] = cpuid_fn;
2656                 vcpu->arch.regs[VCPU_REGS_RCX] = 0;
2657
2658                 ret = svm_invoke_exit_handler(vcpu, SVM_EXIT_CPUID);
2659                 if (!ret) {
2660                         ret = -EINVAL;
2661                         break;
2662                 }
2663
2664                 cpuid_reg = get_ghcb_msr_bits(svm,
2665                                               GHCB_MSR_CPUID_REG_MASK,
2666                                               GHCB_MSR_CPUID_REG_POS);
2667                 if (cpuid_reg == 0)
2668                         cpuid_value = vcpu->arch.regs[VCPU_REGS_RAX];
2669                 else if (cpuid_reg == 1)
2670                         cpuid_value = vcpu->arch.regs[VCPU_REGS_RBX];
2671                 else if (cpuid_reg == 2)
2672                         cpuid_value = vcpu->arch.regs[VCPU_REGS_RCX];
2673                 else
2674                         cpuid_value = vcpu->arch.regs[VCPU_REGS_RDX];
2675
2676                 set_ghcb_msr_bits(svm, cpuid_value,
2677                                   GHCB_MSR_CPUID_VALUE_MASK,
2678                                   GHCB_MSR_CPUID_VALUE_POS);
2679
2680                 set_ghcb_msr_bits(svm, GHCB_MSR_CPUID_RESP,
2681                                   GHCB_MSR_INFO_MASK,
2682                                   GHCB_MSR_INFO_POS);
2683                 break;
2684         }
2685         case GHCB_MSR_TERM_REQ: {
2686                 u64 reason_set, reason_code;
2687
2688                 reason_set = get_ghcb_msr_bits(svm,
2689                                                GHCB_MSR_TERM_REASON_SET_MASK,
2690                                                GHCB_MSR_TERM_REASON_SET_POS);
2691                 reason_code = get_ghcb_msr_bits(svm,
2692                                                 GHCB_MSR_TERM_REASON_MASK,
2693                                                 GHCB_MSR_TERM_REASON_POS);
2694                 pr_info("SEV-ES guest requested termination: %#llx:%#llx\n",
2695                         reason_set, reason_code);
2696                 fallthrough;
2697         }
2698         default:
2699                 ret = -EINVAL;
2700         }
2701
2702         trace_kvm_vmgexit_msr_protocol_exit(svm->vcpu.vcpu_id,
2703                                             control->ghcb_gpa, ret);
2704
2705         return ret;
2706 }
2707
2708 int sev_handle_vmgexit(struct kvm_vcpu *vcpu)
2709 {
2710         struct vcpu_svm *svm = to_svm(vcpu);
2711         struct vmcb_control_area *control = &svm->vmcb->control;
2712         u64 ghcb_gpa, exit_code;
2713         struct ghcb *ghcb;
2714         int ret;
2715
2716         /* Validate the GHCB */
2717         ghcb_gpa = control->ghcb_gpa;
2718         if (ghcb_gpa & GHCB_MSR_INFO_MASK)
2719                 return sev_handle_vmgexit_msr_protocol(svm);
2720
2721         if (!ghcb_gpa) {
2722                 vcpu_unimpl(vcpu, "vmgexit: GHCB gpa is not set\n");
2723                 return -EINVAL;
2724         }
2725
2726         if (kvm_vcpu_map(vcpu, ghcb_gpa >> PAGE_SHIFT, &svm->sev_es.ghcb_map)) {
2727                 /* Unable to map GHCB from guest */
2728                 vcpu_unimpl(vcpu, "vmgexit: error mapping GHCB [%#llx] from guest\n",
2729                             ghcb_gpa);
2730                 return -EINVAL;
2731         }
2732
2733         svm->sev_es.ghcb = svm->sev_es.ghcb_map.hva;
2734         ghcb = svm->sev_es.ghcb_map.hva;
2735
2736         trace_kvm_vmgexit_enter(vcpu->vcpu_id, ghcb);
2737
2738         exit_code = ghcb_get_sw_exit_code(ghcb);
2739
2740         ret = sev_es_validate_vmgexit(svm);
2741         if (ret)
2742                 return ret;
2743
2744         sev_es_sync_from_ghcb(svm);
2745         ghcb_set_sw_exit_info_1(ghcb, 0);
2746         ghcb_set_sw_exit_info_2(ghcb, 0);
2747
2748         switch (exit_code) {
2749         case SVM_VMGEXIT_MMIO_READ:
2750                 ret = setup_vmgexit_scratch(svm, true, control->exit_info_2);
2751                 if (ret)
2752                         break;
2753
2754                 ret = kvm_sev_es_mmio_read(vcpu,
2755                                            control->exit_info_1,
2756                                            control->exit_info_2,
2757                                            svm->sev_es.ghcb_sa);
2758                 break;
2759         case SVM_VMGEXIT_MMIO_WRITE:
2760                 ret = setup_vmgexit_scratch(svm, false, control->exit_info_2);
2761                 if (ret)
2762                         break;
2763
2764                 ret = kvm_sev_es_mmio_write(vcpu,
2765                                             control->exit_info_1,
2766                                             control->exit_info_2,
2767                                             svm->sev_es.ghcb_sa);
2768                 break;
2769         case SVM_VMGEXIT_NMI_COMPLETE:
2770                 ret = svm_invoke_exit_handler(vcpu, SVM_EXIT_IRET);
2771                 break;
2772         case SVM_VMGEXIT_AP_HLT_LOOP:
2773                 ret = kvm_emulate_ap_reset_hold(vcpu);
2774                 break;
2775         case SVM_VMGEXIT_AP_JUMP_TABLE: {
2776                 struct kvm_sev_info *sev = &to_kvm_svm(vcpu->kvm)->sev_info;
2777
2778                 switch (control->exit_info_1) {
2779                 case 0:
2780                         /* Set AP jump table address */
2781                         sev->ap_jump_table = control->exit_info_2;
2782                         break;
2783                 case 1:
2784                         /* Get AP jump table address */
2785                         ghcb_set_sw_exit_info_2(ghcb, sev->ap_jump_table);
2786                         break;
2787                 default:
2788                         pr_err("svm: vmgexit: unsupported AP jump table request - exit_info_1=%#llx\n",
2789                                control->exit_info_1);
2790                         ghcb_set_sw_exit_info_1(ghcb, 1);
2791                         ghcb_set_sw_exit_info_2(ghcb,
2792                                                 X86_TRAP_UD |
2793                                                 SVM_EVTINJ_TYPE_EXEPT |
2794                                                 SVM_EVTINJ_VALID);
2795                 }
2796
2797                 ret = 1;
2798                 break;
2799         }
2800         case SVM_VMGEXIT_UNSUPPORTED_EVENT:
2801                 vcpu_unimpl(vcpu,
2802                             "vmgexit: unsupported event - exit_info_1=%#llx, exit_info_2=%#llx\n",
2803                             control->exit_info_1, control->exit_info_2);
2804                 ret = -EINVAL;
2805                 break;
2806         default:
2807                 ret = svm_invoke_exit_handler(vcpu, exit_code);
2808         }
2809
2810         return ret;
2811 }
2812
2813 int sev_es_string_io(struct vcpu_svm *svm, int size, unsigned int port, int in)
2814 {
2815         int count;
2816         int bytes;
2817         int r;
2818
2819         if (svm->vmcb->control.exit_info_2 > INT_MAX)
2820                 return -EINVAL;
2821
2822         count = svm->vmcb->control.exit_info_2;
2823         if (unlikely(check_mul_overflow(count, size, &bytes)))
2824                 return -EINVAL;
2825
2826         r = setup_vmgexit_scratch(svm, in, bytes);
2827         if (r)
2828                 return r;
2829
2830         return kvm_sev_es_string_io(&svm->vcpu, size, port, svm->sev_es.ghcb_sa,
2831                                     count, in);
2832 }
2833
2834 void sev_es_init_vmcb(struct vcpu_svm *svm)
2835 {
2836         struct kvm_vcpu *vcpu = &svm->vcpu;
2837
2838         svm->vmcb->control.nested_ctl |= SVM_NESTED_CTL_SEV_ES_ENABLE;
2839         svm->vmcb->control.virt_ext |= LBR_CTL_ENABLE_MASK;
2840
2841         /*
2842          * An SEV-ES guest requires a VMSA area that is a separate from the
2843          * VMCB page. Do not include the encryption mask on the VMSA physical
2844          * address since hardware will access it using the guest key.
2845          */
2846         svm->vmcb->control.vmsa_pa = __pa(svm->sev_es.vmsa);
2847
2848         /* Can't intercept CR register access, HV can't modify CR registers */
2849         svm_clr_intercept(svm, INTERCEPT_CR0_READ);
2850         svm_clr_intercept(svm, INTERCEPT_CR4_READ);
2851         svm_clr_intercept(svm, INTERCEPT_CR8_READ);
2852         svm_clr_intercept(svm, INTERCEPT_CR0_WRITE);
2853         svm_clr_intercept(svm, INTERCEPT_CR4_WRITE);
2854         svm_clr_intercept(svm, INTERCEPT_CR8_WRITE);
2855
2856         svm_clr_intercept(svm, INTERCEPT_SELECTIVE_CR0);
2857
2858         /* Track EFER/CR register changes */
2859         svm_set_intercept(svm, TRAP_EFER_WRITE);
2860         svm_set_intercept(svm, TRAP_CR0_WRITE);
2861         svm_set_intercept(svm, TRAP_CR4_WRITE);
2862         svm_set_intercept(svm, TRAP_CR8_WRITE);
2863
2864         /* No support for enable_vmware_backdoor */
2865         clr_exception_intercept(svm, GP_VECTOR);
2866
2867         /* Can't intercept XSETBV, HV can't modify XCR0 directly */
2868         svm_clr_intercept(svm, INTERCEPT_XSETBV);
2869
2870         /* Clear intercepts on selected MSRs */
2871         set_msr_interception(vcpu, svm->msrpm, MSR_EFER, 1, 1);
2872         set_msr_interception(vcpu, svm->msrpm, MSR_IA32_CR_PAT, 1, 1);
2873         set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTBRANCHFROMIP, 1, 1);
2874         set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTBRANCHTOIP, 1, 1);
2875         set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTINTFROMIP, 1, 1);
2876         set_msr_interception(vcpu, svm->msrpm, MSR_IA32_LASTINTTOIP, 1, 1);
2877 }
2878
2879 void sev_es_vcpu_reset(struct vcpu_svm *svm)
2880 {
2881         /*
2882          * Set the GHCB MSR value as per the GHCB specification when emulating
2883          * vCPU RESET for an SEV-ES guest.
2884          */
2885         set_ghcb_msr(svm, GHCB_MSR_SEV_INFO(GHCB_VERSION_MAX,
2886                                             GHCB_VERSION_MIN,
2887                                             sev_enc_bit));
2888 }
2889
2890 void sev_es_prepare_guest_switch(struct vcpu_svm *svm, unsigned int cpu)
2891 {
2892         struct svm_cpu_data *sd = per_cpu(svm_data, cpu);
2893         struct vmcb_save_area *hostsa;
2894
2895         /*
2896          * As an SEV-ES guest, hardware will restore the host state on VMEXIT,
2897          * of which one step is to perform a VMLOAD. Since hardware does not
2898          * perform a VMSAVE on VMRUN, the host savearea must be updated.
2899          */
2900         vmsave(__sme_page_pa(sd->save_area));
2901
2902         /* XCR0 is restored on VMEXIT, save the current host value */
2903         hostsa = (struct vmcb_save_area *)(page_address(sd->save_area) + 0x400);
2904         hostsa->xcr0 = xgetbv(XCR_XFEATURE_ENABLED_MASK);
2905
2906         /* PKRU is restored on VMEXIT, save the current host value */
2907         hostsa->pkru = read_pkru();
2908
2909         /* MSR_IA32_XSS is restored on VMEXIT, save the currnet host value */
2910         hostsa->xss = host_xss;
2911 }
2912
2913 void sev_vcpu_deliver_sipi_vector(struct kvm_vcpu *vcpu, u8 vector)
2914 {
2915         struct vcpu_svm *svm = to_svm(vcpu);
2916
2917         /* First SIPI: Use the values as initially set by the VMM */
2918         if (!svm->sev_es.received_first_sipi) {
2919                 svm->sev_es.received_first_sipi = true;
2920                 return;
2921         }
2922
2923         /*
2924          * Subsequent SIPI: Return from an AP Reset Hold VMGEXIT, where
2925          * the guest will set the CS and RIP. Set SW_EXIT_INFO_2 to a
2926          * non-zero value.
2927          */
2928         if (!svm->sev_es.ghcb)
2929                 return;
2930
2931         ghcb_set_sw_exit_info_2(svm->sev_es.ghcb, 1);
2932 }